GM. This is Milk Road PRO, tracking where AI value lands when the smartest model stops being the scarcest thing in the stack.
The market spent 2024 and 2025 asking who had the best model and the most GPUs. In 2026, the question is who can deliver useful tokens cheaply, at scale, and on infrastructure they actually control.
Intelligence just got cheap. DeepSeek V4 Pro prices output at 1/57th the cost of Fable 5. At the same time, access just got political. Washington gave Anthropic 90 minutes to pull Fable 5 from every foreign national and gated GPT-5.6’s flagship tier to roughly 20 approved partners.
Cheap alternatives from below and political risk from above are squeezing pricing power at the top of the stack. Value is flowing toward the layers that stay scarce: power, memory, packaging, optical interconnect, and data-center capacity. A smarter model solves none of those constraints.
This month’s report looks at where that value is moving, why these shifts matter for investors, and how they’re changing my portfolio.
Here's what's inside today:
- Repricing the Models layer.
- The $700B power bottleneck.
- Three chip squeezes at once.
- Meta, the fourth hyperscaler.
- The MIT proof gap.
- Portfolio impact.
MODELS LAYER: THE OFF SWITCH
Fable 5 launched on June 9. Three days later, the Commerce Department gave Anthropic 90 minutes to comply with an export-control order pulling access from every foreign national on the platform. By Friday evening, both Fable 5 and Mythos were dark worldwide. The lawsuit Anthropic filed over its "supply chain risk to national security" designation from March is still in federal court.
Ten days later, OpenAI got the same treatment in a different flavor. GPT-5.6 launched on June 26 in three tiers: Sol, Terra, and Luna. Under a Trump administration request, the flagship Sol was gated to roughly 20 government-approved partners. OpenAI itself said publicly this is "not our preferred long-term model."
Different mechanism, same message. Frontier AI models are national security assets, and Washington controls the release calendar.
In the meantime, Anthropic brought Fable 5 back globally with new classifiers and deeper U.S. government pre-release testing. Amazon, Microsoft, Google, and other Glasswing partners are drafting the industry consensus framework for jailbreak response.
The switch got flipped back on. The insecurity remains.
That new operating environment landed at exactly the moment pricing power was collapsing from below.
Chinese open-weight models have closed the capability gap enough that companies switch for cost alone. OpenRouter shows Chinese-origin models grew from under 2% of token consumption in late 2024 to more than 50% by June 2026. DeepSeek V4 Pro at $0.87 per million output tokens runs at 1/57th the price of Fable 5. Even GPT-5.6 Sol delivers comparable capability at one-third the output tokens of the prior model.
Enterprise adoption is already reflecting this pressure. Uber exhausted its 2026 AI budget in four months, Cursor shifted customers to usage-based pricing, Microsoft reduced Claude Code access internally, and startup Lindy moved its production workload from Claude to DeepSeek after its CEO said the cost curve had “crashed to the ground.” Goldman Sachs expects agentic AI token consumption to grow roughly 24x by 2030, reinforcing that cost, not capability, is becoming the key constraint.
The result: token cost now decides which model wins the workload. That is the moment a market becomes a commodity.
The second squeeze is political.
China published a 17-point plan this month to push AI into every shop, clinic, and classroom with explicit subsidies for AI consumption.
Put the two forces together: enterprise buyers optimize for cost, governments optimize for sovereignty, and neither purchase turns on absolute capability.
For the majority of workloads, frontier models are effectively interchangeable with cheaper open-weight or vertical alternatives, and any pricing power collapses the moment users have a credible free option.
That is why the market is heading toward a symphony of models rather than a monolith. Most tasks will run on good-enough open or open-weight models chosen for cost or sovereignty. The high-stakes work, the smaller share where marginal capability actually moves the outcome, will run on frontier or vertical-specific models tuned for the job.
The direction is already visible in the deployments. Palantir is integrating NVIDIA’s Nemotron models into agents running on sovereign AI operating system architecture. Eli Lilly's $1B co-innovation lab with NVIDIA, building vertical pharma models on Vera Rubin architecture, is the same story in a different vertical.
Healthcare, finance, and industrial companies are increasingly building their own infrastructure and specialized models rather than routing everything through a frontier lab.
What this means: The squeeze on the Models layer is structural. Regulatory pressure from above, Chinese commoditization from below, and vertical plus open-weight models eating into the middle all push value further down the stack.
Every frontier lab, Chinese competitor, and vertical model still needs compute and energy to run.
That is bullish for companies making inference cheaper and more sovereign.
It is also bullish for Palantir, which sits at the deployment and governance layer where enterprises connect models to proprietary data or operational workflows. As the model layer becomes more fragmented and interchangeable, the ability to deploy AI securely across an organization becomes more valuable, not less.
ENERGY LAYER: THE $25B VALIDATION
Two things happened in the Energy layer this month. A very large institutional check got written behind Bloom, and the macro setup underneath the position quietly improved.
Brookfield expanded its Bloom partnership from a $5B framework signed in Oct. 2025 to $25B on June 30. A fivefold increase, sitting inside Brookfield's $100B AI Infrastructure Fund launched in Nov. 2025. Bloom surged 10-12% on the news.
What this shows: behind-the-meter power for hyperscale AI has graduated into an institutional asset class, with $25B of committed capital lined up behind it.
The real sovereignty trade
BE CEO Sridhar reframed the entire energy thesis in an interview this month and it picks up exactly where the sovereignty discussion in the Models section left off.
His argument: energy sovereignty is more important than model sovereignty.
If Washington can revoke Fable and gate GPT-5.6, the bigger structural problem for every non-U.S. buyer is that the compute stack still runs on power infrastructure that is centralized and controlled by grid operators years behind demand.
Behind-the-meter fuel cells are "power-to-the-edge". They move the constraint from someone else's grid to your own building.
U.S. data center construction hit a $50B annualized rate in April, surpassing public spending on all transport infrastructure combined. The four largest hyperscalers are collectively targeting $700B in CapEx this year. The capital is committed years ahead of the capacity.
Of the roughly 12 gigawatts of new capacity operators promised in the U.S. for 2026, only about 5 gigawatts are actually under active construction.
That is why sovereignty and grid physics converge on the same answer.
Behind-the-meter fuel cells, geothermal wells, small modular reactors, and solar plus storage are four categories that let a buyer bring their own power instead of waiting in the utility interconnect queue.
Bloom sits in the first category with the biggest institutional check written behind it. That is what the Brookfield expansion actually validated.
The other three categories are moving too.
FuelCell Energy signed a 380-megawatt deal with Fit Energy this month and joined the Russell 2000 and Russell 3000 alongside Bloom's promotion.
On geothermal, Fervo Energy raised $1.89B in its May IPO with a $7.2B backlog and Google as counterparty, and Ormat unveiled a 100-megawatt unit engineered for non-volcanic regions.
Small modular reactors will not deliver commercial power until the early to mid 2030s, so Fervo and Ormat are delivering 100-plus megawatts by 2027 while SMRs are still a decade out. The boring option is years ahead of the glamorous one.
Solar also benefited from improving U.S. policy expectations, reinforcing that multiple power technologies are now riding the same AI infrastructure buildout.
Vera Rubin's power problem
What makes these off-grid developments especially relevant is the shift in how much power each rack of AI silicon actually needs.
Rack power requirements are rising dramatically, from roughly 40 kW with Hopper to as much as 1 MW on NVIDIA’s roadmap by 2028. Every increase pushes more demand onto transformers, switchgear, batteries and behind-the-meter power.
Every additional kilowatt is a new load on the transformers, switchgear, batteries, and behind-the-meter power the layer just walked through.
The utility interconnect queue was already 12 months behind. Each generation of the roadmap pushes the required power density per rack materially higher, which pushes any operator that wants to deploy the newer silicon toward on-site generation whether they wanted to be there or not.
Bloom fits directly into that transition.
NVIDIA is standardizing on 800V DC as the power architecture for AI factories, first as an optional configuration on Vera Rubin, then as the standard for Rubin Ultra and beyond. Bloom has been publicly positioning its fuel cell platform against the same 800V DC roadmap. The two directions line up.
What this means: The Energy layer is no longer the speculative part of this trade. Institutional capital is underwriting it at 5x prior scale, passive index flows are lifting it, and the sovereignty reframe gives the thesis a second leg.
The next generation of NVIDIA silicon is engineered around a wattage envelope the public grid cannot serve on the buildout's timeline. The electrons have to come from somewhere.
CHIPS LAYER: SYSTEM SCARCITY
The Chips layer has moved past GPU scarcity. The shortage now sits in everything that has to surround a GPU to make it useful: memory, packaging, optics, and power delivery.
Picture the die itself, the small piece of silicon that does the actual computing, as the middle of a very expensive sandwich. Around it sits the memory that feeds it data. Underneath sits the packaging that bonds those pieces together. Above sits the cooling and power delivery that keeps it alive. Alongside sits the optical wiring that connects it to the tens of thousands of other GPUs in the same cluster. The die is the easy part. Every one of those surrounding layers is getting tighter, more expensive, or more politically contested at the same time.
Three specific squeezes are worth pointing out.
Squeeze one: memory
Modern AI systems stall when the GPU sits idle waiting for data. That is why every one ships with stacks of high-bandwidth memory, called HBM, physically bonded to the die.
Only three companies make HBM in volume. SK Hynix supplies most of it. The next generation, HBM4, is what Vera Rubin needs, and it is already sold out through the middle of 2027.
The bonding process itself is the second bottleneck. It is called CoWoS, and only TSMC does it at scale. Its capacity is spoken for years in advance by a small set of hyperscalers. If you are not on the list, you are not shipping.
The bigger point is that packaging has quietly become as strategic as the silicon it wraps.
TSMC and Amkor signed a 10-year deal on June 16 to build advanced packaging capacity in Arizona. The industry is now investing years in advance to secure the capacity required to turn leading-edge chips into complete AI systems.
Squeeze two: optical interconnect
Packaging is only one bottleneck. Once the GPUs are assembled, tens of thousands of GPUs are wired together so they can train the same model in parallel. At those speeds, ordinary copper wires do not work. The signals bleed into each other. So the wiring is done with light, through components called optical transceivers.
That is where NVIDIA’s investment in Corning fits. The partnership is designed to expand U.S. manufacturing of the optical connectivity needed to link increasingly large AI clusters.
But scaling optical networks creates another point of exposure: the materials inside the components that convert electrical signals into light.
The critical raw material inside those transceivers is a compound called indium phosphide.
China refines roughly 70% of the world's indium. This year it introduced a licensing regime that requires a permit for every export shipment, buyer by buyer. That gives Beijing a tap it can turn down on any specific customer at any time.
NVIDIA has been positioning around this quietly.
Its investment in Lumentum, one of the leading optical transceiver suppliers, fits the same pattern as its earlier moves into Corning and Marvell. The "follow Jensen" trade has consistently meant following him into the parts of the supply chain about to become strategic. Every one of those moves bet on the components around the die rather than the die itself.
Squeeze three: power delivery
The physical stack around each GPU is getting tighter because the wattage envelope keeps climbing, and that is the handoff back to the 800VDC part of the Energy section.
The higher the power density, the more critical the semiconductors that deliver clean, high-voltage power into the chip become. Those are called power semis, and Infineon is the cleanest exposure to them.
Infineon announced its second 2026 price hike on July 1, following an April hike. Roughly 20 chipmakers raised power semi prices 10-25% in the same window. The company told the market that its AI data center revenue is on track to nearly double from €1.5B in fiscal 2026 to €2.5B in fiscal 2027, and committed an incremental €500M to accelerate expansion. Lead times on some products are running as long as a year.
What this means: The chips layer is broadening beyond the GPU. As AI systems scale, performance increasingly depends on whether companies can secure the memory, packaging, optical connectivity and power components required to turn a chip into a functioning cluster.
That changes where value can accrue. NVIDIA remains the central platform, but the scarcity premium is spreading to the suppliers around it. Corning benefits as larger clusters require more advanced connectivity, while Infineon benefits as power delivery becomes a larger and more demanding part of the system.
These inputs are getting more expensive and harder to substitute at the same time. The more complex the AI system becomes, the more strategic the companies that package, connect, and power it become.
INFRASTRUCTURE LAYER: FROM SPACEX TO SOVEREIGN
Public markets staged the loudest moment of the Infrastructure layer this year.
SpaceX went public on June 12 at $135 per share, raising $75B at a $1.77T valuation. The stock peaked at $225 four days later and closed the month at $160.
The interesting number is the gap. Morningstar values the actual operating business at around $780B. The IPO priced roughly $1T above that. Most of the premium is the orbital compute thesis. Days before the IPO, SpaceX unveiled AI1, a satellite as wide as a jumbo jet built to carry AI computing hardware, with announced plans for 1 gigawatt of orbital compute by the end of 2027.
The prospectus tells a different story. It promises launches "as early as 2028" and revenue "by the end of the decade." Elon himself said the 2027 target should be taken "with a grain of salt." The cooling problem, radiating gigawatts of heat into space as infrared light, is unsolved at this scale. The filing also concedes SpaceX cannot currently source enough chips to execute the plan.
Roughly $1T of premium for a buildout the company itself describes as starting in 2028, generating revenue by 2030 if it works, and not currently chippable. Not a valuation I pay directly.
The sovereignty reshuffle
The strategically important SpaceX story sits somewhere else: where the compute physically lives and who holds jurisdiction over it. That question is reorganizing the whole layer.
That is the infrastructure-layer version of the sovereignty argument. Governments and enterprise buyers are asking whether their national AI workloads should run on U.S.-domiciled cloud infrastructure at all.
Japan and South Korea both accelerated domestic AI compute buildouts this month.
In Europe, StackIT, Schwarz Group's sovereign cloud arm, keeps pulling regulated German and EU workloads out of the hyperscalers by pitching data residency and legal jurisdiction as the actual product.
The Palantir-NVIDIA Nemotron partnership from the Models section is the same sovereignty question answered from the vendor side: a stack designed so a customer can run frontier-class AI on their own infrastructure, on their own data, without handing either back to the model provider.
The picture across the layer is coherent even if it is messy.
Models can be gated, power can be interrupted, and compute can sit in the wrong jurisdiction. Every layer of the stack is starting to get priced for that risk, and infrastructure is where the answers show up as physical assets in specific countries.
The fourth hyperscaler
The other infrastructure story that landed this month sets the trade up differently.
On July 1, Meta confirmed it is building a cloud infrastructure business to sell external customers its excess AI computing power, becoming the fourth hyperscaler cloud after AWS, Azure, and Google Cloud.
The direct competition with the incumbents is the headline. The bigger signal is buried in the reasoning.
Meta admitted it has excess AI compute to sell. First hyperscaler admission on the record that some of the $700B of AI CapEx may not clear at internal ROI on the timeline the market has been pricing.
The immediate victims were the neoclouds.
CoreWeave dropped 12% on the Meta announcement. Nebius fell in sympathy.
Both neoclouds had strong momentum coming in. On June 22, CoreWeave and Nebius both joined the Nasdaq-100. Nebius rallied ~240% year to date coming into the debut. CoreWeave carries a $99B contracted backlog anchored by OpenAI ($22.4B) and Meta ($21B).
Meta pivoting to cloud does not kill the neocloud thesis, but a well-capitalized hyperscaler entering the raw-compute-rental market compresses pricing for every pure-play provider.
What this means: The infrastructure layer is splitting into two very different trades.
Raw compute capacity is becoming more competitive. Meta’s entry adds another well-capitalized supplier, increases the risk of excess capacity, and weakens the scarcity premium that pure-play neoclouds have been trading on. Owning GPUs is no longer enough if hyperscalers can bring additional supply to market and compress rental economics.
The more durable opportunity is controlled compute: infrastructure that meets requirements around jurisdiction, data residency and security. That is where sovereign-cloud providers and architectures such as Palantir plus NVIDIA become strategically important.
SpaceX made the loudest noise this month. The durable signal is the migration of value from undifferentiated compute supply toward infrastructure customers can legally, securely, and operationally control.
APPLICATIONS LAYER: THE DEPLOYMENT GAP
The Applications layer is where the strategic frame lands hardest. Three forces are hitting it at the same time:
- Intelligence is becoming easier to access. Chinese open-weight models handle roughly half of OpenRouter traffic. Nemotron 3 opened up frontier-class weights and training data. Vertical models are being built inside enterprises.
- Token cost, not model quality, is now the constraint enterprise buyers are optimizing around.
- Markets have started demanding proof of AI ROI, and not from the hyperscalers but from the operating companies that were supposed to benefit. The hyperscalers can still point to their CapEx as forward-looking. The end users cannot.
The proof gap
The demand for proof of AI ROI has been simmering since the end of 2025, and in 2026 it stopped being a background debate. MIT NANDA quantified it. After $30-40B of enterprise spending on generative AI pilots, 95% of organizations have seen zero measurable P&L impact. Only 5% of pilots actually moved the financials.
The 2026 detail is worse. A study of more than 100,000 GitHub developers this year showed teams generate nearly 10x more lines of code with Claude Code but only about 30% more software actually ships to users. The gains collapse at every stage between writing code and releasing it. Code review, integration, testing, and the human judgment work that does not automate easily.
What this means: The proof gap is the market's new selection framework. The question has moved from who has AI to who can turn it into controlled workflows and measurable economics.
That points to two categories of winner.
The first: companies that make adoption possible through governance, security, data control, and workflow orchestration.
Palantir is the cleanest example in my portfolio.
Maven Smart System became a Pentagon program of record this month, meaning permanent budget and recurring revenue visibility.
The same Nemotron partnership the models and infrastructure sections already flagged doubles as the applications-layer thesis here. It is what turns governed AI into a repeatable enterprise product. The product is the layer that decides which data an AI can touch and thus is what makes AI usable in enterprise workflows and unlocks real business value.
The second: companies that adopt AI deeply enough to actually move revenue, productivity, margins, or asset utilization.
Tesla is that bet in my portfolio on the physical side.
Q2 deliveries came in at 480,126 vehicles, well above the 406,000 consensus. Robotaxi moved to unsupervised operation in limited areas of Austin, with hundreds of Cybercab units reportedly parked in staging lots across multiple cities ahead of broader deployment. Optimus first-generation production is installing in California with a 1M unit annual target.
The bet is on physical-world deployment of embodied AI, which does not need to route through enterprise sales cycles to prove ROI.
PORTFOLIO LENS
Every position in the portfolio is a bet on a piece of the stack that intelligence commoditization does not threaten. Most are pieces it actively strengthens. The month's events tested that thesis but reinforced the existing composition rather than reshaping it.
⚡ Energy
Bloom Energy (BE)
Brookfield 5x'd its partnership to $25B, Bloom joined the Russell 1000 and Russell Top 200, and the CEO reframe around energy sovereignty reset the strategic argument.
Two crosscurrents underneath the headline. The Brookfield expansion strengthens the deployment financing case, and Vera Rubin's wattage profile pulls directly on that vector. But the raw-capacity trade, the reason for the earlier trim, is still under pressure from frontier lab pricing collapse and Chinese peer efficiency gains upstream.
Holding.
🧠 Chips
NVIDIA (NVDA)
The Nemotron 3 release plus the Palantir sovereign OS partnership are the two developments that matter, and they matter together. NVIDIA is expanding from a hardware supplier into a full-stack sovereign AI platform without giving up any hardware pricing power. Every framework that pulls compute demand higher, whether frontier model competition, Chinese buildouts, agentic robotics, or sovereign open-model deployments, still flows through NVIDIA.
The thesis remains intact. The question is now one of relative risk and reward.
Holding, with a potential trim to fund higher-conviction opportunities elsewhere in the stack.
The Chips section covers the two 2026 price hikes, the revenue guide nearly doubling to €2.5B, and lead times running as long as a year.
The position takeaway: scarcity is showing up directly in the price sheet, exactly what the thesis predicted, and IFNNY remains a great power semiconductor exposure into the AI CapEx cycle. The Mistral round and broader European sovereignty signaling add a second tailwind.
Holding, building on weakness.
🏗️ Infrastructure
Corning (GLW)
The Chips layer research above is Corning's thesis in one sentence. Every input that wraps the GPU is getting tighter. Corning sits inside two of the three tightenings, optical interconnect via the indium phosphide chokepoint, and advanced packaging via glass substrates. TSMC-Amkor's 10-year advanced packaging deal in Arizona is a tailwind for the whole category.
Holding.
🤖 Applications
Palantir (PLTR)
Maven becoming a Pentagon program of record and the Nemotron partnership are covered above.
The counterpoint is Europe: France's domestic intelligence agency plans to phase Palantir out for a domestic provider, proof that sovereignty can outweigh capability in sensitive workloads.
Net positive month. Karp's own argument cuts our way: as models get cheaper and more interchangeable, the ontology and control layer gets more valuable.
Holding.
Tesla (TSLA)
The Applications section covers the delivery beat and the Robotaxi and Optimus ramps. The portfolio angle is the SpaceX linkage: the $2B equity stake, the Terafab joint fab in Texas, and the orbital data center collaboration all point toward eventual entanglement, which is the specific outcome I am positioned for.
Starlink is the connectivity layer physical AI runs on, and holding Tesla is the cheapest way to own the combined story without paying the $1T orbital premium directly.
Holding.
Eli Lilly (LLY)
I entered LLY as the portfolio's purest expression of this report's core claim: when intelligence commoditizes, value lands with whoever deploys it against assets nobody else owns.
Lilly points intelligence at proprietary clinical data, trials, and manufacturing scale, so its AI CapEx compounds a moat while hyperscaler CapEx funds a margin war. Same spending line, opposite investment case.
The $1B NVIDIA co-innovation lab (Models section) confirms the path: vertical models on Lilly's own infrastructure, with no frontier lab sitting in the loop between Lilly and its data.
The return shows up as more efficient processes, shortened legal drug processes, more efficient manufacturing and a better pipeline rather than an AI revenue line, which keeps the position insulated from token pricing, export politics, and everything else squeezing the top of the stack.
Holding, adding on weakness.
💰 Economic rails
ETH, BTC, SOL, HOOD
Robinhood’s “The World is Flat” event highlighted the speed of its move into crypto and agentic finance. New launches included Robinhood Chain, expanded stock tokens and DeFi products. More important than any single announcement, HOOD is evolving from a trading app into the interface through which AI agents can trade and manage money.
ETH and SOL remain the settlement rails, BTC the scarcity asset and HOOD the consumer gateway.
Holding.
Got questions about this piece, or any other stocks you like? Hit me up at @WhiteCollarExit on X.
Catch you in the next one. 🫡
AI-GENERATED PODCAST 🤖
We’ve turned this AI PRO report into an AI-generated podcast to make it even easier to digest. You'll find the audio player below. 👇️🎧️
Why Anthropic, OpenAI, and DeepSeek face a pricing collapse
Disclaimer: This podcast was created using AI and is based on the research report above. While we've done our best to ensure accuracy, the audio may contain minor errors, technical glitches, or mispronunciations. Please note that this podcast provides an overview of the report and is not a comprehensive or definitive take on the topic.
