The message that arrived at midnight
Last week, a former colleague from a market-making desk sent me two lines at midnight Copenhagen time. "Anthropic is profitable. Two quarters running. Eighty percent gross margin." Then a single word: "Finally."
I have spent enough of my working life inside cost structures to flinch at that word. In March 2020, I was organizing rapid-response briefings for anxious DAI holders while the peg wobbled, and I learned that panic is a data point you have to measure before you can calm it. In 2022, I was live-streaming cold-wallet audits to fifty thousand traders who had just discovered how elastic the word "reserves" can be. The lesson from both years is simple and uncomfortable: numbers never lie, but they always answer the question they were asked, and rarely the question you meant to ask.
The Financial Times reported that Anthropic, the company behind Claude, has told a small group of shareholders that it posted adjusted operating profit in two consecutive quarters, with gross margin above 80%. The disclosure was private. The source was anonymous. The valuation marker from early 2024 sits near $60 billion, with roughly $4 billion of Amazon money behind it. And the detail that most coverage folded into a subordinate clause is what that 80% leaves out: the revenue share paid to distribution partners such as Amazon, and the cost of training the models.
That is not a rounding error. That is the entire cost thesis of the AI compute economy, and it is being replicated almost line for line in the token market I watch every day. What arrived at midnight was not really an earnings headline. It was a price tag looking for a buyer.
Why this number surfaced now
The known information set is narrow, and the narrowness matters more than the number itself.
Anthropic builds the Claude family of large language models. Amazon has committed approximately $4 billion and distributes Claude through Bedrock, its managed model service. Google is also an investor. The company's last widely reported valuation marker was around $60 billion in early 2024. It has never published an audited financial statement, because it is not yet a public company, and it has never disclosed revenue, customer counts, or cash burn.
Against that backdrop, "two consecutive quarters of adjusted operating profit, gross margin above 80%" is not primarily an earnings disclosure. It is a financing signal, the kind of carefully worded leak that precedes a roadshow. When a private company shares favorable figures with a small group of shareholders and those figures reach a serious financial newspaper, the purpose is rarely archival. It is to establish a reference price for a conversation that has not happened yet.
I have watched this pattern before from the other side of the table. In 2024, I built a comparative matrix of fifteen custodial providers for institutional advisors who did not trust crypto's own risk disclosures. The hardest question any of them asked was never about technology. It was always the same: who audited this, and what did they leave out? Private, unaudited, selectively disclosed numbers fail that test by construction. They are not false. They are simply unaudited, unaudited numbers get revised, and revisions are where reputations die.
Now place this beside the sector I cover. Crypto's AI compute complex spent 2024 pricing in exactly the narrative Anthropic just told. The pitch was clean and seductive: AI compute demand is exploding, centralized clouds are the bottleneck, and token networks can monetize idle silicon at the edges of the world. Names like Bittensor, Akash, Render, io.net and Aethir became the liquid expression of that thesis. They rallied hard, then gave most of it back as the broader market slid into the sideways chop we are living through now.
So the interesting question is not whether Anthropic is profitable. The interesting question is what its accounting choices reveal about the economics that every decentralized compute protocol is quietly betting its treasury on. And there is a second question underneath it: when the same story is available as equity or as a token, which wrapper does the marginal dollar choose?
The arithmetic of 80%
Let me do the math that the headline invites, using a clean revenue base of one hundred dollars. The point is not the exact figure. The point is the direction and the magnitude, and how quickly a comfortable-looking margin becomes a fragile one.
A gross margin of 80% means cost of goods sold of twenty dollars on one hundred dollars of revenue. In most software businesses, that is unremarkable. Enterprise software companies have historically run 75 to 85 percent gross margins because serving the next customer costs almost nothing. The code is already written, the marginal server bill is trivial, and the business scales without a proportional cost line.
AI is not that business. Every token generated is a slice of GPU time, and GPU time is metered, priced, and finite. Inference, the act of running a trained model to answer a query, is not a fixed cost. It is the most variable cost in the entire stack. When you serve a customer, you pay for electricity, for silicon depreciation, and for the opportunity cost of the GPU you are not renting to somebody else that same second.
So when Anthropic's 80% is described as excluding the cost of training models, that exclusion is doing enormous work. Training a frontier model is a multi-billion-dollar capital event. It is arguably the defining cost of being a model company at all. Remove it from the denominator and you have not computed a gross margin. You have computed a contribution margin, which is a measure of whether the last dollar of revenue covers its own direct costs. That is a weaker and quite different claim.
Then there is the second exclusion: revenue share to distribution partners like Amazon. Distribution through Bedrock is not free distribution. It is paid distribution, typically structured as a percentage of revenue that flows back to the channel. Public terms for this kind of arrangement are rare, but industry norms for managed model marketplaces sit in the mid-teens to low-twenties range. Take a conservative mid-point of roughly 20 percent, and on one hundred dollars of gross bookings, twenty dollars leaves the building before it ever reaches a margin line.
Stack the exclusions and a very different picture emerges. On one hundred dollars of end-customer revenue: subtract a plausible twenty for channel share, subtract the amortized cost of training across a multi-billion-dollar program, and subtract the GPU cost of running inference itself. The wider market's estimate for inference compute as a share of AI company revenue sits somewhere between 30 and 50 percent, depending on model size, quantization strategy, batching efficiency and the mix of cheap and expensive requests. Put those pieces together and a business that prints 80 percent gross margin can plausibly land in the 20 to 40 percent range once the full cost of goods is honored. And that is before stock-based compensation, which adjusted operating profit typically excludes entirely.
Stock-based compensation is not a rounding error either. It is real compensation paid to real engineers in the currency of future dilution. Excluding it from an operating profit figure is common practice in technology, and it is also the single most contested adjustment in modern financial reporting, because it flatters cash generation by deferring the bill to existing shareholders. In a company that competes for scarce machine-learning talent, that bill is not small.
None of this means the number is fake. It means the number is a contribution margin wearing a gross margin's suit, and the suit is the story. When I ran cost audits on token emission schedules, I used a simple discipline: name every cost that would exist if the subsidy stopped. Applied here, that discipline turns 80 percent into a range, and the range's floor is uncomfortable.
Inference is the new cost of goods sold
Here is the part that connects directly to the chain.
For two decades, software investors learned one reflex: high gross margin means defensive economics. AI breaks that reflex, because AI introduces a variable cost that scales with usage and does not shrink to zero at scale. That cost is inference, and it is the closest thing the AI industry has to a raw material. Copper for the industrial economy. GPU-seconds for this one.
I want to be precise about why that matters for token networks, because the parallel is exact and rarely drawn carefully.
A decentralized inference marketplace, meaning a network that routes inference jobs to independent GPU operators, has a cost structure that looks like this: the price paid to node operators for compute, the cost of verifying that the work was actually done, the cost of coordination including discovery, reputation and slashing, and the cost of capital that funds all of it while demand matures. The most sophisticated of these networks run token emission schedules to subsidize supply during the gap between launch and real demand.
Notice the symmetry with the Anthropic disclosure. Anthropic excludes training from its margin. Decentralized networks exclude almost everything, because in many cases they do not have a margin at all. They have an emission budget.
Take a representative inference network paying operators in its own token. On paper, the gross margin is whatever fraction of job fees the protocol retains after paying nodes. But the token paid to operators is minted, not earned. Its real cost is dilution borne by every holder. Economically, that is a cost of goods sold recorded in a currency the protocol controls, which is precisely the accounting move that made Anthropic's 80 percent possible, applied more aggressively and with fewer disclosures.
When you subsidize supply with inflation, you are not deferring cost. You are transferring it, from the income statement to the cap table, and from this quarter to the long-term holder. That is not a moral judgment. It is an accounting identity, and I have watched too many projects discover it only when the emissions curve flattened and the node operators left.
The number I want to see is not gross margin. It is the ratio of real, externally paid demand to emissions, a figure I started calling the demand-coverage ratio in my own work after 2022, when so many sustainable-yield products turned out to be circular flows dressed as revenue. For most AI compute networks today, that ratio is the single most important number nobody publishes.
The crypto mirror: same bet, worse arithmetic
Let me make the comparison concrete, because abstraction lets everyone off the hook.
The centralized case for AI compute profitability rests on four pillars: proprietary model quality, hyperscaler distribution, sticky enterprise contracts, and a declining inference cost per token. Anthropic has three of those in some form. It has frontier models that trade near the top of public leaderboards. It has Amazon's distribution into the largest cloud customer base on earth. It has, presumably, a growing base of enterprise and developer usage. What it does not yet have publicly is a demonstrated downward slope on inference cost per token, and that slope is the variable that decides whether this is a software business or a utility.
The decentralized case rests on a different four pillars: idle or underutilized GPU capacity, price competition against hyperscalers, permissionless access, and verifiable computation. The first three are real advantages in theory and partially real in practice. Akash runs reverse auctions that can undercut list cloud prices for interruptible workloads. Bittensor coordinates specialized subnets with incentive mechanisms that reward measurable quality. Render moved some of its rendering capacity toward general compute. io.net aggregated supply quickly and then had to spend a year repairing trust after early verification failures.
The fourth pillar, verifiable computation, is the one that quietly breaks the model. Think about what a decentralized network actually needs to prove. If a node claims it ran a model and returned a correct answer, the network must verify either the computation or the result. Verifying a large model's inference is orders of magnitude more expensive than trusting a well-known operator. This is the same wall I have written about in the zero-knowledge rollup context for three years, and my position there has not changed: proving cost is not a footnote, it is the determinant.
A rollup that pays more to prove a transaction than the transaction is worth has negative unit economics no matter how elegant its cryptography. The identical constraint applies to verifiable AI inference. Optimistic systems reduce the proving cost by assuming honesty and punishing fraud after the fact, but they introduce settlement delays and challenge-game overhead that break latency-sensitive applications. Zero-knowledge approaches produce cryptographic certainty at a price that, today, is still measured in multiples of the underlying job for anything beyond small models. Either way, the verification bill lands in exactly the place Anthropic excluded, which is the cost of goods.
And there is a second, subtler bill, one that the oracle world taught me to respect. Compute networks need price discovery, a reliable feed that tells them what a GPU-hour is worth right now. Weak price feeds, slow price feeds, or feeds sourced from three venues with thin depth produce exactly the failure mode that has burned lending markets repeatedly: a signal everyone trusts until the moment it matters, and then doesn't. A decentralized compute market with a fragile pricing oracle is an arbitrage surface, not a marketplace. That is not a hypothetical. I have watched an oracle lag by forty minutes turn a routine liquidation cascade into a panic.
What the market is actually paying for
So what is priced in?
In a sideways tape, and we are unmistakably in one, capital does not disappear. It rotates into narratives that promise a resolution. AI compute has been the strongest of those narratives for eighteen months. The trade has been straightforward: buy the picks and shovels, because the gold rush is real and the picks are scarce.
Anthropic's leak strengthens the gold-rush half of that story and quietly weakens the picks half, at least for decentralized picks. If the best-capitalized centralized labs can approach profitability, capital has a new and simpler destination. It can buy equity in the lab instead of tokens in the network. Equities carry legal claims on residual cash flows. Tokens often carry claims on governance and, too frequently, on a stream of emissions that dilutes the very holders who are supposed to benefit.
When the same thesis is available in two wrappers, the wrapper with cash-flow rights tends to win the marginal institutional dollar. That matters more than most people in this sector want to admit, because institutional dollars are exactly what crypto's AI compute complex has been courting since the ETF era changed the composition of the bid. The reflexive answer is that tokens offer upside convexity equities cannot. That is true in the tail. It is also true that most of the tokens in question have no mechanism to capture the surplus they are generating.
This is where I part company with the enthusiasm I hear in community channels. The bullish reading is that Anthropic proves the pie is real and decentralized networks will capture the long tail. The bearish reading, which I think is closer to the ground, is that Anthropic proves the pie is real and that capturing it requires hyperscaler distribution, frontier models, and a training budget no token network can match. Those are not long-tail advantages. They are the definition of the head. The long tail gets the fragments that fall off the table.
The contrarian read: a financing event dressed as an earnings event
Here is the angle I have not seen anyone state plainly.
Anthropic's disclosure is not evidence that AI has become a mature, self-funding industry. It is evidence that AI's capital markets have become sophisticated enough to manufacture a favorable earnings narrative ahead of a liquidity event. That is not a scandal. It is standard practice, executed well by people who understand how financial journalists and future public-market investors read numbers. But it should be read as a financing instrument first and an operating result second.
The tell is in the disclosure mechanism itself. Two quarters is not a trend. It is an anecdote with a calendar attached. Private disclosure to a small group of shareholders is not reporting. It is signaling. Anonymous sourcing is not verification. It is deniability. Every one of those choices is rational for a company preparing to speak to public investors, and every one of them is a reason for the rest of us to withhold judgment until a prospectus forces specificity.
The second contrarian point is aimed at my own readers rather than at Anthropic, and it is the one I most want to be uncomfortable.
The crypto industry has spent two years arguing that decentralized compute will capture the surplus of the AI boom. The uncomfortable implication of this report is that if centralized AI becomes genuinely profitable, it has less reason to share that surplus with anyone, least of all with a permissionless network whose node operators are functionally its competitors. Scarcity created the decentralized opportunity. Profitability at the center reduces the incentive to relieve that scarcity. The bottleneck was always the business model.
There is a version of this that turns bullish later. If inference costs keep falling and model quality commoditizes, the head loses its moat and the long tail wins on price and access. That is a real thesis, and I do not dismiss it. But it is a three-to-five-year thesis, and it depends on a variable, cost per token, that Anthropic did not disclose. Which brings me back to my first complaint. The number everyone is celebrating is the one number that answers the least important question.
Community Pulse
I spend part of every week in the channels where retail actually lives, and I want to record what I hear rather than what I wish I heard.
Sentiment across the major AI compute communities is best described as reluctantly patient. The early-2024 euphoria has cooled, funding rates on the associated perpetuals have normalized, and the loudest voices have drifted to shorter-cycle narratives. What remains is a quieter cohort of builders, node operators and mid-sized holders who are asking better questions than they did a year ago. The two I hear most often are: what is our paid demand versus our emissions, and what does it cost us to verify a job? Those are the right questions. The fact that they are now common is, in my view, the most constructive development in this sector this year.
Anxiety is concentrated, predictably, among holders who bought the narrative top. That cohort is not asking questions. It is waiting, and waiting cohorts are the ones that panic on the next leg down. I flag sentiment honestly even when it is unflattering to the assets I analyze, because a market that cannot measure its own fear will keep mispricing it.
Ethical Impact
I keep a standing metric in this column, and I will apply it to both sides of the comparison.
For Anthropic, the ethical question is disclosure. Private, unaudited, selectively released profitability figures create an information asymmetry between a small group of shareholders and everyone else: engineers holding illiquid options, customers negotiating contracts, and the future public-market investors who will eventually price the company. Nothing about that is illegal. Something about it is worth naming.
For the token networks, the ethical question is who pays for growth. When a protocol subsidizes compute with emissions, the subsidy is funded by dilution, and dilution is borne disproportionately by holders who arrive early and sell late. Framing that subsidy as network growth without naming its cost is a disclosure failure in the same family as the one above, even though it happens on-chain and in real time. The ethical pulse of the decentralized economy does not beat faster because the ledger is public. It beats on whether the costs are named.
What to watch
Four signals will tell us more than any single earnings leak.
First, the IPO process itself: underwriter appointments, registration filings, and eventually a prospectus with audited revenue, customer concentration and a real cost of goods sold. That document will settle the debate this leak opened, and it will do so with far less ambiguity.
Second, Amazon's distribution terms. If the revenue share narrows or widens, the true margin moves with it, and nothing in the public conversation has priced that sensitivity.
Third, the slope of cost per token. If Anthropic or a competitor publishes a credible declining cost curve for inference, the economics of the entire sector, centralized and decentralized alike, change in a measurable way. That single variable decides whether decentralized compute is a complement to the head or a casualty of it.
Fourth, verification cost inside the decentralized networks. When proving a job stops costing more than the job, the permissionless model becomes defensible on unit economics rather than ideology. Until then, the decentralized compute story remains a story about promises, and promises are the one thing this market has never had trouble minting.
Building bridges in a fragmented digital frontier means being honest about which bank is solid and which bank is a rendering. Anthropic may well be solid. But the bridge between an 80 percent headline and a sustainable business has not been built yet, and the load-bearing piers are the costs nobody wants to name.
