Hook
The number is absurd on its face—$517 billion, spread across a decade, attributed to Anthropic's cloud and compute commitments. Annualize it and you land near $51.7 billion a year in procurement, a figure that sits in the same order of magnitude as the entire annual capital expenditure of a global hyperscaler. Now hold that against what we can actually verify about the company doing the buying. Anthropic's public revenue, depending on whose estimate you trust, sat in the low billions as recently as 2024 and at best in the low tens of billions in the most optimistic 2025 scenarios.
I have spent much of my career reading term sheets, audit memos, and the fine print that market participants prefer to compress into a single headline. When I first saw a nine-figure-per-year commitment attached to a company whose cash flows are one to two orders of magnitude smaller, my instinct was not excitement. It was forensic. In my experience, when a number is too large to reconcile with a company's balance sheet, it is rarely a number at all. It is a narrative wearing a number's clothes.
That is the anomaly worth interrogating. Not whether Anthropic is serious about compute—of course it is. Not whether AWS and Google benefit—obviously they do. The question is what kind of instrument a $517 billion commitment actually is, who is really funding it, and why the crypto-native playbook I have watched for the last six years suddenly reads like a mirror of the AI infrastructure story nobody wants to call by its name.
Context
To understand the stakes, you have to reconstruct the terrain without the marketing. Anthropic occupies a peculiar position in the frontier-model race. It is a closed-source lab whose brand is built on safety and alignment, whose Claude family has earned genuine enterprise traction in code generation, long-context reasoning, and agentic tool use, and whose distribution runs through two competitors-turned-patrons: Amazon and Google. Both have invested in Anthropic. Both sell its models through their clouds—AWS Bedrock and Google Vertex. Both, conveniently, also sell the compute that trains and serves those models.
That triangular structure is the first thing a financial engineer notices. The investor, the supplier, and the distribution channel are the same entities. AWS has been the loudest proponent of its own silicon, Trainium, and has spent years trying to wean the market off NVIDIA's stranglehold. Google owns its TPU line, now several generations deep. Anthropic, wedged between them, has an obvious incentive to diversify away from NVIDIA's pricing power—and an equally obvious incentive to accept favorable terms that come bundled with dependence.
The article that seeded this analysis was thin: a fast news item noting cloud and compute deals over a decade, intensifying AI infrastructure competition, and a conclusion that the transaction benefits giants and hardware suppliers. That is the consensus read, and it is not wrong. It is simply incomplete to the point of being misleading. The interesting content is in what the headline omits: the counterparties, the tenor structure, the constraints, and the original source of the number itself.
I want to flag my priors here, because they shape everything that follows. I came into crypto through derivatives, not through ideology. My first serious piece of work was a forty-page internal white paper during the dYdX perpetual swap beta, arguing that order-book centralization was the only viable path for institutional capital, precisely because fragmented liquidity destroys price discovery. That experience taught me to distrust any structure whose economics only work when you stop asking where the money actually comes from. When I later watched Terra's algorithmic stablecoin unwind, I saw the same pattern at civilizational scale: recursive claims dressed as reserves, sustained by sentiment until sentiment broke. And when I built out our desk's AI-and-crypto coverage in 2025, analyzing decentralized compute markets like Render and Akash, I learned that the compute layer is where the real economics live—and where the real obfuscation hides.
So when a $517 billion figure appears without a named counterparty or a confirmed structure, I do not treat it as a fact. I treat it as a signal about what someone wants the market to believe.
Core
Let me start with the arithmetic, because the arithmetic is the whole argument.
A $517 billion commitment over ten years implies roughly $51.7 billion in annual procurement. That is not a rounding error; it is a structural claim about scale that Anthropic's revenue cannot currently support.
We do not need exact financials to see the gap. Public reporting places Anthropic's annualized revenue somewhere between a few hundred million and a few billion dollars across 2024, with more aggressive estimates reaching the low tens of billions in 2025 as enterprise API and Claude subscription revenues scaled. Even taking the most generous figure—say $10 billion annualized—the ratio between procurement commitment and revenue is roughly five to one. At the more conservative end, it approaches fifty to one.
No company commits to buying five times its revenue in inputs on a recurring basis using its own cash flow. That is not how capital structures work. It is how leverage works, how vendor financing works, and how circular transactions work. The question is which of these is operating here.
There are three candidate structures, and they have radically different implications.
The first is a genuine take-or-pay contract. In this reading, Anthropic has signed a binding commitment to purchase a minimum volume of cloud and compute capacity regardless of whether its business materializes. This would be a hard liability, a fixed cost that sits above the operating line, and it would require massive external financing—debt, structured equity, or guarantees from strategic investors.
The second is a procurement ceiling or framework agreement. Here the $517 billion represents the maximum value the relationship could reach if every option were exercised at full volume. This is how large enterprise agreements are often described in press materials: the headline number is the notional maximum, the actual spend depends on usage. This structure is commercially weak and mostly a marketing artifact, though it flatters the supplier's backlog.
The third, and in my view the most likely, is a hybrid that blends commitment, ceiling, and circularity. Anthropic commits to consuming capacity. The cloud providers extend credit, equity investment, or volume discounts. The providers then book the anticipated spend as backlog or future revenue, which supports their own capital expenditure plans, which in turn expands the capacity Anthropic is expected to consume. The loop closes—and the loop is the point.
Note: Sentiment turning bearish on L2s. I raise that for a reason. Crypto already ran this experiment. The recursive financing pattern we are now watching in AI infrastructure is structurally identical to the reflexive loops that defined DeFi's worst excesses and Terra's terminal spiral: claims that appear to be external revenue but are actually reshuffled internal capital.
Let me be precise about the mechanics, because precision is what separates analysis from panic. In a circular structure, the same dollar does the work of several. Provider A invests in Company B. Company B commits to purchase $X of services from Provider A. Provider A recognizes $X as contracted revenue and cites it as evidence of demand to justify $Y of capital expenditure. Provider A's stock rises. Company B, now flush with an inflated valuation and access to cheaper capital thanks to Provider A's support, commits to purchase more. Each leg is defensible in isolation. The system as a whole is a hall of mirrors.
This is not speculation about Anthropic specifically. It is the standard anatomy of vendor-financed growth at scale, and it has appeared in telecom, in solar manufacturing, and most recently in crypto. The reason it deserves scrutiny in AI is the sheer magnitude of the numbers involved. When the numbers are small, circularity is a curiosity. When they approach sovereign-debt thresholds, it becomes a systemic question.
Consider the capital-expenditure benchmark. The four largest hyperscalers have been running combined annual capital expenditures in the range of $150 billion to $250 billion in recent years, with AI infrastructure comprising a rising share. Microsoft, Google, Amazon, and Meta each operate at scales where individual annual capex can flirt with $50 billion or more. Against that backdrop, a single AI lab committing roughly $51.7 billion a year to cloud and compute would make it one of the largest AI compute buyers on the planet—larger, on a procurement basis, than most of the hyperscalers' individual AI budgets.
That comparison is the locus of the anomaly. It is possible. It is not impossible. But it requires a financial structure that, if it exists, nobody has described in public.
Let me now decompose the compute side, because this is where my crypto background becomes an asset rather than a liability.
Compute procurement breaks into two distinct workloads: training and inference. They have opposite economic profiles. Training is lumpy, capital-intensive, and concentrated—you build a cluster, you run it hard, you burn enormous power, and you produce a frontier model whose value is uncertain until it ships. Inference is continuous, distributed, and margin-sensitive—the cost of every API call, every agentic tool invocation, every long-context request, sits directly against revenue. As Claude's enterprise adoption grows, inference becomes the dominant cost line, and inference cost is the single largest determinant of gross margin in the API business.
If the $517 billion skews toward inference, the commitment is essentially a bet that Anthropic's API demand will grow fast enough to absorb capacity. If it skews toward training, it is a bet on the continued payoff of scaling—the hypothesis that more compute reliably produces more capability. Both bets are defensible. Neither is guaranteed. And notably, neither is an architecture-level innovation. A nine-figure-per-year compute commitment is a capital expression of the scaling thesis, not evidence of a breakthrough. It tells us where Anthropic is placing financial weight, not where it is pushing technical frontiers.
The chip mix matters for the same reason. If the deals cover NVIDIA GPUs alongside Google TPUs and AWS Trainium, the subtext is diversification away from a single vendor's pricing power. That is infrastructure engineering, not model architecture. It is also strategically loaded: adopting Trainium or TPU at scale means internalizing the quirks of less mature software ecosystems, accepting lower developer familiarity, and trading flexibility for cost. NVIDIA's dominance in training is not an accident of marketing; it is a function of a decade of tooling, libraries, and talent. Breaking that moat is a multi-year, capital-intensive project, and a procurement commitment is one of the few levers that can force it.
I want to bring in the crypto parallel explicitly here, because it is where I add information gain rather than reciting the news.
The decentralized compute markets I covered throughout 2025—Render, Akash, and their peers—exist precisely because centralized compute is expensive, opaque, and politically concentrated. The pitch was always that idle GPU capacity across the world could be aggregated into a cheaper, permissionless alternative. The reality, as I wrote then, is that those networks are excellent for certain workloads—rendering, batch inference, censorship-resistant jobs—and structurally unsuited for frontier training, which demands extreme interconnect bandwidth, tightly coupled memory, and fault-tolerant scheduling that distributed consumer hardware cannot deliver.
That distinction is now the most important frame in the entire AI-compute debate. The $517 billion commitment is a bet that centralized, hyperscale compute remains the only viable substrate for frontier work—and that bet is exactly what the decentralized compute thesis has been shorting. If Anthropic's number is real and the capacity is actually consumed, it validates the centralization of the intelligence layer. If the number is inflated, it reveals that the apparent demand was partly manufactured by the same reflexive loops that decentralized networks were supposed to route around.
Either way, the crypto market is mispricing the signal. Compute is becoming the reserve asset of the AI economy. Whoever controls it controls the cost structure of every downstream application—exactly the way whoever controls settlement controls the economics of DeFi. And just as I argued during the dYdX audit that liquidity depth beats speculative hype, the compute layer is where durability lives and where the token games are most likely to be theater.
Now let me address the financing, because this is where the structure either holds or collapses under inspection.
A commitment this size has four plausible funding sources, and they are not mutually exclusive. The first is equity—diluting existing shareholders through successive mega-rounds, which is precisely what Anthropic has been doing, with valuations climbing from tens of billions to hundreds of billions across 2023 to 2025. The second is debt, including convertible instruments, structured notes, and private credit, all of which become available when a company has strategic backers willing to provide implicit or explicit support. The third is vendor credit—cloud providers extending favorable terms, credits, and capacity guarantees in exchange for commitment. The fourth is circular—investment flowing into Anthropic that is effectively recycled back to the investors as procurement revenue.
Of these, vendor credit and circularity are the ones that most deserve scrutiny, because they are the least transparent and the most flattering to reported metrics. A cloud provider that invests $X in a lab and then books $X in future committed revenue has not created $2X of value. It has created one dollar of capital and one dollar of contingent revenue, and the market is free to double-count them.
The tell is in how these numbers get reported. When a hyperscaler cites a rising "remaining performance obligation" or backlog number, the market reads it as demand visibility. When a lab cites a multi-year compute agreement, the market reads it as commitment strength. Neither party is lying. They are simply describing the same transaction from two ends, and the aggregate impression is larger than the underlying reality.
I have seen this exact pattern in crypto. During the 2021 cycle, exchanges, market makers, and token issuers engaged in tightly coupled arrangements where investment, market-making, and listing revenue formed a closed loop. The metric that measured "ecosystem growth" was partly the metric of the loop. When the loop unwound, the growth evaporated overnight. The lesson is not that loops are fraudulent. It is that loops are fragile, and that their fragility is invisible as long as the direction of capital flow stays positive.
Which brings me to the benchmark that matters most: power.
Chips are a manufactured constraint; electricity is a physical one. A commitment of this magnitude, if genuinely consumed, translates into gigawatt-scale incremental demand. Datacenters at that scale require land, grid interconnection, water for cooling, and years of lead time. The permitting and interconnection queues in major markets are already stretched. This is the real bottleneck, and it is the one that determines whether a contract is executable or aspirational.
Here again the crypto analogy is instructive. When DeFi projects promised transaction capacity they could not deliver, throughput became the constraint that exposed the gap between narrative and reality—the same way oracle feed latency, the Achilles' heel I have written about for years, quietly determines what is actually possible versus what is marketed. In AI, the equivalent is power delivery. Anthropic can sign any number on paper. It cannot sign for electrons that the grid does not have.
If the $517 billion figure is real and binding, the practical constraint is that the capacity may simply not exist on the timeline claimed, which means the contract is either a ceiling with flexible drawdown or a headline with limited operational teeth. If the figure is a ceiling, then its commercial significance is far smaller than the coverage implies. Either way, the number as reported does not survive contact with physical reality without qualification.
Let me turn to the competitive landscape, because the strategic logic only makes sense in context.
Anthropic is not committing this capital in a vacuum. It is competing against OpenAI, which has its own hyperscale partners and its own multi-hundred-billion-dollar infrastructure ambitions; against Google, which owns its own silicon, cloud, and models; against Meta, which builds its own clusters and gives away its models to commoditize the layer above; and against xAI, which has pursued a build-it-yourself approach. In that field, Anthropic's differentiator is not raw capability—all frontier labs trade blows. It is distribution and safety positioning: enterprise relationships, cloud marketplaces, and a reputation for alignment that appeals to regulated buyers.
A mega-commitment to compute supports that positioning by guaranteeing capacity—which is a real advantage when frontier training runs are gated by cluster availability. But it also deepens dependence on the same two partners that are simultaneously its investors, suppliers, and competitors. That is a governance knot, not just a commercial one. When your largest investors also set the price of your most important input, your negotiating leverage is a function of their goodwill, not your market position.
There is also an antitrust dimension that the coverage ignores entirely. Cloud providers that invest in a customer and then capture that customer's procurement are creating exactly the kind of vertically integrated arrangement that regulators scrutinize. If commitments include exclusivity or minimum-purchase clauses tied to a specific provider's silicon, the arrangement edges toward a de facto tie that competition authorities may find interesting. I would not be surprised to see this structure attract attention in the EU or the US within the next eighteen months, precisely because the amounts are now large enough to matter to national compute policy.
Let me consolidate the investment implications, because that is what a market brief is ultimately for.
The value capture in this transaction is almost entirely on the supply side. If the deal lands, the clearest beneficiaries are the cloud providers booking locked-in revenue, the chip vendors filling orders, the datacenter operators signing long-term leases, and the utilities and cooling suppliers whose demand curves shift upward. The lab itself captures optionality and capacity, but at the cost of financial inflexibility. Downstream application companies capture nothing except a higher cost of goods and a steeper competitive moat to climb.
That asymmetry is the thing most readers miss. The headline frames the deal as Anthropic's bold move. Structurally, it is a transfer of value toward infrastructure—which is exactly what happened in crypto when the infrastructure layer captured the economics and the application layer bled margin. In both cases, the durable position is the layer with physical or settlement scarcity. In AI, that layer is compute and power. In crypto, it is liquidity and settlement. Everything else is a rent payer.
For traders, the practical read is that the AI infrastructure trade—chips, datacenters, power, cooling, networking—retains the clearest fundamental support as long as commitments like this keep being signed, even if the underlying lab economics are shakier than advertised. The risk is a sentiment reversal, and sentiment reversals in capital-intensive sectors are violent because they force the cancellation of exactly the backlog that supported the rally. This is the same dynamic as the L2 narrative decay I have been flagging. Note: Sentiment turning bearish on L2s. The lesson generalizes: when a sector's valuation rests on promised future consumption rather than realized demand, it is repriced quickly once the promises stop compounding.
Let me also address the counterparty risk that nobody is pricing. If a meaningful portion of the commitment is vendor-financed or circular, then the cloud providers are, in effect, lending to a customer with limited independent cash flow. That is a credit exposure dressed as revenue. In a downturn—say, a funding-market freeze or a slowdown in enterprise AI adoption—the pressure runs in both directions at once: the lab cannot pay, and the provider loses both the revenue and the value of its equity stake. The correlations stack instead of diversifying. That is the definition of a fragile structure.
I will state my core assessment plainly, because hedging adds nothing here.
The $517 billion figure is almost certainly not a committed cash outflow. It is most plausibly a multi-year framework with a headline number that conflates ceiling, commitment, and circular investment. Treating it as a fixed liability overstates the risk; treating it as pure marketing understates the signal. The truth is that it reveals the direction of capital—toward compute, power, and the suppliers of both—and the fragility of the financing behind that direction.
Contrarian
The prevailing narrative is that this is a story about Anthropic's ambition and the validation of AI infrastructure demand. I think that is the wrong frame, and the wrong frame is hiding the most important fact.
The contrarian read is this: the most important output of a $517 billion deal is not the capacity it buys. It is the permission it grants to everyone else to spend. When a headline of this size circulates, it becomes a reference point that hyperscalers cite to justify their own capital expenditure, that startups cite to justify their valuations, and that investors cite to justify their exposure. The number functions less as a contract and more as a coordination device—a Schelling point around which the entire AI capital stack recalibrates upward.
That is a subtle and dangerous mechanism. In crypto, the same dynamic played out with total-value-locked metrics, which became coordination devices for an entire ecosystem. Everyone pointed to the aggregate number to justify their own position, and the aggregate number was itself assembled from the positions it justified. AI compute commitments are now performing that identical function—except the amounts involved are three to four orders of magnitude larger, and the physical resource constraints are real rather than notional.
The blind spot in the bullish case is that it treats demand as evidenced by commitment. But in a circular, vendor-financed, headline-driven market, commitment is not evidence of demand. It is evidence of coordination. And coordination is indistinguishable from demand until the moment it breaks.
Note: Sentiment turning bearish on L2s. The L2 market taught the same lesson in miniature. A proliferation of promised throughput and announced partnerships drove valuations, and the underlying usage never validated the aggregate. The correction came not when a single project failed but when the market realized the commitments were mutually referential. AI compute, at a vastly larger scale, is now running the same play.
There is a second blind spot worth naming: the assumption that scarcity of compute guarantees the economics of compute buyers. It does not. Scarcity guarantees the economics of compute sellers. A lab that locks in capacity at favorable terms in a tight market is making a smart trade—until the market loosens, at which point it is locked into above-market pricing while its competitors buy at spot. The commitment is a hedge against scarcity that becomes a liability in abundance. And abundance is exactly what every hyperscaler's capex plan is currently engineering toward.
Takeaway
So where does this leave the reader who has to position, not pontificate?
Watch the structure, not the number. The next disclosure that matters is not another headline figure—it is the first official filing that distinguishes committed spend from ceiling, and cash from circular. That filing will reprice the entire AI infrastructure complex in one direction or the other. Until it appears, treat every nine-figure-per-year commitment as a coordination device first and a contract second.
And ask the question the coverage refuses to ask: if the demand for frontier compute were as self-evidently real as the bulls claim, why would it need a $517 billion signpost to be believed?
The answer is the trade.