Table of Contents
The discount that ate the frontier
American companies are quietly routing their AI workloads to China, and the reason is not ideology. It is the invoice.
Chinese-built models are gaining real traction inside U.S. companies as they narrow the capability gap with American frontier systems while staying dramatically cheaper to run, CNBC reported this week. The numbers behind the story are stark. The share of tokens that U.S. companies push through Chinese models on OpenRouter — the routing layer that lets developers switch between hundreds of models with one API — has held above 30% every week since February 8, peaking at 46%. Over the previous twelve months, that average was just 11%. In the first half of 2025, it was 4.5%. In eighteen months, Chinese AI went from a rounding error in American token traffic to nearly half of it.
The anecdotes are more vivid than the aggregates. In June, the AI agent startup Lindy moved 100% of its traffic off Anthropic’s Claude models and onto DeepSeek. “We did it, and you could see that cost curve go down, like, crash to the ground,” CEO Flo Crivello told CNBC, adding that the switch will save the company millions of dollars within months. When Z.ai released its GLM 5.2 model in June, it became the fastest-adopted model Vercel has tracked all year: daily token volume grew roughly 27x in its first full week, and the number of customers using it grew about 80x, according to Vercel’s head of agentic infrastructure, Harpreet Arora. His diagnosis was blunt: “Price is doing the work here.”
That sentence deserves to be read twice, because it marks a phase change in how enterprises buy AI. For three years, model selection was a capability race — you bought the smartest thing available and treated the bill as the price of relevance. Now it is a procurement decision. Teams route each task to the cheapest model that clears the quality bar, and for a growing share of tasks, the model that clears the bar ships from Beijing, Hangzhou, or Shenzhen. Open-weight releases from DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, and Z.ai have compressed the frontier premium for the median workload to nearly zero, while the price of American frontier access keeps climbing.
The timing sharpens the stakes. The shift is accelerating in the same season that U.S. labs are raising effective prices and Washington is tightening its grip on the most capable American models. At the end of June, OpenAI agreed to limit the rollout of a new set of models at the government’s request, and export controls on Anthropic’s Mythos and Fable models were lifted only after a tense standoff with the administration — a fight I unpacked in America Just Put AI Models on a Leash. The juxtaposition is uncomfortable: the more friction and cost accumulate around American frontier models, the more attractive the unrestricted, nearly-free alternative becomes.
“Chinese AI models are particularly attractive to American companies now as AI costs skyrocket,” Kyle Chan, a fellow at the Brookings Institution’s John L. Thornton China Center, told CNBC. “Where previously U.S. companies were prioritizing AI adoption regardless of model, now they’re getting more cost-conscious.” Cost-consciousness is precisely what I argued was coming when the token binge ended and enterprise AI entered its rationing era. What I underestimated was how quickly the rationing impulse would acquire a geopolitical dimension. The cheapest good-enough token now has a flag on it.
That is the thesis of this piece. The AI price war is no longer a contest between OpenAI and Anthropic pricing pages. It is a structural arbitrage between two national AI economies — one optimizing for frontier capability and revenue per token, the other optimizing for adoption and marginal cost — and American enterprises are the arbitrageurs. What is at risk is not just lab revenue. It is the assumption, embedded in a trillion dollars of infrastructure commitments, that American companies would keep paying American prices.
Do the token math, then do it again
Start with the sticker prices, because they explain almost everything else.
DeepSeek’s V4 Flash lists at $0.09 per million input tokens and $0.18 per million output tokens. OpenAI’s GPT-5.5 lists at $5 and $30 respectively, according to OpenRouter’s analysis of DeepSeek V4 adoption. That is not a discount; it is a different order of magnitude. Higher up the quality curve, Data Gravity’s analysis puts DeepSeek-V4-Pro at roughly $0.435 per million input tokens — about 3x below Google’s Gemini 3.1 Pro and roughly 12x below GPT-5.5 at comparable benchmark performance. Justin Summerville, who works on data and analytics at OpenRouter, told CNBC that open-source Chinese models run “60% to 90% cheaper” than leading Anthropic and OpenAI models.
| Model | Input $/M | Output $/M |
|---|---|---|
| DeepSeek V4 Flash | $0.09 | $0.18 |
| GPT-5.5 | $5.00 | $30.00 |
The gap compounds because of what modern workloads look like. Agentic systems — the tool-calling, multi-step pipelines that now dominate serious deployments — burn roughly 15x more tokens per request than conversational use, per OpenRouter. Here is the arithmetic that should keep pricing teams in San Francisco awake. Take a modest agentic deployment consuming two billion tokens a month, split 80/20 between input and output. On GPT-5.5, that is $8,000 of input and $12,000 of output — $20,000 a month. On V4 Flash, the identical volume costs $216. That is a 93x gap, roughly $237,000 a year per workload, and a mid-size engineering organization runs dozens of such workloads. The agentic era did not just increase AI spend; it multiplied the dollar value of every percentage point of price difference by an order of magnitude.
Adoption data confirms the math is being acted on. DeepSeek’s V4, launched April 24, became OpenRouter’s top model by mid-May, with V4-Flash comprising 70% of DeepSeek’s agentic token flow within a month; the company’s overall share of OpenRouter tokens doubled from about 10% in January to 18% by June. Data Gravity calculates that Chinese open-weight models captured roughly 61% of all tokens consumed on OpenRouter by May, with Xiaomi’s MiMo alone at ~21% of routed traffic and DeepSeek at ~17.6% — ahead of Anthropic’s 15.4%. Meta’s Llama, the erstwhile open-weight standard-bearer, has fallen below 1%. Alibaba’s Qwen, meanwhile, passed 1 billion cumulative Hugging Face downloads, and roughly 40% of all new model derivatives on the platform are now Qwen-based. OpenRouter’s own State of AI report, built on over 100 trillion routed tokens, frames the platform as the closest thing the industry has to a neutral demand census — which is what makes these shares hard to dismiss as niche.
Now look at the other side of the ledger: what American AI actually costs in mid-2026. Industry analyst Josh Bersin documented the squeeze in May: hyperscaler capital spending is headed from roughly $370-410 billion in 2025 toward $650 billion in 2026, and his back-of-envelope math suggests the industry needs more than $1 trillion in new annual revenue to earn a return on that buildout. Someone has to pay, and it is the customer. Bersin recounts CIOs startled by “high Claude Code costs”; PagerDuty’s CIO told him he was “preparing myself to be surprised” by upcoming bills. An Investing.com analysis of the token pricing crisis adds the casualties: Uber burned through its entire 2026 AI budget in four months as Claude adoption jumped from 32% to 84% of its 5,000 engineers, with per-engineer API costs running $500 to $2,000 a month; GitHub’s June 1 shift of Copilot to usage-based billing left one developer projecting a jump from €67 to €966 a month — roughly 1,340%.
The pricing pressure lands on labs that can least afford restraint. OpenAI generated $5.7 billion in first-quarter revenue while Anthropic approached $4.8 billion, and Anthropic’s current fundraising round values the company at up to $950 billion against OpenAI’s $850 billion, per the same Investing.com analysis. Valuations like those are promises about future revenue per token. Every workload that defects to a model priced at $0.09 per million tokens is a small breach of that promise — and the defections are no longer small.
Capability was supposed to be the moat that justified those prices. It is eroding at the margin. CNBC notes that GLM 5.2 landed within a percentage point of Anthropic’s Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost. Brookings’ Chan estimates Chinese models sit “six to nine months” behind the American frontier — a real gap, but one that only matters for tasks that need the frontier. “The new open source models are performing well and prove capable for all but the most complex LLM tasks,” Summerville said. For the email-drafting, code-completing, document-classifying bulk of enterprise AI, six months behind the frontier is indistinguishable from the frontier. This is the same dynamic that let Chinese labs flood the market with releases in the year after DeepSeek’s breakout: they do not need to win the race; they need to stay within drafting distance while charging a tenth of the price.
All the ways the discount dies
Every arbitrage attracts a regulator, and this one has attracted two committees.
In April, the House Homeland Security Committee and the House Select Committee on the CCP opened a joint investigation into Airbnb and Anysphere, the maker of the Cursor coding platform, over their use of Chinese AI models — Semafor broke the story. Airbnb had built its customer-service agent on Alibaba’s Qwen, which CEO Brian Chesky cheerfully described as “fast and cheap” — a phrase that, as Forbes noted, has since become Exhibit A in congressional letters. Anysphere’s Composer 2 model, marketed as performing comparably to top OpenAI and Anthropic models at a fraction of the cost, turned out to be built on Moonshot AI’s Kimi. The committees’ own announcement frames the probe explicitly as a national-security matter, and the letters demand details on model choices and communications with Chinese providers, per Nextgov. The message to every CIO watching: the discount comes with a subpoena risk.
The security argument deserves to be taken seriously rather than dismissed as protectionism. When a company calls a Chinese-hosted API, its prompts — which may contain customer data, proprietary code, or strategic plans — transit infrastructure subject to Chinese law. The standard mitigation is to run open-weight models on U.S. infrastructure, which is exactly what most serious enterprise deployments do; weights downloaded from Hugging Face and served on domestic GPUs send nothing to Beijing. But that distinction, however technically sound, is doing a lot of political load-bearing. If Congress declines to honor it — if “uses a Chinese model” becomes reputationally or legally equivalent to “sends data to China” — the arbitrage closes by fiat, regardless of where the inference runs. Lawmakers are already probing whether DeepSeek, Moonshot, and MiniMax trained on illicitly harvested American model outputs, which would taint the weights themselves, not just the hosting.
The second failure mode is economic: popularity is not a business. What is striking in the field reporting is how unsentimental the buyers are. “The output quality, to be honest, I can’t tell the difference,” Stu Clott, a San Diego developer who switched to Chinese models, told Rest of World. Lindy’s Crivello was pithier about reserving frontier models for frontier problems: “You don’t need God to write your email.” But Rest of World’s reporting also captured the paradox — DeepSeek’s share of gateway tokens on Vercel jumped from under 1% to 17% in May, while its share of revenue stayed around 1%. Developers like Ruben Garcia Jr. pay $500 a month to Anthropic and OpenAI for the hardest tasks and just $200 for the Chinese models that handle 90% of their volume. That ratio is wonderful for buyers and brutal for sellers. If Chinese labs are pricing below cost to buy share — and their token economics suggest many are — the discount is a subsidy with a shelf life. Anyone who has watched a subsidized market knows how it ends: consolidation, price normalization, or a state backstop. DeepSeek’s $7.4 billion first external raise suggests even the flagship needs capital to keep playing.
The third counterweight is that the frontier still matters where the money is. Regulated industries remain firmly on American models: Claude and ChatGPT still dominate usage on LaunchLemonade, an agent platform serving regulated sectors, even as GLM 5.2 cracks its top five, founder Cien Solon told CNBC. Anthropic did not overtake OpenAI in enterprise adoption by being cheap; it did so on trust, compliance, and capability at the hard end of the distribution. And benchmark parity claims deserve their usual haircut — “within a percentage point on one agentic benchmark” is a marketing sentence, not a deployment guarantee. Hugging Face’s Yacine Jernite put the structural worry best in the CNBC piece: users risk being “stuck having to choose between performant but expensive US proprietary models whose price and accessibility can quickly fluctuate, or using Chinese models as the only feasible alternative.” A market with only those two poles is fragile from both directions — one side can reprice you, the other can be legislated away.
Buy the barbell, meter everything
The likeliest future is not Chinese models winning or losing. It is the barbell hardening into standard architecture.
On one end: American frontier models for the tasks where capability, liability, and compliance genuinely bind — the 10% of workloads that produce the majority of value and nearly all of the risk. On the other: open-weight models, disproportionately Chinese for now, self-hosted or served by U.S. inference providers, absorbing the high-volume, low-stakes bulk. In the middle sits the new strategic control point — the router. Whoever owns the layer that decides, request by request, which model is good enough owns the customer relationship that OpenAI and Anthropic thought subscription pricing had secured. OpenRouter’s rise from plumbing to industry census is the tell; Vercel’s gateway data becoming a market indicator is another.
Expect three responses. First, American labs will segment harder — cheaper distilled tiers, aggressive caching, capability-gated pricing — because losing the bulk tier to a 93x price gap is survivable only if you keep the frontier tier’s margins. Google is already signaling this direction with each Flash release priced an order of magnitude below the frontier. Second, Washington will keep escalating; with the committees’ investigations live and model export politics already inflamed, some restriction on Chinese models in sensitive sectors before the midterms looks more likely than not. Third, Chinese labs will keep converting open weights into ecosystem gravity rather than direct revenue — Moonshot’s Kimi doubling from a $100 million to $200 million run-rate in a single month shows monetization is starting to follow usage, just slowly.
The honest uncertainty is sequencing. If regulation lands before U.S. labs reprice, enterprises get squeezed from both sides — Jernite’s bad binary. If repricing lands first, the Chinese share gains stall at the compliance frontier and the market settles into a durable two-tier equilibrium. Either way, the era in which an American enterprise could ignore where its tokens came from is over.
For operators making decisions this quarter, the checklist:
- Meter before you migrate. Instrument per-task token consumption and cost now; you cannot arbitrage a spend you cannot see, and agentic workloads hide 15x multipliers inside innocuous features.
- Route by task tier, not by loyalty. Define explicit quality bars per workload class and let an eval harness — not a vendor relationship — decide which model clears each bar.
- If you use Chinese models, self-host or use U.S. inference providers. Never send regulated or proprietary data to foreign-hosted APIs; the technical distinction between weights and endpoints is your only defensible position.
- Write the disclosure memo before Congress writes the letter. Airbnb and Anysphere were caught explaining after the fact. Document why each model was chosen, where it runs, and what data touches it.
- Negotiate with the arbitrage, not just the account rep. A credible, benchmarked ability to move 60% of volume to open weights is worth more in an Anthropic or OpenAI renewal than any procurement tactic.
- Stress-test the subsidy. Model your unit economics at 3x current Chinese-model prices. If the workload only pencils at $0.09 per million tokens, it is a bet on someone else’s balance sheet.
- Keep a frontier lane open. The hardest 10% of tasks — and the newest capabilities — will stay American and expensive for the foreseeable future. Budget for the barbell, not the average.
In other news
Anthropic brings Claude Code and Cowork to government — Anthropic launched Claude for Government in public beta, delivering Claude Code and Claude Cowork through a FedRAMP High authorized environment with agency-grade admin controls, tamper-evident audit logs, and spend governance. Anthropic acts as the contracted billing party, so agencies need no separate cloud-provider relationship (Claude blog).
Brussels fuses AI and cybersecurity policy — The European Commission presented an EU Action Plan on Cybersecurity and Artificial Intelligence, pairing measures to defend against AI-enabled attacks with programs to deploy AI in cyber defense across member states (European Commission).
Baseten raises $1.5 billion at $13 billion — The AI inference platform closed a Series F led by Altimeter Capital, Conviction Partners, Spark Capital, Sands Capital, and Wellington Management — its fourth raise in 18 months, underscoring how inference infrastructure keeps outdrawing model labs for growth capital (Crunchbase News).
Agility Robotics heads to public markets via SPAC — The humanoid-robot maker plans to merge with Churchill Capital Corp XI, positioning it as the first pure-play humanoid robotics company on U.S. exchanges, though CEO Peggy Johnson is tempering expectations about home robots anytime soon (TechCrunch).
Kling AI courts $2 billion at an $18 billion valuation — Kuaishou’s video-generation unit is in talks for a General Atlantic-led round after its annualized revenue run-rate quintupled to $500 million in twelve months, a rare U.S. growth-capital bet on Chinese AI ahead of a planned Hong Kong IPO (Analytics Insight).