Dario Amodei has raised roughly $7.5 billion across multiple funding rounds and has never posted a profitable quarter. That is not an accusation. It is an accounting fact. So when the CEO of Anthropic is reported to have questioned the AI industry's spending narrative, the correct first move is not to quote him. It is to check the ledger on both sides of the sentence.
Here is what the reporting actually delivered: three bullet points, no verbatim quotes, no venue, no transcript, no context. A crypto vertical covering an AI capital-expenditure story, which means the signal arrived downstream of at least one translation layer and probably two. The raw material was thin enough that any honest analyst would flag it as unusable for decision-making.
That should bother you more than the spending number does.
Because the real story is not that Amodei doubts the capex cycle. The real story is that an asset class worth tens of billions of dollars — the AI-token complex — is currently underwritten by second-hand paraphrases of what technology executives say in interviews. That is a narrative supply chain, and it has no audit trail. No signed receipts. No provenance. Just a headline, a screenshot, and a bid.
Audit the code, not the pitch.
Context: Two Balance Sheets, One Borrowed Story
Anthropic was founded in 2021 by a group of former OpenAI researchers, with AI safety as its stated differentiator. Its technical signature is Constitutional AI — a training approach that substitutes a written rule set for a portion of human preference labeling. Its commercial signature is Claude, a frontier model family that has earned real enterprise adoption without ever displacing GPT-4's default status.
The financing record matters more than the product record for this discussion. Google committed roughly $2 billion in 2023. Amazon committed up to $4 billion in the same window, tied to AWS as the training substrate. A 2025 Series E valued the company near $61.5 billion. Later rounds pushed the headline valuation into the $180 billion range. Cumulative capital raised sits in the neighborhood of $7.5 billion, with substantially larger rounds reported in negotiation.
Amodei's public statements have been consistent for over a year. He has described a race to the top in training spend. He has floated the possibility of individual model training runs costing $100 billion before the end of the decade. He has said the situation worries him. That is the context the crypto coverage collapsed into three bullets. A CEO who is simultaneously a beneficiary and a critic of the capex cycle made a nuanced statement, and the nuance did not survive the trip.
I have watched this pattern for twenty-seven years. Every cycle, a genuinely difficult technology gets reduced to a ticker. In 2017 it was sharding and throughput. In 2020 it was yield and composability. In 2021 it was JPEGs and utility. In 2024 and 2025 it is GPUs and inference. The technology changes. The narrative arbitrage does not.
On this side of the fence, the AI-token complex has spent two years borrowing the credibility of that same capex cycle. Bittensor's TAO. Render's RNDR. Akash's AKT. The Fetch/ASI merger. io.net. Grass. Dozens of smaller GPU-marketplace tokens with names that sound like datacenter hardware and whitepapers that read like pitch decks. Aggregate market capitalization has swung between roughly $20 billion and $50 billion depending on the month, with the top clustering around peak AI sentiment.
The bridge concept is decentralized compute. The pitch is that idle GPUs around the world can be aggregated into a training and inference substrate that competes with hyperscale clouds, and that the token is the coordination and payment layer for that substrate.
That pitch deserves the same scrutiny Amodei is applying to his own industry. So let me apply it.
Core: A Systematic Teardown
Start with the arithmetic that makes the AI spending debate real, because it is also the arithmetic that makes most AI-token claims impossible.
Frontier model training is measured in floating-point operations. A top-tier model trained in 2024 sits somewhere between 10^25 and 10^26 FLOPs. The hardware that produces those operations is a cluster of tens of thousands of accelerators — H100s, H200s, B200s — each costing between $25,000 and $40,000, wired together with specialized interconnect, drawing power at the scale of a small city, and cooled with industrial infrastructure. The capex for a single frontier training campus now runs into the billions.
The cost structure is not like software. In classical SaaS, the marginal cost of serving one more user is close to zero. That is where the 80% gross margin came from, and it is the reason software companies got rewarded for growth. AI breaks the pattern.
AI is the first software business whose gross margin deteriorates as adoption increases.
Every additional token of inference consumes GPU-seconds. Every additional user consumes more GPU-seconds than the last, because usage scales with value delivered. There is no version of the business where serving more customers costs nothing. There is only a version where cost per token falls faster than volume rises. If it does not, the company is buying revenue with capex, and the capex is funded by equity holders who have not yet been told.
This is why the capex debate is not a debate about sentiment. It is a debate about whether the cost curve bends in time. And it is exactly why crypto's AI tokens, which sell a story about compute as an abundant, commoditized, near-free resource, are exposed. If compute were near-free, Anthropic would not need $7.5 billion.
Now examine the compute claim itself.
There are three classes of machine learning workload, and they have radically different network requirements.
The first class is embarrassingly parallel batch work. Rendering frames. Transcoding video. Fine-tuning a 7B model on a private dataset. Running inference on a small open-weight model. These workloads tolerate high latency, variable throughput, and node churn. They can run on a consumer GPU or across ten thousand GPUs in a dozen countries without anyone noticing the difference.
The second class is latency-sensitive inference. Chat completions, agent loops, real-time translation. These need predictable response times — usually under a second to first token, often under 200 milliseconds. That requires geographic proximity and stable channel quality, not raw aggregate FLOPs.
The third class is frontier pre-training. And this is where the pitch collapses.
Pre-training a frontier model is a synchronous collective operation. Every training step requires every GPU in the cluster to compute its gradients, then participate in an all-reduce — a collective communication pattern where all ranks exchange and sum their gradients before the next step can begin. That all-reduce happens thousands of times per training run.
At that scale, the binding constraint is not compute. It is interconnect bandwidth and latency.
Inside an NVIDIA NVLink domain, GPUs exchange data at roughly 900 gigabytes per second with microsecond latency. Between nodes inside a datacenter, InfiniBand moves data at 400 gigabits per second with latencies in the low microseconds. Between a GPU in Frankfurt and a GPU in São Paulo, over the public internet, you get maybe a gigabit per second on a good day, with 200 milliseconds of round-trip latency, plus jitter, plus packet loss, plus whatever the transit provider decides to do at three in the morning.
Run the numbers. A single all-reduce across a geographically distributed cluster with a 200-millisecond RTT takes a full second per round trip at minimum. Thousands of steps. The training run that takes three months in a datacenter takes decades across the internet. And that is before accounting for stragglers — the slowest node gates the entire synchronous step, and in a heterogeneous consumer network, the slowest node is a laptop on hotel Wi-Fi.
There are research directions that relax the synchrony requirement. DiLoCo and its descendants reduce communication frequency by training local copies and averaging periodically. They show real promise. They also degrade model quality at current scales, and nobody has demonstrated them at frontier parameter counts. Hybrid approaches — centralized pre-training, decentralized fine-tuning — are plausible. That is a fine-tuning market, not a training market, and it is a fraction of the total value.
So the honest version of the decentralized-training claim is this: it might work someday, for some model sizes, with significant quality tradeoffs. The dishonest version, which appears on token whitepapers, is that it works now.
Distributing FLOPs is easy. Agreeing on the gradient is hard.
That should sound familiar. It is the same structural problem blockchains solved for state. Sharding is easy; consensus is hard. Any system can split work across machines. The difficulty is making the machines agree on a single correct answer, cheaply, under adversarial conditions. The AI version of that problem has not been solved, and the crypto projects selling it have mostly renamed it rather than solved it.
Which brings us to verification — the actual unsolved problem.
If I rent you a GPU and pay you to compute gradients, how do I know you computed them correctly? In a datacenter, I know because I own the machines and the network. In a permissionless network, I do not. You could return random numbers, or subtly corrupted gradients, or gradients computed on a smaller batch to save power.
You can verify this in three ways, all of them expensive. You can re-run the computation redundantly, which doubles or triples cost, and which fails against colluding redundant nodes. You can use cryptographic proofs of ML work — zkML — which for training-scale operations currently carry overheads in the thousands of times range. Or you can use trusted execution environments, which require specific hardware, still leak side-channel information, and introduce a vendor you have to trust — which is precisely the thing you were trying to avoid.
There is no fourth option. There is no cheap, trustless, general-purpose verification of ML training work. This is not an engineering detail. This is the load-bearing beam of the entire decentralized-compute thesis, and it is missing.
Complexity hides risk, and the missing verification layer is where the risk lives.
Now look at how the tokens actually route around this.
Bittensor is the most intellectually serious of the group, so it deserves the most careful treatment. Its design is a subnet architecture: miners produce model outputs, validators score them, and the protocol emits TAO to participants proportional to their measured contribution. The subnet mechanism is a genuine innovation — an incentive layer that pays for open model quality rather than for token speculation alone.
Here is the forensic question. What fraction of TAO emissions funds work that a paying customer would purchase with cash?
The emission schedule is fixed and continues regardless of downstream demand. That means the network's economics are subsidy-driven. Participants mine because the token has value. The token has value because the network appears to be doing AI work. The AI work is funded by the token. It is a closed loop with no external revenue anchor.
For subnets where the output is a genuinely useful open model, the loop can be defended as public-goods funding. For subnets where the output is a leaderboard position, the loop is a subsidy program denominated in a volatile asset, and the subsidy is paid by whoever bought the token last.
Render and Akash are different. They have real paying customers. Render started in rendering and extended toward inference; Akash runs a reverse auction for containerized GPU workloads. Both have actual revenue.
So do the arithmetic. Take the network's annualized fee revenue. Compare it to the token's fully diluted market cap. The implied multiple will be somewhere between absurd and insulting, especially against listed comparables like CoreWeave, which trades on actual revenue and actual contracted backlog.
A token market cap is not a revenue statement. It is a claim about the future, priced by people who may not have read the past.
Now the structural danger, and this is where my own history is relevant.
In 2022, I spent six months modeling the seigniorage mechanics of UST before the peg broke. I published the analysis months in advance on liquidity-depth metrics, and the on-chain data validated it. The core defect was a circular dependency: the stablecoin's stability was maintained by a sister asset, whose value derived entirely from confidence in the stablecoin's stability. One direction of the loop worked. The other direction worked faster.
Token-funded compute has the same topology.
The project issues a token. It sells the token to buy GPUs. The GPUs deliver services that support the narrative. The narrative supports the token price. The token price funds more GPUs.
Reverse the arrow and the whole thing unwinds. Token price falls, the project cannot fund GPU purchases, service quality degrades, the narrative weakens, the price falls further.
I have watched this movie. The liquidity-depth metric I used on UST applies unchanged here. Ask what percentage of the token's market cap can be liquidated in twenty-four hours without moving the price more than a few percent. If the answer is under one percent, the financing loop is a rounding error away from failure.
Any system that finances its own demand with its own issuance is a seigniorage loop, and seigniorage loops fail in exactly one direction.
Add the depreciation problem, which nobody models. A datacenter-grade GPU has an economic life of roughly three to five years before it is obsolete for frontier work. A decentralized network's node operators amortize that hardware on their own schedules, with their own cost of capital, at their own power prices. When the token reward falls below the operator's marginal cost of electricity and depreciation, the node goes offline. There is no contractual obligation to keep serving. There is no service-level agreement. The network's capacity is a function of a token price, and the token price is a function of the narrative.
Now the stablecoin layer, which is the part of this stack that gets the least attention and deserves the most.
Decentralized compute networks price their services in stablecoins. That is the only sane choice for a marketplace where job costs fluctuate and operators need predictable settlement. But it creates a dependency that the industry has been remarkably quiet about. USDC's compliance-first architecture means Circle can freeze any address on its blacklist, typically within twenty-four hours. For a human user, that is an inconvenience. For an autonomous agent running inference jobs on a decentralized network, it is a kill switch.
Consider the actual flow. An agent takes a job, posts collateral in USDC, receives compute, delivers output, is paid in USDC. If any address in that chain gets flagged — through a sanctions match, a counterparty risk score, or a heuristic that nobody audits — the funds freeze mid-job. The compute was delivered. The payment did not settle. The operator is out the electricity.
This is the quiet absurdity of the entire decentralized compute thesis. The compute layer is permissionless. The payment layer is not. And the payment layer is controlled by a company in Boston that can unilaterally decide your node does not get paid.
The regulatory picture makes this worse, not better.
Under MiCA, a token that exists to pay for compute has an ambiguous classification. It is not a financial instrument in the classic sense. It is not an asset-referenced token, because it does not peg to anything. It could be an e-money token if it functions as payment, which triggers reserve and licensing obligations. It could be a utility token, which is the classification most projects claim and the one that shrinks toward meaninglessness the moment the token is marketed with yield.
Then there is the CASP regime. Providing crypto-asset services in the EU now requires authorization, with capital requirements, governance standards, custody separation, and ongoing reporting. Compliance costs in the mid six figures to low seven figures annually, before legal fees. For a network doing a few million dollars in compute revenue, that is a business-ending line item.
MiCA gives Europe apparent clarity while concentrating the market into whoever can afford the paperwork. Clarity is a public good. Compliance cost is a regressive tax, and it lands hardest on exactly the small, technical teams that were supposed to benefit from regulatory certainty.
And there is a crossover that nobody in either industry has properly priced. The EU AI Act sets a compute threshold — 10^25 FLOPs — above which a general-purpose AI model is presumed to carry systemic risk and attracts additional obligations. The threshold is, in effect, a capex threshold. It only binds firms that spent enough to cross it.
Now consider a decentralized network. Individually, no participant crosses 10^25 FLOPs. Collectively, the network might. Where does the obligation attach? To the protocol? To the subnet? To the largest miner? To the token holder who funded the cluster? The AI Act was drafted for a world of discrete corporate actors with balance sheets. It has no theory of a permissionless network. Neither does MiCA.
That gap is currently filed under unsolved. Unsolved regulatory questions have a way of being resolved retroactively, with penalties.
Finally, the primitive that is actually missing: verifiability of compute.
If a project claims to have delivered a petabyte-scale training run, it should be able to publish signed job receipts — cryptographic attestations that this GPU computed this work at this time, under this operator, for this customer. Proof-of-reserves for compute. Nobody does this. Dashboards show aggregate stats that are self-reported. Whitepapers show benchmarks that are unverifiable. No auditor signs off.
I have audited collateral in DeFi. In 2020 I found an oracle manipulation vector in a Chainlink feed for KNC tokens that forced MakerDAO to adjust thresholds before the exploit could be triggered. The lesson generalizes: if you cannot independently verify the input, you are not pricing the asset. You are pricing the story about the asset.
Trust no one. Verify everything. Including, and especially, the numbers a protocol publishes about itself.
Contrarian: What the Bulls Actually Got Right
The bear case above is not complete without the parts that survive scrutiny.
Compute concentration is a real systemic risk, and inefficiency is sometimes the price of resilience. Three companies control the majority of frontier training capacity. One company controls the accelerator supply. The most advanced fabrication sits in a small number of geographic locations with a geopolitical fuse attached. Any serious supply-chain analyst looks at that and sees a single point of failure. Decentralized compute, even at a large efficiency discount, buys optionality. Options that are never exercised still have value.
The privacy and censorship-resistance angle is not a marketing overlay. Running an open-weight model inside a trusted execution environment on hardware you do not own, without the operator being able to read your prompts, solves a problem no hyperscaler will solve for you, because the hyperscaler's business model depends on seeing your data. That demand exists today, it is legal today, and it is small today. Small and real beats large and fictional.
Idle GPU monetization is genuinely non-zero. There is a long tail of accelerators — consumer 4090s, workstation A6000s, older A100s retired from datacenters — whose opportunity cost in a decentralized marketplace is above zero. Aggregating them is not a frontier training play. It is a spot market, and spot markets have value.
And the point that cuts against everything above: Amodei's critique is also a positioning move. He has raised $7.5 billion and intends to spend more. A CEO who has already signed multi-year compute contracts and locked in supply is precisely the actor who benefits from a narrative in which his spending is rational and other people's is not. "The industry is over-investing" is, from a sufficiently funded incumbent, a moat statement wearing a warning label.
Both things are true. The AI spending narrative is genuinely unstable, and the person pointing at it has the least to lose by pointing.
Takeaway
Here is the one metric I would put on every AI-compute token dashboard, and the one nobody publishes: attested GPU-hours delivered to paying customers in the last quarter, with signed receipts, cross-verified by an independent auditor.
Everything else — emission schedules, subnet leaderboards, benchmark tables, partnership announcements, podcast appearances — is narrative. Narrative does not survive a balance sheet, and it does not survive a freeze order.
The capex cycle will resolve one way or another. If the cost curve bends in time, the deficit spenders win and the crypto rails are a rounding error. If it does not bend, the correction will not spare the protocols that borrowed the narrative without owning the compute.
So the question is not whether Amodei is right. The question is whether any of these networks can show you a signed receipt — or whether the entire sector is priced on a paraphrase of a paraphrase of something a CEO said, once, at a conference nobody attended.
