Token Volume Is Not Revenue: What Vercel's Open-Model Telemetry Actually Proves

$0.27.

That is what a DeepSeek V3-class open-weight endpoint charges per million input tokens on a cache miss, against roughly $1.10 on the output side. A Claude Sonnet-class closed endpoint bills about $3.00 in and $15.00 out. Same unit, same meter — an elevenfold spread at input, closer to fourteenfold at output, and beyond thirtyfold inside long-context agent loops where a 12,000-token system prompt is replayed forty times per task.

Against that backdrop, a flash item crossed my desk claiming open models are "dominating token volume" over closed ones on Vercel's AI Gateway. No percentage. No time window. No model list. No definition of "open." No absolute volume. Just a present participle doing the work of a dataset.

I have spent enough nights in block explorers to recognize that shape. It is a transaction hash with no receipt attached — and in my line of work, tracing the gas trails back to the root cause, the receipt is the entire argument. The code does not lie, but the auditor must dig.

A Router Is Not a Market

Vercel AI Gateway is a routing layer. It sits between an application and a set of model providers, normalizes their APIs, and sends each request to whichever endpoint is cheapest or fastest under a policy the developer configures. Model portability and cost optimization are the product. In the most literal sense, it is a load balancer for intelligence — which makes it a magnificent instrument and a terrible oracle.

The population routing inference through a Next.js-ecosystem gateway is not the population buying inference for a bank. It is front-end and full-stack developers, prototype builders, indie founders, and small teams with a documented allergy to vendor lock-in: price-sensitive, iteration-driven, self-selecting. Of course it prefers open weights. That is the point of choosing a gateway that advertises open weights as a feature.

Enterprise procurement runs through different pipes — AWS Bedrock, Azure AI Foundry, Google Vertex, direct vendor contracts with negotiated SLAs, indemnification, audit trails, residency guarantees. Those requests never touch Vercel's telemetry, and in those channels closed models hold a far larger share, because the buyer is purchasing accountability as much as intelligence. An aggregator's volume describes the aggregator's users. An L2's transaction count describes its incentive program. The sampling frame is the finding.

One clarification the headline skipped: what it calls "open models" are, with rare exceptions, open-weight models. Weights download; training data, data recipes, and full training code generally do not. That distinction changes what can be reproduced and governed, and it was dropped because it complicates the story. Four numbers would have salvaged the claim — a percentage, a window, a model list, a token taxonomy. The item contained none. A dominance claim without a denominator is not a measurement. It is a mood.

Reading the Telemetry Like a Ledger

The denominator problem

"Token volume" is not one metric. It is at least four, and they price differently.

prompt_tokens:          12,400
cached_prompt_tokens:   12,200
completion_tokens:         380

That is one plausible turn inside an agent loop. Output is 3.2% of tokens; cached input is 98% of input. Sum all four categories into one headline and what you measured is prompt replay — the cost of an agent re-reading its own context — not a preference between model families. A routing policy that keeps one cheap endpoint warm in cache dominates that metric mechanically, without any user expressing a preference.

Output tokens are the expensive ones, running three to five times input on both sides of the aisle. If open weights handle high-volume, low-output work — classification, extraction, retrieval summarization — while closed models handle long-form generation and multi-step reasoning, output share and input share can point in opposite directions while the aggregate looks decisive. Gas used, gas price, and fees paid in dollars are three different rankings on-chain, and nobody serious argues about which chain is "bigger" without naming the unit first. Ranking models by token count while ignoring the category split is the error we stopped making about chains in 2021.

The conversion nobody performs

If the price wedge between the two buckets is tenfold to thirtyfold, a token-share figure becomes a revenue-share figure only after a currency step the article never performs. Open weights at 70% of token volume and one-tenth the unit price land at roughly 10% to 20% of dollar volume. Same telemetry, two entirely different stories, depending on whether you count units or money. Token volume is not revenue. It never was.

In May 2022 I spent two weeks reverse-engineering the LUNA/UST peg — specifically the seigniorage logic inside Anchor's contracts — because the on-chain metrics everyone cited looked magnificent right up until they were arithmetically impossible. Supply, deposit counts, TVL: all enormous, all meaningless once you saw that the same coins were recycled through a mint-and-burn loop that only worked in one direction. In the chaos of a crash, the data remains silent, but the arithmetic had been screaming for months. A share metric with a tenfold price wedge inside it cannot become a market-share claim without an explicit settlement step. Publishing the ratio and withholding the conversion is a rounding error with a headline.

Dust that never settles

A second category inflates token counts while producing exactly zero dollars of API revenue: self-hosted inference. A team running a quantized 70B model on rented GPUs generates tokens at the marginal cost of electricity and depreciation and pays no API fee to anyone. On-chain analytics has a name for this — dust. Counting dust as adoption is the oldest mistake in the discipline, and it is being repeated in an adjacent field with a different unit of account. The sharper question is who pays. A gateway routes to an endpoint that may be the model's creator, a third-party host, a cloud marketplace, or a lab absorbing losses. The token count is identical in all four cases. The economics are not.

The squeeze is on the middle

The framing everyone reaches for is open versus closed. The actual structure is a barbell. The head of the market — frontier reasoning, multi-step agentic coding, long-horizon execution, formal mathematics — is not being commoditized; it is being repriced upward, because inference-time compute scaling means the most capable systems deliberately spend more tokens to think. The tail — bulk classification, translation, extraction, first-line support, code completion — collapses toward marginal cost, and open weights are the mechanism. What dies in a barbell is the middle: capability good but not frontier, priced above commodity but below frontier. The real question is not whether open weights beat closed models. It is which closed model is priced at a capability tier that no longer has a buyer.

Where this becomes a ledger problem

Inference is becoming a metered, machine-purchased, machine-consumed resource, and agents are already paying for their own compute. The rails under construction — HTTP-native micropayments, agent registries, machine identity standards — converge on one requirement: a machine must be able to pay for a computation and verify the computation happened as billed. That is an accounting problem before it is a cryptographic one.

In 2025 I led a research initiative designing a decentralized identity framework for AI agents operating on-chain, using zero-knowledge proofs to let an agent prove its computational work without disclosing its proprietary algorithm. The hard constraint was never identity. It was proof of work performed — because the moment an agent pays for inference, the invoice becomes a consensus object, and the parties to that consensus are the provider and the consumer, whose interests are directly opposed.

Three approaches exist, with cost profiles as different as optimistic rollups are from validity rollups. Hardware attestation inside a TEE is cheap to verify, expensive to trust, and inherits a vendor's threat model. Optimistic verification during a dispute window inherits the exact latency trade-off I dissected in a 5,000-word analysis of Optimism's first-generation state commitments — you cannot challenge what you cannot re-execute, and nobody re-executes a 400-billion-parameter forward pass at 3 a.m. on a laptop. ZK proofs of inference are where three months inside StarkNet's recursive proof system taught me the threshold that matters: when verification cost drops below execution cost, verification stops being overhead and becomes a product.

An inference market that cannot prove execution is an oracle problem wearing a machine-learning costume. A metered economy settles on a number someone asserts. If that number comes from the party being paid, you have not built a market. You have built an invoice.

The router is a sequencer

A router sits between many execution environments and one user. It batches, orders, selects, and decides which endpoint receives which work under which policy. That is not a metaphor for a sequencer. It is a sequencer. Which makes the routing layer's value a function of fragmentation: the more models exist, the more the aggregation layer is worth, and the more it becomes the point of control. Vercel, OpenRouter, Cloudflare, and every cloud model catalog are fighting to be the default front door to inference.

Which is exactly why their telemetry needs a disclosure note. A router publishing data showing that open models are cheaper and more popular is not reporting on the market. It is reporting on behavior its own policy engine was built to produce. The observation is real. It is also endogenous, and the source never said so.

The Blind Spots Nobody Billed For

Two omissions outweigh everything the item contained. The first is the absolute baseline. Every structural conclusion — open weights winning, closed models losing relevance, a reallocation underway — collapses if the gateway's throughput is a rounding error against global inference. If Vercel routes a fraction of a percent of the world's tokens, the correct reading is not "the market shifted" but "one aggregation layer's clientele behaves the way that clientele's incentives predict." No additional percentage fixes this. Only an absolute volume does.

The second is geography, and it is why this is a procurement story dressed as a technology story. The open-weight models accumulating share — DeepSeek, Qwen, GLM, Kimi, MiniMax — are disproportionately Chinese in origin, and in a US enterprise context that is not a footnote. It is data residency, supply-chain review, government-market restrictions, and the awkward fact that a checkpoint is a data artifact crossing borders whether or not anyone approved the transfer. The upstream framing is democratization; the compliance officer's framing is vendor risk. Both can hold at once, and only one generates a legal budget line. What the flash item omitted was not nuance. It was liability.

Nor does anyone ask what "open" licenses. Not reproducibility — the data recipe is private. Not auditability — the training code is private. What it licenses is deployment freedom at the inference layer, which is genuinely valuable and also precisely what makes alignment guarantees removable with a LoRA adapter and a weekend. The property that makes a model cheap to serve makes it cheap to strip.

A closing note on the source. A crypto-native outlet covering AI infrastructure is fine, but "token" carries two unrelated meanings across two industries, which is a search-engine magnet, and the item had the rhythm of a template fill: striking verb, institutional citation, no numbers. Shifting the consensus layer, one block at a time is what publication looks like when the headline is written before the data.

The Takeaway

Watch three signals as a falsification suite rather than a scoreboard. Whether Vercel publishes a methodology note with an absolute throughput figure — without it, the structural reading stays anecdotal. Whether OpenRouter's monthly rankings move the same way, confirming a channel property rather than one vendor's artifact. And whether the next enterprise LLM spend survey shows dollar share migrating, not token share — because if dollars stay put while tokens move, we have learned that cost optimization at the developer tail and budget expansion at the enterprise head are two markets sharing a vocabulary.

The first nine-figure failure in the agent economy will not be a jailbreak. It will be a settlement error — an unverifiable inference invoice, a routing policy that silently degrades the model behind a stable endpoint name, or a proof system whose verification cost quietly exceeds the value of what it verifies. Nobody audits the meter when the meter is the product.

When your agent pays for its own thinking, whose ledger do you trust — the one that ran the model, or the one that charged for it?