Capex Is the New Tokenomics: AWS, Anthropic, and the AI Liquidity Mirage
In the first week of May 2025, Amazon's stock posted its largest single-day gain in over a decade. Fifteen percent, gone vertical in a single session. The trigger was AWS. Its cloud division reported an annualized revenue run rate above $115 billion, an operating margin holding near 37 percent, and a generative AI business compounding at triple-digit rates. Management, in the same breath, raised the 2025 capital expenditure forecast to $145–160 billion. The market's verdict was instantaneous: validation. AI capital expenditure, the story went, has graduated from faith-based investment to profit-producing reality.
I have audited this geometry before.
In 2018, I spent four hundred hours inside EtherDelta's source code and found an integer overflow that no stress test had surfaced — one arithmetic flaw that could have turned a running exchange into a drain. That vulnerability was a single faulty line inside a trading engine. The AWS trade is the same kind of arithmetic flaw, inflated to market scale: analysts are summing revenue, contractual commitments, and forward expectations into one figure and labeling it “AI validation.” The inputs are not wrong. The input-output mapping is.
The code doesn't lie. Contracts do.
Business context first. Andy Jassy, Amazon's CEO, called AI the largest technological shift since cloud computing itself, framed it as a hundred-billion-dollar revenue opportunity, and disclosed that AWS's AI business was growing at triple-digit year-over-year percentages. More importantly, he named the constraint. It was not demand. It was supply.
“We don't have enough accelerator capacity to meet customer needs,” Jassy told the earnings call — the kind of statement that changes how infrastructure is read in a market cycle. Demand vacuums are promotional fictions. Supply constraints are physical facts. When a vendor says the problem is chips and not customers, the market is entitled to believe customers exist. And the market did believe: the stock's reaction pushed AWS's valuation logic from “mature cloud franchise” to “core AI infrastructure asset” in five trading hours.
The backdrop is enormous. Microsoft, Google, and Amazon — the three hyperscalers — are projected to deploy more than three hundred billion dollars in combined capital expenditure in 2025, most of it aimed at AI compute. That scale forced an unusually binary market argument: bubble versus trend. Bulls read AWS's simultaneous revenue acceleration and margin expansion as the strongest evidence yet that the trend thesis is correct. Bears read the same numbers as a six-year overbuild waiting for a single credit cycle to break it. The AWS print, on its surface, gave the bulls their proof.
But price signals are not proofs. I learned that watching the DeFi cycle — specifically, the way a token's price ratcheted upward as total value locked climbed, with that TVL itself fed by emissions of the same token. The loop closed cleanly on paper. It survived for multiple quarters. It did not survive contact with one question: who wants this, and are they paying with capital that came from outside the loop?
I intend to ask that same question of AWS's AI revenue. Layer by layer.
The first item on the audit schedule is composition. AWS does not disclose how its AI revenue splits between two very different categories: contractual obligations from strategic partners, and organic consumption from enterprises running workloads on Bedrock, SageMaker, or other AI services. The distinction is not academic. It is the difference between revenue and invoicing.
The largest known contributor to the first category is Anthropic. The model company has committed to spend tens of billions of dollars on AWS compute over a multi-year period. That commitment was foundational for both sides: it gave Anthropic guaranteed capacity without owning a data center, and it gave AWS a contracted floor under its AI revenue line. But a commitment is not consumption. It is a liability until drawn down. It does not prove that enterprise users want AI inference at scale. It proves that a well-capitalized model lab chose a financing structure.
In DeFi, we called this liquidity mining: paying for activity metrics. The metrics were real. The activity was circular. In early 2022, I analyzed three lending platforms that had borrowed short to lend long, and I published a model forecasting a 30 percent drawdown in total value locked within six weeks. Their revenue charts looked excellent right up until the moment borrowers stopped rolling positions. The TVL was real. The collateral underneath it was not.
The equivalent question for AWS: how much of that triple-digit growth is Anthropic's committed drawdown, and how much is a mid-sized enterprise discovering that Bedrock is cheaper than standing up its own cluster? Growth from a contracted base is not growth. It is invoicing. The accelerator shortage Jassy cited — which made capacity itself the product — actually favors the contractual model, because it permits AWS to sell compute at premium rates to whoever signs first. But it also means the revenue curve is, in part, a function of contract signing dates rather than genuine workload emergence. Aave and Compound's interest-rate curves are arbitrary in the same way: they track internal parameters, not real supply and demand. AWS's AI growth is priced as a demand signal. Structurally, it is closer to an amortization schedule.
The second item is the loop itself. AWS's validation triggered a stock rise. The stock rise validated the capital expenditure. The capital expenditure builds capacity. The capacity books future revenue. That revenue — if it arrives — validates the stock again. This is tokenomics. It has the same mathematical anatomy as the flywheel that drove DeFi's most spectacular collapses: token price inflates the treasury, the treasury emits more tokens, the emissions inflate TVL, and the TVL inflates the price.
The AWS version of this loop is not fraudulent. It is structurally analogous. The difference is that AWS has a real, 37-percent-margin legacy business underneath the flywheel — actual customers generating actual cash. That is an edge no DeFi protocol ever possessed. But it also means the market's confidence in the AI narrative can borrow against that edge, and that leverage is exactly where the danger lives.
Consider the margin arithmetic. AWS reported an operating margin near 37 percent, and the market treated it as proof that AI infrastructure pays for itself. But there is a delay between capital expenditure and depreciation. The $145–160 billion capex program will weigh on income statements for years. The margin on existing, largely depreciated data centers is flattering the blended number. The margin on new accelerator builds — purchased at premium prices, commissioned into a market where inference pricing is falling — is the number that matters, and it is not disclosed.
This is how I read protocol exploits too. A system shows a beautiful steady-state economics model, and the exploit lives in the transition — in the period when the system grows faster than its accounting can track. A 50 percent year-over-year increase in AI workload volume, combined with a 50 percent year-over-year decline in inference unit prices, produces a revenue curve that is already a miracle of volume. The market saw the headline revenue acceleration. Nobody audited the gap between capex and ramp.
The third item is hardware. AWS's margin story depends on a piece of information Amazon does not provide: the deployment scale of its in-house Trainium and Inferentia accelerators. The logic is inescapable. If AWS ran inference purely on NVIDIA GPUs, the cost structure would be brutal. The 37 percent operating margin could not survive a market where compute is rented at falling gpu-hour prices. So the margin is either a signal that custom silicon is carrying a meaningful share of inference workloads, or a signal that the margin will erode over the next two years as the GPU-heavy capex cycle hits depreciation. Both readings are bearish for the narrative that the margin will hold.
In early 2025, I co-audited a zero-knowledge proof system designed to verify AI inference. The hardest part was not the cryptography. It was the hardware efficiency claims. Every vendor has a benchmark. Few vendors disclose the utilization rates behind the benchmark. One team's “10x improvement” turned out to rely on speculative execution gains that required a memory layout no production system could sustain. The lesson stuck: when a company withholds its hardware mix, the margin is the closest thing to a confession.
AWS's own engineering output tells the same story obliquely. The company publishes material on quantization, speculative sampling, KV-cache optimization, and batch inference — all aimed at one objective: making inference cheaper. This is excellent engineering. It is also a warning. Every such publication is a memo to the market that the unit price of AI completions is about to fall. AWS's strategy is not to compete on frontier model quality. It is to win the cost curve, the way Amazon always has. That wins customer lock-in. It does not maximize revenue per unit. The bottleneck isn't the infrastructure. It's the accounting.
The final item is centralization. This is where a decade in crypto has ruined me for optimism.
After the fourth Bitcoin halving, miner revenue collapsed and hash power consolidated toward a handful of pools. The decentralization consensus turned out to be a narrative wrapper over a physical reality: whoever owns the cheapest electricity and the fittest hardware writes history. The same logic now governs AI. The hyperscalers plus NVIDIA are the three largest pools. Their capital expenditure is locking up a significant share of global advanced-chip capacity, which systematically raises every other entrant's marginal cost of compute. Independent model labs — those not bound to a cloud provider — face a double bind: they compete for venture capital against the cloud providers' own AI bets, and they buy compute at prices the hyperscalers largely set.
The industry analysis I read frames AWS's strategy as a “neutral multi-model platform” — Bedrock offering Anthropic, Meta, Mistral, and Amazon's own models side by side. That neutrality has a name in crypto: a fee market. Offering many models is not decentralization. It is a toll booth with a choice of lanes. The DAO governance debates taught me that “code is law” is fiction when three multisig signers hold the upgrade keys. AI's equivalent fiction is the belief that model choice protects users, when the platform controlling accelerator allocation holds the more decisive switch.
I am not arguing that AI is a bubble. I am arguing that the word “validation” is doing too much work. What the market validated was a structure — one where the largest winners are infrastructure providers, where the marginal model lab is a cost center with pricing-taker positioning, and where the decentralization rhetoric crypto once exported is nowhere to be found. The market did not validate organic demand. It validated a toll road.
Now the blind spots. The first is power. Chip supply improved through 2025 as NVIDIA's production ramped, but the physical complex around it did not move at the same speed. Grid interconnection queues, substation construction timelines, and power purchase agreements are now the true constraint on AI capacity in the regions where hyperscalers actually build — Northern Virginia, Oregon, the Ohio Valley. Three hundred billion dollars in capex can buy GPUs. It cannot buy a faster grid queue. And the marginal data center that powers up in 2027 will be paying premium prices for electricity that no current margin model includes.
The second blind spot is the loop's fragility. The self-reinforcing dynamic I described is not a perpetual motion machine. It depends on continuous positive feedback. A single quarter where AI revenue misses — or where Anthropic renegotiates its commitment at renewal time — does not merely dent the multiple. It breaks the loop. Markets that price a compounding story do not correct gradually. They unwind.
And then there is the dependency the sources barely mention: the AI startup financing cycle itself. Cloud AI revenue is partly funded by venture capital flowing into model companies, which then purchase cloud capacity. If that pipeline stalls — if AI startup funding tightens — the contractual commitments get drawn down more slowly, and the “validated” revenue curve flattens into what it always was: deferred capex with a nice label. I have run enough audits to know the difference between a well-capitalized project and a well-capitalized story. The code doesn't lie. But the market often does — to itself.
The market did not validate AI fundamentals. It validated a financing structure: hyperscaler capex, fed by committed contracts, supported by the assumption that enterprise consumption will eventually fill the void. Watch the next two quarters. If AWS begins disclosing a committed-versus-consumed split, read it the way you would read a proof-of-reserves. If Anthropic renegotiates on renewal, treat it as a canary. And if power-constrained data-center buildouts slip by six months, that is not a supply-chain footnote. That is the physical ceiling beneath the narrative.
AI's winter will not arrive because models fail. It will arrive because the venture pipeline that funds the commitments stalls, the contracts run out, and the power constraint nobody budgeted turns a growth curve into a wall.
Resilience isn't audited in the winter. It's revealed there.