- Menlo estimates Chinese models take 89% of spending on open-source models, extrapolated from a US survey. The South China Morning Post reports proprietary models still took most model budgets.
- SCMP gives Cursor as one example of a US product built on a Chinese open model. This story argues buyers should ask what base model sits under each AI product.
- Ask each AI vendor which base model powers its product, where it runs, and what happens if its maker changes terms.
A product may run on a model you never chose
A contract names a vendor. It does not always say which AI model the product runs on. That gap is the question this story raises. The argument is the publication's own, not Menlo's.
One case shows how the gap can open. SCMP describes a US coding product built on a Chinese open model. The case proves the route exists. It does not show how common the route is.
What Menlo estimates, and how narrow it is
The South China Morning Post reported on Menlo's latest enterprise AI analysis. Menlo estimates that Chinese models account for 89% of spending on open-source models. Menlo uses "open-source" as a broad label for open-weight models.
The figure covers one slice of the market. SCMP reports that proprietary models still took most model budgets. So the 89% describes a small category, not the whole market.
What the estimate rests on
Menlo says its global figures were extrapolated from a September survey. The survey reached close to 400 US enterprise decision-makers. Together they manage more than US$4 billion in AI budgets. Menlo also used public filings and third-party reports.
That makes the 89% an estimate built on an American sample. It is not a count of invoices worldwide. The SCMP article does not give the dollar size of the open-model category.
Menlo is also an active investor in Anthropic and in several AI infrastructure providers examined in its report. Readers should weigh the estimate with that in mind.
How an open model becomes someone else's product
A closed model is reached through the provider's hosted service. The provider keeps the model's internal settings, called parameters. An open-weight model lets developers download those parameters and change them.
Fine-tuning means training a finished model further on chosen examples. It adjusts the model rather than replacing it. So a tuned model starts from the base model's learned behaviour, and the base model is part of what a buyer is evaluating. A buyer who does not know it is missing part of the picture.
SCMP gives Cursor, a US AI coding platform, as one example. SCMP says Cursor's technical report from March states that its Composer 2 was built on Moonshot's Kimi K2.5. Per SCMP, Cursor trained the Chinese model further on data heavy in code. It then used reinforcement learning, which rewards good results, to improve software engineering performance.
Usage and the supply-chain framing
SCMP adds a usage figure. OpenRouter is a platform that aggregates models. On it, Chinese models' token usage topped that of US models for a 27th week running, from October 26 to November 1. A token is a small chunk of text a model reads or writes.
SCMP describes a split in the global AI supply chain. US firms lead in selling closed-source APIs. Chinese open models act as foundations that developers abroad adapt into their own services. That framing is SCMP's.
The lesson: ask what the product stands on
Manufacturers ask what parts go into a finished product. Software buyers learned to ask the same about open-source libraries. This shows AI products invite the same question. A vendor may be an assembler as well as a builder.
Nothing in the source says an open Chinese model is unsafe or unfit. The point is narrower. The next four points are the publication's own reasoning, not findings from Menlo or SCMP. Each may depend on the base model and on the vendor's licence and hosting setup. The source does not test any of them.
Licensing may be one. The base model's maker sets its licence terms, and a product built on it may depend on them. Support may be another. Fixes to the base model may come from a different party than the vendor. Changed terms may be a third. If the maker changes them, the vendor may have to respond.
Data handling may be the fourth, and it may turn on where the model runs. A vendor that hosts open weights itself may keep your data in its own systems. Ask where the model runs rather than assuming.
What to ask this quarter
Menlo puts coding tools and enterprise chatbots at 76% of the global enterprise AI application market, which it sizes at US$77.3 billion. Check whether tools of these kinds sit in your contracts. Then put three questions to each vendor.
First, which base model powers this product, and who built it? Second, was it adapted from an open-weight model, and which one? Third, if that base model changed or became unavailable, what would you do, and how fast?
Then ask your own security and procurement teams a fourth question. Do your AI vendor reviews record the model underneath, or only the company on the invoice? An AI product is a stack. Know what it stands on.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: South China Morning Post.
