Menlo Ventures just finished counting the world's AI business, and the number that matters isn't the one they lead with. The lead is scale: enterprise AI deployment spending reached $177 billion by September 2026, up 5.4x from $33 billion a year earlier — roughly a 100x climb in three years, from $1.7 billion in 2023. For every dollar going into training, $1.70 of deployment revenue comes back. That's the headline everyone will quote.
The number that matters is buried in a single chart deeper in the report. Of the money enterprises spend on open-source AI systems, 89 percent goes to Chinese models. Moonshot's Kimi family takes 61 percent of it. DeepSeek takes 15. Z.ai's GLM takes 11. Alibaba's Qwen takes 3. The South China Morning Post, working off the same data this week, rounds it to "about 90 percent." Read either way, the conclusion is the same: the open AI stack that enterprises on Earth run is, dollar for dollar, Chinese.
The Number the Report Doesn't Sit With
"2026: The State of Enterprise AI" is a venture firm's annual census, and in most of it the story is still a very American one. Frontier US labs command 69 percent of deployment dollars, up from 50. Anthropic has climbed from 40 to 55 percent of enterprise API spend; OpenAI from 27 to 31; Google has fallen from 21 to 9. Anthropic and OpenAI together hold 86 percent of enterprise LLM API spend, up from 67 a year ago. The Chinese near-frontier models — Kimi K3, DeepSeek V4, GLM-5.3 — "approach frontier performance at a fraction of the price," in Menlo's words, and still manage just under 5 percent of that closed-API money.
On the closed side, it's a US contest. That framing is the one that will make the rounds, and it isn't wrong. But it isn't the whole report, and the part where it stops being a US story is the part the export-control crowd needs to read twice.
The Channel the Embargo Can't Close
Open-source spend isn't paid to anyone for the model. The weights are free, or nearly free. What enterprises actually buy is inference: the GPUs to self-host, the engineering to keep a thousand of them running, the platform bills that wrap around them. A secondary analysis by FourWeekMBA puts the multiplier at about 5.4x — the same token costs roughly five and a half times more through a closed API than through an open-weight deployment. So when Menlo says open-source is 19 percent of enterprise token usage but only 4 percent of spend, that gap is the value US closed-API providers capture. And inside that 4 percent, 89 cents of every open-source dollar is being spent hosting Chinese code.
Here's the geopolitical part. The chip embargo stopped American silicon going to China. It did not stop Chinese models going everywhere else. Open weights travel as a file, not a container, and the file now carries 89 percent of the open-source enterprise spend on the planet. Two years ago we called this the Trojan horse — China is giving away trillion-parameter models while the West celebrates its own demise. Since then the blitz has kept coming: open model after open model, with US makers watching in panic, Moonshot shipping K2.6 and then K2.7 as open weights at prices that undercut the American frontier on code, and Minimax going further with its 2.7-trillion-parameter open model. Then Moonshot leaked K3.1 with a 1-million-token context window and its reasoning benchmarks ahead of anyone else. What had been missing until now was the outcome of all that shipping: who is actually paying, and hosting, and running it in production. Menlo's 89 percent is the bill for all of it.
The Context Behind the Number
Context matters, because the 89 rides on top of a much larger machine. Agents are the engine of it. They now generate 80 percent of all tokens on OpenRouter — about 130.9 trillion per week, roughly 7x human-driven usage. Agent traffic overtook human traffic in February and has grown 37x since, against 5x for humans. The inflection was Anthropic's Opus 4.5 in November 2025, the first model that could execute multi-step work instead of just answering questions. The plumbing rebuilt fast: Harvey and Legora re-architected contract review into multi-agent systems, Sierra launched Horizon for agents that work over days and weeks, Factory's Droids doubled revenue month over month for six straight months. The labs' own coding agents grew from $1 billion to $24 billion in a year, and coding apps plus chatbots make up 76 percent of the $77.3 billion independent app market — $34 billion and $25 billion respectively — because they're the environments where agents can act, learn, and be verified.
The demand side is outrunning the plan. Enterprises spent 4.7x their 2025 levels against a 1.9x they had planned; adoption rose from 55 to 65 percent; and 55 percent of the incremental spend is net-new budget, not money shuffled out of software or labor line items. One quarter of AI's money is pure new money. That's the "SaaSpocalypse" fear, answered with data: the market is expanding, not cannibalizing.
What It Means
1. The one channel the embargo can't close just became mainstream. Every enterprise in Singapore, Lagos, or São Paulo that self-hosts a Kimi or DeepSeek model is running state-of-the-art Chinese intelligence on hardware they own, with no American approval, no American telemetry, and no American kill switch. The 89 percent is hard evidence that this isn't a niche of tinkerers; it's the default open channel for the majority of the world. The chip embargo closed the hardware pipe and the models flowed through the software one. Open weights are how Chinese AI capability reaches every enterprise on Earth, and it's now doing the work.
2. The value-capture gap is where the next war gets fought. Open-source models hold 19 percent of usage but 4 percent of spend, with the 5.4x price multiplier doing the capturing. That gap is the entire economics of why the American frontier labs are winning the money while Chinese models are winning the install base. One of the two numbers has to move. Either closed pricing collapses toward open, or open usage stops growing — and with agent traffic up 37x in a year, the usage side has momentum. Watch Menlo's next report: the 4 percent is the line that will move first.
3. One startup just became the largest open-source company on Earth. Sixty-one percent is not a market share. It's Kimi, from a five-year-old Beijing startup, outspending Meta, Mistral, Google's open families, and everything else combined in enterprise open deployments. The non-Chinese remainder of the open market — 11 percent — is the whole rest of the world's open-source AI industry, and it's the minority position. The open stack has a capital city, and it's Beijing.
4. The agent shift rewires who gets paid. When 80 percent of tokens come from machines doing work rather than people asking questions, spend attaches to tasks, not seats. That's why coding and chatbots — the verifiable agent environments — take 76 percent of the app market, and why net-new budget funds over half of the growth. The next pricing wars will be priced per outcome, and whoever runs the agent layer decides what the outcome costs.
🔥 Hot Takes
1. The biggest frontier collapse in the report is a US-vs-US story, and nobody is covering it. Google fell from 21 to 9 percent of enterprise API spend in a year. That's a faster fall than the entire US-vs-China dynamic, and it has nothing to do with Kimi. The value is concentrating in two labs — Anthropic and OpenAI at 86 percent combined — while the third American pillar quietly halves. If you're only watching the China flank, you'll miss the American defection inside the American camp.
2. "Open source" is doing marketing work it can't support. Eighty-nine percent of open enterprise spend doesn't flow to a public good. It flows to the weights of a handful of VC-backed Chinese startups — first and foremost Moonshot — which is a commercial distribution strategy that happens to ship free. Enterprises hosting this stack think they bought vendor neutrality. The number says they consolidated onto a single company's judgment about what its models will and won't do, with no recourse, no telemetry, and no one to call when the model misbehaves. Open weights were sold as the anti-lock-in. By this arithmetic, they've never been more locked in.
3. The 4 percent is today's number, not the durable one. The price moat only holds if the 5.4x multiplier stays fixed while usage share stops growing. It isn't. Open usage jumped from 11 to 19 percent in a year, and the demand engine — agents — is running at 7x human token velocity and growing 37x since February. At that trajectory, usage share closes the spend gap by arithmetic alone within a couple of report cycles. The labs' 86 percent of API spend is the last wall of the current pricing regime, and it's the wall agents will press first.
The Bottom Line
$177 billion is the number you put on the slide. 89 percent is the number you take home. The world's enterprises just spent their first real open-source AI money, and the overwhelming share of it went to models written in China — led by one startup's Kimi family — because open weights were the one channel the embargo never reached. The US keeps the closed-API cash for now, concentrated in two labs with a third quietly breaking. But the gap between 19 percent of usage and 4 percent of spend is the fault line, and on the agent-velocity curve it's the usage side that moves. The Trojan horse isn't in the stable anymore. It's doing most of the work.