1. A Warning About the Cost of Building
Christopher Wood, the global head of equity strategy at Jefferies, has cautioned that the surge in American artificial intelligence spending could finish with large losses for investors. He set out the argument in the October 9, 2026 edition of his weekly GREED & fear newsletter, where he described the likely result as massive capital destruction across the US AI sector.
The reasoning is notable for what it leaves out. Wood does not argue that demand for AI services is fading. His concern is economic: inexpensive, openly available models developed in China are, in his view, gaining market share while the prices customers pay for AI output keep falling. If that continues, the returns that American companies expect from their enormous infrastructure budgets may not appear.
The note adds to a broader debate about whether current investment levels can be justified by future revenue. Wood's contribution is to frame the risk around pricing and competition instead of technological capability, a distinction that changes which warning signs investors ought to monitor.
Wood is a well-known voice among market strategists, and his weekly commentary is widely read by institutional investors who follow global equity allocation. His latest warning therefore carries weight in discussions about how much risk is already priced into technology shares, even though it remains a single analyst's judgment and not a consensus view.
2. The Scale of Hyperscaler Capex
At the center of the argument is the capital expenditure of the largest US cloud operators, often called hyperscalers. These are companies that rent computing capacity to customers and therefore build and run vast fleets of data centers. Their combined capex guidance for 2026 stands at about $695 billion, and the figure is projected to climb to roughly $870 billion in 2027.
Capital expenditure in this context covers the physical and technical foundations of AI: the buildings that house servers, the chips that run models and the networking and cooling hardware that keeps them operating. Such spending is paid up front, while the revenue it is meant to generate arrives over many years, which makes the investment sensitive to any change in expected pricing.
Wood characterizes a large share of this outlay as malinvestment. The term refers to money committed to projects that are unlikely to earn back what they cost, usually because the assumptions behind them prove too optimistic. Applying it to hundreds of billions of dollars of planned spending is a strong claim, and it rests on the view that the economics of AI delivery are shifting against the builders.
To put the numbers in perspective, the step from $695 billion to $870 billion represents an increase of about a quarter in a single year. Spending that grows at that pace assumes that customers will pay for the capacity as soon as it comes online. If pricing weakens at the same time, the gap between what has been built and what can be earned could widen quickly.
3. Borrowing Raises the Stakes
A second theme in the note is how the buildout is being paid for. Wood's analysis says the industry has moved substantially toward debt financing instead of covering expansion from its own cash. That change matters because the consequences of a poor investment depend on how it was funded.
When a company spends internal cash on a project that disappoints, the cost shows up as lower profit and a weaker return on capital. When it borrows to do so, the same disappointment can also strain its ability to service obligations. Lenders and bondholders then become exposed, and a problem that began as a strategic misjudgment can turn into a credit issue.
This is the channel through which Wood sees capital destruction becoming a market problem. Heavy borrowing tied to assets whose earning power is in doubt can produce losses that spread beyond shareholders, and it can set the stage for a broader correction if confidence in the underlying projects slips.
None of this means a downturn is certain. Companies with strong cash generation can absorb setbacks better than those that are highly leveraged, and the note describes a risk scenario rather than a fixed outcome. The point is that the margin for error narrows as borrowing rises.
4. Chinese Models and the Token Count
To show that competition is already material, the note points to usage data. Chinese AI models processed 36.39 trillion tokens in the week ending July 19, 2026, compared with 7.39 trillion for leading American models in the same period. Tokens are the small units of text that a model reads and produces, so they serve as a rough measure of how much work the models are actually doing.
The comparison comes from a single week, and the underlying data provider is not identified in the summary of the note, so it should be viewed as one data point rather than a complete picture of the market. Still, a gap of that size is the kind of evidence Wood uses to argue that Chinese systems are being adopted at scale.
Usage figures of this type are also imperfect for another reason. Token counts say how much text was processed but not how much revenue it earned, and cheaper models can handle more volume at lower margins. The numbers therefore support the claim about adoption more directly than the claim about profitability.
5. Price Is the Deciding Factor
Pricing is the strongest part of the case. Models such as Moonshot AI's Kimi K3 and Zhipu AI's GLM-5.2 are reported to cost roughly one quarter as much per token as some American alternatives. For businesses that process large volumes of text, a price difference of that magnitude can outweigh small gaps in performance.
Wood draws a direct line from these facts to the spending outlook. If the price per token continues to fall and Chinese models keep taking share, he argues, the capex that US hyperscalers have projected becomes difficult to sustain. Fewer customers paying less would leave the new capacity underused and its cost harder to recover.
Open-source distribution strengthens the effect. Companies can adopt such models widely without the licensing terms attached to closed systems, and they can run them on their own infrastructure. When comparable output is available for much less, the premium that supports huge infrastructure budgets is harder to defend.
Lower prices also change customer behavior. When a service becomes cheap enough, businesses that once limited their use begin to apply it more widely, which can lift total demand. Wood's argument is that this increase in volume does not necessarily compensate for the fall in price, especially for providers who have already committed capital on the basis of higher expected rates.
6. A Race Measured in Cost
Public discussion of AI often treats it as a contest over capability, with each new release judged by what it can do. Wood's framing is different. He asks who can deliver usable AI at the lowest cost per token, a measure closer to how commodity industries compete than to how technology showcases are judged.
This shift in perspective has practical implications. In a market decided by cost, the leader is not necessarily the developer with the most advanced system but the one that offers acceptable quality at the lowest price. Producers with large fixed costs then face pressure, because they must recover their investment from a falling unit price.
The argument does not claim that American developers lack skill. It says that technical leadership may not translate into high returns if rivals can match results cheaply. That is why Wood places the risk on the investment side rather than on the technology side.
Historical parallels are often drawn with earlier infrastructure booms, in which large networks were built ahead of demand and later repriced after competition intensified. Whether AI follows that path is uncertain, but the comparison explains why investors pay attention when a senior strategist questions the return on such spending.
7. Implications for Investors
The most immediate pressure point is the balance sheet. Strategies that rely on heavy borrowing to fund growth could attract closer scrutiny if conditions worsen, and lenders may demand more evidence that the projects can pay for themselves.
Profit margins are a second concern. Cheaper Chinese alternatives could squeeze profitability for American AI providers and add to volatility across the technology sector, which has become a large part of major stock indexes.
Wood points to semiconductor companies as possible beneficiaries and favors picks-and-shovels investments, meaning businesses that supply the tools for a boom instead of betting on which developer wins. There is a tension in that view, however. The spending he calls malinvestment is revenue for those suppliers, so a pullback in capex would also hit the hardware chain, not just the hyperscalers themselves.
For long-term investors, the practical question is how much of the technology sector's valuation depends on today's pricing holding up. A portfolio concentrated in the largest AI spenders carries a different risk from one that includes the companies that supply components, and the note encourages readers to think about that difference.
8. Signals to Watch
The first indicator is future capex guidance. Any reduction to the $695 billion planned for 2026 or the $870 billion projected for 2027 would suggest that companies are taking the warning seriously or have revised their expectations of demand.
The second is usage data. If Chinese models continue to process very high token volumes while per-token prices keep falling, the pricing pressure on US providers will be harder to dismiss. A reversal in either trend would weaken the case.
The third is the financing mix. Whether hyperscalers keep turning to debt markets will show how exposed the sector is to the credit risk Wood describes. As with all strategist warnings, the forecast is an opinion based on selected data, and the actual outcome will depend on how revenue, prices and funding costs develop over the coming quarters.