If the direction of the US ten-year Treasury yield is one key issue facing the US stock market, another is the issue of the AI capex arms race and whether the continuing sell-off in memory stocks means the market starting to sniff out that a deceleration in hyperscaler capex could be coming much earlier than what investors have been assuming.
The estimates for the four major hyperscalers’ capex this year and 2027 have been running at US$720bn and US$1.04tn respectively.
The capex guidance has been one of the key issues investors have been focused on with the hyperscalers’ recently reported third quarter results.
At the start of the second quarter of this year, this writer had been of the opinion that there was significant risk that this year could mark the peak in AI capex because the view was that investors would start to ask hard questions about where the returns will come from, given that the AI capex arms race is now more than three and a half years old and given that the nature of these former quintessentially asset light businesses has changed dramatically.
In the event, the hyperscalers have broadly maintained their capex guidance at 99% of their forecast operating cash flow this year.
Microsoft lowered its 2026 capex guidance slightly from US$175bn to US$170bn, though it expects capex to accelerate in 1H27 as the delayed capacity kicks in. Meanwhile, Amazon and Alphabet have maintained their 2026 capex guidance at US$220bn and US$195-205bn respectively. As for Meta, it raised its capex guidance slightly from US$130-145bn to US$135-145bn.
The latest hyperscalers earnings season has undoubtedly added to the nervousness on AI capex.
This is less because all four major hyperscalers again increased their capex in 3Q26 while three of the four maintained or increased their 2026 capex guidance.
But more because, importantly, the market chose to focus on growing evidence, in the cases of Alphabet and Amazon, that AI is beginning to cannibalise their legacy businesses in terms of search and e-commerce.
Google’s search revenue rose by 12% YoY to US$63bn in 3Q26, down from 19% YoY in 1Q26 and 17% YoY in 2Q26, while Google Cloud revenue surged by 103% YoY to US$30.7bn in 3Q26.
As a result, search revenue has for the first time fallen below half of Alphabet’s total revenue, declining from 52.8% in 2Q26 to 49.7% in 3Q26, while Google Cloud’s share rose from 20.7% to 24.1% over the same period.
As for Amazon, revenue from online stores rose by only 2% YoY to US$68.5bn in 3Q26, down from 15% YoY in 2Q26.
By contrast, Amazon Web Services’ (AWS) revenue rose by 39% YoY to US$46bn in 3Q26.
As a result, online stores’ share of total revenue declined from 37.4% in 3Q25 to 33.9% in 3Q26, while AWS’ share rose from 18.3% to 22.8% over the same period.
Yet both Alphabet and Amazon plan to continue to spend more money even as free cash flow is running out.
Amazon and Alphabet free cash flows declined to a negative US$35bn and US$9bn respectively in 3Q26.
So the questioning of the returns on AI capex has now begun and could well intensify in the forthcoming earnings season.
Investor Doubts About AI Returns are Growing
In this respect, the base case of this writer has always been that the most likely longer term outcome of the AI story in market terms will be massive capital destruction in the US with market share going to cheaper open-source Chinese large language models.
In this respect, the key issue has always been one of timing in the sense of when the market will start to worry about the returns on that investment while in the meantime the view has been since the start of 2023, when the AI story began in stock market terms, that the best AI-related investment was the picks and shovels trade, in terms of the companies who are profiting directly from the hyperscalers’ spending.
The first hint of a reassessment of the AI story, and a renewed questioning of the returns, came in May with the reaction against so-called tokenmaxxing previously discussed here (see The AI Pause Nobody Asked For, 24 September 2026).
This would have led corporates to cut back on their previous practice of positively encouraging employees to play around with AI models almost regardless of the cost.
One issue which can be revived at any time is the perceived security threat from Chinese models.
This writer has no view on the security threat but certainly has a view on the growing business threat from China.
The latest data from OpenRouter shows that Chinese models continue to grow in popularity.
The top Chinese AI models processed 99.55tn tokens on the global aggregator platform OpenRouter in the week ended 1 November, up from 4.37tn in late April.
This compares with 20.68tn tokens for the top US models, based on the weekly usage of the top nine models on OpenRouter.
This is the latest manifestation of a story which began with the so-called DeepSeek moment in late January 2025, in terms of the growing popularity and competitiveness of China’s open-source models.
Initially the focus was on the cheapness of China models and the related commoditisation threat facing large language models. But now there is also a growing realisation that China has become a technological peer to the US in AI, as well as in so many other areas.
The question now is whether all of the above has been priced fully into markets.
This writer does not think so, though it has clearly started to be with the China AI Semiconductor Top 10 Index down 50% since peaking at the end of June.
This index tracks the performance of 10 leading listed Chinese companies operating across the semiconductor, memory, and server segments of the AI value chain.
Debt is a Growing Risk to the AI Buildout
One important reason is because it is also important to understand that the AI story also faces growing credit risks.
In this sense, the AI capex spending has over the past year increasingly been funded less by hyperscalers’ cash and more by debt.
The hyperscalers have now issued more investment-grade debt than the energy sector in the US and indeed are now the largest single issuer having raised US$296bn of debt year to date compared with, for example, US$67bn of investment-grade issuance in the energy sector.
It is also interesting to note that the spreads on their debt have begun to widen.
The 10-year bond yield spreads over Treasuries of Amazon, Alphabet and Meta have risen from 61bp, 57bp and 87bp respectively on 3 July to 89bp, 81bp and 133bp.
Meanwhile, Oracle, which is sometimes viewed as the fifth hyperscaler, albeit a much more leveraged one, saw its US$130bn of outstanding debt downgraded by S&P to BBB- on 9 July.
It is now only one notch above junk. Its 10-year bond yield spread has risen from 176bp to 306bp since the downgrade. .
The reason credit risk has become more of an issue is that it can no longer be assumed that large language models will ever be profitable given the related ongoing collapse in token pricing.
This is extremely relevant for the hyperscalers since they have invested in data center capacity on the assumption that the compute will be purchased by large language models.
Will the Market Start Questioning the Value of “RPOs”?
In this respect, there is a 2008 real estate credit angle to this AI boom in the sense that the hyperscalers have invested in data centers on the assumption of a promise by the assumed tenants “of future payment in exchange for future compute”.
There is around US$2.5tn in aggregate contracted backlog across the four big platforms (Microsoft, Alphabet, Amazon and Oracle as Meta is not included as it is not yet in the cloud business).
What is the form of these contractual commitments?
The credit is extended in the specific form of so-called remaining performance obligations (RPOs).
What are these RPOs?
They are forward contractual commitments, or “a promise of future payment in exchange for future compute”.
These RPOs totalled US$2.5tn at the end of 3Q26 and have grown by 106% or US$1.28tn over the past four quarters from US$1.21tn at the end of 3Q25.
So far investors, for the most part, have been viewing the rising RPOs as evidence of demand to justify the hyperscaler capex and related data center build-out.
But clearly there is a lot of scope for these assumptions to be questioned.
The views expressed in Chris Wood’s column on Grizzle reflect Chris Wood’s personal opinion only, and they have not been reviewed or endorsed by Jefferies. The information in the column has not been reviewed or verified by Jefferies. None of Jefferies, its affiliates or employees, directors or officers shall have any liability whatsoever in connection with the content published on this website.
The opinions provided in this article are those of the author and do not constitute investment advice. Readers should assume that the author and/or employees of Grizzle hold positions in the company or companies mentioned in the article. For more information, please see our Content Disclaimer.