Hook
$517 billion. Ten years. That is the reported aggregate compute commitment Anthropic has assembled across Amazon, Google, and β since November β Microsoft and Nvidia. Read the number twice, then read what it is not. It is not a payment. It is not revenue booked by anyone. It is a ceiling on an obligation, denominated in a commodity that depreciates faster than any asset in modern corporate history.
The market treated the headline as a validation event. DePIN tokens ticked up. AI-adjacent tickers traded green. That reaction is the tell: nobody priced the structure. Everyone priced the number.
Context
Anthropic's compute stack is now a three-cloud arrangement with equity attached to each leg. Amazon holds roughly $8 billion of equity and is delivering Project Rainier, a Trainium2 cluster approaching half a million chips. Google, already an investor, signed capacity for up to one million TPUs and more than a gigawatt of power by 2026, described as a multi-tens-of-billions commitment. Microsoft and Nvidia arrived in November with a $30 billion Azure commitment, alongside an Nvidia investment of up to $10 billion and a Microsoft investment of up to $5 billion.
Map the direction of the cash. Equity flows into Anthropic. Anthropic commits to buy compute back from the same parties. That is not a scandal β it is a financing structure, and it is common in capital-intensive industries. But it changes what the revenue means. When a hyperscaler reports AI cloud revenue that originated from a company the hyperscaler funded, free cash flow conversion is not the same as organic demand. This is the exact pattern that made investors nervous about Nvidia's arrangements with OpenAI. Anthropic's version is larger and spread across three balance sheets.
The equity side is the collateral that makes the commitment credible. Anthropic's September 2025 round priced it near $183 billion, with later reporting pointing to conversations at multiples of that. Valuation, not revenue, is currently underwriting the tab.
Core
Here is the arithmetic that matters, and it is not in the press release.
Anthropic's annualized revenue run-rate entered late 2025 in the single-digit billions, with 2026 projections clustered β and they are projections β in the $20 billion to $26 billion range. A $517 billion commitment over ten years implies roughly $50 billion per year of compute spend. For most of that horizon, the obligation exceeds the revenue that would fund it. That gap is the actual story.
Three things follow.
First, the commitment is not a liability in the accounting sense, but it is a liability in the economic sense. If Anthropic's growth stalls, the capacity does not disappear β it is still contracted, or it is renegotiated at a penalty. Take-or-pay structures exist precisely because someone has to eat unused capacity. I have watched this movie before. In 2022, the algorithmic stablecoin complex collapsed not because the peg math was wrong, but because the collateral backing the promise had a duration mismatch with the promise itself. Long-dated commitments funded by short-dated revenue is the same disease wearing different clothes. Strategy is the art of surviving your own leverage β and a decade-long compute commitment is leverage, whether or not it sits on a balance sheet.
Second, depreciation. Hyperscaler AI capex is depreciated over roughly five to six years. Frontier training hardware is economically obsolete far sooner than that β not physically dead, but irrelevant for the workloads that command premium pricing. If a meaningful share of the reported $517 billion is recognized at list prices on hardware with a shorter true economic life, the reported revenue overstates the durability of the underlying economics. A compute commitment is a forward on a depreciating asset, and the depreciation curve, not the headline price, determines whether the deal was a good one.
Third β and this is where crypto gets it wrong β the order flow never touched the chain. Look at where the dollars land: AWS capex, Google Cloud capex, Azure capex. Compare that to the market caps of decentralized compute networks, which repriced on the same narrative. Across the twelve months to early 2025, hyperscaler AI capex guidance moved from roughly $150 billion toward $400 billion annually, while the aggregate market cap of decentralized compute tokens stayed in the low single-digit billions. That ratio is the signal. I have tracked GPU utilization dashboards for decentralized compute since early 2025. Utilization was real, growing, and almost entirely inference and fine-tuning. Frontier pretraining of the kind Anthropic is buying never appears on those networks. The tokens captured the narrative. The hyperscalers captured the cash flow. Arbitrage is just patience wearing a math mask β here, between a narrative traded on-chain and an order book settled in data center interconnect.
Contrarian
The consensus bull case for decentralized compute is that scale will close the gap. It will not, and the reason is physically specific: interconnect bandwidth, not raw FLOPs.
A frontier training cluster is not a pile of independent accelerators. It is a fabric. Google's TPU pods and Nvidia's rack-scale systems move data at intersocket bandwidth measured in hundreds of gigabytes per second with microsecond latency. A thousand distributed GPUs connected by best-effort consumer internet are not a small training cluster β they are a thousand separate computers that happen to run the same code. Checkpointing overhead destroys the training loop long before compute becomes the binding constraint. This is not a hardware problem three years of progress solves. It is a topology problem.
The second blind spot is liquidity. Everyone is talking about tokenizing compute. I have looked at what a compute contract actually is: a non-transferable, capacity-constrained, counterparty-specific right with delivery windows and SLAs. There is no secondary market. There is no depth. Value it the way I value any illiquid claim β at a discount to the nearest liquid substitute, and the nearest liquid substitute is renting spot capacity on the same clouds at market rates. Liquidity doesn't forgive. If you cannot exit the claim in a week without moving the price, you do not own an asset. You own a commitment β someone else's.
Takeaway
The number to watch is not $517 billion. It is the slope of Anthropic's revenue against the slope of its commitment schedule, disclosed quarterly through hyperscaler capex guides and depreciation lines. Two of those three curves are public. The third is not, and that asymmetry is where the risk sits. Impermanence is the only permanent yield β and in AI infrastructure, the impermanence is the hardware, while the contract is forever.