Menlo Ventures Report: 89% of Enterprise Open Source Model Spending Goes to Chinese Models, Kimi Accounts for 61%

In the early hours of this morning, US stocks were slaughtered.

The Nasdaq index plunged 5.52%, its biggest single-day drop since April 2025, officially entering technical correction territory. The Philadelphia Semiconductor Index crashed 12.3%, its worst day since the pandemic crash of March 2020.

Nvidia fell 9.26%, wiping out nearly $450 billion in market value in a single day; AMD fell 17%, SanDisk fell 22%, and Broadcom fell 10%.

Media estimates put the one-day loss across US stocks at roughly $2.8 trillion!

Yet just two days earlier, on November 3, Menlo Ventures—a 50-year-old Silicon Valley venture capital firm and an early investor in Anthropic—released its "2026 State of Enterprise AI Report," and the numbers inside pointed in a completely opposite direction from the stock market:

Enterprise AI deployment spending in 2026 soared from $33 billion to $177 billion, a 5.4x surge in a single year. Three years ago this market was only $1.7 billion—a 100x increase in three years.

On one side, $2.8 trillion goes up in smoke; on the other, $177 billion in real money is being poured in.

The contrast is absurdly stark. Wall Street is questioning whether supply-side capital expenditure can earn back its cost, while on the demand side, CTOs are adding orders until their budgets spiral out of control.

The report also contains two more intriguing threads:

Chinese models captured 89% of enterprise spending on open-source models, with Kimi accounting for 61%; at the same time, enterprise money is accelerating toward more expensive flagship models, with the spending share of frontier models rising from 35% to 63%.

Put these two things together, and that's where the real show begins.

Chinese Models Are "Topping the Charts" on American Companies' Books

In enterprise spending on open-source models, the Chinese contingent occupies the top four with overwhelming dominance: Kimi 61%, DeepSeek 15%, GLM 11%, Qwen 3%—89% combined. Meta's Llama, the West's most presentable open-source contender, took only 7%.

In other words, when American companies run open-source models in production environments, nine times out of ten the underlying model comes from China.

But what's more interesting is that Chinese models are no longer just "cheap alternatives"—they are starting to "grow" into other people's products.

Just yesterday, November 5, Thinking Machines Lab released a new model, Inkling 2 Small. This company needs little introduction: founded by former OpenAI CTO Mira Murati after her departure, it raised $2 billion in a single seed round, and its most recent round was discussed at a $400 billion valuation—one of the highest-caliber star labs in Silicon Valley.

The model itself is no slouch either: a 594-billion-total-parameter, 17-billion-active MoE, scoring 31 on the Artificial Analysis Intelligence Index.

But none of that is the point. The point is hidden in an inconspicuous corner of the technical blog—

In the latter stages of pretraining, most of the data fed to the model was not scraped from the web, but generated by having several earlier models run through it first. The first one named: Moonshot AI's Kimi K3.

What do you call this? You call it Kimi serving as a teacher to a star Silicon Valley lab.

Being a teacher requires one prerequisite: the answers you write are ones others are willing to copy. A year ago, this role belonged almost exclusively to GPT and Claude.

Being a teacher is only one use case. Kimi K3 is also serving as a foundation for American companies.

In August this year, legal AI unicorn Harvey—valued at over $10 billion, with OpenAI still a shareholder—released its first in-house model, Tenet.

What was the base? Not shareholder OpenAI's GPT, but Kimi K3.

Harvey used only about 150 Nvidia B300s and spent two months doing post-training for legal tasks on Kimi K3.

As a result, Tenet nearly doubled the number of tasks completed on its own legal agent benchmark, and on multiple complex legal tasks it traded blows with the flagship models of Anthropic, OpenAI, and Google.

Model builder Thinking Machines uses it as a teacher; application builder Harvey uses it as a foundation. From the lab to the application layer, Kimi has squeezed its way in.

For these companies, open weights mean getting a knife they can sharpen themselves, not just a cheap knife.

Chinese models have thus entered the production chains of top overseas applications—from being called upon to being used as teachers and foundations.

And the momentum is still accelerating.

Kimi K3.1 Pro, released just on October 23, self-reported a Terminal-Bench 4.0 score of 65.7%; then on November 5 it landed on Microsoft Foundry—

This is the first open-weight model on Microsoft's cloud to adopt a "revenue-sharing" model. Simply put: every time an enterprise calls Kimi on Microsoft Azure, Moonshot AI gets a cut.

And before Microsoft, Kimi had already landed on Amazon AWS. Microsoft and Amazon, the world's two largest clouds, have successively issued "entry tickets" to Chinese open-source models.

Looking only at usage, the open-source momentum is indeed fierce: token share rose from 11% to 19%, and monthly processing volume surged 22.5x to about 2.9 quadrillion.

But the money? Spending share actually fell from 6% to 4%.

The reason is sobering: over the past year, most of the open-source camp has been grinding away at the same thing—building smaller, faster, cheaper "flash"-tier models and fighting a relentless price war.

In one year, the average price of open-source tokens fell 3.3x, now just $0.13 per million tokens; closed-source models average $0.70, more than five times as expensive.

The result of the price war: twice the work done, yet less money collected—all the incremental gains were eaten up by price cuts.

The open-source Flash models fighting the price war won the buzz; the flagship models won the wallets.

Enterprise budgets don't care about open source versus closed source—only about who can get the job done. It's just that in the open-source camp, the players good enough to qualify as flagships can be counted on one hand.

This is the most counterintuitive finding in the entire report: the spending share of frontier flagship models soared from 35% to 63%, and their token share doubled from 18% to 42%.

Last year, enterprises' mainstream choices were Claude Sonnet 4.5 ($3 per million input tokens) and GPT-5 ($1.25); this year they switched to Claude Opus 5.5 ($4) and GPT-6 Astra ($10). Prices multiplied several times over, yet users didn't flee—they flocked in.

Where did the money flow? Anthropic rose from 40% to 55%, OpenAI from 27% to 31%, and the two together captured 86% of enterprise LLM API spending. Google fell from 21% to 9%, falling behind.

Meanwhile, players like Kimi K3, DeepSeek V4, and GLM-5.3, whose performance is already approaching flagship level, together account for less than 5%. The capability is there, but enterprises still haven't dared to hand over their lifeline.

The worst off is the layer stuck in the middle: small closed-source models cut prices by over 80%, yet their token share still fell from 29% to 13%. Above, flagships siphon off complex tasks; below, open source grabs the simple work—squeezed from both ends, hit from both sides, with nowhere to belong.

So why are enterprises willing to spend more?

Imagine this: having AI fix a production outage—read the code, check the logs, pinpoint the problem, modify files, run tests. A cheap model goes off track every few steps, and you have to explain, check, and redo it over and over, and in the end you still have to do it yourself. You save on the token bill, but you burn an engineer's entire afternoon.

So the ledger enterprises keep in their heads isn't really about a few dollars per million tokens—it's about how much it costs to get the whole thing done from start to finish. Model invocation fees are just a rounding error; people watching, waiting, and fixing—that's where the real money burns.

Cheap is certainly good, but there's a precondition—the work actually has to get done.

But then again, can open source not handle big jobs?

Take Kimi K3.1 Pro, for instance: a 2.8-trillion-parameter MoE architecture, supporting a 1-million-token context window, scoring 65.7% on Terminal-Bench 4.0, surpassing GPT-6 Astra.

On the financial research benchmark OfficeQA Pro v2, it scored 64.4, ranking first globally. Its hallucination error rate was pressed down from 23.4% in the previous generation to 6.85%.

This is no longer "good enough to get by"—this is arm-wrestling with closed-source flagships.

What open source has always lacked is not intelligence, but enterprises' confidence in it—compliance, after-sales support, long-term roadmap; these scores "outside the exam room" are the final hurdle.

Ultimately, today's U.S. stock market plunge and the data in this report are two sides of the same story.

Wall Street is watching the supply side: massive capital expenditure, data center approvals, power costs—every one a sword hanging overhead.

Meanwhile, the Menlo Ventures report is watching the demand side: for every $1 spent training a model, enterprises spend $1.70 using it—downstream demand is underwriting upstream investment.

Earlier this year, Wall Street had a panic round over "AI eating SaaS." But the report shows: 55% of new AI spending comes from entirely new budgets, and only 9% was shifted from traditional software budgets to AI-native vendors.

Tim Tully, a partner at Menlo Ventures, which wrote this report, said something very honest:

Flagship vs. open source. Flagship models took 69% of deployment spending, and the first-mover advantage of frontier labs is amplified in the Agent era.

But the report also said something meaningful—"the scarcity of capability won't last long."

Whenever open source catches up to flagships on a certain type of task, the original premium has to be re-examined.

Closed-source flagships have no permanent right to a premium.

Usage vs. revenue. Chinese open source has proven its capability with an 89% share, but between its 19% token usage and 4% spending lies a commercialization test.

Being called in large volumes is one hurdle; reliably handling complex tasks and keeping customers paying is the next.

The cases of Harvey and Thinking Machines point to one path—not relying on being cheap, but on becoming others' teacher and others' foundation; Microsoft's and AWS's revenue-sharing model points to another—partnering with cloud giants, getting a cut every time the model is called.

Stock price vs. bill. The contrast between Wall Street and enterprises isn't actually that contradictory.

Stock prices calculate expectations; bills record reality. $2.8 trillion evaporating in a day is expectations pulling back; a 5.4x rise in a year is reality accelerating. Both can be true at the same time.

But looking further out, what this report reveals may be more than just a business cycle—it may be a redistribution of industrial power.

The main workforce has changed too: 80% of tokens are burned by Agents themselves.

Whose model can fight, who trains the models, who helps you run the models—the division of labor along the entire chain is being rewritten by China's open-source forces.

English text is being prepared.