What happened
The Register reports that Jefferies is questioning whether large language models can become profitable as prices fall. The firm argues that Chinese, open- and smaller models could pressure the prices charged by OpenAI and Anthropic, whose demand is important to major cloud providers. The note cites OpenRouter data showing sharply increased token processing by Chinese models and estimates that Moonshot AI’s Kimi K3 performs at about 95 percent of Anthropic’s Claude Fable 5. These claims and comparisons have not been independently verified here.
According to The Register, Jefferies estimates that about half of the roughly $2.3 trillion in contracted backlog at Microsoft, Alphabet, Amazon and Oracle is associated with OpenAI and Anthropic. The firm argues that lower-cost Chinese, open- and small language models could weaken the pricing power and future revenue prospects of those model providers.
The Register cites OpenRouter data showing Chinese models processed 76.12 trillion tokens in the week ending September 27, compared with 4.37 trillion in late April and 14.2 trillion for the four leading US providers. The article also reports Jefferies’ claim that Moonshot AI’s Kimi K3 is competing at the frontier with an estimated 95 percent of Claude Fable 5’s performance.
Jefferies, as quoted by The Register, questions whether the major hyperscalers can justify projected infrastructure spending of $600 billion to $800 billion in the current year and asks whether they will spend $993 billion on capital expenditure in 2027. The article presents these as analyst estimates and questions, not confirmed spending outcomes.
Source details: theregister.com ↗
Why it matters
The argument directly challenges the economic assumptions behind the AI infrastructure buildout. If model capabilities become cheaper faster than demand expands, cloud providers could face difficulty recovering hundreds of billions of dollars in data-center spending. The Register also cites Bain’s estimate that AI may need $6 trillion in annual revenue by 2031 to support expected infrastructure demand, leaving a large gap even under optimistic assumptions. The figures are forecasts and analyst interpretations, not demonstrated outcomes.
The source’s central issue is the potential mismatch between falling prices and the enormous capital required to build and finance AI data centers. A cheaper model ecosystem could expand usage while simultaneously reducing the revenue earned per , making the economics of the infrastructure investment harder to assess.
The Register also cites Bain’s 2026 Global Technology Report, which estimates that AI would need $6 trillion in annual revenue by 2031 to fund anticipated infrastructure demand. Bain reportedly estimates $1.8 trillion from existing applications and $1.5 trillion from emerging categories, leaving a $2.7 trillion gap to be filled by products and uses that do not yet exist.
The source does not independently establish Jefferies’ accounting interpretation that all of the reported earnings increase at four hyperscalers came from other non-operating income, nor does it demonstrate that investment gains will reverse. Those points remain important but unconfirmed claims in the reported analysis.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Watch whether prices, model usage and paid demand develop as Jefferies expects, and whether cheaper open- models materially change enterprise purchasing. Also watch hyperscaler capital spending, debt levels and revenue from AI-related services. The source does not establish that a reset is imminent, that any provider is currently unprofitable, or that the cited models are broadly available under specific access or pricing terms.
There is no product launch or newly documented access change in this report. Pricing, availability and usage terms for Kimi K3 and the other models are not provided, so their practical accessibility cannot be assessed from the source.
Useful evidence would include independently measured model performance, sustained -volume data, actual -price changes, cloud-provider AI revenue and disclosed customer demand. The source does not provide enough information to determine whether Chinese or open- models are displacing paid US model usage.
The main risks to the Jefferies thesis are that lower prices stimulate much greater demand, new AI applications create additional revenue, or hyperscalers find profitable uses for capacity beyond the model providers named in the article. The source offers these possibilities but does not resolve them.