There is a split happening in AI right now and today's news makes it impossible to ignore. On one side, the two most famous American labs are racing toward trillion-dollar public offerings backed by historic revenue. On the other, the actual work, the raw tokens being processed by developers worldwide, is increasingly running on cheap Chinese open-weight models. The valuations are American. The volume is Chinese. That gap is the most important thing in AI this week, and almost nobody is framing it honestly. Here are the three stories that define it.

Databricks CEO Ali Ghodsi disclosed this week, citing OpenRouter data, that Chinese AI models grew from roughly 1% of usage in 2024 to more than 60% by May 2026. OpenRouter is the largest third-party model-routing platform for developers, which makes its data one of the cleanest real-world signals of where developer usage actually goes, not benchmark marketing.

Here is the honest detail behind the headline. The 61% figure is token consumption among the top-ten models. The top three slots are all Chinese: MiniMax, Kimi, and Zhipu's GLM line. The reason is not that Chinese models are beating Claude or GPT-5.5 on raw capability. It is cost. DeepSeek's V3.2 runs around $0.28 per million input tokens. GLM-5.1 completed a 10-evaluation benchmark suite for $544 versus $4,811 for Claude. That is not a small discount. That is a roughly 9x cost gap on comparable work.

OpenRouter's COO put it plainly: Chinese open-weight models captured share because they are disproportionately heavy in agentic workflows run by US firms. Translation: American developers are using cheap Chinese models as the default workhorse for high-volume agent tasks, and calling expensive American models only when the job actually requires the top tier. Ghodsi calls it the "advisor model." Cheap open models as the default layer, premium US models as the occasional specialist.

The thing to understand: this is a usage story, not a capability story. The US still leads the frontier. But the frontier is not where most of the tokens are. Most of the tokens are in bulk agentic grunt work, and that work is going to whoever is cheapest. Right now that is China.

OpenAI is preparing to file a confidential S-1 with the SEC as early as today, targeting a September 2026 listing at a valuation of roughly $852 billion to $1 trillion. Anthropic, which has also engaged investment banks, is raising at a $900 billion valuation and targeting an October listing. Prediction-market data implies an 85% probability that OpenAI lists before Anthropic.

This caps the IPO wave we have been tracking all month. Cerebras went public and nearly doubled. SpaceX filed for what could be the largest IPO in history. Now the two biggest pure-play AI labs are both heading for the public markets within weeks of each other. The honest caveat: OpenAI's filing is confidential and described as "as early as today," so the accurate framing is that OpenAI is preparing to file, not that it has gone public. Trading would be months away.

The reason this matters beyond the headline number: going public means these labs have to disclose real financials and answer to public shareholders for the first time. The era of raising private billions on vibes and vision is ending. The spreadsheets are about to become public record.

ending June 2026, up 130% from $4.8 billion in Q1. The company says it is on track for its first-ever quarterly operating profit. Reporting described it as the most dramatic single-quarter revenue jump in AI history.

This is the number that justifies the $900 billion valuation Anthropic is raising at. A 130% quarter-over-quarter revenue increase at this scale is almost unheard of. It also reframes the SpaceX detail from earlier this month: Anthropic is paying SpaceX $1.25 billion every month through May 2029 for GPU compute. That monthly bill only makes sense if you believe revenue keeps compounding at something near this rate. Anthropic clearly does.

Where the US-China AI race actually stands today, by what matters:

The pattern: the US owns the ceiling and the cap table. China owns the floor and the volume. The fight is over everything in between.

Today's story is that cheap models do bulk work 9x cheaper. Here is how to actually use that.

If you run any repetitive AI task at volume, summarizing, tagging, drafting, or classifying, route it through a cheaper model and save your premium model calls for the hard stuff. On OpenRouter you can A/B the same prompt across DeepSeek V3.2 and Claude side by side in two minutes. Run your real workload through both, compare the output quality, and you will usually find the cheap model is good enough for 70 to 80% of your tasks. Keep Claude or GPT-5.5 for the 20% that actually needs frontier reasoning. That is the "advisor model" Databricks' CEO described, and it can cut an API bill by more than half.

Three stories, one split. Chinese models hit 61% of top-model token usage on OpenRouter, up from 1% in 2024, driven by roughly 9x lower cost, doing the bulk of the world's agentic grunt work. Meanwhile OpenAI is preparing to file a confidential S-1 as early as today at up to a $1 trillion valuation, with Anthropic close behind at $900 billion targeting October. And Anthropic projected a $10.9 billion June quarter, up 130%, its first operating profit ever. The valuations are American. The volume is Chinese. The US owns the frontier and is monetizing the top of the market. China owns the usage and is winning on price. Both bets are now public, and the next year decides which one was right.

Stay building. 🤖