Anthropic is priced at $965 billion and still cannot buy power without a guarantor. That is the most underreported sentence in AI infrastructure right now. It gets buried beneath the celebratory noise of a term sheet, but it is the entire story.
Sit with the contradiction for a moment. $965 billion. That figure exceeds the GDP of most sovereign states. It was manufactured by a pricing mechanism that assumes Claude's capabilities โ bundled into API calls and enterprise subscriptions โ can be converted into revenue at a trajectory justifying a nine-hundred-and-sixty-five-billion-dollar capitalization. But a valuation, no matter how large, cannot compile into physical surface area. You cannot stake a multiple inside a transformer. You cannot deploy a P/E ratio into a substation. And you certainly cannot convert it into 150,000 accelerator cards and 1.6 gigawatts of behind-the-meter natural gas.
Valuation is a consensus. Power is physics. And in the grassy flatlands around Hubbard, Texas, Anthropic did something that revealed more than a thousand pitch decks could: it signed a deal that separates the physical machine from the financial claim on it. The structure is elegant. The structure is dangerous. And if you have spent years auditing smart contracts, the smell is unmistakable.
This is not a story about model quality. It is not a story about alignment or benchmarks. It is a story about what happens when a frontier AI lab discovers that its balance sheet is not a balance sheet at all โ it is a narrative, and narratives cannot collateralize a gas turbine.
Reversing the stack to find the original intent: the original intent was to make Anthropic look light while making Google heavy. That is the abstraction leak. Let me trace it.
The Anatomy of a Machine the Size of a City
Start with the physical facts, because everything else is commentary built on top of them.
The project sits on 2,800 acres in Hubbard, Texas, roughly equidistant from the kind of grid infrastructure that used to attract data centers and now repels them. The power envelope is 1.6 gigawatts. That number has a history: in March 2026, the Financial Times reported a $5 billion project with 612 megawatts. It has since scaled to $150 billion in total project value and 1.6 gigawatts, a tripling of ambition in months. That kind of trajectory is not planning. That is a war footing.
1.6 gigawatts is enough to power roughly 1.2 to 1.6 million American homes at average loads. It is enough to run approximately 150,000 to 300,000 of the latest generation accelerator cards, depending on whether you assume two kilowatts or one kilowatt per chip under sustained load. That scale places the project in the top tier of American data center builds โ alongside the OpenAI/Microsoft Mount Pleasant complex in Wisconsin at roughly 2.4 gigawatts and ahead of xAI's Colossus in Memphis, which measured around 300 megawatts at its peak. The project is larger, in pure power terms, than most national supercomputing centers on earth.
But the power is not the most interesting part. The most interesting part is how the power is sourced. The site includes a behind-the-meter natural gas power plant, which means Anthropic is not plugging into the Texas grid and waiting for an interconnect queue that runs three to five years. It is building a private utility. It will burn its own gas, generate its own electrons, and route them directly, in the most literal sense, to its own accelerators.
The economic logic is transparent. Natural gas power generation, at current Henry Hub prices, produces electricity at roughly $40 to $60 per megawatt-hour. Industrial grid purchase prices in the United States currently run $50 to $80 per megawatt-hour. Behind-the-meter generation also eliminates transmission losses and demand charges. On paper, this is a COGS reduction play disguised as a real estate deal.
The carbon angle is the giveaway. At a 50% capacity factor, 1.6 gigawatts of natural gas generation produces roughly 3 to 4 million tons of CO2-equivalent per year in Scope 1 emissions. That is a direct, on-the-books climate liability for a company whose public brand is AI safety. No amount of carbon offset accounting can obscure a physical fact that is visible from orbit. This will become a regulatory and reputational exposure that no term sheet can contain.
And then there is the silicon. This is the part that most market commentary has under-weighted. The project's chips are custom tensor processing units, co-designed by Google and Broadcom, and procured under a supplier financing arrangement rather than purchased outright. Let that settle: Anthropic โ a company that emerged in the public imagination as the anti-Google, a certified B-Corp whose founding principle was an alternative to the entrenched AI powers โ is now building its compute runway on Google-designed ASICs, manufactured through Broadcom's design flow, fabbed on TSMC's advanced process nodes, and financed, in effect, by a housing loan from the landlord.
The structural choice is not neutral. Google's TPU line has been in production since 2015, moving from inference-focused v1 to the training-capable v4, v5, and the v6 series that currently powers much of Gemini's training load. Anthropic's custom TPU is not a generic GPU rental; it is a bespoke ASIC whose instruction set, memory hierarchy, and interconnect fabric will be optimized for Claude's specific workload characteristics. That is what the term 'custom' means in this context. It means Anthropic is committing its model architecture, its optimization stack, and its long-term cost structure to a silicon strategy designed by the very company it competes with for frontier model superiority.
This is not a small commitment. It is a five-year, deeply technical vows-of-obeisance to the Google hardware ecosystem. And it was not done for fun. It was done because the alternative โ buying NVIDIA GPUs on the open market โ would have been either impossible at this scale or economically ruinous.
Layer Zero: The Financial Stack
Now let me apply the analytical technique I developed auditing the 0x protocol in 2017. In that v0.9.9 audit, I found three critical unsigned integer overflow vulnerabilities in the fillOrder function. The bug was not visible at the level of the user interface. It was embedded two layers down, where raw arithmetic interacted with token transfer ordering. To find it, you had to read the code as a stack, not a story. Same discipline applies here.
The Anthropic financing structure is a multi-layer stack, and each layer is designed to quarantine risk from the layer above. Let me enumerate them.
Layer 0 is the physical asset: the land, the gas plant, the data center shells. This is the layer where concrete meets dirt. It is also the layer where Nexus Data Centers โ a developer with, and I quote myself from a previous life as a smart contract auditor, 'limited public track record in hyperscale projects' โ operates. Nexus is the general contractor, the entity responsible for delivering a 1.6-gigawatt AI campus on schedule.
Layer 1 is the project financing vehicle. This is a special purpose vehicle, an SPV, that holds the physical assets and their associated debt. The genius of the SPV structure is that Anthropic does not own this vehicle. The debt โ the $14 billion bridge financing and the subsequent term loans that will refinance it โ sits in the SPV. If the project fails, creditors have recourse to the SPV, not to Anthropic's general balance sheet. This is standard project finance. It is also standard English-law obtuseness: the debt exists, it is enormous, and it is guaranteed, at least in part, by Google.
Layer 2 is the operational layer: the supplier financing agreement for the custom TPUs. Here, Anthropic does not pay cash for the chips. Instead, the chip supplier โ an entity closely tied to Google and Broadcom โ receives a long-term stream of payments, secured by Anthropic's commitment to purchase a minimum quantity of compute. This is, to use the precise technical term, a compute mortgage. It converts capital expenditure into operating expenditure. It converts 'we built a machine' into 'we rent a machine.' And it does so without telling you, in any line item, what the interest rate is.
Layer 3 is the commercial layer: Anthropic's API business and the revenue it generates from selling tokenized intelligence. This is the layer where the model is expected to generate enough gross margin to service all the obligations beneath it.
The three layers are designed to be separable. Chips can depreciate at a different rate than the building. The building can carry a 25-30 year depreciation schedule while the silicon turns over every 12-18 months. The financial architecture is built to avoid the classic mistake of coupling fast-moving compute to slow-moving concrete. This is clever. It is also precisely the kind of cleverness that a deterministic failure mapper will immediately attempt to attack.
The problem: the layers are not actually independent. Chips without power are bricks. Power without chips is a gas plant with a light bill. And the offtake contract, the agreement where Anthropic commits to buying a certain amount of compute, ties the entire stack to a usage forecast. If Anthropic's API demand does not grow at the forecast rate, the minimum purchase obligations still stand. The fixed charges still come due. Failing that forecast, the tidy separation of layers collapses into a single cascading default.
Google's Five-Hat Problem
Abstraction layers hide complexity, but not error. And the most consequential error in this entire structure is the role of Google, which now wears five hats simultaneously.
Hat one: shareholder. Google holds approximately 14% of Anthropic, a position accumulated through multiple funding rounds. Hat two: chip designer. Google's TPU team is co-designing the custom silicon. Hat three: guarantor. Google is providing billions of dollars in lease and power purchase agreement guarantees to the project's lenders. Hat four: landlord. Google has taken a 20% equity stake in the data center and power project vehicle, making it a direct beneficiary of the project's long-term appreciation. And hat five: competitor. Google's Gemini models compete directly with Anthropic's Claude across nearly every product surface, from enterprise APIs to consumer chatbots.
In a smart contract context, this is equivalent to the validator, the sequencer, the bridge operator, the oracle provider, and a competing L1 all being controlled by the same entity. You would flag that on day one. You would call it, in the clearest possible terms, a centralization risk. But in the context of a venture capital term sheet, the same structure is called 'strategic alignment.'
Now, let me be fair to Google. From its perspective, this structure is optimal. By guaranteeing the project and taking an equity stake, Google achieves three objectives. First, it locks Anthropic into the Google Cloud ecosystem for the next decade, preventing the lab's migration to Azure or AWS. Second, it secures a financial return on the infrastructure boom that will occur regardless of whether Claude or Gemini wins the model wars. And third, it places Google in a position where, in the event of an Anthropic failure, it can acquire the physical assets and the compute contracts at a distressed price. This is not malice. This is capital working as capital does. The only surprising thing is that the market is surprised.
What does Anthropic get? It gets lower upfront capex, it gets a AAA-rated credit wrapper that reduces the project's financing costs by perhaps 200 basis points, and it gets access to Google's TPU supply chain. The cost is strategic autonomy. The cost is that the landlord knows your rent, knows your chip architecture, knows your training pipeline, and, as a competitor, knows exactly which models are being trained on which dedicated hardware.
For comparison, consider the OpenAI-Microsoft relationship. OpenAI runs on Azure and Microsoft is building its own Maia chips, but Microsoft's role is primarily vendor and partner, not guarantor, not co-designer, not landlord with a 20% equity stake. The Anthropic-Google architecture is one level deeper in every dimension. It is not a partnership with a cloud provider. It is a joint venture between a competitor and its tenant.
The Off-Balance-Sheet Ledger
There is a deeper financial question that the market has only begun to ask: what does Anthropic's balance sheet actually look like after this deal?
The design intent is explicit. By separating the physical infrastructure into an SPV and financing the chips through supplier agreements, Anthropic keeps the most volatile, capital-intensive assets off its own books. This is the same project finance logic that built airports and pipelines. When you build a pipeline, you do not put it on the oil company's balance sheet; you put it in a vehicle that sells its throughput to the oil company under a take-or-pay contract. The oil company gets the pipeline without the depreciation, the construction risk, and the leverage.
This worked for pipelines because the offtake was diversified across many shippers and the underlying asset had a long, predictable life. Neither condition is true here. The lifespan of AI accelerator hardware is 12 to 18 months before the next generation makes the prior one commercially obsolete. The offtaker is a single lab with a revenue base that, based on available estimates, is on the order of $1 billion to $2 billion annualized โ a fraction of the $150 billion infrastructure commitment.
This is where I draw on my Terra/Luna post-mortem. In that collapse, the core error was not the algorithm itself. It was the assumption that a feedback loop could sustain itself without an external source of value. UST relied on LUNA's market cap as its collateral. LUNA's market cap relied on UST's stability. The system worked until it did not, and the moment the feedback loop inverted, it became mathematically irreversible. Here, the same pattern is visible. Anthropic's valuation relies on its compute capacity. Its compute capacity relies on its access to capital. Its access to capital relies on Google's willingness to guarantee. And Google's willingness to guarantee relies on Anthropic's valuation. If any node in that loop breaks โ a missed milestone, a regulatory intervention, a model failure โ the loop does not just slow down. It flips sign.
What does that mean specifically? It means the $150 billion of infrastructure obligations, currently sitting in entities that Anthropic does not consolidate on its books, will at some stage need to be either repaid, refinanced, or serviced by operating cash flow. It is certain that the AP I cost structure of Claude โ the cost per token โ includes an implicit rent for the off-balance-sheet assets. What the market does not know is the magnitude of that implicit rent. What it does know is that Anthropic's gross margins will be structurally lower than they look because the rent is hiding in cost of goods sold rather than in depreciation.
This is where the Enron analogy forces itself into the room. Enron's fatal flaw was not the SPVs. It was the use of SPVs to obscure the economic risks of the underlying business while making the income statement look materially better than cash flows. When the market eventually did the forensic work โ tracing the actual cash flows โ the equity evaporated in weeks. I am not predicting fraud here. I am predicting a mismatch between reported and economic reality that will be resolved at the moment of maximum stress. The structure does not make the risk disappear. It only moves it across the line of sight.
In a smart contract, I call that a reentrancy risk misreported as an interface issue. The code says 'safe,' but the state transition you did not inspect is the one that drains the treasury.
The Critical Path Problem
Enough about money. Let me talk about engineering, because the most likely failure mode is not financial. It is temporal.
AI infrastructure is now a story about three parallel critical paths that must converge with millisecond-level (or at least month-level) precision. Path one is the construction of the data center and power plant. Path two is the design and volume production of the custom TPU. Path three is the training schedule of the model that will consume the compute.
Each path has its own latency. A natural gas power plant takes 24 to 36 months to bring online, from groundbreaking to first fire, especially when it is behind-the-meter and requires pipeline interconnects for fuel supply. A hyperscale data center, in the optimistic scenario, takes 12 to 18 months. A custom ASIC โ designed, tape-out, validated, volume-produced โ takes 18 to 36 months. These are not additive delays. They interact. If the power plant is late, the data center waits. If the chip is late, the power plant waits. If both are on time but the model architecture changes, the entire compute stack might be mismatched.
This is the point where the Nexus Data Centers execution risk becomes the most consequential variable in the entire deal. The 1.6-gigawatt project has grown from 612 megawatts in a matter of months. That level of scale-up velocity is precisely what precedes schedule slips in large infrastructure. Add a builder with limited hyperscale track record, and the probability of a 12-24 month schedule overrun is not a tail risk. It is the base case.
What happens when the schedule slips? First, Anthropic's model training timeline โ which is directly tied to its product roadmap and IPO narrative โ takes a hit. Second, the guaranteed obligations held by Google are triggered, causing Google to step in and restructure. Third, the creditors face a project that generates no revenue while accumulating interest. In an environment where the entire AI capex cycle is financed on confidence, one high-profile delay will repricing every other AI infrastructure project in the pipeline. This is the fragility that deterministic failure mapping is designed to expose. The market is pricing a portfolio of call options on completion. The underlying assets are corridors of risk.
The chip supply chain adds another layer of vulnerability. The custom TPU is being manufactured by Broadcom, which routes advanced logic to TSMC. Broadcom's capacity at TSMC's 3nm and 4nm nodes is finite and shared. In that queue sits Google's own TPU v7 production for Gemini's next generation. The same fab can produce Anthropic's chips or Google's chips. When events conflict, whose tape-out gets the allocation? The question answers itself. Anthropic's compute schedule is now subordinate to Google's internal priorities at the exact point where the physical silicon is created.
And then there is the power supply itself. A behind-the-meter gas plant running at full load daily creates a dependency on gas pipeline delivery. During Texas winter storms, gas pipeline capacity is historically unreliable. The plant has no backup power source. The site has no redundant grid interconnection. In a severe weather event, the entire 1.6-gigawatt campus goes dark. This is a single point of failure, embedded in the design, accepted because the alternative (grid interconnection) has a multi-year queue. The tradeoff between speed and resilience is real. But let me note: a project designed to avoid the grid queue is also a project designed to fail against the grid.
The cooling requirement is another unstated constraint. At 1.6 gigawatts with a PUE of 1.2, the site requires the dissipation of roughly 320 megawatts of waste heat. That implies an enormous direct liquid cooling system that needs water, which is a scarce resource in central Texas during drought cycles. The cooling infrastructure and its water supply have not been publicly detailed. In an auditor's mind, that is a disclosure gap, and disclosure gaps are where risks hide.
The Competitive Landscape: NVIDIA's Loss, Broadcom's Game
Anthropic's pivot from rented NVIDIA GPUs to custom Google TPUs is frequently described as a 'diversification' strategy. That is not what it is. It is a concentration strategy.
Look at the competitive map. OpenAI runs primarily on Azure while designing its own Maia silicon. xAI built Colossus on NVIDIA GPUs, at record speed. Meta self-built its own clusters, including a planned 240,000-H100 fleet, wholly owned and operated. Google DeepMind trains Gemini on its own TPUs, fully vertically integrated. Every frontier lab is moving toward integration in one form or another. Anthropic was the last independent holdout, renting compute from hyperscalers and maintaining neutrality.
With this deal, Anthropic abandons neutrality. It signs a long-term lease on Google's infrastructure, Google's silicon, and Google's supply chain. The vertical integration is real, but it is Google's vertical integration, not Anthropic's. The lab has effectively offshored its ability to decide its own hardware future. Custom TPUs are not a commodity. Once Claude's neural architecture is optimized for TPU instruction sets, transferring to another hardware platform becomes a multi-quarter engineering effort with massive sunk costs. The switching cost grows every single day the company trains on that hardware. Anthropic has hardwired its own future to Google.
NVIDIA will feel this as a lost customer for frontier-scale GPU clusters. But the real beneficiary is Broadcom, which cements its role as the arms dealer of custom AI silicon. The company is already co-designing TPUs with Google, XPUs with Meta, and increasingly, custom accelerators for every major cloud player except NVIDIA. This deal further entrenches the 'design-once, sell-many' ASIC model that threatens NVIDIA's blanket-market dominance.
But there is an underappreciated longer-term risk. If every major lab moves to custom silicon tied to a specific fabrication partner, the industry loses its secondary market. NVIDIA GPUs retain residual value because they are fungible and widely deployable across many users. A custom TPU designed expressly for Anthropic has essentially no secondary market. If Anthropic fails, its custom silicon is worth scrap. This is the financial definition of asset specificity. And asset specificity is exactly what a lender does not want to see when it underwrites a project.
The Market Context: When Valuations Meet Physics
The timing of this transaction is everything. A $965 billion private valuation on a company with maybe $2 billion of revenue is a multiple that assumes extraordinary continuation. The market is effectively pricing in that Anthropic will become the most important software company on earth. That may well be true. But a multiple cannot buy a transformer. A multiple cannot build a gas plant.
This is why the structure of the financing matters more than the amount. Anthropic is not spending its own cash to build; it is borrowing against its future. Google's guarantee makes that borrowing cheap. But the guarantee is a form of endorsement that carries a heavy implicit condition. Google now holds 20% of the physical asset. If the IPO is delayed, if the training schedule slips, if the revenue forecast fails, Google has the right to convert distress into ownership. The term sheet does not say 'I love you.' The term sheet says 'I will ensure you exist so that I can own your residual value.'
The IPO plan for October 2026 adds yet another layer of pressure. The entire infrastructure transaction had to be structured before the IPO to allow the balance sheet to look as clean as possible. But 'looking clean' and 'being clean' are not the same. When the S-1 is finally filed, the minimum purchase obligations, the related-party transactions with Google, and the contingent liabilities embedded in the supplier financing will all be disclosed. Analysts will then take out their spreadsheets and compute the real economics. The first-time convergence of those numbers will be the most important moment for Anthropic's valuation since this transaction was announced.
The Regulatory Blind Spot
In the same way that NFT collectors believed their metadata lived on-chain when it actually resided on centralized IPFS gateways, the market is assuming that this infrastructure deal is a neutral, charitable act by Google to support a competitor. It is not. The structure is now so dense that it begs for regulatory review. The FTC has already looked at Microsoft-OpenAI. It is difficult to see how the FTC ignores Google holding 14% of Anthropic, guaranteeing billions in debt, supplying it with custom silicon, and owning equity in its power plant. That is not close to a vertical merger. It is adjacent to a full merger of interest.
There is a credible scenario where the FTC forces Google to either divest its stake in the project vehicle or relinquish its board influence at Anthropic. This would invalidate the core financing structure, pushing significant obligations back onto Anthropic's balance sheet right as it is preparing to go public. That alone justifies assigning a substantial probability to the regulatory denouement of this deal.
The Contrarian Angle: The Bull Case Is the Bear Case
The usual response to all of this is: 'But Google's involvement makes the deal safe.' The market treats the guarantee as de-risking. I consider that inverted.
Google's guarantee does not reduce the risk that the infrastructure fails. It only changes who holds it when it does. The risk of a late power plant, a slow tape-out, or a demand shortfall remains exactly as large as it was. The guarantee merely transfers the loss from Anthropic's shareholders to Google's, and then, via Google's tacit status as a quasi-systemic institution, onward to the broader market. The guarantee does not shrink the risk. It hides it, like off-chain state committed without a proof.
The stronger version of the contrarian case is something that makes me uneasy precisely because I enjoy it too much. Here it is: this deal is the best possible insurance for Anthropic in the event that the AI bubble bursts. If the market cracks, Anthropic's off-balance-sheet obligations do not vanish. But they do become negotiable. Google, as guarantor, has every incentive to renegotiate the terms rather than let the asset sit idle. The 20% equity stake gives Google an incentive to keep the project alive even if the original economics have collapsed.
That incentive makes the entire structure surprisingly stress-tolerant in a mild downturn. The real failure mode is a scenario where the demand for AI compute collapses so rapidly that even the guarantor cannot justify maintaining the asset. The banks would then hold a 1.6-gigawatt gas plant and a semi-bespoke accelerator fleet with no marginal buyer. The collateral would not cover the debt.
This is the exact scenario that metastasizes into a systemic event. At $150 billion of infrastructure, with Tesla- or bank-grade dominoes behind it, this asset class has become too big to fail and too illiquid to rescue. The financial instrument is structurally similar to subprime CDOs in 2006. The inputs look safe because they are each individually rated. The system, however, has never been tested at scale under a simultaneous peak-power, peak-debt, peak-competition environment.
What I Watch, As An Analyst
Practical guidance follows. I have spent three-quarters of my career auditing code and economic models, and I have learned that the leading indicators of collapse are always the unglamorous ones. Watch the milestones. Watch the gas pipeline interconnect agreement. Watch the date on which the TPU tape-out is announced. Watch the hiring announcements at Nexus Data Centers. If a 1.6-gigawatt project needs 500 engineers to execute and the LinkedIn headcount does not move, the schedule is already slipping.
And watch the IPO filing. When Anthropic's S-1 is published, the single most valuable piece of analysis will be a line item, not a paragraph: the amount of 'future minimum lease payments' and 'unconditional purchase obligations' that appear in the footnotes to the financial statements. Those numbers will tell you the real leverage. Everything else is marketing.
Another overlooked data point is the pace of Google Cloud's TPU expansion. If Google's own TPU v7 allocation is publicly discussed in supply chain reports before Anthropic's custom TPU enters mass production, you know that Anthropic is second priority. Similarly, watch the natural gas futures curve. If long-dated gas forward prices rise, the entire LCOE advantage of behind-the-meter generation evaporates. A plant that saves money at $3/MMBtu becomes a money-loser at $6/MMBtu.
The Takeaway
I will end where I started. The $965 billion valuation is not the story. The story is that the most valuable private AI company in the world cannot secure its own power, cannot design its own silicon, and cannot finance its own physical growth. It must pass through the perimeter of an even larger incumbent. It trades autonomy for capital. That trade is perfectly rational; rational trades can still end in ruin.
The next time you read that the race for AGI is a race of algorithms, remember the 2,800 acres in Texas. The algorithm is not the bottleneck. The bottleneck is a gas turbine with a construction schedule. And the company that owns the turbine also owns the lender, the chip designer, and the competitor.
Truth is not consensus; truth is verifiable code. In this case, the verifiable reality is that Anthropic has taken a loan against its own future, and the collateral is not its models.
It is its freedom.