On November 3, 2026, Menlo Ventures published its annual State of Enterprise AI report. The Silicon Valley venture firm, an early investor in Anthropic, surveyed nearly 400 U.S. enterprise AI decision-makers responsible for more than $4 billion in combined AI spending. The report's headline was a number that had been building for two years and had finally become impossible to frame as an anomaly.

Of every dollar U.S. enterprises spent on open-source AI models in 2026, 89 cents went to a Chinese model.

The breakdown was more specific than the headline. Kimi, developed by Moonshot AI, held 61 percent of U.S. enterprise open-source spending. DeepSeek held 15 percent. Zhipu's GLM series held 11 percent. Qwen held 3 percent. Meta's Llama — the most widely used American open-source model — held 7 percent.

The report also measured the broader market. Annualized U.S. enterprise AI deployment spending reached $176.8 billion by September 2026, up from $33 billion a year earlier — a 5.4-fold increase in twelve months. Model development spending was estimated at $102.5 billion. For every dollar spent developing models, roughly $1.70 was spent deploying them. The center of gravity in enterprise AI spending had moved from building models to using them.

That shift is the context for the 89 percent figure. The question is why American enterprises, deploying AI at a scale that dwarfs model development spending, are choosing Chinese open-weight foundations to deploy on.

What 89 Percent Measures — and What It Doesn't

The Menlo figure measures spending, not usage, downloads, or benchmark performance. It is a procurement statistic. It describes where enterprise budget lines are allocated, not which models developers experiment with.

That distinction matters because the U.S. enterprise market is not a monolith. A large share of enterprise AI spending still flows to closed American models — OpenAI, Anthropic, and Google. The 89 percent applies to the open-source segment of enterprise spending, which is a subset of the total. The report's broader finding is that closed models still dominate total enterprise spend, while open-source models are growing faster and are disproportionately Chinese.

The spending pattern also reflects a structural feature of how enterprises buy AI. A closed model is purchased through an API contract. An open-weight model can be downloaded, fine-tuned, and deployed on the enterprise's own infrastructure or a cloud provider of its choice. The spending that flows to a Chinese open-weight model may be zero in licensing fees — the weights are free — and appear instead as compute, integration, and support costs. The Menlo figure captures the total economic activity associated with deploying those models, not just the license line.

The 89 percent is best read as a measure of where enterprises are building, not where they are merely experimenting.

Why Kimi, and Why 61 Percent

The concentration of spending in a single model family is the most striking part of the data. Kimi holding 61 percent of U.S. enterprise open-source spending is not the pattern of a market evenly distributed across many options. It is the pattern of a market that has converged on a default.

Kimi K3, released by Moonshot AI in July 2026, is a 2.8-trillion-parameter open-weight model with a 1-million-token context window. It was benchmarked as second only to Claude Fable 5 and GPT-5.6 Sol among participating models. Its licensing terms are permissive for most enterprise uses, with a commercial agreement required for companies above a revenue threshold.

The combination of frontier-class capability, a permissive license, and a large context window is the profile that enterprise buyers select for production deployment. An enterprise that needs to process long documents, maintain multi-turn context, and run inference at scale is not choosing between Kimi and Llama on the basis of brand. It is choosing on the basis of whether the model fits the workload and whether the license permits the deployment.

Kimi's 61 percent share suggests that, for a large share of U.S. enterprise workloads, it does.

What Llama's 7 Percent Reveals

The most consequential data point in the report may be the one that is smallest.

Llama held 7 percent of U.S. enterprise open-source spending. That is a remarkable outcome for a model family that defined the open-weight movement. Meta released Llama to build an ecosystem, to commoditize the model layer beneath its application business, and to establish an American open-weight alternative to closed models. By the spending metric, that strategy has not produced the enterprise adoption that the release was designed to generate.

The reasons are partly technical. Llama 4 underperformed expectations. Meta's subsequent models, including Muse Spark, were released as closed models before the company pivoted back to open weights. The Llama family's most recent releases did not match the frontier-class capability of Kimi K3 or the cost efficiency of DeepSeek. Enterprise buyers choosing a foundation for production deployment selected the models that met their requirements, not the models that carried the American open-source banner.

The 7 percent does not measure Llama's quality. It measures the gap between the model that a company releases to build an ecosystem and the model that an enterprise chooses to deploy. Those are different selection criteria, and the enterprise market applied the second one.

The Cost and Capability Arithmetic

The spending pattern is consistent with the cost data that has accumulated over the past year.

A U.S. enterprise deploying an open-weight model at scale is making a total-cost-of-ownership calculation. The model weights are free. The costs are compute, integration, fine-tuning, and operations. A model that delivers frontier-class capability at a lower inference cost reduces every one of those line items.

DeepSeek's V4 Flash model, for example, is priced in the range of $0.09 to $0.14 per million input tokens and $0.18 to $0.28 per million output tokens, depending on the provider and caching configuration, according to public API rate cards. A comparable American closed model charges multiples of that rate. For an enterprise running millions of queries per month, the difference is not a marginal saving. It is the difference between a viable unit economic and an unviable one.

Kimi's pricing is higher than DeepSeek's but still below the closed American frontier. Its 61 percent share suggests that enterprises are willing to pay a premium over DeepSeek for capability, but not a premium that approaches closed-model pricing.

The enterprise market is not choosing Chinese models because they are Chinese. It is choosing them because the cost-per-unit-of-capability is lower, and the license permits the deployment the enterprise needs.

The Policy Disconnect

The 89 percent figure lands in a policy environment that has been organized around the assumption that American enterprises should not depend on Chinese AI.

The U.S. government has restricted the export of advanced chips to China, proposed investment restrictions in Chinese AI and quantum technologies, and sent letters to 35 Pax Silica signatory countries warning them against adopting Chinese AI infrastructure. The Treasury Department's proposed outbound investment rules, the Commerce Department's entity list additions, and the State Department's Pax Silica letters all point in the same direction: Chinese AI is a supply-chain risk, and American enterprises should treat it as one.

The Menlo data shows that American enterprises have reached a different conclusion. They are not treating Chinese open-weight models as a geopolitical risk. They are treating them as a procurement decision. And the procurement decision, by a margin of nearly nine to one, favors Chinese models.

That disconnect is not a failure of the enterprises to understand the policy. It is a failure of the policy to account for the economics. An enterprise that can deploy a frontier-class open-weight model for a fraction of the cost of a closed American alternative is not making a political statement by choosing it. It is making a budget decision. The policy framework has not offered a cost-competitive alternative that changes that decision.

What the 89 Percent Means

The Menlo report does not say that Chinese models are better than American models. It says that U.S. enterprises, when spending their own money on open-source deployment, are choosing Chinese models by a wide margin.

That is a market outcome. It reflects the combination of capability, cost, licensing terms, and deployment flexibility that enterprise buyers weigh when they commit budget to a foundation model. The Chinese models that won the spending share won it on those criteria.

The policy debate has focused on restricting the supply of Chinese AI. The Menlo data suggests that the demand is the more consequential variable. American enterprises are not buying Chinese models because they are forced to. They are buying them because the models work, the licenses permit the deployment, and the economics are better.

The 89 percent measures that demand. It also measures how far the policy framework is from changing it.

Sources: Menlo Ventures, State of Enterprise AI 2026, released November 3, 2026, via South China Morning Post (November 6, 2026), 36Kr (November 6, 2026), Wallstreetcn (November 6, 2026), Jiemian (November 6, 2026), and Guandian (November 6, 2026); 163.com on Llama's 7 percent share (November 6–7, 2026); Menlo Ventures report coverage on $176.8 billion enterprise spend via Eastmoney (November 7, 2026) and 163.com (November 6, 2026); Kimi K3 technical documentation via Moonshot AI (July 2026); DeepSeek V4 Flash pricing via BenchGecko and Vercel API rate cards (2026); U.S. Treasury proposed outbound investment rules via Treasury Department press release (October 22, 2026); State Department Pax Silica letters via Reuters (August 2026).

Disclaimer

The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.

Limitations

The 89% figure measures enterprise spending on open-source models, not total enterprise AI spend, where closed American models still lead.

The survey covers roughly 400 U.S. enterprise AI decision-makers and is self-reported, so the spending split carries survey sampling error.

Licensing terms described are as published by the model developers and may differ in practice for specific enterprise agreements.

API prices are drawn from public rate cards and vary by provider, region, and caching configuration.

Reporting standards differ between Chinese and Western outlets, and figures cited here reflect the sourcing available at the time of writing.

Market size and shipment estimates vary by methodology; where estimates conflict, this article cites the figure used by the original source.

The sectors discussed are changing quickly, and some figures may be revised after publication.


Sources

  1. Menlo Ventures, State of Enterprise AI 2026, released November 3, 2026, via South China Morning Post (November 6, 2026), 36Kr (November 6, 2026), Wallstreetcn (November 6, 2026), Jiemian (November 6, 2026), and Guandian (November 6, 2026)
  2. 163.com on Llama's 7 percent share (November 6–7, 2026)
  3. Menlo Ventures report coverage on $176.8 billion enterprise spend via Eastmoney (November 7, 2026) and 163.com (November 6, 2026)
  4. Kimi K3 technical documentation via Moonshot AI (July 2026)
  5. DeepSeek V4 Flash pricing via BenchGecko and Vercel API rate cards (2026)
  6. U.S. Treasury proposed outbound investment rules via Treasury Department press release (October 22, 2026)
  7. State Department Pax Silica letters via Reuters (August 2026).

The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.

Limitations: The 89% figure measures enterprise spending on open-source models, not total enterprise AI spend, where closed American models still lead.; The survey covers roughly 400 U.S. enterprise AI decision-makers and is self-reported, so the spending split carries survey sampling error.; Licensing terms described are as published by the model developers and may differ in practice for specific enterprise agreements.; API prices are drawn from public rate cards and vary by provider, region, and caching configuration.; Reporting standards differ between Chinese and Western outlets, and figures cited here reflect the sourcing available at the time of writing.; Market size and shipment estimates vary by methodology; where estimates conflict, this article cites the figure used by the original source.; The sectors discussed are changing quickly, and some figures may be revised after publication.