Earnings per share came in at $2.03 adjusted vs. $1.96 expected and, most importantly, the company confirmed its revenue guidance for the upcoming year at $90B. By 2030, the company wants to reach $225B, setting highly ambitious goals.

A vital metric for this is Remaining Performance Obligations, or RPO, which ended the quarter at $638 billion, up 363% USD year-over-year and up $85 billion sequentially from the end of Q3. This indicates massive demand for data centers that shows zero signs of slowing down; in any other market, the stock wouldn’t be dropping 7%, it would have already taken off. Just bad timing. At the same time, investors have grown more cautious about infrastructure plays due to high debt levels.

Just the day before, Crusoe reported that a client asked to pause data center construction, though sources say it was actually due to resistance from local residents, and another company is rumored to be taking over the project. In short, the sentiment ahead of the print was highly anxious. Meanwhile, Oracle plans to bring 1GW online in Q1’27 alone, compared to 1.2GW delivered across the entire fiscal year 2026.

All GPUs are taken

On the call, CEO Clay Magouyrk dropped a key note, stating that the Global GPU utilization rate is sitting at 97.5%, even though only about half of the 59 clients renewed their contracts for 35,000 GPUs. Those who stay are taking capacity to the absolute maximum, meaning the data centers face zero idle time. Moreover, roughly $20-25B comes from client prepayments, so Oracle itself will only need to borrow $40B externally, $20B of which is the already announced ATM equity issuance, while total Capex hits $70B. Contracts where clients either bring their own GPUs or pay upfront have reached $75B.

On top of that, the market has started cutting borrowing rates for infrastructure companies; just recently, Applied Digital secured financing at 7% instead of the 10% demanded a few months ago. As another growth vector, Oracle highlighted the MultiCloud database segment, which surged 404% in Q4. Oracle Health is also projected to post double-digit growth in the upcoming fiscal year, signaling that the $28B Cerner acquisition is finally starting to pay off. Oracle is also teaming up with OpenAI to make frontier models and Codex directly available to OCI clients.

However, the stock still tumbled 10% because the market is fundamentally unready for good news. High inflation, a looming war in Iran with zero signs of diplomatic progress - are negotiations even happening? - have poisoned the tape. It is easy to imagine how much more volatile the reaction would have been if there had been any actual mentions of construction delays.

The price to pay

Ultimately, this aggressive pivot means the company is paying a substantial price to secure its future. Pushing Capex to $55.6B has compressed Free Cash Flow to negative $23.7B for the fiscal year, requiring continuous external capital raises and a temporary pause in meaningful share buybacks alongside the upcoming $20B ATM dilution.

Furthermore, while Oracle successfully defended a strong 45% non-GAAP operating margin this quarter, management openly noted that operating margins will head lower next year as the revenue mix shifts from legacy software to heavy hardware infrastructure.

Yet, with a 97.5% GPU utilization rate keeping the current footprint highly efficient, this massive capital investment is a necessary and calculated move

Death of the Software Subscription

Oracle is fundamentally reshaping enterprise software monetization by abandoning traditional seat-based licensing in favor of two distinct models.

The first is outcome-based pricing. Instead of charging a fixed monthly fee per user regardless of utility, Oracle now ties its revenue directly to measurable business outcomes delivered by its autonomous agents. In healthcare, billing is tied to patient throughput.

In HR, fees are based on the number of candidates screened and moved to final stages. In sales, hospitality and retail assistants take a percentage cut of closed upsell transactions. This lowers the barrier to adoption, as corporate clients only incur costs when the software generates verifiable operational savings or revenue.

The second model is universal token bundles, which provide enterprise buyers with strict budget predictability. Instead of buying separate software modules, clients purchase a predictable pool of tokens and allocate them dynamically across application suites, shifting capacity to HR agents during a hiring surge or to ERP agents during end-of-quarter financial closing.

Right now, looking at Oracle’s standard 10% Cloud Applications (SaaS) growth completely misses the revolution. These new token bundles are still in a limited rollout phase with only 33 major corporate clients, including Aon and Liberty Energy, participating in Q4. Because clients buy these tokens in advance, the billions in cash prepayments do not show up in current revenue, instead, they sit locked in an unprecedented $85 billion quarterly surge in RPO.

This creates a massive cash spring. As these initial 33 clients - and the thousands waiting in line - gradually burn through their tokens to run live AI agents over the coming quarters, the current 10% growth rate is mechanically set to accelerate significantly over time.

While SaaS used to monetize through flat-rate add-on fees, Oracle is shifting toward utility billing and performance-sharing, securing highly predictable, high-margin revenue pipelines

Buying Opportunity

Another positive sign for investors is that the company won a major government contract for HR optimization. My valuation of the company remains unchanged—I believe it should be worth at least $250. All of my calculations can be viewed here:

The earnings report proves once again that there is massive demand for compute, and GPUs remain highly sought after by clients who actually know how to monetize them. In other words, differentiation is taking place in the market, which is a highly positive factor.

I also see a clear positive in the fact that the company is operating in an environment of fierce competition against the Hyperscalers despite having significantly smaller financial resources, yet it continues to grow at 20% a year even with a stagnating SaaS segment.

Another major advantage for the company is its ability to offer a full stack of services simultaneously; they possess a complete cloud infrastructure, and the more GPU clients they acquire, the greater their monetization opportunities become.

Oracle also promises higher earnings from its healthcare business, an industry where, as we know, serious money changes hands. All in all, today’s sell-off can be used as an opportunity to add to the position. Given the new bombings of Iran, the price could drop further, so the macro landscape must be taken into account.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author is long Oracle.

The irony is that $HYLN is essentially a zombie company that just threw in the towel and killed off its legacy powertrain business for heavy trucks. They are currently touting a $400 million backlog, but if you actually crack open their financial filings, this entire “pipeline” consists of non-binding LOIs, meaning entirely unbacked promises. Commercial revenue is flat zero, and the business is keeping the lights on solely through R&D life support thrown their way by the Pentagon.

And yet, in a market where software moats are evaporating and AI image generators are eating Adobe’s lunch, investors are desperate for a moat in the physical world. $HYLN promises the KARNO linear engine, which pipes 800V DC power straight into AI racks and handles multi-fuel switching from gas to diesel on the fly.

Let’s dive in and see whether there is real, tangible hardware here, or if we’re just looking at another venture-backed simulacrum living on government subsidies.

History

The story of Hyliion is a textbook saga of a venture-backed startup smashing face-first into the brutal reality of heavy manufacturing, only to slickly rebrand itself just in time to ride the AI wave.

The company stumbled onto the public markets in 2020 via a SPAC merger at the absolute peak of the “green tech” bubble, back when retail investors were throwing cash at any EV truck maker capable of coasting down a hill. Founder Thomas Healy. then not even 30 years old, promised Wall Street a revolution: electrified and hybrid powertrains for Class 8 heavy-duty trucks. The game plan was simple: take existing rigs from Peterbilt or Kenworth, rip out the diesel engines, drop in their proprietary battery packs and software, and sell them to logistics fleets.

But reality proved far more cynical than their pitch decks. Over the next eight years, Hyliion hit every conceivable wall: capital expenditure hell, regulatory bottlenecks, and a total refusal from fleet operators to buy heavy batteries that slaughtered payload capacity and required a charging infrastructure that flat-out didn’t exist.

By late 2023, the writing was on the wall. The project was dead. Having burned through hundreds of millions of investor dollars and locking in massive losses, management made the only logical move left, they completely liquidated the powertrain business. They spent the first quarter of 2026 scrapping old machinery and selling off legacy assets, leaving them with nothing but a tiny, non-recurring credit on the balance sheet.

In any rational universe, this is where the story ends in chapter 11. But Healy had one final card to play. Back in 2022, they had quietly scooped up the rights to the KARNO technology from GE Additive for pennies on the dollar, a compact linear generator originally designed for aerospace applications.

When Silicon Valley suddenly woke up to the realization that training massive LLMs was about to break the US power grid, Hyliion pulled off an emergency pivot. They repackaged General Electric’s aerospace tech into a sleek enclosure, slapped the label “modular power plant for data centers” on it, and marched back to retail investors with a fresh mantra: “Forget trucks. We are here to feed NVIDIA chips.”

Investors took the bait, and the zombie stock rose from the grave.

Read more

There will be plenty of players in this race, but NVIDIA has already figured out how to make the most money. They are pulling off their signature move, the exact same playbook that made them the kings of computer graphics and deep learning.

The mechanics are simple: Developing a robot from scratch is brutally expensive and time-consuming. Every manufacturer has to spend years writing code just to teach a machine how to walk without falling over at the slightest push. NVIDIA stepped in and said: “Here is a ready-to-use operating system and baseline AI models. Take them and save your R&D money”.

As a result, robot makers - from American startups to Chinese giants - are happily adopting this software to skip billions of dollars in development costs. But the moment they do, they step right into the trap. Once the entire industry builds its applications on top of NVIDIA’s software, switching off it will be virtually impossible. NVIDIA will effectively become the main toll booth and infrastructure owner for the global robotics market.

The Bifurcated Market: China’s Hardware vs. Western Brains

The global robotics race isn’t going to have just one winner. Instead, the market is splitting into two completely different camps, each playing a totally different game.

Camp 1: China (The Kings of Cheap Hardware) China is treating humanoids exactly like they treated electric vehicles: push massive supplier ecosystems, cut production costs to the bone, and flood the market. Companies like Unitree are already selling robots for $10,000 to $13,500, the price of a cheap used car. They can do this because they are reusing China’s massive EV supply chain to stamp out motors and mechanical joints at lightning speed. The catch? On their own, these cheap robots lack advanced safety features and high-end processing power.

Camp 2: The West (The Kings of Complex AI and Safety) US and European giants like Tesla (Optimus) or Figure aren’t trying to build cheap toys. They are building expensive, vertically integrated machines designed to pass strict safety audits so they can legally work right next to human factory workers. They build their own custom actuators and proprietary AI brains, keeping tight control over the entire system.

NVIDIA’s Double Win

This market split is exactly why NVIDIA wins twice. They sit right in the middle of this geopolitical divide, collecting a toll from both sides:

• Cheap Chinese hardware needs to get smart fast, so manufacturers plug Unitree’s low-cost bodies straight into NVIDIA’s Jetson onboard chips and GR00T software.

• High-end Western developers need to test their custom setups without breaking million-dollar prototypes, so they use NVIDIA’s advanced simulators to train their models in a digital world before moving to the real factory floor

No matter who wins the battle for the physical robot body, NVIDIA owns the infrastructure powering it

Sourcing the Muscle: How NVIDIA Captured China’s Hardware Fleet

If you want to understand how this supercycle plays out on the global stage, ignore the political grandstanding about semiconductor bans. While Washington blocks NVIDIA’s top-tier data center GPUs from entering Beijing, NVIDIA has quietly executed a brilliant flank maneuver right through China’s robotics backdoor

They aren’t trying to sell high-end server chips to Chinese tech giants; they are shipping edge-compute brains (the Jetson Thor platform) and handing out software keys directly to China’s fastest-moving hardware startups.

Look no further than Unitree Robotics

China’s playbook for humanoids is an exact carbon copy of its electric vehicle strategy: exploit an ultra-dense, localized electromechanical supply chain to compress hardware costs to the absolute bone. Unitree shocked the industry by putting a public price tag of $10,000 to $13,500 on its G1 humanoid, the price of a cheap commuter car. They can pull this off because they sit in the middle of a massive manufacturing cluster, grabbing motors, bearings, and structural frames at fraction-of-a-cent margins.

But cheap hardware has a fatal flaw: it’s profoundly stupid. Left to their own devices, these dirt-cheap Chinese prototypes lack the complex vision systems, real-time balance, and basic safety guardrails required to step foot on an active assembly line without causing a disaster

Enter NVIDIA’s asymmetrical trade. They NVIDIA didn’t bother trying to build a cheaper mechanical body than the Chinese. Instead, they took Unitree’s H2 Plus chassis, loaded it with their high-performance Jetson AGX Thor T5000 board, and wrapped it in the GR00T cognitive software stack to create their official Reference Humanoid Robot.

The mechanics of this partnership are purely transactional:

• For Unitree: They instantly wipe out billions in software R&D costs and skip years of trial-and-error by plugging into a world-class AI brain.

• For NVIDIA: They ensure that the largest, fastest-scaling hardware fleet on earth is being built on their proprietary software architecture.

NVIDIA doesn’t care if geopolitical tensions freeze out traditional silicon markets. By partnering with Unitree and giving China the software tools to commoditize the robot body, NVIDIA guarantees that whenever a cheap Chinese robot walks out of a factory, it operates on an unskippable NVIDIA tax

The US Landscape: Real Capex, Pure Plays, and the Hidden Infrastructure Cash

While China focuses on scaling cheap hardware, the US ecosystem is where the heavy institutional capital is drawing the line between laboratory science fiction and actual corporate balance sheets. If you want to allocate capital here without getting burned by pre-revenue vaporware, you need to separate the asset classes.

1. The Logistics Overlords: Amazon and Tesla

Don’t buy Amazon (AMZN) or Tesla (TSLA) thinking they are “robotics startups.” They are the ultimate end-users and data collectors.

Take Amazon: they are committing over €10 billion to expand and upgrade their European fulfillment network with advanced automation. They aren’t waiting for a perfect human clone. They are deploying STARK, a dedicated robotic tote-handling system, across 15 European sites by 2027, and scaling Vulcan, a system built explicitly with the high-end tactile sensing capabilities that legacy automation lacks. Their upgraded Proteus autonomous floor robot doesn’t just sit in the loading docks anymore; it calculates its own routes, priorities, and timing across the entire warehouse floor. Amazon is building the structural demand; they are buying efficiency, not hype.

Tesla plays a similar game with Optimus, vertically integrating their automotive FSD compute stack into a custom bipedal frame to cut out third-party suppliers entirely. They are their own first customer.

2. Lidars: Ouster and Aeva

If you don’t want to guess whether Amazon’s or Tesla’s proprietary brain wins, you buy the eyes. You cannot move a multi-hundred-pound metal agent fenceless alongside human workers without flawless, low-latency 3D spatial awareness.

This is where the LiDAR plays become critical. has already hardcoded themselves into NVIDIA’s ecosystem. Their new Rev8 digital LiDAR family runs on custom L4 Silicon featuring native-color fusion. By baking color data directly into the 3D point cloud at the chip level, they eliminate the clunky, power-hungry sensor-fusion software layers that drain robot batteries.

NVIDIA natively supports Rev8 inside their JetPack and Isaac ROS middleware, turning Ouster into a structural gatekeeper of physical AI perception.

follows a similar high-assurance trajectory with FMCW technology, capturing velocity data along with depth.

3. The “Anti-Humanoid” Smart Money: Teradyne

Here is the ultimate institutional hedge. If you think humanoids are still too early, too clumsy, and too financially risky for mass adoption, you buy Teradyne owns Universal Robots (the undisputed global king of collaborative robotic arms, or cobots) and Mobile Industrial Robots (MiR) (autonomous warehouse vehicles). Teradyne doesn’t give a damn about building a robot that can walk up stairs or fold laundry. They care about factory throughput today.

While humanoid startups spend billions trying to figure out how to keep a bipedal robot from falling over when a battery dies, Teradyne’s stationary cobots and wheeled platforms are already generating billions in high-margin industrial revenue. They are the pragmatic alternative for investors who want automation capex without the humanoid execution risk.

The South Korean Cluster: Industrial Execution vs. Valuation Traps

South Korea has become the explosive epicenter of the current Physical AI investment supercycle. When NVIDIA’s Jensen Huang visited Seoul, it triggered a massive rally across Korean robotics stocks on intense speculation over deep hardware-software cooperation. Retail investors are chasing anything with “robot” in the name, but if you look past the headlines and dive into the actual Q1 2026 financial balance sheets, the underlying realities of these companies could not be more different.

If you want to allocate capital in the Korean cluster without buying into pre-revenue vaporware, this is how you separate the actual winners from the valuation traps:

1. The Top Pick: Doosan Robotics (KS: 454910)

Doosan is the rare player successfully escaping “pilot purgatory” and transforming into a true global system integrator. They aren’t just selling bare-metal robotic arms; they are deploying complete, turn-key automated packaging cells straight into premium Western markets.

Following their acquisition of US automation firm ONExia, Doosan legally launched Doosan Robotics Americas, shifted its headquarters to Pennsylvania, and approved a new 90,000-square-foot facility to aggressively scale up North American production. The financial pivot is already visible on the tape: Doosan’s Q1 2026 revenue exploded by +189.7% year-over-year, driven by rapid US expansion and domestic automation demand.

The NVIDIA Connection: Doosan is a premier technological ally of NVIDIA. Their proprietary Agentic Robot OS officially embeds NVIDIA’s newly launched Cosmos 3, the open frontier foundation model for physical AI, alongside the cuMotion path-planning library. This enables their collaborative robots to dynamically recognize obstacles, read product labels, and instantly calculate motion paths on a live production line without any manual programming.

Yes, they are still reporting a temporary operational loss due to aggressive front-loaded investments in AI developers and US hiring, but with a massive post-IPO cash cushion of 245 billion won and practically zero debt, Doosan is the cleanest commercial execution play in Asia.

2. The Valuation Trap: Rainbow Robotics (KOSDAQ: 277810)

On paper, Rainbow Robotics looks like an institutional darling. Samsung Electronics recently executed its options to push its stake to 35%, effectively absorbing the company into its corporate контур and turning it into a captive, in-house automation provider for Samsung’s semiconductor fabs, logistics hubs, and shipyards. Backed by Samsung’s massive volume, their Q1 2026 revenue jumped a solid +117%.

But look at the actual market math. Rainbow is currently trading at a jaw-dropping, completely detached 380x to 540x LTM Price-to-Sales. You are paying for roughly 500 years of current revenue today. Their entire growth trajectory is trapped inside the capital expenditure budget of a single customer.

To make matters worse, the stock is carrying severe corporate governance risks: Seoul prosecutors recently raided their headquarters over an insider trading probe, alleging that management used non-public info about the Samsung buyout to front-run the stock for illicit profits. At these extreme multiples, with toxic legal clouds overhead, you stay far away.

3. The R&D Money Pit: Robotis (KOSDAQ: 108490)

Robotis occupies a highly critical infrastructure niche. They manufacture DYNAMIXEL smart actuators, the highly precise mechanical “joints” that other companies buy to make their humanoids and robotic arms move. LG Electronics is a major strategic backer, securing a steady component sales channel.

The problem? Robotis is trapped in a brutal, capital-destroying R&D arms race. Their Q1 2026 earnings revealed a catastrophic consolidated operational loss of 11.84 billion won on flat revenues. Why? Because they had to nearly quadruple their quarterly R&D spending to 9.42 billion won just to upgrade their actuator tech and prevent aggressive, lower-cost Chinese component makers from eating their market share. Robotis has to burn almost every dollar it makes just to stand still, historically resulting in negative returns on equity and constant shareholder dilution. They are a vital supplier, but a terrible vehicle for generating free cash flow.

4. The Micro-Cap Alternative: SBB Tech (KOSDAQ: 389500)

If you want to hedge against the near-monopoly of Japanese gearbox makers like Harmonic Drive, SBB Tech is South Korea’s homegrown alternative. They specialize in localizing the manufacturing of precision bearings and harmonic reducers, aiming to dramatically slash component costs for domestic robot makers. It’s a high-risk, micro-cap bet on supply chain independence, but if they can capture even a fraction of the market pivoting away from high Japanese hardware premiums, the upside is highly asymmetric.

Crucially, SBB Tech supplies harmonic reducers and actuators directly to Hyundai for their Robotics Lab and smart manufacturing platforms. This puts SBB Tech in a highly strategic position: Hyundai happens to own Boston Dynamics, widely recognized as the most advanced humanoid robotics company on the planet. By embedding its hardware into Hyundai's automation and R&D pipelines, SBB Tech acts as a high-risk, asymmetric micro-cap play on domestic supply chain independence. If they can capture even a fraction of the hardware volume moving through Hyundai's massive industrial ecosystem, the upside is massive.

The Japanese Throttle: Hardware Monopolies

If you want to move away from the software layers and the corporate end-users, you have to look at the cold, hard laws of physics. The single largest cost block of any humanoid robot is actuation, the motors and gearboxes that allow joints to move, accounting for a staggering 40% to 60% of the entire bill of materials.

You can write the most brilliant AI software in Silicon Valley, but if your robot doesn’t have a high-torque, precision gearbox in its knee or elbow, it’s just a very expensive, motionless metal statue. And when it comes to precision mechanical joints, Japan owns the entire global infrastructure.

If you want a pure-play investment on the literal nuts and bolts of this industry, you look at two companies:

1. Harmonic Drive Systems (Tokyo: 6324)

This is the ultimate bottleneck asset. Humanoid joints require incredibly compact, light, and zero-backlash gearboxes. The industry standard for this is the strain-wave drive, and Harmonic Drive Systems is the undisputed, historical gatekeeper of this technology.

Scaling up production for these gearboxes is capital-intensive, precision-bound, and requires long qualification cycles. Unlike software, you cannot just spin up more server capacity to build a gearbox; it requires heavy industrial machinery and decades of metrology expertise. As every humanoid maker tries to scale production from dozens of pilots to thousands of commercial units, they all rush to the exact same Japanese doorstep. It is a structural choke point.

2. Nabtesco Corporation (Tokyo: 6268)

While Harmonic Drive dominates the micro-precision joints like wrists and fingers, Nabtesco is the king of heavy lifting. They specialize in cycloidal drives, which are preferred for larger, high-impact robotic joints that require extreme durability and torque density like hips and knees.

If a robot falls over, backdriving forces can instantly shatter a cheap gearbox. Nabtesco’s cycloidal setups are built to survive industrial factory floors. While alternative gearbox tech exists, Nabtesco remains deeply insulated from competition due to massive barriers to entry in precision engineering.

While the US and China fight a loud geopolitical war over AI models and software standards, Japanese component suppliers are quietly sitting back, waiting to collect their hardware tax from every single robot that gets assembled anywhere on Earth.

Many of these companies have already seen significant runs, but they still hold massive upside as the broader market scales. Because of this, the smart play is to build your positions gradually, especially since pre-profit tech stocks are notoriously volatile. By diversifying your exposure, buying both the final robot manufacturers and the core actuator component suppliers, you can significantly de-risk your portfolio.

Personally, I see a very strong combination in pairing Hyundai with Doosan Robotics, though I am still patiently waiting for a cleaner entry point on Doosan before pulling the trigger.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author hold positions in Hyundai and NVIDIA

Naturally, everyone wants a piece of this growth. But if you look at the US public markets, direct pure-play options are extremely limited. Investors are mostly left with foundational enablers like or hardware infrastructure proxies like lidar manufacturers . The prominent humanoid startups are either still private like Figure or tucked away inside massive tech conglomerates

South Korea offers a fundamentally different dynamic. The local ecosystem features pure-play robotics companies that are already publicly traded, allowing investors to gain direct exposure through the open market.

At the center of this landscape is Doosan Robotics (KRX: 454910).

The company represents a compelling case study in structural business model transformation. While traditional financial frameworks often flag its consecutive years of net deficits, forward-looking market participants are focusing on its shift from an industrial hardware vendor to a high-margin “Physical AI” software platform.

The Family

Doosan Robotics is already a top-5 global player in collaborative robots (cobots). They have established manufacturing facilities, a global distribution footprint, and active commercial contracts.

In Q1 2026, the company reported a 189.7% year-over-year revenue surge. While operating margins remain negative due to aggressive R&D and expansion costs, the underlying volume is scaling rapidly, driven by U.S. commercial backlog and the integration of their newly acquired systems integrator, ONExia.

Right now, Doosan is undergoing a radical transition into software-defined automation. Their proprietary Dart-Suite ecosystem moves the business toward no-code, drag-and-drop AI task planning, drastically lowering integration costs for end-users. At CES 2026, their vision-based AI solutions earned top innovation awards, showcasing machines that adapt to unmapped environments in real time without pre-defined CAD data.

This software-defined foundation serves as their direct springboard into humanoids. Doosan Robotics has officially established a product roadmap targeting the commercial launch of industrial humanoid platforms by 2028. They are executing this strategy as a core member of South Korea’s state-backed K-Humanoid Alliance, while simultaneously collaborating with Nvidia to train their models via the Isaac Sim pipeline and co-developing custom on-device AI semiconductors.

Ultimately, Doosan offers a unique equity vehicle: a business with a proven hardware moat, zero debt, a 6-year cash runway, and an embedded call option on the physical AI revolution.

We will break down the mechanics of their balance sheet, the strategic importance of their U.S. footprint, the structural impact of their recent corporate restructuring, and how they stack up against global peers like Universal Robots and Fanuc.

To evaluate Doosan Robotics properly, it must be viewed through the lens of its corporate architecture. The company functions as the high-tech vanguard of the Doosan Group, one of South Korea’s oldest and most deeply capitalized chaebols (conglomerates).

Established in 1896, Doosan Group has pivoted across multiple industrial cycles over the last century. Today, the conglomerate generates between $13B and $15B in annual revenue, commanding structural market positions in heavy infrastructure, nuclear power components, clean energy, and semiconductor testing. When an industrial parent of this scale establishes a dedicated robotics division, the investment horizon is measured in decades, backed by significant corporate treasury and deep engineering synergies.

The structural origin of Doosan Robotics began in July 2015. Faced with accelerating demographic headwinds and labor shortages across industrial economies, particularly in East Asia, Doosan Group formed the subsidiary to capture the shifting economics of factory automation.

Cobots

Rather than entering the highly commoditized market for traditional, high-payload industrial robots that operate in isolated safety cages, Doosan focused its initial three-year R&D cycle entirely on collaborative robots. The primary technical objective was mastering force control; the company developed and standardized high-precision torque sensors across all six joints of its robotic arms, a feature typically omitted by lower-cost competitors.

Read more

You might have spent weeks digging through 10-Ks, reading research reports, modeling forward projections, and debating in the comments. You invested your time and your intellect into the idea. You entered the position, lived with it, married it - and then, out of nowhere, the market trend snaps you over its knee.

Right then, your brain slams the mental trap shut. Clicking “sell” means officially signing off on your own stupidity. How could this happen? The stock was trading at $100, top-tier research firms had a target of $155, and yet here you are, submissively bagholding for six months at $70.

The market is littered with these exact setups: SoFi, Robinhood, Lemonade, or a dozen other software names that retail investors camp out in for years, praying for a miracle.

When the Magic Dies: Adobe, Nike, and the Power of Perception

A textbook example playing out right now is Adobe. The company drops solid earnings. Every single quarter, they post record revenue, strong metrics, and a clean balance sheet. Yet, the market refuses to buy it. Why? Because the magic is gone. Investors no longer see a moat. Any generic chat-bot can spin up a flawless image in three seconds via a simple text prompt. You don’t even need to argue about the nuances of their enterprise business model or ARR stability. The only thing that matters is that the market’s perception of the company has fundamentally flipped. It went from being an unchallenged, innovative monopolist to a defensive incumbent with hungry startups camping on its lawn.

An even more brutal example from traditional retail is Nike. A legacy brand with a decades-old moat, or so everyone thought. But management decided to play distribution geniuses: they aggressively cut ties with wholesalers and traditional retail storefronts to force everything through their own website and app (the Direct-to-Consumer model).

The result? They voluntarily handed over premium shelf space to young, hungry competitors like On Running and Hoka. The brand lost its obsessive focus on innovation, the product lines got stale, and the wholesale channels were left shattered. The core investment thesis - ”a permanent premium brand that dictates market terms” - broke. Retail investors who blindly bought the dip simply because “it’s Nike, it always bounces” are now sitting on massive, bleeding losses because the underlying business fundamentally degraded.

Sometimes the setup is different: a company is doing fine operationally, but all the liquidity in the market is rotating into semiconductors and AI infrastructure. Capital is flooding into names like AOSL or UMC, stocks most retail investors hadn’t even heard of three years ago back when everyone was blindly allocation-stacking the Magnificent 7. Today, the landscape has shifted.

The market is always right. You have two choices: remain an stubborn passenger in a losing position to prove a point to your monitor, or cut the loss, free up the capital, and go make money where the actual trend is running.

The Investor’s Dilemma: Price vs. Thesis

We need to make a critical distinction here. If you are trading as a pure speculator using Mark Minervini’s playbook, your life is simple: the price breaks a technical stop-loss at 7% to 8%, you cut it, forget it, and hunt for the next chart setup. But if you position yourself as a long-term investor, blindly copying that approach will chop your account to pieces on random market volatility.

For a long-term investor, a 30% or 40% drop in share price is not, on its own, a reason to sell. The only legitimate reason to sell is when your original investment thesis breaks.

When you bought the stock, you shouldn’t have just chased a green chart. You should have had a specific playbook written down: “I am buying this company because they own a unique product, their margins are expanding by 5% annually, and their primary competitor is lagging.” When the stock takes a hit, look at that playbook, not your unrealized P&L:

1. Thesis Intact, Price Dropped: If margins are stable, the product is clearing out, management is executing, and the stock is getting hammered purely because the broader sector is in a macro down-cycle - do not sell. This isn’t a loss; it’s a clearance sale. This is exactly where elite long-term cost bases are built. Look at Micron: it cascaded from $150 down to $80, testing the resolve of every investor on the roster, right before the market realized that global AI progress is physically impossible without their HBM chips.

   Or look at how the media buried Google during the early OpenAI rollouts. The panic was so intense it looked as though the search giant was about to repeat Microsoft’s trajectory from 2010, when they completely slept on the mobile market and went stagnant for years. And make no mistake, that risk was entirely real. But instead of drifting into irrelevance, Google’s management shifted into emergency mobilization.

   They declared an internal "Code Red," frantically overhauled their AI models, integrated AI overviews straight into the core search layout, and aggressively scaled up Google Cloud infrastructure. The case didn't recover on its own, investors stepped back in only when they saw that leadership was ready to fight tooth and nail. Institutional funds mapped out the forward cash flows under this new reality and bought the bottom. The business survived because management executed a timely pivot and protected their cash-printing machine.

2. Thesis Broken, Price Dropped: But if margins are cratering, competitors have shipped a better, cheaper product, and management is stumbling through earnings calls without answers, your thesis is dead. The business has changed. Sitting there waiting for the stock to “get back to $100” just to save your pride is financial suicide.

Look at what is happening at the top of Nike following their disappointing guidance that sent shares into a tailspin. During an internal all-hands meeting, CEO Elliott Hill flatly told employees: “I’m so tired, and I know you are too, of talking about fixing this business. I want to move to inspiring and driving growth and having fun.” He openly admitted his strategy is lagging, taking “longer, way longer than I’d like.” Simultaneously, CFO Matthew Friend noted on calls that performance isn’t meeting expectations, results are mixed, and the trajectory for the business is “stepping down.”

When the C-suite of a mega-cap giant explicitly tells you the business trajectory is stepping down and the CEO is openly exhausted from trying to patch up broken internal processes, that is your exit siren. The magic is dead, the moat is breached. Holding through that hoping to break even is a delusion.

The Sleep Test and the 3-Month Window

The ultimate red flag that a position is toxic is your sleep quality. If you are waking up in the middle of the night to frantically refresh pre-market tickers, your position sizing is wrong or you are trapped in a bad trade.

An actual portfolio cannot consist exclusively of multi-baggers. If your brain is wired to demand that every single line item in your terminal stays green so you can compete with anonymous avatars on X, you are on a fast track to blowing up your capital. Your overall P&L needs to be green over time. Period.

So when do you pull the trigger without mercy? The primary indicator is a busted earnings report combined with the clear realization that for the next three months, no institutional money is coming to buy this stock.

That three-month window (one quarter) is critical. It’s the timeframe management gets to show the street how they plan to navigate the wreckage. Strong management teams weaponize this window to go on the offensive and reclaim market confidence:

• Credo responds to a vicious sell-off by instantly executing an acquisition of an Israeli tech startup to expand their hardware stack.

• Marvell snaps up Celestial assets to lock down their footprint in the AI data center pipeline.

  

• Robinhood ruthlessly fires their head of crypto after revenue drops

If you see structural moves, M&A execution, or a sharp pivot within those three months (like Nike scrambling to repair wholesale partnerships), you can give management room to work. But if the earnings report is a disaster, the thesis is broken, and executives are just shrugging and asking for patience, cut the position.

Admit the mistake, reclaim the cash, and pivot. The market serves up fresh opportunities every single day, but they are only accessible to players who keep their capital liquid and their heads completely free of illusions.

Actionable Blueprint: The Evacuation Protocol

If a specific ticker in your portfolio flashed through your mind while reading this article, stop hoping. Engage cold logic and run this three-step protocol right now:

• Step 1: The Clean Sheet Test. Ask yourself: “If I did not own this position today, but had the exact same amount of cash sitting in my account, would I buy this stock right now at current prices?” If the answer is an honest no, move to Step 2.

• Step 2: Scale out via Micro-Sizing. If cutting the entire block at once triggers too much emotional resistance, do it in tranches. Sell exactly 30% of the position today. The moment you hit the button and free up that first block of capital, your psychological bias resets. You realize the world didn’t end, and your brain regains objective clarity.

• Step 3: Immediate Re-allocation. Do not leave that cash sitting idle on your balance sheet, or you’ll be tempted to buy another falling knife. Immediately re-route that capital into a high-conviction, structural trend where margins are scaling, earnings are acceleration-primed, and institutional volume is actively buying. Make your money work for your P&L, not for your ego.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author holds positions in Robinhood, Micron, Credo and Google

In this post, we want to look past the hype, understand who is actually building these robots, and find the smartest ways to make money from this trend.

The Main Roadblock is the Math

To understand where the market is going, we have to look at the numbers. According to a recent report by McKinsey, the biggest roadblock right now is the cost of components. Building a single humanoid robot currently costs anywhere between $30,000 and $150,000. For this technology to become truly mass-market, the cost of the hardware needs to drop below $20,000.

Why are robots so expensive? The industry is in a “pre-modular” phase. There are simply no standardized, off-the-shelf parts available. Robotics companies are forced to do almost everything themselves through vertical integration, or custom-design unique joints and mechanisms for every single robot. Under these conditions, scaling production rapidly is physically impossible.

The Automotive Shortcut

However, robotics has a powerful accelerator: it can leverage the existing capacity and technologies of adjacent sectors. First and foremost, the electric vehicle (EV) and semiconductor industries.

To move and perform tasks, robots need the exact same batteries, sensors, chips, and motors as modern electric vehicles. This is precisely why China is moving faster than anyone else, they are building robots on the foundation of their highly developed EV manufacturing base.

Essentially, autonomous vehicle tech and robotics are merging into a single ecosystem:

• Infrastructure and Brains: Nvidia is building the foundation for the entire industry. Their Isaac and Jetson platforms provide a ready-made environment where robots can train in digital simulations before deploying into the real world.

• Robotic Vision: Autonomous driving tech is transitioning directly into robotics. Lidar manufacturer Ouster integrates its new Rev8 sensors with the NVIDIA Jetson and DRIVE platforms. This gives robots high-precision 3D vision and spatial awareness, technology originally built for self-driving cars.

• Robots at Work: While humanoids are still in testing, other companies are already making billions on automation today. This includes Teradyne (TER), which owns Universal Robots (the market leader in collaborative arms), and Symbotic (SYM), which completely automates massive warehouses for giants like Walmart.

There are several ways to play this trend: through software and sensor manufacturers (like Nvidia or Ouster), established industrial automation plays (Symbotic, Teradyne), or unexpected backdoors in the traditional automotive industry, where the future giants of robotics are hiding.

The Industrial Scaling of Boston Dynamics

Boston Dynamics remains the industry benchmark for mechanical engineering and physical agility. Kia leadership recently indicated that the company is targeting an IPO by 2028 with an estimated valuation starting at $20-30B. Instead of waiting for the public offering and paying an investment-banking premium, investors can build exposure today through Hyundai Motor, which controls a 30% stake in Boston Dynamics.

This proxy provides a unique entry point because Hyundai is actively solving the project economics that sink most robotics startups. The structural flaw in the humanoid business model is the chicken-and-egg problem of volume: without massive orders, component manufacturers cannot tool up for cheap mass production, keeping robot costs high.

Hyundai is bypassing this issue by acting as its own anchor customer. The conglomerate is establishing a dedicated manufacturing system in Savannah, Georgia, with the capacity to produce over 30,000 Atlas units annually by 2028. Rather than selling these unproven machines to third parties, Hyundai plans to deploy over 25,000 Atlas robots directly onto its own and Kia’s automotive assembly lines.

This captive deployment model guarantees the volume necessary to force component suppliers to lower their prices. It also provides Hyundai with millions of hours of proprietary operational data, allowing them to refine the control algorithms in a structured, real-world environment.

Operating a humanoid robot also requires massive burst-power to handle heavy lifting and rapid stabilization. To meet these demands, the electric Atlas utilizes custom-engineered 46-series cylindrical NCM battery cells supplied by LG Energy Solution. This partnership highlights the direct advantage of leveraging scaled automotive battery manufacturing for robotic applications.

Beyond the operational integration, Hyundai Motor has historically traded at a deep discount, often referred to as the Korea Discount, due to its complex circular holding structure.

However, taking Boston Dynamics public on the NASDAQ in 2027 will require a major simplification of Hyundai’s corporate governance. This restructuring is highly likely to unlock significant trapped value for Hyundai Motor shareholders.

Dogs and human(oid)s

Spot

To understand the commercial viability of this enterprise, investors must analyze what these machines actually do in the field. Boston Dynamics has deliberately split the market into two distinct commercial forms, each addressing a specific operational pain point. The first is Spot, the quadrupedal robot dog that has been in serial production since 2020.

Spot is not designed for complex manual manipulation; its primary function is autonomous industrial inspection and anomaly detection in high-risk environments. Equipped with thermal cameras, radiation sensors, and acoustic imagers, thousands of Spot units are already deployed to patrol offshore oil rigs, electrical substations, and deep-mining operations where human presence carries a high insurance and safety penalty.

Read more

These financial results belong to the sovereign ruler of computing infrastructure. NVIDIA is actively engineering its own demand, laying the foundation for a $4 trillion market, and systematically dismantling the business models of legacy competitors.

Capital Dominance: How to Engineer Your Own Buyers

NVIDIA is using its balance sheet as a first-strike financial weapon. In a single quarter, the company deployed $40 billion in direct equity investments into partner companies and specialized infrastructure funds. This capital went straight to the key players in the Neo-Clouds: Nebius, CoreWeave, and IREN

This is a ruthless strategic calculation with a triple-threat effect:

• Guaranteed supply absorption. The recipients of these investments are contractually obligated to build their data factories exclusively on the CUDA architecture. Huang has created a loyal pool of buyers, completely insulated from any attempts by Big Tech to transition to in-house silicon.

• Monopolizing energy grids. The ultimate bottleneck in the AI industry is access to electricity. Independent cloud providers use their long-term NVIDIA chip delivery contracts as high-grade collateral. Using this leverage, they secure billions in debt financing and lock down grid access commitments from global utility companies before anyone else can.

• A distribution channel for frontier developers. Pure-play AI clouds operate without the heavy legacy infrastructure of traditional web services. They deploy computing capacity in record time, making them the preferred partners for elite AI labs like Anthropic and Perplexity.

Breaking the Big Tech Monopoly

For a long time, NVIDIA’s primary vulnerability was customer concentration—relying on just four or five massive tech companies for nearly all of its revenue. The Q1 report officially marks the end of this risk. Now, datacenter revenues are split into two comparable segments: traditional Hyperscalers and ACIE (Alternative Clouds, Industrial, & Enterprise).

The revenue splits have achieved near-parity:

The growth of the ACIE segment ($37.38 billion, +31% QoQ) is impressive, but it must be viewed as a defensive buffer against the inevitable loss of traditional Hyperscaler volume ($37.87 billion, +12% QoQ). Big Tech is actively building a parallel supply chain:

• The TPU and ASIC Insurgency: Google is aggressively scaling its TPU v5p and v6e architectures. Amazon is deploying Trainium2 and Inferentia, while Microsoft scales its Maia silicon. These chips do not need to outperform NVIDIA on raw specs; they only need to offer superior price-to-performance for specific internal workloads. Every workload migrated to in-house silicon permanently erodes NVIDIA’s addressable market.

• The Blackstone-Google Alliance: Hyperscalers are bypassing NVIDIA’s land-and-energy grab through massive direct partnerships with infrastructure capital. Google’s collaborative data center play with Blackstone and other energy-infrastructure giants bypasses the Neo-Cloud proxy network entirely. By building massive, customized data center campuses optimized specifically for in-house TPUs, hyperscalers are securing gigawatt grid allocations

• The Software Decoupling (XLA and PyTorch): The real moat has always been CUDA. However, the industry is aligning around open-source compilers like Google’s XLA (Accelerated Linear Algebra) and PyTorch 2.0. These frameworks allow developers to compile models directly to custom ASICs with minimal code modification, steadily diluting the developer-lock that historically kept customers tied to NVIDIA hardware.

Evicting x86: The Vera CPU Conquest

NVIDIA is claiming the traditional server CPU market with its new Vera processor. Management officially confirmed visibility to roughly $20 billion in standalone sales for Vera this year alone.

The scale of this move is massive: the annualized sales of this single product immediately match AMD’s entire datacenter segment. With a total addressable market for server CPUs exceeding $200 billion, NVIDIA has immense room to expand. This is an architectural necessity for the era of Agentic AI.

But it comes with a price. NVIDIA’s Vera is built on the ARM architecture, but hyperscalers have been optimizing their own custom ARM CPUs for years. Google’s Axion, AWS’s Graviton4, and Microsoft’s Cobalt 100 are already deployed at scale. These custom CPUs handle general-purpose cloud workloads and node orchestration at a fraction of the cost of merchant silicon. Introducing a premium, third-party CPU like Vera faces direct internal resistance within hyperscaler datacenters

Intel and AMD still control the vast majority of legacy enterprise server infrastructure. Enterprise database migration from x86 to ARM is notoriously slow, costly, and complex. While Vera may succeed in pure-play AI environments, capturing general-purpose enterprise CPU workloads requires displacing entrenched legacy architectures with decade-long software support cycles.

Physical AI: The Brain for the Real World

While venture capitalists argue over the capabilities of text-based language models, NVIDIA is taking over the physical world. Physical AI and robotics have become the third core pillar of the company’s long-term strategy to embed the CUDA architecture into cars, factories, medical instruments, and drones.

When I talk about physical AI, and I talk about how the rest of the $100 trillion industry that has not been impacted by IT in the last 30 years, it’s about to be impacted by AI. That is the segment that I’m talking about. The second cluster is growing incredibly fast. Our share of that, of course, is very large. We’re fairly unique in our ability to be able to serve this industry. Our platform is built like it’s vertically integrated so that everything works. Jensen Huang

This segment transitioned out of research experiments long ago and is already generating over $9 billion in annualized revenue. The growth is fueled by long-term software integration contracts and massive infrastructure partnerships, including the deployment of Uber’s robotaxis across 30 cities by 2028. By controlling the simulation software (Omniverse) and the physical processors that run the machines, NVIDIA guarantees its leadership in the next wave of the technology cycle when the hype around generative software cools down.

The Loop

NVIDIA is generating an unprecedented $48.6 billion in free cash flow per quarter. These resources are being deployed to lock out any moves by potential competitors.

The company has locked down $145 billion in outstanding commitments, prepayments, and inventory. NVIDIA is buying up advanced packaging slots at TSMC and high-bandwidth memory (HBM3e/HBM4) supply capacity years in advance. Other players are physically blocked from scaling up alternative hardware production.

At the same time, proxy clouds like Nebius, CoreWeave, and IREN are leveraging their guaranteed chip allocations as collateral to reserve grid capacity from utility companies globally, totaling approximately 15 Gigawatts not counting 10 GW from Oracle

This is a flawless, self-sustaining loop: NVIDIA injects capital into partners ➔ partners leverage their hardware supply contracts to lock down scarce global energy and buy more chips ➔ the revenue loops back into the ACIE segment ➔ NVIDIA grows its free cash flow and pre-emptively buys out more factory capacity.

Investors are making a mistake by evaluating NVIDIA through legacy semiconductor metrics. Jensen Huang controls a $145 billion supply chain runway, directly distributes 25 gigawatts of scarce global electricity through proxy cloud networks, dictates new server architecture standards via the Vera CPU, and is capturing physical robotics.

Jensen Huang has built an aggressive, capital-recycled ecosystem to defend his market share. However, this strategy is balanced against a highly motivated hyperscaler cartel that is weaponizing custom silicon, forming direct energy alliances with global capital providers, and building open-source software bridges to bypass the CUDA moat. NVIDIA’s sovereign cloud cartel has secured a commanding lead, but the capital-intensive war for the future of global computing has only just begun.

Personally, I am gonna hold NVDA and will not sell the industry leader, even if it means missing out on gains from other stocks. In my view, these shares form the foundation of a core portfolio, and the earnings report once again confirms that the company is not standing still, but is actively shaping new markets around itself. Chasing maximum possible returns never ends well, so I hold common stock and purchase LEAPS on pullbacks (170–180) or breakouts (200+).

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author holds positions in NVidia, IREN and Nebius.

However, looking past the headline numbers into the actual financial mechanics of this quarter reveals a highly questionable setup.

The primary catalyst for the post-earnings optimism was a narrowed net loss of $23.7 million. But the quality of this beat is incredibly low. A quick look at the operating expenses shows that R&D spending was aggressively cut from $11.1 million down to $7.0 million, a 37% drop year-over-year.

Bloom Energy pours $186M a year!

For an unprofitable tech company trying to commercialize next-gen energy infrastructure, starvation of R&D is a structural red flag. Even worse is the specific allocation of these cuts. Management explicitly states the decrease was driven by “lower spending on commercial development efforts related to our solid oxide power generation and electrolysis platforms.”

Essentially, to print a cleaner quarterly loss statement for Wall Street, the company gutted funding and cut heads from the exact solid oxide and electrolysis programs they are currently hyping to secure data center contracts.

It’s a textbook corporate optics play: pumping the current stock price by cannibalizing the technology needed for long-term survival.

Deconstructing the AI Narrative Through Physics and the Balance Sheet

As the alternative energy sector undergoes a tectonic shift driven by the colossal power demands of AI data centers, investors are frantically searching for any proxy to play this trend. Capitalizing on this momentum, FuelCell Energy ($FCEL$) is aggressively attempting to market its own Blackwell-tailored narrative. In its May corporate presentation, management rolled out a 100 MW data center concept promising to solve the two biggest pain points for hyperscalers: the distributed baseload power deficit and server cooling via an integrated Absorption Chilling system.

Before we dive under the hood of their financial reports, let’s break down exactly what the company is promising the market and how their technology is supposed to work “on paper.”

AI Utopia

Officially, FuelCell Energy’s core business is the manufacturing and servicing of industrial-grade fuel cells. Unlike traditional diesel generators that combust fuel, fuel cells generate electricity via an electrochemical reaction. The system takes natural gas or biogas, mixes it with water and air, and outputs direct current (DC) electricity.

FuelCell Energy basically builds large-scale fuel cells that turn natural gas or biogas into electricity through a chemical reaction instead of burning it. To get a piece of the AI boom, they’ve recently rolled out modular 12.5 MW power blocks. The main idea is to set these up right next to data centers to give them a steady, uninterrupted power supply while using the leftover heat to run cooling systems for the servers.

It sounds like a great plug-and-play solution on paper, but in reality, these fuel cells operate at a constant speed and can't handle sudden spikes in computer workloads. Because of this, they still need to be paired with heavy battery setups and a constant, high-pressure natural gas supply to actually work.

For AI data centers, the company pitches a comprehensive “tri-generation” or co-generation layout built around three main pillars:

1. Continuous 24/7 Baseload Power: AI clusters packed with Nvidia Blackwell chips require stable, uninterrupted, ultra-high-density power. Wind and solar are fundamentally unsuited for this - they are too intermittent.

   FuelCell Energy promises to deliver megawatts of continuous power directly on-site, allowing developers to bypass the years-long waiting lists to connect to overextended regional grids.

2. “Free” Cooling (Absorption Chilling): The electrochemical reaction inside these fuel cells generates massive amounts of byproduct heat, with exhaust temperatures reaching roughly 370°C. FCEL engineers offer an elegant pitch: place specialized absorption chillers right next to the units.

   This setup captures the “waste” heat and uses a chemical cycle to convert it into chilled water to cool those exact same boiling server racks. Management promises a significantly improved Power Usage Effectiveness (PUE) and millions of dollars in utility savings.

Read more

Now, you’re sitting there doing the “regret math”, calculating how many cars you could have bought or how much your life would have changed if you had just put $10,000 into the trade. But staring at what could have been is the fastest way to lose what you currently have.

1. The Laggard Trap (Symmetrical Expectation)

The biggest pitfall for a smart analyst is the urge to buy the “underperformer.” When AMD hits $400 and photonics names are soaring in parallel, there’s a strong temptation to find a third-tier company that hasn’t moved yet.

• The Catch: In a powerful bull market, capital is hyper-selective. Leaders rise because they have the actual contracts, the chips, and the foundry allocations. Those lagging behind often lack access to these critical resources.

• The Outcome: When the leaders eventually cool off, these “laggards” are usually the first to collapse, often without ever having staged a meaningful rally. Never mistake a low-quality asset for a “discounted” entry.

2. The Hero Short (Fighting the Parabola)

When an asset covers the distance to a major psychological level in days, the analytical mind screams: “This is irrational! The multiples are insane.”

• The Catch: During parabolic moves, the market is driven by liquidity and forced liquidations (Short Squeezes), not fundamentals. Attempting to short such a trend is like standing in front of a high-speed train simply because your calculations suggest it’s overspeeding.

• The Outcome: Market irrationality almost always outlasts your margin. Overbought conditions are a statement of fact, not a timing signal for a trade.

3. Kill the Anchor: The Reality of the New Range

One of the hardest psychological hurdles is letting go of “yesterday’s price.” If Marvell was $90 last month, that price no longer exists in the current market reality.

• The Logic: A sharp movement often signals a regime change. You must live in the reality of the existing price range. While nothing is guaranteed - rates could rise or competitors could emerge, the chart has moved.

• The Approach: It is often safer to buy a stock on a breakout and move higher than to “catch a falling knife.” Buying at $200 on confirmed momentum is a better trade than waiting for $170 while the train leaves the station.

• The Tactical Play: If you can’t stomach the current price but want to participate, sell a Cash-Secured Put (CSP). It puts you in the game, allows you to get paid to wait for an entry, and keeps your skin in the game without chasing the peak.

4. The “First High-Quality Base” Strategy

The hardest part of missing a move is accepting that the current price offers no “mathematical” entry. Buying in the middle of a vertical move means you have no logical place for a stop-loss.

• The Approach: Instead of chasing, you must wait for the first consolidation phase, the “High-Quality Base.” This is where the stock stops going up, moves sideways, and allows moving averages (like the 10 or 20MA) to catch up.

• The Logic: Even if this new base is $10 higher than where you wanted to buy, it is a safer trade because it provides a clear level of invalidation. You aren’t trading the price; you are trading the structure.

5. The “Anti-FOMO” Micro-Position (Buying Objectivity)

Sometimes the psychological pressure of being “out” is so high that it clouds your judgment. This is where you use a tactical trick: the Information Position.

• The Tactic: Buy a tiny, almost symbolic amount of the asset (e.g., 5-10% of your usual size). Just enough to “be in the game,” but not enough to hurt your portfolio if it drops.

• The Logic: The moment you own a piece of the asset, your brain switches from “regret mode” to “management mode.” It satisfies the FOMO itch and allows you to watch the price action objectively. You can always average up or down once the structure clarifies.

6. Pivot: The Market Always Gives a Second Act

If you’ve missed the “Brain” (Chips) and the “Nervous System” (Photonics), stop living in the past. The market is a rotating machine. The smart move is to analyze the supply chain for the next logical beneficiary.

• The Case for Robotics: We are seeing the hardware cycle evolve. If the compute is ready, the “Body” is next. This is why I have started forming positions in companies like Hyundai (via Boston Dynamics), they are already producing the physical vessels for this AI.

• The Philosophy: Missing a profit is not the same as taking a loss. “Paper gains” you didn’t make don’t exist; only the capital currently in your account is real. Focus on protecting that capital and deploying it where the risk-reward is still in your favor, not where the party has already peaked.

Missing a trade is the cost of having a disciplined system. The real mistake isn’t being on the sidelines; it’s trying to force a logic onto a market that has already moved on.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author holds a positions in Hyundai at the moment of publication.

Since its IPO in 2016, the stock has essentially been flatlining, stuck on a multi-year plateau while the rest of the tech world moved through various cycles. That stagnation ended abruptly this year. But to understand why the market is suddenly paying attention, we first need to look at what they actually do.

The Tech: What makes Everspin different?

Everspin is the global leader in Magnetoresistive RAM (MRAM). Unlike traditional DRAM (which is fast but loses data when power is cut) or NAND Flash (which saves data but is slow and wears out), MRAM is a “universal memory.” It uses magnetic states to store data, making it non-volatile (it remembers without power), incredibly fast, and virtually indestructible in terms of write endurance.

The Use Case: Where is it actually used?
Currently, they dominate high-stakes niches where failure isn’t an option:

• Aerospace & Defense: Their chips are radiation-hardened, making them the standard for satellites and military hardware.

• Industrial Automation: Used in factory robots and power grids that need to save their “state” instantly during a power failure.

• Data Center Buffering: In high-end storage arrays, MRAM acts as a “write cache,” protecting critical data logs before they are moved to slower, permanent storage.

While the 2016-2024 era was about survival and perfecting the tech in these small niches, the 2026 breakout suggests the market is betting on MRAM finally moving into the AI mainstream, either as a replacement for SRAM in AI chips or as a critical component in the U.S. government’s push for domestic semiconductor sovereignty.

The Data Center Connection: Riding on IBM’s Shoulders

If Everspin has any footprint in the high-end data center market, it’s not because Amazon or Google are buying from them directly. Instead, Everspin acts as a sub-component supplier for storage giants, most notably IBM.

In IBM’s FlashSystem storage arrays, MRAM serves a very specific, unglamorous purpose: it’s a write cache.

• The Problem: Traditional DRAM is volatile. If a data center loses power, any data sitting in the DRAM cache is vaporized before it can be written to the SSD.

• The Old Solution: Massive, expensive, and unreliable battery backup modules (BBU) or supercapacitors that keep the DRAM alive just long enough to dump data to a disk.

• The MRAM Solution: IBM uses Everspin’s chips because they are non-volatile. You can pull the plug, and the data stays exactly where it is. No batteries required.

This is a legacy solution. While it’s technically “data center technology,” it’s focused on storage reliability, not the high-speed AI processing that is driving the current market mania.

Read more

The AI revolution has entered a phase where capital is no longer the primary constraint, physical infrastructure and strategic energy access are. NVIDIA is moving past being a mere component supplier to Alphabet and Amazon. By backing players like IREN, CoreWeave, and Nebius, Jensen Huang is effectively building a “Shadow Cloud” designed to strip power from Google Cloud and AWS.

This is a defensive and offensive masterstroke. While the “Big Three” attempt to bypass NVIDIA with proprietary chips (TPUs, Trainium), NVIDIA is counter-attacking by creating a rival ecosystem.

By funneling $2.1B into IREN, NVIDIA ensures that the most advanced Blackwell hardware stays within its own “DSX” architecture, building a private army of infrastructure providers that cannot be held hostage by the software preferences of the legacy hyperscalers.

The “Holy Trinity” and the 600,000 GPU Ambition

NVIDIA has established a global axis of power to commoditize the cloud:

1. CoreWeave: The North American specialist.

2. Nebius: The European AI-native cloud.

3. IREN: The massive-scale utility powerhouse.

The defining factor for IREN is its target scale: 600,000 GPUs. To put this in perspective, this capacity would place IREN among the largest AI compute clusters globally, rivaling the internal infrastructure of the hyperscalers themselves. For NVIDIA, having 600,000 of its high-end GPUs controlled by a partner with its own 5GW power pipeline is the ultimate insurance policy against Hyperscalers and other chipmakers.

For IREN, the move into Spain via the Ingenostrum acquisition is a geopolitical pivot. By seizing 490MW in Europe and 5GW of total grid-connected capacity, IREN sidesteps US power constraints. This “golden ticket” access to Blackwell chips allows IREN to offer AI training to European enterprises faster than American giants can clear local power grid delays.

Helping hand

The most critical takeaway from the recent IREN earnings call is the depth of their integration into the NVIDIA ecosystem. IREN has moved beyond the phase of complex preliminary negotiations; they are now in a position to selectively “pick and choose” which Tier-1 clients to serve.

As Co-CEO Dan Roberts noted, negotiations are significantly streamlined when a provider has a proven deployment record and a transparent timeline. This marks a massive strengthening of IREN’s engineering capabilities, a sector where they previously lagged behind incumbents like CoreWeave and NEBIUS.

Dan Roberts: “This, again, is part of the close working relationship we’ve got with NVIDIA. You know, we’ve spent a lot of the last fortnight in their San Jose office working through how we service all types of customers, all the way from the trillion-dollar hyperscalers through to the emerging AI scale-ups, where a lot of this innovation and development is taking place. It’s funny, speaking to someone the other day, you don’t need a sales team in this market, particularly when you’ve got NVIDIA. "

Closing The Gap

IREN has historically lagged behind AI-native players like CoreWeave and NEBIUS in engineering depth. However, the recent earnings call signaled a massive strengthening of these capabilities:

• The Engineering Gap: The acquisition of Mirantis and the deployment of the first GPU clusters indicate that IREN is closing the gap. As Co-CEO Dan Roberts noted, negotiations are significantly easier when you have a proven deployment record and a clear timeline.

• The Competition: While IREN’s recent $3.4B deal is a strong signal, it must be viewed against HUT’s recent $10B deal and Nebius’ deals with META. IREN is operating in a crowded environment where “land and power” are no longer enough; they must prove they can win against incumbents securing even larger contracts.

• Revenue Reality: The $33.6M in current AI revenue is a small fraction of the projected $3.7B ARR. The transition depends entirely on a flawless roll-out of the Blackwell clusters. Despite the headline $247.8M net loss in Q3 FY26, primarily a result of non-cash impairments on legacy mining hardware, IREN maintained an Adjusted EBITDA margin of 41%.

  This proves that the core infrastructure business remains profitable even amidst a capital-intensive structural transition. However, the primary interest for institutional investors lies in the comparison with pure-play AI-Cloud providers. Currently, CoreWeave operates at 56% EBITDA margin. This 15% (1,500 bps) gap represents the immediate upside:

  • Normalizing the Mix: The 41% margin is “blended,” weighed down by legacy mining. As AI revenue becomes the dominant share, the margin is structurally positioned to converge toward the 56% benchmark.

  • Vertical Leverage: By deploying hardware on proprietary infrastructure rather than leased space, and utilizing the Mirantis software layer, IREN can optimize unit-level OpEx. Closing this 15% gap on a multi-billion dollar revenue base is a massive lever for earnings growth.

Capex is the big concern

The lack of explicit CAPEX guidance for the full 600k GPU build-out is a primary concern for institutional investors.

• Capital Gap: Reaching the $3.7B ARR goal will likely require $3.5B–$4.2B in additional liquidity.

• ATM Program: The $6B ATM program creates a persistent “overhang” on the share price. We model a fully diluted share count expansion to 450M-500M shares, which significantly dilutes the “per-share” value of the future ARR.

Re-evaluating the “NVIDIA Signal”

NVIDIA’s option to invest at $70 per share should be viewed as a Strategic Validation Signal rather than a guaranteed floor:

• Warrants vs. Commitments: This is a “Godfather offer” designed to ensure IREN’s loyalty to the NVIDIA stack. It confirms that NVIDIA values the potential of IREN’s 5GW capacity.

• Risk/Reward: If IREN fails to deliver on its engineering milestones or thermal management of 600k GPUs, even NVIDIA’s endorsement won’t protect the stock from a significant re-rating.

Forward Valuation (FY 2027 Estimates)

We apply a Forward P/S multiplier to the $3.7B ARR, adjusted for the projected 450M fully diluted share count.

However, the path to the $70-80 range depends on IREN proving it can move beyond “land and power” to become a reliable operator without drowning its shareholders in new stock issuances. Investors must weigh the $70 strategic anchor against the massive execution risks and aggressive moves by competitors like HUT and Nebius.

We are no longer in a chip war; we are in a global energy and infrastructure land grab. IREN’s 5GW pipeline and 600,000 GPU ambition make it a valuable asset to NVIDIA’s “Shadow Cloud” strategy.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author holds positions in IREN, Google, Amazon, Nebius and NVIDIA at the moment of the publication.

This rapid doubling of capacity will not be felt evenly; instead, it is creating a stark geographical disparity. While constrained grids in the Mid-Atlantic and Midwest face rising utility costs and reliability crises, energy-rich markets like Texas and the Southeast are emerging as critical “safe harbors.”

For the market, this signals the end of stability and the beginning of deep volatility. As the grid strains under this massive infrastructure shift, power is ceasing to be a uniform commodity, it is becoming a scarce, location-specific asset where geography dictates survival.

Texas is emerging as a more stable environment for data center construction, particularly as the political climate intensifies during the midterm elections. The strategic distribution of sites and lower projected grid strain compared to other regions make infrastructure expansion in Texas less likely to become a focal point for political friction.

With gasoline prices climbing to $4.50 per gallon due to the ongoing conflict with Iran, electricity costs have become a critical factor for industrial operations. While this could be a point of tension, hyperscalers are expected to compensate for these costs to prevent any slowdown in construction.

Furthermore, companies are increasingly turning to alternative power solutions, such as on-site gas turbines, to ensure energy independence.

This shift is mirrored by industry leaders like who are prioritizing rapid data center expansion to meet demand, even if it means temporarily pivoting from previous green initiatives following recent earnings reports.

To manage public perception and avoid unnecessary scrutiny, there is a growing trend of anonymity in the sector. Crypto-miners-turned-infrastructure-providers are becoming increasingly secretive about their partnerships.

For instance, and have recently secured major deals while keeping the identity of their clients undisclosed. I compared their deals to understand why the HUT 8’s agreement is so much more favorable.

When you strip away the press release marketing and look at the engineering physics, the premium Hut 8 secured is driven by three technical factors: the definition of usable power, the cost of Nvidia-specific infrastructure, and grid location.

1. The Trap: “Gross” vs. “Critical” IT Load

This is the single largest distortion in the comparison.

• Cipher (CIFR): Their 300 MW figure is often cited as Total Site Capacity. In a traditional air-cooled facility (PUE ~1.25), a significant portion of that power runs fans and chillers. The actual power available for servers (IT Load) is likely closer to 240 MW.

• Hut 8 (HUT): Today’s announcement explicitly specifies 352 MW of IT Load. This is “pure” power available for the chips, exclusive of cooling overhead.

  

  Result: The “premium” shrinks from ~50% to ~21%. This is a rational market increase, but there are still two other drivers.

  2. The “Nvidia DSX” Factor (Capex Recovery)

  Hut 8 isn’t just leasing a shell; they are effectively financing a custom factory.

  • Cipher: Built a standard Tier-3 shell. It’s efficient, but generic.

  • Hut 8: The deal explicitly mentions a redesign for NVIDIA DSX specifications. This requires high-density Direct Liquid Cooling (DLC) loops and reinforced floor loading for heavier racks (100kW+ per rack).

  • The Lease Logic: In a Triple-Net Lease, higher upfront construction costs result in higher monthly rent. Hut 8 spent more to build a “Ferrari” garage, so the tenant pays a higher lease rate than Cipher charges for a “Ford” garage.

3. Grid Physics: Congestion vs. Connectivity

The value of a megawatt in Texas depends entirely on where it sits on the ERCOT map.

⚡️ The “Grid Congestion” Tax
Cipher’s Black Pearl sits in the “West Zone,” an area famously known as a renewable energy bottleneck.

• The Problem: The region generates more wind and solar power than the transmission lines can carry.

• The Consequence: When lines are clogged, ERCOT forces users and generators to shut down (Curtailment) or prices crash to negative levels.

• The AI Impact: Hyperscalers like AWS demand “five nines” (99.999%) reliability. They pay less rent in West Texas because they are effectively accepting a higher risk of grid instability and downtime.

🥶 The “Cooling Advantage”
Hut 8’s Beacon Point sits in Corpus Christi, a major industrial port.

• The Advantage: Access to industrial-scale water infrastructure is critical for the next generation of AI. Nvidia’s Blackwell and Rubin chips require liquid cooling loops that are incredibly water-intensive.

• The Premium: In the desert (Wink), water is a scarce luxury. On the coast (Corpus Christi), it is a utility. Hut 8 can support higher-density liquid cooling more cheaply and reliably than a desert facility, justifying the higher lease rate.

This imbalance has driven CIFR to conduct an experiment at a small site in Reveille, where they aim to become a co-owner of the compute—though the $700M on their balance sheet will admittedly not be enough for this. Nevertheless, we are seeing a clear differentiation among miners, and despite a 40% drop in U.S. data center construction, clients are demonstrating a willingness to pay a premium for quality.

IREN, for instance, was compelled to acquire Mirantis, although this move was obvious a year ago. Tomorrow, we will see the investors' reaction and find out when the next deal will take place.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author hold positions in IREN and CIFR

It is worth writing about not just because its stock has quadrupled. It is important as a pattern: demonstrates how a classic manufacturing base, accumulated since its 1998 IPO, can be transformed into a high-tech R&D hub under the right management.

If the AI boom teaches us anything, it is that the exponential development of technology constantly reveals new bottlenecks. The winners are those ready to clear these bottlenecks in real-time.

We previously covered Vicor, which solves the critical issue of chip power by converting current right at the foot of the processor. Now, we are moving further along this technological value chain. Celestica is the logical next step because it integrates disparate components, chips, power, and cooling, into a single, functioning system.

Network Breakthrough: The 400G -> 800G -> 1.6T Evolution

To understand the surge in Celestica shares, one must look at server ports in data centers rather than stock charts. Until 2023, Celestica was just one of many companies assembling standard network equipment. But as AI models began to grow exponentially, the network became the primary bottleneck for progress.

2023: The 400G Era and the Awakening

In 2023, the market realized that NVIDIA H100 chips are useless if they are connected by slow wires. A massive transition to the 400G standard (400 gigabits per second) began.

Why this helped CLS: At this exact moment, Celestica began aggressively offering its ODM solutions. Hyperscalers (primarily Google) realized it was more profitable to order designs directly from Celestica than to overpay for the Cisco brand. This provided the first powerful boost to revenue that we saw on the stock chart.

2024–2025: The Great 800G Transition

We are currently in a phase where 400G is no longer enough. Training models like Llama 3 or Gemini requires 800G.

The problem that emerged: At 800G speeds, ordinary signal transmission over copper cables begins to decay at a distance of a couple of meters. The signal simply falls apart. Celestica’s solution: The company brought its flagships to market (like the DS series seen in the photos).

They were the first to learn how to build systems where the signal remains clean even at such insane data density. The 69% growth in the Communications segment in the last quarter is a clear victory for 800G programs.

1.6T: The Physical Limit and New Architecture

Right now (late 2025 – early 2026), Celestica is launching 1.6T programs (1.6 terabits per second). Here, engineers have hit a wall that creates a moat around the company’s business.

The main problem - Thermal Apocalypse: At 1.6T speeds, switch chips (such as Broadcom Tomahawk 5/6) generate so much heat that air can no longer dissipate it. The server turns into a blast furnace. The technological response: Celestica developed the DS6001 model (seen in the screenshot).

This is a hybrid solution: CPO (Co-Packaged Optics): Optical fibers are brought close to the chip, replacing copper. Liquid Cooling: Liquid circulates inside the switch. Why this is important: Making such a piece of hardware reliable at a scale of thousands of units is a task that only 2 or 3 companies in the world can solve. Celestica is one of them.

The turning point came when Celestica's management realized the destructiveness of the "EMS trap", a business model where you compete with other factories only on who can solder someone else's chip to someone else's board for the lowest price. To break this cycle of penny-thin margins, the company made a strategic decision to transform into an ODM player by establishing the HPS (Hardware Platform Solutions) division.

Read more

The market reassessed the moment BTC hit its historical maximum and entered a phase of turbulence. While the stock climbed on the enthusiasm of the past year, many investors, myself included, fell into a classic trap: Anchoring Bias.

We often evaluate an asset’s potential by looking at its past peaks, subconsciously believing the price is “obligated” to return to those levels. But in investing, it is not where the price was that matters, but rather the factors driving it now.

From Favorites to Pariahs

In the market, there is a concept of a company changing its “state of matter.”

• Phase of Incredible Strength: This is what we are currently seeing in the semiconductor sector (AMD, ARM). The market forgives any degree of overvaluation because it believes in exponential growth.

• Phase of the Market Pariah: As soon as a business model shows a crack, the asset instantly loses its status. Today’s 8% drop in META is a prime example; the market wasn’t just looking for good numbers, it expected perfection. When a company ramps up CapEx on AI without an immediate return, investors slam on the brakes.

As Benjamin Graham famously said:

“In the short run, the market is a voting machine, but in the long run, it is a weighing machine.”

Today, the “scales” are weighing against Robinhood, and one must have the courage to recognize that.

Robinhood’s 15% revenue growth in Q1 2026 looks weak against the industry average of 21% and the company’s own past performance of 50%+. Total revenue reached $1.07 billion, falling short of the projected $1.13–$1.17 billion.

The primary weight on these numbers was the crypto revenue crash: cryptocurrency transaction revenue plummeted 47% year-over-year to $134 million as trading volumes dried up.

The company is now facing a classic Innovator’s Dilemma. The very features that built its empire - commission-free trading and crypto speculation - are now dragging HOOD down. Meanwhile, new initiatives like prediction markets haven’t gained enough traction to compensate for the decline in core speculative revenue.

The Fundamental Gap: Why HOOD is Not IBKR

If you strip away the hype and look at the raw numbers, you see a chasm in business quality. Robinhood’s “commission-free trading” model is a double-edged sword.

Structural Comparison: Robinhood (HOOD) vs. Interactive Brokers (IBKR)

The 8x income gap is explained by the quality of the audience. During drawdowns, retail players, Robinhood’s core base, are the first to exit the market. This is why a stock bought at $70 requires significantly more time to recover: it doesn’t just need to “grow,” it needs to reinvent its profitability in a high-interest-rate environment.

The company is not standing still. During the earnings call, Tenev announced a record-breaking April. They are also partnering with Susquehanna (SIG) to launch Rothera, a dedicated platform for prediction markets. We will return to this partnership later. However, Robinhood’s primary source of pride isn’t just innovation- it’s traditional banking products.

Instead of calling it “copying JP Morgan,” they label it as building a “Super-App”.

The Super-App Strategy by the Numbers:

• Retirement Accounts: Assets Under Custody (AUC) grew 90% YoY to $27.4 billion. This represents a fundamental shift from “speculative” capital to long-term assets.

• Gold Ecosystem: The number of subscribers increased by 36% to 4.34 million. Gold users hold 5 times more assets than regular users, creating a stable financial foundation.

• Lending: Gold Card balances surged 400% to $1.1 billion. Robinhood is effectively converting traders into long-term borrowers.

• Banking Status: 40% of active customers now use Direct Deposit. The platform is becoming a primary place for storing salaries rather than just an “app for stocks”.

Historical Parallels: Cisco and the 2000s Trap

History teaches us that great companies can go years without returning to their peaks.

• In March 2000, Cisco was the most valuable company in the world at $80 per share.

• It was a magnificent company that was literally “building the internet.”

• Despite Cisco’s business continuing to grow for decades, the stock only returned to those levels 20 years later.

The problem wasn’t that Cisco was a bad company. The problem was that in 2000, investors paid a price that priced in 100 years of future growth. The situation with the SaaS sector and Robinhood today shares these traits: the market overestimated their “bright future,” and now the hangover has arrived.

This is where Robinhood demonstrated an agility that few can match, revealing the Edge we are reporting on first. We spent a long time trying to understand why Robinhood specifically secured millions of American accounts under the Trump program.

The answer likely lies in the connection with Susquehanna (SIG), their partner for prediction markets. The head of Susquehanna, Jeff Yass, is a longtime associate of Trump and a major donor to the Republican Party. Furthermore, as a co-owner of TikTok, Yass was instrumental in helping the platform avoid a ban.

When you connect these dots, it becomes clear why Robinhood specifically received Trump’s backing. The $100 million invested in a platform for these users is a small price to pay, one that is poised to yield a massive, multi-fold return on investment.

The “Trump Accounts” program is a massive government-scale contract targeting 60 million potential users. To put this in perspective, this is more than double Robinhood’s current client base of 27.4 million.

Scale and Acquisition Challenges

Acquiring an audience of this magnitude through standard methods is nearly impossible for several reasons:

• Customer Acquisition Cost (CAC): In the fintech industry, acquiring one active client typically costs between $50 and $100. For 60 million people, this would require a marketing budget of $3–$6 billion, which is comparable to the company’s entire annual revenue.

• Administrative Leverage: This program grants Robinhood direct access to users through government mechanisms, effectively bypassing market competition with giants like Schwab or Fidelity.

• Infrastructural Challenge: Robinhood has allocated $100 million to develop the platform for this influx. In the context of acquiring millions of users, this is a bargain, as the CAC within this program drops to a record low of approximately $1.60.

The Susquehanna and Jeff Yass Connection

As noted, the key Edge here is political capital:

• Jeff Yass (Susquehanna) is more than just a partner for the Rothera platform. He is one of the largest Republican donors and has direct access to Trump.

• His success in lobbying for TikTok (where he holds a significant stake) proves his ability to navigate national legislation.

• For Robinhood, this partnership served as the “entry ticket” to a program that allows them to skip decades of organic growth

Economic Impact

If Robinhood successfully converts even 20–30% of these 60 million into active users:

1. Assets Under Custody (AUC): Even with an average account of $5,000 (lower than the current $11,200), it would add $60–$90 billion to client capital.

2. ARPU Arbitrage: At the current revenue per user of $157, this mass of users would allow the company to finally break through its revenue ceiling and justify a “Super-App” valuation.

This program is the strategic lever that transforms Robinhood from a “meme” app into a systemically important financial institution, tightly woven into the U.S. political elite.

Strategy: Opportunity Cost

But all of that is about future prospects; an investor is interested in a simple question: what do you do right now, when is up 10% and you are just sitting there waiting for crypto to break $80K?

An investor’s greatest mistake is turning from a trader into a “forced long-term investor” simply because a position went into the red.

George Soros once said:

“It’s not whether you’re right or wrong that’s important, but how much money you make when you’re right and how much you lose when you’re wrong.”

If your capital is frozen in a stock that will take years to recover, you are committing a crime against your own portfolio. That money could be working in sectors showing “incredible strength,” such as Big Tech or semiconductor, there are many opportunities in the market

To act like a professional, answer these three questions:

1. The “Clean Slate” Test: If you didn’t have a position in Robinhood today, would you buy it at the current price for the same amount? If the answer is “no,” you are holding it only because of an emotional attachment to your loss.

2. Margin/Leverage Assessment: If you are being squeezed by options or margin requirements, cut the position immediately. Time is working against you (Theta decay).

3. The Reallocation Horizon: Set a deadline (e.g., 3 months). If the earnings reports don’t show an explosive growth in ARPU by then, reallocate into alternative instruments.

Sometimes the best strategy is inaction. But that inaction must be a conscious calculation, not paralysis. Always remember: the market doesn’t know what price you bought at. It doesn’t care. Your task is not to “get your money back,” but to earn new profit.

If you believe in Vlad Tenev, who has saved the company once before, then the stock is worth holding without dwelling on the fact that it once traded at $150. This might even be a good opportunity to average down your position. However, you must realize that if crypto takes a downturn and investors shift their focus toward AI stocks, this could become a quarter of missed opportunities. Personally, I’m not selling yet; I’ll monitor the price action, as my cost basis allows me to stay in the game.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author holds a position in HOOD

The Stealth Architect of the AI Revolution

In today’s stock market, success is often measured by the quality of a CEO’s keynote or the slickness of a company’s website. By those standards, is a complete anomaly. Their website remains a relic of the early 2000s, and their investor presentations are stripped of every possible marketing fluff.

This lack of polish hides a financial powerhouse. The stock has surged 400%, reaching a 12 billion dollar market cap. Vicor is the ultimate “quiet genius”, a company that skips fancy designers because its primary clients, like NVIDIA, prioritize physics over fonts.

Behind this radical shift stands Dr. Patrizio Vinciarelli, a former CERN physicist who founded Vicor in 1981. For over four decades, Vinciarelli has operated more like a scientist than a traditional CEO, holding over 150 patents and maintaining a 44% stake in the company.

His “power component methodology” - treating energy modules like building blocks - is exactly what allowed Vicor to stay ahead of the industry's exponential demand for power density.

Vinciarelli’s commitment is also reflected in his “skin in the game.” As of early 2026, he remains the primary stakeholder, owning roughly 44% of the company. While he has recently executed planned share sales under a 10b5-1 plan, his massive direct holding, over 8.8 million shares, aligns his interests strictly with long-term shareholders. For an investor, this represents a rare “engineering fortress” where the leader is more concerned with the laws of physics than short-term earnings beats.

The Power Problem: Why Vicor Exists

Vicor solves the most critical bottleneck in modern computing: the “Power Wall.” Every high-end AI processor requires massive amounts of electricity, but this energy must be delivered at incredibly low voltages and with extreme precision.

To understand what Vicor does, think of a massive power plant trying to charge a smartphone directly. The mismatch is too high. Vicor acts as the master translator. They take high-voltage electricity and step it down through a unique two-stage process.

First, they stabilize the voltage, and then they multiply the current at the exact moment it enters the chip. This specialization has turned them into a vital partner for anyone building the brains of the AI age.

The Vertical Shift: Why the Moment is Now

For decades, the industry lived in a “Lateral” world. Companies placed power modules around the processor and pushed electricity across the surface of the motherboard. This worked fine when chips were less hungry. But as AI chips began consuming over 1000 watts, the laws of physics intervened. Pushing that much current across a board creates massive heat and wastes a huge portion of the energy.

Vicor’s moment has arrived because they pioneered Vertical Power Delivery. Instead of forcing power to travel across a crowded motherboard, which creates heat and takes up valuable space, Vicor figured out how to deliver power from directly underneath the chip. It was a radical and expensive bet that they made years ago.

Now that AI chips have become power-hungry monsters, the world has realized that Vicor is the only one with a solution that prevents these systems from melting down.

Mapping the Invisible Infrastructure

Vicor is a vital piece of a much larger and more complex puzzle: the global semiconductor supply chain. To understand the AI revolution, we must look beyond the processors themselves and analyze the “infrastructure of the invisible.”

This deep dive into Vicor is the second step in an investigation of the semiconductor ecosystem after . We are identifying the high-potential companies that hold the keys to the future of computing.

Vicor has proven that in the age of AI, mastery of physics creates a powerful monopoly. This is only one of the fascinating companies with significant potential that we will explore as we break down the chain.

The Infrastructure Hunt: Why Vicor is the “Invisible” Beneficiary of the AI Boom

When hyperscalers like Microsoft, Google, and Meta announced CapEx in the hundreds of billions, the market initially flocked to chipmakers. However, investors eventually realized that for NVIDIA’s Blackwell architecture to even function, it required an infrastructure that simply didn’t exist before.

Read more

Today, shares plunged 17%, retreating to levels not seen since late 2022. On the surface, the headline numbers looked solid: a beat on top and bottom lines, raised guidance, and a narrative centered on AI. But the market isn’t buying the “growth at any cost” story anymore. Here is why this crash is a wake-up call for the entire software sector.

The Illusion of Acceleration: A Currency Mirage

ServiceNow reported subscription revenue growth of 22% for Q1 2026. While this looks like an acceleration from the 21% seen in the previous quarter, the investor presentation reveals a different reality.

The company benefited from a massive 300 basis point (3%) currency tailwind. Without this FX gift, organic growth in constant currency actually slowed to 19%. For a company priced for “forever growth,” slipping below the 20% mark, once you strip away the exchange rate fluctuations, is a psychological breaking point for institutional investors.

Buying Growth: The $7.75 Billion “Armis” Tax

The most jarring detail was the acquisition of Armis for $7.75 billion in cash. This is a defensive move of massive proportions. Management admitted that Armis (and new AI capabilities) will contribute roughly 125 basis points to the full-year growth.

Essentially, ServiceNow is “buying” its growth numbers to keep them above 20%. But this growth comes at a steep price:

• Operating Margin: Integration is expected to drag margins down by 75 bps.

• Free Cash Flow: FCF margin is hit by 200 bps.

• Q2 Outlook: Operating margins are projected to drop to 26.5%, a significant decline year-over-year.

The “Devil in the Details”: The SBC Racket

While a forward non-GAAP P/E of 24x after 70x+ might seem to align with industry standards, the devil, as they say, is in the details. The company spent $570 million on Share-Based Compensation (SBC) this quarter alone. Essentially, every sixth dollar earned goes toward employee bonuses. That’s roughly $2 billion a year.

It is important to realize that if these metrics were not constantly sanitized via non-GAAP adjustments, adjustments that exclude these massive SBC costs, the valuation multiples would not look reasonable. They would still look exceptionally bloated.

And when the company touts “share buybacks,” they aren’t returning value to shareholders, they are merely reducing the number of shares in circulation because they had to print new shares for SBC.

In the past, when revenue grew at 25%+, investors turned a blind eye to this dilution. Those days are over. Companies will now have to slash SBC or conduct massive layoffs, as Oracle has already done. This “shareholder tax” must end, or be paid in actual cash, which would further tank FCF.

The RPO Paradox

One “bright” spot was the Remaining Performance Obligations (RPO), which reached $27.7 billion. However, even this was distorted by 200 bps of currency tailwinds. Furthermore, the company admitted to a 75 bps headwind from deal delays in the Middle East. When a giant like ServiceNow starts citing geopolitical “deal delays,” it usually suggests that the sales cycle for expensive SaaS is lengthening across the board.

The New Reality: Multiple Compression and the IBM Warning

We are witnessing a structural shift. Analysts are no longer slapping 50x multiples on SaaS companies just because they grow at 20%. The emergence of OpenAI and Anthropic - who are already reporting revenues that rival established leaders - has changed the game.

If you want to see the terminal stage of this lack of focus, look no further than IBM. It stands as the most egregious example of corporate inefficiency in the sector. Growing at a pathetic 4-5% a year, is a “zombie” weighed down by a bloated consulting arm that acts as a structural anchor. They have failed to define a clear “moat” in the AI era, and their existence serves only as a cautionary tale: this is what happens when a tech leader stops innovating and starts merely “managing decline.”

Unless companies like ServiceNow undergo radical transformation, cutting bloated SBC and trimming the fat, they risk following the same path.

Bottom-fishing in SaaS remains a low-probability trade. The threat from LLM-native competitors is far from exhausted, and I prefer sectors with superior structural clarity. For now, I’m staying on the sidelines

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author may hold positions in the securities mentioned.

But does this strategy actually move the needle?

To put this into perspective: according to the latest filings, is burdened with $165 billion in total long-term liabilities. Even a $10 billion cost-cutting program fails to address even 10% of that debt mountain.

The company has maintained its characteristic lack of transparency, declining to comment on the restructuring. However, TD Cowen analysts suggest the primary target is the Healthcare division, a unit Oracle acquired for $28 billion that has yet to deliver, as major clients reportedly migrate to competitors like Epic.

With the stock having surrendered nearly half of its peak market capitalization, we need to determine if Oracle is a deep-value play or if the market offers better risk-adjusted alternatives. We won’t bore you with a summary of Oracle’s business model; instead, we are diving straight into the balance sheet of the most leveraged infrastructure play in the sector.

Growth at Any Cost

Oracle’s Q3 2026 earnings print presents a masterclass in financial engineering and narrative distraction. On the surface, the numbers look robust: total revenue climbed 22% YoY to $17 billion. Dig deeper, and the cracks in the legacy model become craters. Half of that top line, $8.9 billion, is now categorized as Cloud Revenue, up 44%. However, investors must realize this isn’t organic expansion; it is growth aggressively subsidized by a ballooning debt load.

The engine behind these numbers is the massive build-out of data centers saturated with GPUs. This segment is exploding at an 84% clip, yet it remains the most capital-intensive part of the business. This is precisely why Oracle is burning through cash and leveraging its balance sheet to the hilt. While cloud infrastructure specifically for databases - Exadata and Autonomous Database - posted a respectable 35% gain, it is increasingly overshadowed by the “dumb pipe” hardware business.

This aggressive pivot is reflected in the staggering 66% surge in Cloud and Software operating expenses, which hit $4.7 billion. This is the price of admission for the AI era: Oracle is front-loading massive costs and accepting short-term margin pressure to ensure it doesn’t get left behind. It is a ruthless reallocation of capital, funded by a “clearing of the decks” that includes 30,000 layoffs, all to finance the infrastructure required to stay relevant.

On the software side, Cloud Applications (SaaS) are growing in the mid-teens. While this outpaces peers like , it is fundamentally insufficient to sustain a premium multiple in an AI-driven market. Oracle’s reporting remains intentionally convoluted, designed to highlight growth silos while masking the decay of the core.

The sobering reality is that Oracle’s entire software portfolio - including databases, legacy applications, and support - eked out a meager 3% increase. In constant currency, that’s effectively stagnant. Wall Street, currently gripped by a “SaaS Apocalypse,” is dumping Oracle alongside the likes of and CRM. This is a massive failure of Investor Relations. The company has failed to articulate the reality that it is no longer a software firm, it is undergoing a desperate, high-stakes structural pivot.

The transformation is as dramatic as it is existential: Oracle is being forced to abandon its software roots to survive an era where Anthropic and other AI pioneers are making traditional enterprise code obsolete. They aren’t just changing; they are fleeing a dying industry.

David vs. Goliath: A Fight Oracle Can’t Outspend

The shift in Oracle’s financial engineering is nothing short of drastic. For the past decade, the company was an ATM for its shareholders, deploying $123 billion toward share buybacks. Today, that trend has inverted. To fuel its massive AI infrastructure ambitions, Oracle has pivoted to a $20 billion ATM equity offering, a move that directly dilutes shareholders and pressures future EPS.

The scale of this ‘all-in’ bet becomes clear when compared to the industry titans. While Google Cloud’s quarterly revenue of $17.66 billion matches the entirety of Oracle’s top line, the financial breathing room between the two is vast. As shown in the data, Google maintains a disciplined 22.7% CapEx-to-revenue ratio, funding its expansion through its massive operational cash engine. In contrast, Oracle is operating at a breakneck 75% ratio.

By spending three-quarters of its total revenue on Capital Expenditures, Oracle has effectively traded its status as a reliable dividend-and-buyback play for a high-stakes seat at the AI table. It is competing with Big Tech resources it simply doesn’t have on the balance sheet, forcing a total overhaul of its fiscal policy to keep pace in a race where the entry fee has never been higher.

For the first time in 30 years, Oracle has plunged into negative Free Cash Flow territory. With a TTM negative FCF of $24.74 billion - equivalent to roughly two full years of the company’s net income - the market’s anxiety is becoming visible. Oracle’s 5-year Credit-Default Swap (CDS) recently spiked to 125 basis points, its highest level in three years, signaling that the cost of insuring Oracle’s debt against default is rising alongside its ambitions.

Read more

At the center of this stack sits Applied Materials ($AMAT).

Applied does not design chips. It does not run foundries. And it does not sell GPUs.
Instead, it builds the machines that make modern semiconductors physically possible.

Every advanced chip requires hundreds of process steps. Thin films must be deposited atom by atom. Materials must be etched with angstrom-level precision. Structures only a few nanometers wide must be inspected, repaired and optimized before the wafer ever leaves the fab.

That entire process is driven by semiconductor equipment.

This is where Applied Materials operates.

The company is one of the largest process equipment suppliers in the world, alongside ASML and Lam Research. But while ASML dominates lithography, Applied sits at the center of materials engineering, which has quietly become one of the most critical bottlenecks in modern chip manufacturing.

As transistor architectures become more complex and packaging technologies evolve, the number of materials steps continues to grow. New transistor designs like gate-all-around, advanced DRAM structures used in HBM, and increasingly complex chiplet packaging all require new deposition, etch and materials engineering solutions.

That means the economics of AI are not driven only by chip designers.

They are increasingly driven by the companies that enable chips to exist in the first place.

In this deep dive we will break down:

• how the semiconductor manufacturing stack actually works
• why Applied Materials sits in a structurally advantaged position inside that stack
• how AI data center demand is reshaping wafer production and memory architectures
• and why the industry could reach $1 trillion in semiconductor revenue years earlier than expected

Supply Chain

To understand the AI boom, it is not enough to look at GPUs.

You have to understand the machines that build the chips.

The semiconductor industry is often described as a complex ecosystem, but the basic structure is actually straightforward.

Everything begins with raw materials. Ultra-pure silicon is processed into silicon wafers. These are the circular substrates on which chips are built. Along with wafers, the process requires specialty chemicals, industrial gases and photoresists.

Those wafers then move into the semiconductor fabrication stage, where chips are physically manufactured. This is where foundries like TSMC, Samsung and Intel operate massive fabrication plants.

But these factories cannot function without extremely specialized tools.

That is where companies like Applied Materials, ASML and Lam Research enter the picture.

These firms produce the equipment that actually builds the chip layer by layer on the wafer.

• ASML provides lithography systems that project chip patterns onto the wafer using extreme ultraviolet light.
• Lam Research provides etching tools that carve microscopic structures into the silicon.
• Applied Materials provides deposition and materials engineering tools that add the thin films and structures that ultimately form transistors and interconnects.

In other words, before a single NVIDIA or AMD chip exists, an enormous amount of capital equipment must first build the structure of that chip atom by atom.

Once the wafers are processed, they move to advanced packaging and testing, where individual chips are cut out of the wafer and assembled into final products.

Only then companies like NVIDIA, AMD or Broadcom integrate them into GPUs, accelerators and networking chips that power AI infrastructure.

From an investor perspective, companies like Applied Materials sit in a critical position in this chain.

Read more

In this case, IREN also structured GPU purchases on a post-shipment payment basis, meaning the company does not need to fund the entire hardware order upfront. From a financial standpoint that matters a lot. In the current AI infrastructure market, GPUs, networking equipment and even certain memory components remain constrained. In that kind of environment, securing supply early can be more important than waiting for perfect financing conditions.

Another argument circulating online is that IREN is buying “outdated GPUs.” That claim seems exaggerated. The company is ordering Blackwell Ultra (B300) systems, which are expected to remain one of the main AI training platforms through at least the 2026–2027 cycle. Even hyperscalers are deploying large Blackwell clusters today. Rubin will represent the next architectural step, but it will require different infrastructure, significantly higher rack power density and a new generation of systems. In other words, the current deployment looks much more like a timing decision than a technological mistake.

That said, one of the more nuanced criticisms deserves attention. A significant portion of the new GPU fleet appears to be deployed in existing air-cooled data centers, including facilities in British Columbia and parts of Texas. Air cooling has clear advantages from a capital perspective. Building liquid-cooled hyperscale facilities is extremely expensive, and estimates suggest that using air-cooled infrastructure can save roughly $5–7 million per megawatt of infrastructure cost. When companies scale hundreds of megawatts of compute capacity, those savings become meaningful.

The trade-off is rack density. Air-cooled racks support significantly less power per rack than modern liquid-cooled AI clusters. That means fewer GPUs can be installed per rack, and over time this may result in less efficient infrastructure compared with extremely dense liquid-cooled deployments. If the industry shifts toward very high-density clusters for next-generation training workloads, some air-cooled sites could become less optimal. In other words, this strategy prioritizes speed and capital efficiency today, but potentially sacrifices some long-term density advantages.

The Microsoft contract provides another important reference point. IREN signed a $9.7 billion multi-year AI cloud agreement expected to generate approximately $1.9 billion in annual run-rate revenue once fully deployed. The agreement also includes a 20% prepayment, which effectively provides meaningful upfront capital to support the infrastructure build-out. At the project level, the economics appear extremely strong, reflecting the pricing power that can exist when AI compute supply is constrained.

However, it is important not to misinterpret those economics. Strong project-level margins do not automatically translate into extremely high margins for the entire company. Once GPU depreciation, infrastructure costs, financing expenses and corporate overhead are included, company-level margins will inevitably look much lower.

This is where comparisons with other AI infrastructure companies become useful. Michael Intrator from CoreWeave has indicated that once their data centers stabilize and reach full utilization, platform margins could settle somewhere in the mid-20% range. That type of margin profile is actually typical for capital-intensive infrastructure businesses. It also illustrates how very strong unit economics at the GPU cluster level can translate into more conventional margins once the entire platform cost structure is considered. And let’s take into account that IREN is vertically integrated company with lower costs.

Another interesting signal about demand came recently from NBIS, whose management suggested that demand for AI compute is strong enough that providers can raise prices significantly and still receive full prepayment from customers. If that statement reflects broader market conditions rather than a single company’s experience, it reinforces the idea that the real bottleneck in the industry may not be software or models, but rather access to GPUs, power capacity and operational infrastructure.

One aspect of the IREN story that deserves more attention is the company’s infrastructure architecture. The Microsoft deployment is clearly described as being built in liquid-cooled data centers at the Childress campus, delivering approximately 200 MW of IT load across Horizon 1–4. This infrastructure is designed specifically for hyperscale deployments and supports extremely high rack densities required for large AI training clusters.

By contrast, the newly announced GPU expansion appears to target existing air-cooled facilities in British Columbia and parts of Texas. These deployments are architecturally different from the hyperscale Microsoft cluster.

Taken together, this suggests that IREN may be building two distinct layers of infrastructure.

The first layer is hyperscale infrastructure, built around liquid-cooled data centers designed for very large long-term contracts with hyperscale partners.

The second layer appears to be IREN’s own AI cloud capacity, which can be deployed more quickly using existing air-cooled data centers. This infrastructure allows the company to scale compute capacity faster and at lower capital cost while serving a broader range of AI workloads and customers.

In practice this creates a two-tier model.
Liquid-cooled clusters support hyperscale training deployments and large strategic contracts, while air-cooled capacity allows the company to expand its own AI cloud platform more quickly.

From a valuation perspective, the company recently increased its long-term AI Cloud ARR outlook by roughly $300 million. How much market value that additional revenue translates into depends heavily on how investors classify the business.

If the company is valued primarily as infrastructure similar to traditional data center operators, multiples could fall in the 6–8× ARR range, implying roughly $1.8–$2.4 billion of incremental value. If investors begin to view the platform more like a high-growth AI infrastructure provider, multiples could reach 12–15× ARR, implying perhaps $3.6–$4.5 billion of additional valuation.

In that context, the additional revenue potential could theoretically offset a meaningful portion of the dilution over time.

There is also a longer-term technology risk that cannot be ignored. If the industry rapidly transitions to significantly more efficient GPU architectures with dramatically higher rack densities, some Blackwell-based infrastructure built today could become relatively less competitive. That does not necessarily make the current investment obsolete, but it is a factor investors will need to watch closely over the next several years.

Ultimately, IREN appears to be attempting something ambitious: transitioning from a smaller operator into one of the larger independent AI compute platforms. If the company successfully deploys the planned GPU fleet and demand for AI compute remains strong, the resulting platform could become a meaningful player outside the hyperscalers.

At the same time, the strategy clearly carries risks. Heavy capital requirements, dilution, infrastructure design trade-offs and the need to secure additional large customers will all shape the outcome.

Which brings us back to the question that still hangs over the entire story.

If demand for AI compute is as strong as many industry participants suggest, the next thing investors will want to see is additional large customer contracts that confirm the newly built capacity will be fully utilized.

Until those contracts appear, the debate around the company will likely continue.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author may hold positions in the securities mentioned.

NVIDIA has officially solidified its transition toward a vertical trust model, utilizing CoreWeave as a vehicle for direct expansion into the cloud sector. This $2B investment, relative to CoreWeave’s annual revenue of $6B, appears to be an attempt to stabilize the balance sheet of the industry’s most highly leveraged player and restore creditor confidence ahead of a massive CapEx cycle. However, the underlying financial math is far more ruthless. When considering the ambition to construct AI factories with a total capacity of 5 GW, at a current cost of $40 billion per 1 GW, the total project budget reaches $200–250 billion. Of this amount, approximately 60% ($120–150 billion) is attributed to NVIDIA’s own chips and networking architecture.

Building 1GW costs roughly $15B, meaning $75B for the entire build-out,a figure that is much more grounded and aligns with the investment Huang promised OpenAI but has yet to deliver. Outfitting these factories with chips is a matter of both financial and standard engineering, which is certainly feasible. The task seems entirely manageable, especially since building from scratch isn’t the only option. They could acquire IREN/Cipher/HUT8, which already have sites ready to go, or they could lease capacity and pass the costs onto the consumer through service fees. With $500B in revenue, there are plenty of ways to make it work. Meanwhile, under Intrator, CoreWeave has lost all independence and is forced to look for a buyer unless this entire play was designed from the start as a blunt-force market grab at any cost.

By launching Vera, NVIDIA erases this barrier. The Vera CPU and the Rubin GPU now reside on a unified NVLink 6 bus with bandwidth that standard server solutions cannot touch. The investment in CoreWeave creates a proving ground where this synergy operates in a vacuum. NVIDIA needs to prove that their “closed stack” is superior, and doing so through a controlled provider is the fastest path to market dominance.

A Direct Counter to Microsoft’s Silicon Ambitions

The debut of Vera was timed specifically to blunt the impact of the Microsoft Maia 200. Microsoft, Google, and Amazon have been aggressively developing internal silicon to de-risk their reliance on NVIDIA and relegate them to a mere component supplier. If NVIDIA had remained only in the GPU segment, it risked becoming an “appendage” in Big Tech clouds that control the central processor and system logic.

Vera changes the rules: NVIDIA now dictates the entire server configuration. Through CoreWeave, they are rolling out a “factory standard” that excises Intel and AMD entirely. This puts Microsoft in a defensive position. While the Maia 200 is an impressive step forward featuring custom silicon optimized for high-bandwidth sub-network processing and improved FP8 efficiency for inference, it remains an attempt to optimize a specific part of the chain within a legacy cloud framework. In contrast, NVIDIA is building a purpose-built environment where the training and execution of Physical AI are seamless. Huang is telling the market: “If you want to build the future of robotics, you don’t need a general-purpose cloud; you need our specialized factories.”

The Economics of Control and Competitor Displacement

The financial logic is even more pragmatic. In the construction of a 1GW facility, a significant portion of the margin previously leaked to server platform and CPU manufacturers. With Vera, NVIDIA captures that entire pie. In the bill of materials for a CoreWeave data center, almost every line item is now an NVIDIA product.

This allows for maximum financial engineering. NVIDIA can subsidize Vera shipments to freeze competitors out of the data center or offer bundled long-term lease packages that are economically impossible for anyone assembling servers from disparate vendors. Ultimately, Vera is the lock on the door of these 5GW facilities: entering with third-party hardware will be technically suboptimal, leaving NVIDIA’s integrated stack as the only viable path for next-generation compute.

The Microsoft Strategy: Defensive Optimization

Microsoft is using the Maia 200 to survive the “token war.” By building their own silicon, they avoid the “NVIDIA tax” for their internal services (Bing, Office 365, GPT-5.2 inference). However, the Maia 200 remains a discrete component. It still has to communicate with a host CPU, creating a latency floor that makes it suboptimal for the massive, real-time physical simulations required by the next generation of robotics. Microsoft is essentially building a “cheaper typewriter” for the AI era.

The NVIDIA Strategy: Offensive VerticalizationThe Vera-Rubin stack, deployed via the 5GW CoreWeave build-out, is designed to make general-purpose clouds like Azure obsolete for high-end developers. By integrating the Vera CPU directly onto the same bus as the Rubin GPU, NVIDIA has eliminated the “PCIe bottleneck.”

• Market Leverage: In a 5GW “Factory,” NVIDIA can process simulations 5-10x faster than a Maia-based cluster.

• Financial Engineering: Because NVIDIA owns the entire bill of materials (CPU, GPU, Networking), they can manipulate the “rental price” in CoreWeave to underprice Microsoft’s Azure offerings while maintaining higher net margins.

The 5GW Reality CheckFor a founder building Physical AI, the higher hourly cost of a Vera-Rubin instance is irrelevant compared to the time-to-market. If NVIDIA’s stack can train a humanoid robot’s motor skills in 2 weeks while Microsoft’s Maia cluster takes 8 weeks due to CPU-GPU latency, the “expensive” NVIDIA chip is actually the cheaper option. This is the “Industrial Trap” NVIDIA is setting: they are building the only infrastructure capable of handling the sheer scale of 2026-era Physical AI.

The $20 billion acquisition of Groq, finalized in late 2025, serves as the cornerstone of the 5GW architecture. While the Vera CPU was engineered to eliminate latency between the processor and the GPU, Groq’s technology is the missing link required to drive Physical AI with the speed of a human reflex.

Groq LPU: Why It Is Mission-Critical for Robotics

Traditional GPUs, including the latest Rubin architecture, excel at parallel computing but struggle during the inference stage due to the “Memory Wall.” To execute a single movement or generate a specific token, a GPU must constantly access external HBM (High Bandwidth Memory). This creates micro-delays—jitter—that are negligible in text generation but fatal for an 80kg humanoid robot operating in the physical world.

Groq’s LPU (Language Processing Unit) architecture, built on SRAM, solves this problem fundamentally through:

• Determinism: Groq chips lack dynamic scheduling. The compiler knows exactly which nanosecond data will pass through a specific transistor. This is “hardware with a predictable future.”

• Speed: While a standard GPU might output 100–200 tokens per second, the Groq LPU hits 1600+. In the context of Physical AI, this means a robot can “think through” its actions ten times faster than they occur in reality.

The Vera-Rubin-Groq Synergy: Birth of “Instinctive AI”

NVIDIA’s decision to acquire Groq—its largest deal since Mellanox—was driven by the need to integrate deterministic inference directly into the Vera-Rubin stack. Within a single rack of a 5GW CoreWeave data center, the labor is now divided to mimic biological systems:

1. Rubin GPU: Handles the heavy lifting—foundational training and massive physical simulations in Omniverse.

2. Vera CPU: Manages high-level logic and path planning.

3. Groq LPU (Now NVIDIA IP): Acts as the “reflex arc.”

This tri-part architecture transforms the robot from a lagging machine into an agent with biological reaction speeds. By swallowing Groq, NVIDIA neutralized the only competitor capable of offering Microsoft or Google an architecture for ultra-fast inference. If you want a robot to react to a moving object in less than 100ms, you are now physically tethered to the NVIDIA stack.

Market Capture: From “Smart Words” to “Smart Movements”

The Groq acquisition is Jensen Huang’s admission that the era of the “general-purpose GPU” is over. Physical AI demands specialized, real-time inference.

• Microsoft, with the Maia 200, is playing a defensive game, optimizing for standard cloud efficiency and text-based workloads.

• NVIDIA, by absorbing Groq, is building Real-Time AI.

In the context of the 5GW build-out with CoreWeave, these data centers are evolving into the “central nervous system” for millions of autonomous agents. NVIDIA has effectively “seeded” its future factories with technology that renders any third-party silicon—including Microsoft’s Maia or Google’s TPU—obsolete for real-world physical applications. NVIDIA has moved beyond being a chip vendor; it has become the architect of the physical world’s latency standards.

CoreWeave = Operational Arm

The 5GW roadmap signals a permanent shift: CoreWeave can no longer be viewed as an independent cloud provider, but rather as the operational arm of NVIDIA’s vertical monopoly. The $2B investment is a down payment on a future where CoreWeave is the exclusive showroom for the Vera-Rubin-Groq stack. By providing the equity to de-risk these massive infrastructure builds, NVIDIA has effectively turned CoreWeave into a “captured” entity—a strategic proxy that allows NVIDIA to control the cloud layer without the regulatory friction of a direct acquisition.

For small-cap miners like and , this new reality is a double-edged sword. On one hand, their valuation is no longer tied to the volatility of Bitcoin, but to their status as “Grid Gatekeepers.” In a world where NVIDIA is scrambling to secure 5GW of power, a miner with 500MW of “ready-to-plug” capacity and a 20-year Power Purchase Agreement (PPA) is more valuable for its substation than its mining rigs. They have become the land-owners in an era of industrial AI expansion.

Independence is dead

However, the “independence” of these smaller players is under immediate threat. As NVIDIA builds out its 5GW empire through CoreWeave, it creates a massive “gravity well” for power and talent. IREN and WULF and others now face a stark choice: either they become high-margin infrastructure partners for the NVIDIA-CoreWeave machine, or they become hostile takeover targets.

The reality of 2026 is that independence is dead. These smaller players are being forced into a consolidation deathmatch. They will either be compelled to merge with one another to gain the multi-gigawatt scale required to negotiate with chip titans, or they will inevitably fall under the “protective” wing of Google or Microsoft.

We are seeing the formation of infrastructure protectorates: CIFR is already leaning into Google’s orbit via the Fluidstack deal to secure TPU v6 access, while IREN is positioning itself as Microsoft’s strategic partner. Any attempt to remain a “neutral” operator is now a suicide mission; without a seat at the table of the big three, these companies lose access to the next-generation silicon and the capital required to stay relevant.

This publication is for educational and informational purposes only and does not constitute financial, investment, or trading advice. Readers are solely responsible for their own investment decisions. The author may hold positions in the securities mentioned.