The number arrived without a denominator. A short item published by a crypto-asset outlet reported that Anthropic's token call volume had grown 3.6 times — no reference period, no definition of what "token volume" means in that sentence, no baseline against which the 3.6 was measured, and no link to a primary source. Within a day the figure had been absorbed into a much larger trade. AI-agent tokens were bid, decentralized-compute narratives refreshed, and a chart of "inference demand" circulated as though a mood had acquired a decimal point.
I have spent enough years inside disclosure documents to be suspicious of ratios that arrive clean. Real operational metrics are rarely clean. They come with footnotes about what was excluded, and the footnotes are usually where the business actually lives.
Context: a vocabulary imported from the wrong ledger
Crypto learned the word "token" in a context where it meant something countable, auditable, and priced — gas units, transfer amounts, the discrete atoms of an on-chain transaction. Porting that vocabulary onto model inference imports a false sense of precision. A model token is not a unit of work, a unit of value, or even a stable unit of cost. It is a fragment of a byte-pair encoding vocabulary, and its economic meaning changes entirely depending on whether it is an input, an output, a cached read, or an internal reasoning step that the user never sees.
That matters because the outlet that carried the 3.6x is an asset-price publication, not an AI trade journal. Its readership holds adjacent exposure — AI-agent tokens, decentralized compute networks, GPU-adjacent equities — and its editorial incentive is to convert technical news into a directional signal. I do not say this as an accusation. I say it because I watched the same mechanism operate on total value locked between 2019 and 2021, when I was doing governance risk work around the DAI stablecoin and had to reconcile self-reported collateral figures against what the contracts actually held. TVL was not a lie. It was a metric selected because it always went up. Token volume is now auditioning for that role.
Meanwhile the underlying product change is real and well documented. Anthropic's models moved from single-turn conversation toward extended thinking and agentic execution — the company's own product line, from Haiku through Sonnet to Opus, is now sold less as a chat interface and more as an executor that plans, calls tools, and iterates. That shift has a mechanical consequence for token accounting that has nothing to do with how many people signed up.
Core: decomposing 3.6x
Any volume figure decomposes into requests times tokens per request. The two halves have opposite implications.
If 3.6x came from request count, the story is adoption: more users, more seats, more integrations. If it came from tokens per request, the story is task complexity — same customers, radically heavier workloads. The first is a market-expansion story with improving operating leverage. The second is an ARPU story with inference costs rising in lockstep. The article provides no way to distinguish them, which means every downstream conclusion is a guess wearing a suit.
The most credible explanation sits almost entirely in the second half. Extended thinking generates internal reasoning chains of thousands to tens of thousands of tokens before a single visible output token is emitted. A modest rise in the share of requests using that mode can amplify total volume by five to twenty times with no change in user count at all. Agentic workflows push further: a coding agent executing a multi-step task re-transmits its accumulated context on every tool call, so a single task that a human perceives as one action can consume ten to a hundred times the tokens of a single-turn exchange. These are not speculative mechanisms. They are the documented shape of the product.
Then there is the accounting distortion that almost nobody discusses. Cached input reads are billed at a fraction of the standard input rate — roughly a tenth, depending on the tier — but they are still counted as tokens processed. A rising cache hit rate therefore inflates volume while deflating revenue per token. The metric and the business move in opposite directions, and the metric is the one that gets published.
Which brings us to the scenario that the source material cannot rule out. Same 3.6x, four very different businesses:
| Scenario | Unit price change | Volume | Implied revenue | Trigger | |---|---|---|---|---| | A. Pure demand | 0% | +260% | +260% | Sonnet-tier, list price unchanged | | B. Mixed | −50% | +260% | +80% | Enterprise mix, cache and batch discounts | | C. Price-led | −67% | +260% | +19% | Flagship price cut | | D. Price war | −80% | +260% | −28% | Unit economics deteriorate |
One number, and the revenue outcome ranges from a tripling to an outright contraction. That is not a nuance. That is the entire investment question, and it is absent from the reporting.
There is a further structural point that crypto readers are unusually well equipped to understand, because they have lived it. When a business prices by seat — Pro, Max, Team, Enterprise — token consumption is a pure cost line. A subscriber who runs an agent all day generates more volume and identical revenue. Under seat-based pricing, growth in token volume is not evidence of commercial success; it is evidence of margin pressure. The only regime in which volume maps cleanly to revenue is pure metered API billing, and that is precisely the regime that every model vendor is trying to escape, because metered billing is volatile, deflationary, and hostage to price competition.
None of this makes the volume growth meaningless. It makes it a different signal than the one being traded. The correct frame is Jevons — the observation that falling unit cost increases total consumption rather than reducing it. Unit token prices have fallen sharply over eighteen months; speculative decoding, quantization, and prefix caching have all cut cost per token; and total volume has accelerated anyway. That combination is the strongest empirical rebuttal available to the argument that efficiency gains will shrink the compute footprint. Inference is now the dominant marginal consumer of accelerators, and the bottleneck has migrated from raw FLOPS toward memory — long-context agent workloads are bounded by high-bandwidth memory capacity and interconnect bandwidth long before they are bounded by arithmetic throughput.
The constraint that follows is physical rather than financial. A large inference cluster can be ordered in quarters; the electricity to run it takes years to interconnect, and the queue in most major markets runs two to four years. If token demand compounds at anything like the reported rate, power procurement — not accelerator allocation — becomes the binding constraint within twelve to twenty-four months, and data-residency rules fragment utilization on top of that. This is where the crypto-adjacent trade actually has a defensible leg, and it is not the one that rallied: it is the power, interconnection, and memory supply chain, not the application tokens.
Step back and the more interesting event is not the growth but the metric rotation. For three years the competitive narrative was benchmark scores; it is now call volume, because volume is the one figure that always trends up and never has to be reconciled against a denominator. The moment one lab publishes it, competitors are conscripted into the same disclosure race, which is precisely what happened to TVL in decentralized finance — a metric that began as a rough proxy for liquidity and ended as a marketing instrument that protocols engineered around. In crypto specifically, the read-through is uncomfortable: AI-agent tokens are being repriced off a line item that appears in the vendor's cost of revenue, not its revenue. Every token is a vote for a future we haven't priced.
I learned this the slow way. In 2018 I spent three months reading the 0x v2 contracts line by line and submitted seven edge-case findings, including a reentrancy path in the filler function. None of those findings were visible in aggregate usage statistics. A protocol can process a rising volume of trades while carrying a structural flaw that only appears under a specific call sequence — the aggregate number is silent on integrity, and it is silent in a way that feels like confirmation. Token volume disclosures have the same property. They measure throughput and say nothing about whether the throughput is producing anything.
What a serious disclosure would contain is not exotic: revenue per processed token, gross margin trend across two or three quarters, net revenue retention, and the ratio of inference cost to revenue. Those four numbers would resolve the A-through-D question in a single filing, and a fifth — cost per completed task — would replace token price as the honest efficiency benchmark. Their absence is not an oversight; each one is a variable that could move in the wrong direction. Until they appear, any conclusion drawn from volume alone carries a substantial discount, and the discount belongs before the narrative, not after it.
Contrarian: the growth may be evidence of waste
Here is the reading that the bullish case cannot metabolize. A 3.6x increase in tokens consumed per completed task and a 3.6x increase in tasks completed are indistinguishable in a volume disclosure. Agent loops burn tokens on retries, on re-reading context they have already read, on reasoning branches that terminate in nothing. Call it idle burn. When I spent six months after the 2022 collapse auditing the governance failures behind algorithmic stability — an exercise I never published, because it was for my own model of risk rather than for a market — the lesson that survived was that systems optimized for a headline number will produce that number at the expense of the thing the number was supposed to represent.
The incentive structure reinforces it. A vendor that makes its model dramatically more token-efficient reduces its own revenue under metered pricing. There is no public evidence that any lab is throttling efficiency, and I am not alleging it. But the absence of an incentive to optimize is itself information, and it means efficiency gains will arrive on the schedule that competition forces, not the schedule that customers would prefer. Every token is a vote for a future we haven't audited.
The crypto-specific contrarian point is sharper. The assets that rallied on this figure are the ones most exposed to the mechanism behind it. Decentralized inference marketplaces and compute networks benefit from volume growth in the same way hyperscalers do — but they sell into a market where the reference price is set by firms with better utilization, cheaper capital, and custom silicon. Their token-per-dollar output is the only variable that matters, and a demand tailwind that arrives alongside relentless unit-price deflation is a treadmill, not a lift.
What to watch instead
Three signals will resolve this faster than any commentary. Whether a second lab begins disclosing call volume — if the metric standardizes, it becomes comparable, and comparability is what turns a marketing figure into an analytical one. Whether gross margin language appears alongside volume in the next disclosure cycle; a company confident in its unit economics eventually has to talk about them. And whether the binding constraint on inference turns out to be chips or electricity — the interconnection queue runs two to four years while accelerator lead times run quarters, and the first constraint to bind determines where value accrues across the entire compute chain.
Every token is a vote for a future we haven't measured. The question worth holding through this consolidation is not whether Anthropic's usage tripled. It probably did. The question is why the industry has chosen to report the one number in its income statement that cannot be audited, and what it expects us to stop asking in exchange.
