Another rug pull? Or just another myth?
A headline crossed my feed last week through Crypto Briefing, a crypto outlet reporting on AI infrastructure: "Vercel AI Gateway shows open models dominating token volume over closed ones." No percentage. No time window. No model list. No methodology note. Just a verb — dominating — carrying an enormous amount of unpaid freight.
I have a bad habit. When a headline hands me a verb without a number, I go hunting for the denominator. Here the denominator is Vercel's own telemetry, captured at a router whose core product promise is routing cost-optimized traffic away from expensive models. A cost-optimization instrument reported that traffic is being cost-optimized. That is not a finding. That is a mirror.
What made me sit up wasn't the AI angle. It was that we have run this exact experiment before, under a different acronym, and it cost a lot of people a lot of money. Crypto natives should have caught this in the first three seconds. Most of them didn't, because the word token does a lot of optical work on our side of the fence.
So let me lay out what Vercel AI Gateway actually is, because the sampling frame is the entire story.
It sits between an application and roughly a hundred model endpoints. It handles fallback, caching, per-request cost attribution, and provider abstraction. If you are building on Next.js — and a very large fraction of modern product teams are — it is the dignified way to avoid hard-coding a vendor key into a repository you will eventually open-source by accident.
Its population is the point. Vercel's user base is frontend engineers, indie builders, small SaaS teams, Jamstack shops. Price-sensitive. Prototype-driven. Structurally allergic to vendor lock-in. This cohort has a revealed preference for open-weight models the way a bicycle commuter has a revealed preference for not buying a pickup truck. It is a real preference. It is not a census.
Enterprise inference procurement runs on entirely different plumbing: AWS Bedrock, Azure AI Foundry, GCP Vertex AI, plus direct enterprise agreements with OpenAI and Anthropic. None of that traffic transits a Vercel router, and on those channels closed models hold substantially more share for reasons that have nothing to do with capability — compliance, SLAs, liability allocation, audit trails, and someone to call at 3 a.m.
Then there is the dollar data. Menlo Ventures' enterprise generative-AI spend survey — mid-2025 vintage, and I flag this as a magnitude reference rather than a citation — placed Anthropic and OpenAI combined still holding the dominant share of enterprise LLM API spend, with Anthropic overtaking OpenAI for the first time in that channel, and open-weight models landing somewhere around the 20% range. I run a dual-metric reporting framework for a Geneva wealth management client that tracks narrative strength against realized flows, and the discipline that framework enforces is simple: never let a unit metric and a value metric tell you the same story without checking whether they should.
Here they are telling wildly different stories.
One more term needs pinning down before the arithmetic. DeepSeek, Qwen, Llama, GLM, Kimi — these are overwhelmingly open-weight, not open-source by any OSI-compatible definition. Training data, data recipes, and full training pipelines are mostly unpublished. The media's habitual elision of this distinction systematically overstates how reproducible and how governable the open ecosystem actually is. Code speaks, but culture listens — and the culture here has decided that open means free, which is a different word entirely.
Now the arithmetic, because the arithmetic is where this headline dies.
Claude Sonnet-class inference runs roughly $3 per million input tokens and $15 per million output. GPT-4o-class sits near $2.50 and $10.00. DeepSeek V3-class sits near $0.27 on a cache miss and $1.10 out. That is a ten-to-fifteen-times unit gap in ordinary workloads, and in some agentic patterns it stretches past thirty.
Run the implication. If open-weight models hold 70% of token volume and the blended unit price is roughly one-twelfth of the closed frontier, their share of revenue lands somewhere between 10% and 20%. The token figure isn't wrong. It is answering a question nobody asked.
Token volume measures consumption. Revenue measures willingness to pay. Those two numbers have never been the same thing, and in AI inference right now they are diverging by an order of magnitude.
I have watched this movie. In 2020 I ran three separate threads dissecting the yield mechanics of the early Compound and Aave forks while my peers chased farms, and the thing I kept coming back to wasn't the APY — it was the denominator. Total value locked is denominated in dollars deposited, and in a recursive lending market most of those dollars are the same dollars counted repeatedly. Protocols with four billion in TVL and two hundred thousand in annualized fees were routinely described as dominant. Then incentives stopped, and the TVL was gone in eleven days.
Vercel's number is a TVL metric. It counts units moved through a pipe. It does not count value captured at either end of the pipe. The crypto industry spent 2022 learning that TVL without fee capture is a marketing instrument, not a measurement. The AI infrastructure conversation is about to relearn it from scratch, in public, with worse guardrails.
There is a second distortion stacked on the first, and it is the one that gets ignored because it is unglamorous: a substantial share of open-weight inference is self-hosted. Nobody pays an API bill for it. Marginal cost is GPU electricity plus operations labor. I have run enough weekend experiments on a rented A10 with a quantized 7B model to know the shape of this — the token counter climbs impressively, the dashboard looks busy, and exactly zero dollars flow to any model provider anywhere in the world.
You cannot fold self-hosted tokens into a commercial share calculation. It is like adding your home-cooked dinners to a survey of restaurant market share and concluding that restaurants are losing. The token volume metric here contains an unknown, possibly enormous, quantity of usage that was never a commercial transaction in the first place and never could have been.
Then there is the structural version of the problem, which is more elegant and less forgivable.
Vercel Gateway's entire value proposition is free switching, cost optimization, and lock-in avoidance. A router designed to steer requests toward the cheapest acceptable model will, by construction, over-index toward the cheapest acceptable models. This is not fraud. It is an instrument changing the behavior it measures. In markets we call it reflexivity; in physics we would call it a bad experiment. The dashboard is not a window onto the market. It is a steering wheel with a rearview mirror bolted to it, and the driver has been told the mirror shows the road.
So what is actually true underneath all of this?
Something real is happening, and I want to be precise about its boundaries. On the developer long tail — batch translation and localization, bulk summarization and classification, first-line support responses, code completion, unit-test generation, data cleaning and annotation assistance — open-weight models have genuinely crossed the usability threshold. In those buckets the driver of substitution is cost, not capability, because a tenfold unit-price gap overwhelms a modest quality gap. My working estimate is 40% to 70% substitution across those categories inside six to eighteen months. That is not a prediction I make lightly; it is the shape of every prior cost-driven substitution I have watched in this industry, from cloud migration to API aggregation.
But the picture at the top of the stack is not the same picture. Multi-file agentic refactoring, long-horizon debugging across a monorepo, formal mathematics, high-reliability tool-calling chains, enterprise long-context document analysis — these remain closed-model territory, and their buyers are far more sensitive to the capability ceiling than to the price floor. A firm doing cross-file migrations does not switch models to save four hundred dollars a month. It switches when the cheaper model stops breaking imports.
This is not a war between paradigms. It is a pair of scissors. Capability tiers upward and gets more expensive; commodity tokens fall toward marginal cost. And the thing being cut in the middle, the thing the headline never mentions, is the mid-tier closed API — the mid-capability, mid-price product that now has no defensible position at either end. That is the most valuable observation available in this dataset, and the source article does not touch it.
Follow the cost collapse forward and you arrive somewhere more uncomfortable. If inference unit cost is genuinely falling by an order of magnitude, then a whole business-model class is being sentenced in absentia: the thin wrapper that resells a closed API at markup. Crypto had a version of this in 2024, when a landing page, a system prompt, and a ticker were sufficient to raise capital. When the underlying compute approaches free, the markup approaches zero. The article's real implication is a death sentence for model arbitrage, and it is being reported as a horse race.
The downstream effects are quieter. Data annotation value migrates from headcount scale toward expert scarcity — the frontier labs need mathematicians and multi-language specialists, while the open-weight trainers lean harder on synthetic data and distillation. And the labor effect is a repricing of skills rather than a net subtraction of jobs: demand rises for people who can design, orchestrate, and evaluate AI systems, and falls for people whose job was executing single steps an AI can now execute.
I keep circling back to 2022, because that was the last time I got to watch a cost collapse from the inside. While most analysts fled, I spent weekends in Celestia's Discord arguing about data availability sampling and sharding economics, and produced a case study arguing that modularity could cut transaction costs by roughly 40%. The point wasn't the percentage. The point was that cost collapse at the bottom layer creates new design space at the top, and that space is invisible during a bull market because nobody is looking for it.
The crypto-native mirror of this trend is decentralized inference — Akash, io.net, Aethir, Bittensor subnets, Hyperbolic, and Render on the visual flank. All of them sell a version of the thesis Vercel's data gestures at: inference is commoditizing, and the winner is whoever can rent verifiable GPU time cheapest. Here is the honest part that the sector's own dashboards will not tell you. Selling compute below hyperscaler pricing is a business with a long negative-margin history, and a large fraction of these networks' reported utilization has the same measurement texture as the Vercel number — units moved through a pipe, priced as if they were revenue captured.
The infrastructure consequences are more determinate than the market-share story. Fragmented inference buyers — Together, Fireworks, Groq, DeepInfra, Nebius, plus cloud-hosted inference and enterprise self-hosting — are structurally more price-sensitive than centralized labs, and they prefer price-per-token efficiency over absolute throughput. That creates deterministic demand for the inference optimization stack: vLLM, SGLang, TensorRT-LLM; INT8, INT4, and FP8 quantization along with AWQ and GPTQ; continuous batching; prefix caching; speculative decoding; KV cache optimization. The open-weight ecosystem depends on these techniques far more than closed labs do, because unit cost is the competitive axis rather than a line item. Meanwhile training-side demand concentrates. If 7B-to-70B distilled models satisfy most commercial needs, frontier pretraining capital consolidates into a handful of camps, and that is structurally neutral-to-negative for training compute demand.
Underneath the price tags there is a technical truth worth naming. The open-weight cost advantage traces to four recognizable sources: mixture-of-experts sparse activation trading total parameters for active parameters, FP8 mixed-precision training with aggressive parallel and communication optimization, distillation-based post-training, and aggressive quantization for deployment. Architecturally, nearly all of it is compositional and engineering-level work. MoE is a pre-2017 paradigm. Multi-head latent attention is a module-level contribution. The open-weight advantage is an engineering-efficiency victory, not a theoretical one — and that distinction determines its half-life.
If the advantage flows from distillation and deployment cleverness, its ceiling is set by the frontier closed model. When the frontier jumps a generation — say, toward a new inference-time-compute paradigm — the catch-up cost rises non-linearly. And there is a measurement illusion layered on top: MMLU and HumanEval are contaminated, public open-weight scores are systematically optimistic, and the real enterprise usability gap is consistently wider than the leaderboard gap. I have audited enough smart contracts to know what it looks like when a system performs beautifully on the test and fails on contact.
One more layer, and it is the one the source article entirely avoids. The open-weight models leading token volume are disproportionately Chinese: DeepSeek, Qwen, GLM, Kimi, MiniMax. In a United States enterprise context that is a procurement compliance question before it is a model-quality question, and several government and defense-adjacent environments explicitly restrict DeepSeek. "Open source is winning" carries an unstated supply-chain security footnote that aggregation telemetry is structurally incapable of seeing. Add a regulatory vacuum — EU AI Act obligations for general-purpose AI land primarily on model providers, and self-hosting makes the user the provider — and you get a predictable next move. Within twelve to twenty-four months, expect compliance responsibility to be pushed downstream by rule. And add the unpriced costs that never appear in a token comparison: audit, security hardening, legal risk provisioning, and the absence of the IP indemnification that closed vendors write into contracts.
Now the part where I disagree with where most people are about to go.
The contrarian take is not that open models are overhyped. It is the opposite of what the next forty-eight hours of commentary will conclude. Everyone will read this headline and buy the inference-commoditization trade, which in crypto means decentralized compute tokens. That will be the consensus position almost immediately, and consensus positions entered on a single unquantified data point have a specific and reliable failure mode.
The real contrarian read is this: a collapse in inference cost is bearish for anyone whose product is inference. In crypto, that is the entire DePIN compute sector, whose pitch is cheaper GPUs than the hyperscalers. If API-layer token prices are approaching marginal cost, the price spread decentralized networks monetize compresses toward zero. That thesis requires the spread to widen. It does not survive the spread narrowing. Those two propositions cannot both be true, and the sector has not yet noticed that it is holding both.
And the meta-point, which is where I actually want the argument to land: a cost-optimizing router producing cost-optimized traffic is evidence about routers, not about markets. The only honest test is a dollar-denominated, cross-channel measurement. If Menlo's next report shows open-weight dollars above 35%, I will revise everything above. If it shows dollars still near 20% while token share climbs past 75%, then what we have is not a trend — it is a pricing structure, and pricing structures can be reversed by a single product launch.
The Cassandra complex is real. I predicted the 2022 yield trap and was told I was early and tedious, and both of those things were true. Being right about a mechanism and wrong about a timeline is a specific kind of professional loneliness. The mechanism here — unit-cost collapse redistributing margin up the stack — I believe. The timeline is a guess. Do not let anyone sell you the first as proof of the second.
So the question worth carrying out of this is not whether open models are winning. It is who captures value when inference costs fall by an order of magnitude, and who gets eaten while everyone is watching the percentage.
My answer: application layers with genuine workflow integration collect a margin windfall they have not had since 2021. Routing and observability middleware builds a real moat, because fragmentation is the moat — the more models exist, the more the aggregator matters, and that is the only line in this story that is structurally self-reinforcing rather than structurally fragile. Inference optimization stacks collect deterministic demand. And thin wrappers, mid-tier closed APIs, and every decentralized compute network whose entire pitch is a price spread get squeezed from both ends at once.
Code speaks, but culture listens. The code says open-weight inference is cheap and getting cheaper. The culture — a Vercel dashboard, a crypto media headline, a ticker — is already narrating winners and losers that the numbers do not yet support.
There is a number missing from that headline. Until someone publishes it, alongside a time window and a model list and a definition of what counts, we are all reading tea leaves in somebody else's telemetry — and paying for the privilege in somebody else's tokens.