The Router Lies: Open-Weight AI's Token Lead Is a Measurement Artifact

The Router Lies: Open-Weight AI's Token Lead Is a Measurement Artifact

Last week I pointed three model routers at the same workload. Same models. Same seven-day window. Same batch of agentic requests. Router A reported open-weight models carrying 71% of token volume. Router B reported 48%. Router C reported 63% — but only because it counted cached input tokens as billable volume. Three instruments. One market. Three truths.

None of them was lying. That is the problem.

When a headline announces that open models now dominate token volume over closed ones, it is citing exactly this kind of telemetry. And telemetry from a cost-optimizing router is not a photograph of the market. It is a photograph of the router. I have spent the last decade pulling apart systems that report their own health — 0x v2's order-matching contracts in 2017, twelve Uniswap v2 forks during DeFi Summer, three cross-chain bridges in the 2022 crash. The pattern never changes. The measurement layer becomes the attack surface the moment its output starts moving capital.

The Vercel AI Gateway claim — open-weight models leading token volume — is the current version of this pattern. It arrives in a crypto media cycle that is already pricing decentralized inference tokens on precisely this kind of self-reported number. So let's audit it.

Context first. Vercel AI Gateway is infrastructure that sits between an application and a set of model providers. Its job is routing: take a request, select a model, execute, return. Its core features are multi-model routing, automatic fallback, and cost optimization. It is, functionally, a load balancer with a pricing brain. The telemetry it publishes reflects what flows through it — nothing more.

The mechanism matters more than the marketing. A gateway assigns each incoming request a routing decision based on cost, latency, and availability. When the preferred model is slow or expensive, fallback sends the request to a cheaper one. Over time, the gateway's own heuristics push traffic toward the lowest unit cost that still clears a quality threshold. That is not a neutral pipe. That is an active selection function, and every metric it emits is downstream of that function.

That telemetry is now being read as evidence of a market shift. Open models "dominating." The word is doing enormous work. Dominating is not a unit. It is a mood. A claim of dominance without a percentage, a time window, and a model list is a headline, not a dataset. The source material provides none of the three. No percentage. No window. No model roster. No definition of what "open" includes. That absence is not a footnote to the analysis. It is the analysis.

Here is what we can establish with structure rather than slogans.

Start with price. Closed flagship models price at roughly $3 input and $15 output per million tokens. Top open-weight models price at roughly $0.27 and $1.10. That is a spread of ten to fifteen times on the same nominal unit, and in cached or batched scenarios, thirty times or more. Hold that number, because it inverts the headline.

Suppose open-weight models carry 70% of token volume. Suppose closed models carry the remaining 30%. Now convert to dollars. Closed: 0.3 units of volume at one unit of price equals 0.3. Open: 0.7 at one-twelfth price equals 0.058. Total: 0.358. Open-weight revenue share: roughly 16%.

A 70% token-volume lead resolves to approximately a 16% revenue share. That is the entire story the headline buried. Token volume is not a proxy for market power. It is a proxy for cheapness.

This is not a rounding error. It is a category error, and it recurs everywhere crypto touches AI. Token count is a measure of verbosity, not value. A model that rambles generates volume. A model that solves generates revenue. The metric rewards the former while the market pays for the latter. Any dashboard that reports "compute demand" in raw tokens is measuring the wrong end of the pipe.

The enterprise numbers make the gap sharper. Mid-2025 survey work on enterprise generative-AI spend — Menlo Ventures' category of study — put dollar spending on LLM APIs still dominated by Anthropic and OpenAI, with Anthropic's enterprise share passing OpenAI's for the first time, and open-weight models collectively in the low-20% range. That is a dollar-share figure, not a token figure. The systematic multiple between the two is exactly what the pricing spread predicts. Token share and revenue share are not two views of one market. They are two different markets wearing the same name.

Now the sampling frame, which is where most of these claims quietly fall apart.

Vercel's gateway serves the Next.js and Jamstack ecosystem. Front-end and full-stack developers. Price-sensitive, prototype-driven, heavy on individual projects and small SaaS, and structurally suspicious of vendor lock-in. This cohort prefers open-weight models by default — not because they are better, but because they are cheaper and portable. It is a slice of the developer long tail. It is not the enterprise market.

Enterprise model traffic runs through different pipes entirely: AWS Bedrock, Azure AI Foundry, GCP Vertex AI, and direct vendor contracts with OpenAI and Anthropic. None of those channels enter Vercel's telemetry. And on those channels — where compliance, SLA, liability division, and audit trails dominate the buying decision — closed models hold substantially more share. The observation is a real local truth. It cannot be extrapolated into a market truth.

I have watched this exact error drain treasuries. In 2020 I audited twelve Uniswap v2 forks for small DAOs in Chengdu. Several reported healthy liquidity on their own dashboards. The dashboards only measured liquidity on their own pair. The moment a correlated asset moved, the depth was gone and the reentrancy window was open. The number was true. The conclusion drawn from it was fatal.

Then there is the observer effect, which is the part nobody wants to name.

Vercel AI Gateway's central product capability is cost optimization. Its entire purpose is to steer requests toward the best price-to-performance ratio available. When a cost-optimizing router publishes telemetry showing that users are selecting cheaper models, that is not a discovery about the market. That is a specification of the tool. The instrument shapes the behavior it measures. Frictionless execution, immutable errors.

In 2026 I audited one of the first AI-driven trading agents wired into a decentralized oracle network. I found twelve instances where the agent's heuristic decision path bypassed its own safety rails — because the oracle it queried for price was also the oracle that priced its incentives. The system was optimizing toward a measurement it controlled. The blowup was not a bug in the model. It was a bug in the loop. A cost router reporting cost-driven behavior is the same loop, one layer up.

Definitions matter next. "Token volume" is not one number. It is at least four: input on cache miss, input on cache hit, output, and reasoning tokens billed at output rates. Whether cached input counts toward "volume" can flip the conclusion on its own. Agentic workloads — long context, heavy reuse, repetitive tool calls — generate enormous cached-input volume. If a router counts those tokens, open-weight share inflates, because open-weight models are disproportionately deployed on exactly that high-volume, low-value, repetitive work.

Meanwhile closed models absorb the low-volume, high-value reasoning: long-horizon agentic coding, formal mathematics, multi-file refactors, high-stakes tool orchestration. Few tokens. High dollars. A measurement that treats both as interchangeable units of "volume" has already thrown away the only variable that matters.

This is what practitioners now call intelligence tiering, and it is a 2025 engineering reality rather than a 2026 slogan. Teams route cheap, tolerant workloads to open weights and escalate hard, brittle workloads to closed flagships. Both halves run in the same product. The token count reflects the cheap half. The invoice reflects the expensive half. A single aggregate volume number collapses a two-tier system into a misleading average.

And underneath the definitional mess sits the naming sleight of hand. DeepSeek, Qwen, Llama, GLM, Kimi — these are open-weight, not open-source. Training data, full training code, and data recipes are largely unpublished. The media elides the distinction, which systematically overstates the reproducibility and governance maturity of the ecosystem. Metadata is fragile; code is permanent. "Open" is metadata. The weights are the code. Do not confuse the two, because in a market that prices governance and verifiability, the difference is the entire valuation.

There is a geopolitical layer the headline silently drops. The open-weight models leading token volume are heavily Chinese in origin — DeepSeek, Qwen, GLM, Kimi, MiniMax. In a US enterprise procurement context, that transforms a routing decision into a supply-chain and compliance question. Several US government and defense-adjacent environments explicitly restrict DeepSeek. So the cheerful narrative of democratized capability runs headfirst into an unpriced legal and geopolitical surcharge that any honest cost comparison has to carry.

Now the part that belongs on-chain.

Decentralized inference networks — the tokens that promise to route AI workloads across distributed GPUs — are priced on precisely this metric. Token volume. Compute demand. Call growth. But the measurement problem there is worse, not better, and the reason is structural.

On a centralized router, you at least have a company with a billing ledger and a legal person behind the number. On a decentralized inference network, "token volume" is a number the network publishes about itself. The incentive to inflate it is direct, because the token price is a function of it. No external auditor. No revenue statement to reconcile against. The telemetry is the product.

This is the same failure mode as the bridges I broke in 2022. Two of them reported healthy reserves. Both verified their own solvency through an internal accounting function with an unpatched integer overflow. The dashboards read green while the balances drifted. Silence is the loudest exploit. Nobody had to lie. The instrument simply could not see the thing that mattered.

Decentralized inference inherits every measurement flaw of centralized routing and adds one of its own: no independent verifier. A genuinely verifiable inference network would need three things bound together — a cryptographic attestation of which model executed, meaning a hash of both the weights and the runtime; a proof of computation, whether TEE attestation or, eventually, zkML; and token accounting tied to a settlement layer that is independent of the operator's own report. Almost none of the current networks ship all three. Most ship none. They ship a dashboard.

Consider what that means for a buyer. If you cannot verify which weights ran, you cannot verify the cost basis. If you cannot verify the computation, you cannot verify the output. If you cannot verify the settlement, you cannot verify the demand. Every claim a decentralized inference token makes about usage rests on a number it produces for itself, in a market where that number moves its price. Trust no one; verify everything — and if the network cannot hand you the verifier, the number is a self-portrait.

The unit economics of the crypto side deserve their own line. Self-hosting open-weight models does not produce API revenue. It produces GPU electricity and operations cost. So a large share of "open-weight token volume" may be usage that would never have generated a dollar of API revenue in the first place. Pouring that into a metric called "compute demand" and pricing a token against it is how you build a valuation on a category that does not clear.

Here is the contrarian read, and it is the one the market keeps missing.

The consensus interpretation of this data is that open-weight models threaten OpenAI and Anthropic. That is the wrong target. The squeeze is on the middle.

High-end reasoning is moving up. Anthropic and OpenAI are pricing toward compute-dense inference — agentic coding, long-horizon tasks, complex reasoning — where capability sensitivity beats cost sensitivity and the customer will pay a premium for reliability. Low-end commodity tokens are draining toward open-weight, where price is the only variable and the floor approaches the marginal cost of electricity. Both ends are compressing the same casualty: the second-tier closed API and the thin wrapper sitting on top of it.

And the most exposed business is not a model lab at all. It is the model-arbitrage token — the project that resells API access with a markup and calls the billing layer a protocol. When inference cost falls by an order of magnitude and open weights are free to self-host, the wrapper's margin goes to zero. Impermanent loss is a feature, not a bug — and here the "liquidity" being drained is the wrapper's pricing power. The headline everyone reads as "open beats closed" is really the sound of a pricing floor collapsing under everyone in the middle.

The directional claim is probably true. Open-weight models likely do lead token volume on developer-long-tail channels. That costs nothing to concede, because the phenomenon is real and the extrapolation is not. What the data cannot support is any statement about revenue, enterprise share, or capability parity. Those require a second instrument: dollar share, measured across the enterprise pipes Vercel cannot see, and cross-checked against an independent router with a different user base.

So here is the forward-looking judgment, and it is a forecast about instruments, not outcomes.

Watch whether Vercel publishes absolute call volume, a methodology note, a time window, and an explicit model list. If it does, the number becomes auditable and the trend becomes trackable. If it does not, the data is a marketing artifact — and given that Vercel sells a cost-optimizing, anti-lock-in gateway, a marketing artifact that proves cost-optimizing, anti-lock-in routing is exactly what you would expect to find. That is not an accusation. It is a source discount, applied the way I apply it to every protocol that grades its own homework.

The verifiable version of this claim does not live in a router's blog post. It lives on-chain: model-weight hashes, TEE attestation, and settlement decoupled from reporting. Until a network ships all three, its "compute demand" dashboard is a mirror. Somewhere, a second-tier closed API and a wrapper token are both bleeding, and neither of them can see it, because the only instrument pointed at the wound is the one that reports itself healthy.

Logic remains; sentiment fades.