Cheap Chinese and open source LLMs could trigger trillion-dollar AI infrastructure 'reset'

Jefferies warns that cheaper Chinese and open-source AI models could pressure pricing and challenge hyperscalers’ ability to monetize datacenter investments. Microsoft, Alphabet, Amazon and Oracle have about $2.3 trillion in contracted backlog, roughly half tied to OpenAI and Anthropic, while the four are expected to spend $600–$800 billion on datacenters this year and Jefferies questions whether they will spend $993 billion on capex in 2027. The note also says non-operating income accounted for all of the group’s $222 billion annualized earnings increase to $581 billion through 2Q26; Bain estimates the AI industry needs $6 trillion in annual revenue by 2031, with a potential $2.7 trillion gap after identified applications.

Analysis

The key risk is not simply cheaper models; it is a mismatch between long-lived data-center capacity and a customer base whose willingness to pay per unit of compute may fall faster than usage grows. If lower inference prices expand workloads, hyperscalers can still win through utilization, networking, storage and bundled cloud services—but token counts alone do not demonstrate profitable demand or capacity absorption. The most exposed economics are therefore model-provider rents and incremental returns on capacity, not necessarily all cloud revenue.

Near term (days), this is a sentiment and positioning risk for MSFT, GOOG, AMZN and ORCL; a durable repricing needs evidence in guidance, capex plans or credit spreads. Over 1–3 months, track cloud growth against capex, utilization commentary, backlog conversion and financing costs. Over 6–18 months, the key test is whether AI workloads broaden beyond a small set of model builders and support returns above the cost of capital. The cited non-operating earnings contribution warrants diligence, not a conclusion that operating earnings are overstated: inspect filings for investment gains, cash-flow conversion and segment performance.

The contrarian case is credible: cheaper inference can stimulate adoption and shift value toward cloud platforms and application owners. But that volume upside may accrue to customers rather than infrastructure owners if competition passes savings through. The thesis weakens if cloud growth and cash generation accelerate while capex intensity stabilizes; it strengthens if capex rises while backlog conversion, cloud margins or financing metrics deteriorate.

Key Decisions for Investors

• Treat the four hyperscalers as a capex-return risk basket, not as a uniform short. Avoid adding to AI-infrastructure exposure solely on token-growth metrics; require evidence of paid workload growth and capacity utilization.
• For existing concentrated exposure, consider a modest, defined-risk hedge on MSFT or a diversified technology position rather than an outright sector short. Initiate only if upcoming guidance shows capex rising without corresponding cloud-revenue or cash-flow improvement; no price target is justified by the supplied data.
• Set a 1–3 month diligence trigger: compare each company’s capex and cloud growth guidance, operating cash flow, backlog conversion and debt funding costs. Reassess if credit spreads widen alongside weaker conversion; remove the hedge if monetization and cash generation improve while capex intensity peaks.
• Verify the composition and cash realization of 'other non-operating' results in company filings before using reported earnings growth as an operating multiple denominator. A material reversal in investment gains would be a downside catalyst; recurring operating growth that offsets it would falsify the earnings-quality concern.
• Watch for beneficiaries beyond model providers: lower inference costs may help software and consumer businesses deploy AI, but there is no basis here to identify a specific winner. Wait for adoption-linked revenue or margin evidence before rotating into application names.