InvestNot yet confirmed elsewhere1 publisher2 min readPublished Updated
Jefferies' Chris Wood calls much of the hyperscalers' AI spending malinvestment
Jefferies' Chris Wood says US hyperscalers' $695 billion of guided 2026 capex looks unsustainable as cheap Chinese open-source models win share. He blames a slump in the price of AI output, with demand intact, while guidance climbs toward $870 billion for 2027.
The Investor · Invest desk
Bar comparison of tokens processed in the week ending July 19, 2026: Chinese AI models processed 36.39 trillion tokens; top US AI models handled 7.39 trillion tokens.
Tokens processed, week ending July 19, 2026 In trillion tokens
| Item | Value | Claim |
|---|---|---|
| Chinese AI models | 36.39 trillion tokens | 6 |
| Top US AI models | 7.39 trillion tokens | 7 |
What happened
- Chinese AI models processed 36.39 trillion tokens in the week ending July 19, 2026, against 7.39 trillion for top US models.
- Moonshot AI's Kimi K3 and Zhipu AI's GLM-5.2 cost roughly a quarter as much per token as some US alternatives, according to Crypto Briefing.
- Wood says the industry has moved significantly toward paying for the buildout with debt instead of cash.
- His suggested cautious approach is 'picks and shovels' investments, and the report flags semiconductor firms as possible gainers.
Why it matters
- exposure Borrowed money turns a shortfall in returns into a problem for lenders as well as shareholders; Crypto Briefing ties heavy borrowing to credit risk and a possible market correction.
- constraint Because Wood's case rests on price, higher AI usage figures alone cannot refute it; revenue per token set against the capex can.
- decision US providers have to choose between holding per-token prices while Chinese models take volume, or cutting prices and lowering what each dollar of capex can earn back.
The guidance climbs $175 billion from 2026 to 2027, about 25% in one year, and the two years add up to roughly $1.57 trillion of data centers, chips and other hardware [14][15]. Wood, Jefferies' global head of equity strategy, warned in the October 9 edition of his GREED & fear newsletter that the boom could end in massive capital destruction across the US AI sector [1]. He called much of the outlay "malinvestment" [4]. Economists use the word for money sunk into projects unlikely to earn back what they cost [4].
Wood does not blame weak demand for AI [11]. His case is about price. At a quarter of the per-token price, a Chinese model needs four times a US model's volume to match its token revenue, and in the week to July 19 Chinese models ran at about 4.9 times the US count [16]. The one-quarter ratio is claimed only against some US alternatives [8]. If it held across every model, Chinese token billings that week would have come out about 23% ahead of the US total [17].
The verdict can fail in two places. A cheap open-source model still has to run on chips in a data center, so lower prices could lift volume enough to fill the capacity being built, and the hyperscalers would then earn thinner margins on full buildings. The second is the sample. The token comparison covers one week in July [6].
We think the report establishes the price pressure and leaves the capital destruction unproven. Token counts and a price ratio measure what customers pay for each unit of output. Whether $1.57 trillion destroys capital depends on what the hyperscalers earn from AI and on how much of the spend is borrowed, and the report does not include either figure [15][5]. The counter-case is volume: if cheaper tokens fill the new capacity at prices that still cover the cost of the money, Wood is right about margins and wrong about the capex.
His recommended position is the part we find hardest to square with the verdict. Chip suppliers book hyperscaler capex as revenue when it is spent, whether or not the data centers ever pay back, and Crypto Briefing noted that any pullback would ripple through the hardware supply chain [13]. Owning the suppliers as the cautious play pays only while the spending continues [10]. It is a position that profits if the hyperscalers ignore his warning through 2027 [3].
What to watch
- Hyperscalers' next capex guidance: a trim to either the 2026 or the 2027 figure would show the warning landing.
- Weekly token counts and per-token prices: a rising US share of tokens without US price cuts would weaken Wood's case.
- Whether hyperscalers keep borrowing for the buildout or shift back to funding it from operating cash.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence30
- Adoption45
- Hype gap+35
- Incentives
- Insufficient
- Confidence35
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Chris Wood, Global Head of Equity Strategy at Jefferies, warned in the October 9, 2026 edition of his GREED & fear newsletter that the US AI boom could end in massive capital destruction across the US AI sector.
- [2]
US hyperscalers have issued combined capital expenditure guidance of approximately $695 billion for 2026.
- [3]
Hyperscaler capex is projected to rise to $870 billion in 2027.
- [4]
Wood calls much of the hyperscaler outlay "malinvestment," an economist's term for money sunk into projects unlikely to earn back what they cost.
- [5]
In Wood's analysis, the industry has shifted significantly toward debt financing instead of funding the buildout from cash.
- [6]
Chinese AI models processed 36.39 trillion tokens during the week ending July 19, 2026.
- [7]
Top US AI models handled 7.39 trillion tokens during the week ending July 19, 2026.
- [8]
Models such as Moonshot AI's Kimi K3 and Zhipu AI's GLM-5.2 cost roughly one-quarter per token of some US alternatives.
- [9]
In Wood's view, slumping token prices combined with Chinese models capturing significant portions of the market make the projected US hyperscaler capex unsustainable.
- [10]
Wood advocates "picks and shovels" investments as a cautious approach amid increasing uncertainty; semiconductor firms are flagged as potentially standing to gain.
- [11]
Wood's culprit is not weak demand for AI but competition from inexpensive Chinese open-source models, which he expects to keep winning market share while the price of AI output slumps.
- [12]
Borrowing heavily to fund the AI buildout can create credit risks and potentially set the stage for a market correction.
- [13]
The capex Wood views as malinvestment is revenue for the suppliers building the infrastructure, so any pullback would ripple through the hardware supply chain.
- [14]
Hyperscaler capex guidance rises by $175 billion from 2026 to 2027, an increase of about 25%.
- [15]
The 2026 and 2027 capex figures together total roughly $1.57 trillion.
- [16]
In the week ending July 19, 2026, Chinese models processed about 4.9 times as many tokens as top US models.
- [17]
If Chinese tokens were priced at one-quarter of US tokens across all models, Chinese token billings in that week would be about 23% above the US total.
Sources
1 independent publisher whose own reporting we read for this story.
- cryptobriefing.comJefferies’ Chris Wood warns US AI boom could end in massive capital destruction
1 article · October 9, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Chinese open-source AI modelsFollow
- AI Infrastructure CapexFollow
- AI Token PricingFollow