Jefferies Says Cheap Chinese Models Could Reset AI Infrastructure: Claimed vs Verified

An analyst note says the AI build-out could be repriced by models that cost a fraction as much to run. On October 13, 2026, The Register reported on a Jefferies research note arguing that cheap Chinese and open-source LLMs could trigger a trillion-dollar AI infrastructure reset. It is a forecast, not an event, and several of its numbers are the analysts' own estimates. This post separates what is claimed from what can be checked, and then turns it into decisions for people who build with or pay for AI.

We are not covering stock moves here. The reason this note is worth a builder's time is the mechanism it describes: token prices falling faster than the contracts that were signed to serve them.

TL;DR: claimed vs verified

What the note actually argues

According to The Register's summary, the Jefferies note makes a chain of arguments.

1. Concentration. A large share of cloud providers' contracted backlog is owed by two labs, OpenAI and Anthropic. If those labs cannot keep their prices, the backlog's quality is in question.
2. Price pressure. Cheap Chinese and open-weight models undercut what the two US labs charge. Jefferies says it "can no longer be assumed" that LLMs will ever be profitable given the collapse in token pricing.
3. Capex doubt. The note asks whether the four hyperscalers will really spend $993 billion on capex in 2027.
4. Pricing risk. It argues that markets may not have fully priced these risks.

The Register adds context from Bain and Company's 2026 Global Technology Report: the AI industry would need $6 trillion in annual revenue by 2031 to fund the infrastructure expected, existing applications might produce $1.8 trillion, and a $4.2 trillion shortfall would remain. Even with $1.5 trillion from emerging uses such as search, ads, autonomous vehicles and simulation, a $2.7 trillion gap remains, which Bain suggests might be filled by "new products and uses that don't exist today."

Notice what is and isn't in evidence. The note presents a risk scenario. It does not say a reset has begun, and The Register does not report lab revenue, margins or cancelled contracts.

The OpenRouter numbers, and their limits

The most concrete datapoint is the token volume. Per Jefferies, Chinese models processed 76.12 trillion tokens in the week ended September 27, against 4.37 trillion in late April, a roughly seventeen-fold increase, while the top four US providers processed 14.2 trillion. You can browse live rankings on OpenRouter.

Three caveats matter before you quote this:

• It is one marketplace. OpenRouter reflects the traffic of developers who route through it. Enterprise contracts direct with OpenAI, Anthropic or Google do not appear. A model that is free or deeply discounted can dominate token counts without dominating revenue.
• Tokens are not value. Agent loops and cheap bulk jobs consume enormous token counts. A trillion tokens of low-priced batch work is not equal to a trillion tokens of frontier reasoning.
• Promotions distort it. Free tiers and launch pricing can inflate a model's weekly share. We covered how free open-weight models get pushed through coding tools, which is one way volume rises without matching revenue.

This is the sort of statistic worth citing with its source rather than as a fact about the whole market.

Is "95 percent of Fable 5" believable?

Jefferies, per The Register, calls Moonshot AI's open-source Kimi K3 "competing at the frontier" with an estimated 95 percent of Claude Fable 5's performance. The direction matches what we have seen from the gap narrowing on public leaderboards: our write-up of Fireworks Ember 1 serving Kimi K3 with reasoning tokens covers how cost and reasoning-token behavior differ in practice. But "95 percent" is a composite that hides where models differ. Frontier models still tend to win on the hardest long-horizon agent tasks, while open models close the gap on routine coding, extraction and classification.

The useful test is your own workload. Coinbase's account of a fine-tuned small model beating much larger ones on fraud detection is a good template: Coinbase's Qwen3.5 9B fraud agent beat Opus 4.5. That was a measured task with a measured result, which is exactly what a headline percentage is not.

Why the "reset" scenario is plausible, and why it is not certain

A feather resting on a low stack of coins, illustrating falling token pricing behind the AI infrastructure reset argument

Why plausible. Inference cost per unit of capability has been falling for years. Smaller and sparse models, better serving stacks and competition from open weights all push in one direction. Our look at small models and the economics of Luna makes the practitioner case: many tasks do not need the biggest model. If buyers move work to cheaper models, a lab with fixed compute commitments has to fill capacity at lower prices.

Why not certain.

• Demand elasticity: cheaper tokens often increase total usage, and agents multiply it. Falling price per token does not require falling revenue.
• Capability premium: the top of the frontier keeps commanding prices for tasks where reliability matters. Recent coverage shows Opus 5.5 topping the Epoch Capabilities Index, a reminder that frontier labs are still moving.
• Hardware constraints: demand for chips and power remains tight. We tracked Nvidia's share in China against Huawei, and supply is still a bottleneck for everyone.
• Forecast risk: analyst notes are written to provoke attention to risk. The IMF made a related warning on circular financing and share-price crashes in its stability report, which we summarized in the IMF financial stability coverage; neither document is evidence that any single contract will fail.

What this means for what you build or pay

This is the section that justifies the story for a builder.

1. Assume prices keep dropping, but do not assume they drop evenly. Budget on cost per completed task, not price per million tokens. A cheaper model that needs three retries may cost more.
2. Stay model-agnostic. Put a routing layer between your product and any single vendor, and keep prompts, tools and evals portable. If a Chinese or open-weight model passes your evals at a fraction of the price, you should be able to switch in an afternoon.
3. Run your own bake-off. Pick 50 real tasks from your product, run the incumbent and two open-weight candidates, and compare success rate and cost. Our guide to running a coding-agent eval on real repos is a starting template.
4. Check the terms and data path. Cheap hosted access to a foreign-lab model raises data-residency and compliance questions. Self-hosting open weights avoids some of those and adds operations work.
5. Watch contract structure. If you sign committed-spend deals with a lab or cloud, ask for flexibility on model mix and price step-downs. Today's committed price may be tomorrow's premium.
6. Don't treat this as a trading signal. Whether a bubble pops is not what this post can tell you; what you can control is how exposed your product is to one supplier's price.

A prompt for your own planning: "List the five most expensive model calls in our product, the accuracy each needs, and one cheaper model that could plausibly meet it. Propose an eval to test that."

How to check these claims yourself

• Read the original report at The Register and note that the Jefferies note itself is a client research product; we did not see it.
• Check OpenRouter's public rankings for model share and compare weekly trends, remembering the marketplace caveats above.
• Compare lab pricing pages against open-weight host pricing for a model you could substitute.
• For the backlog claim, the numbers to look for are the remaining performance obligations disclosed in each hyperscaler's quarterly filings, and any named customer concentration. We have not matched the Jefferies "half" figure against those filings.
• Treat any single percentage as a hypothesis until it survives your own evals.

What we could not confirm

We did not find primary documents for the Jefferies note, and The Register does not link them. We did not independently verify the $2.3 trillion backlog, the $993 billion 2027 capex figure, the earnings decomposition or the Kimi K3 95 percent estimate. We also did not find any lab statement responding to the note. If OpenAI, Anthropic or the hyperscalers respond, or if the note becomes public, we will update this post.

Bottom line

The note's honest core is simple: the cost of intelligence is falling, and long-dated infrastructure commitments assume it will not fall too fast. Whether that produces a "reset" is an open forecast. For builders, the practical takeaway does not depend on who is right: design for a world where a good-enough model gets cheaper every quarter.

Related reading

• Ultramarine 1 Flash: a post-trained GLM-5.3 Flash claimed Fable-class at 60x lower cost
• Fireworks Ember 1 and Kimi K3 reasoning tokens
• Coinbase's Qwen3.5 9B fraud agent beats Opus 4.5
• Small models have arrived: the economics of Luna
• IMF warning on AI share crash and circular financing
• Nvidia vs Huawei in China AI chips: claimed vs verified
• Claude Opus 5.5 tops the Epoch Capabilities Index

Figures are analyst estimates reported on October 13, 2026 and may be revised.