Bianca Botes, Director at Citadel Global
United States (US) Artificial Intelligence (AI) stocks have entered a new phase and the market is starting to notice. For two years the investment thesis was simple: anything with AI in the narrative was a buy, capital expenditure (capex) was a signal of commitment rather than a cost and investors trusted that the revenue would follow. That consensus is cracking. The question being asked in earnings calls, analyst notes and boardrooms this quarter is no longer about technological ambition. It is about arithmetic.
US hyperscalers
The numbers that have surfaced during the second quarter’s earnings season are stark. The five largest US hyperscalers will spend roughly $760 billion on AI infrastructure in 2026 while expensing only $211 billion of it on their income statements. The remaining $549 billion is deferred – spread across 20-year to 30-year depreciation cycles on data centres and three-to-five-year cycles on graphics processing unit (GPU) clusters (networked groups of computers). Alphabet reported second quarter earnings with headline earnings per share (EPS) growth of 294% – strip out a $99 billion paper gain on stakes in Anthropic and SpaceX and core EPS came in at roughly $2.85 against a $2.88 estimate, with the underlying business delivering solid but ordinary growth. In the same quarter, Alphabet spent $44.9 billion on capital projects – more than double a year earlier – while free cash flow swung to negative $5.9 billion. Morgan Stanley has described the current period as a “golden window where everybody looks good.” The window, however, is narrowing.
Investment bank, Bank of America, calculated that hyperscaler capex is already consuming 94% of operating cash flow after dividends and buybacks. Goldman Sachs estimates depreciation and amortisation will climb from 7% of revenue in 2022 to 12% by 2027, converting upfront spending into fixed costs that will compress margins for years. The capex-to-revenue divergence is running at approximately 46% – already exceeding the 32% divergence recorded during the 2001 telecom overbuild cycle, a period that ended in a multi-year correction. These comparisons are not cherry-picked. They are the ones being circulated inside the institutions pricing these stocks.
Chinese models outplaying the US?
Let’s bring China into the discussion. Moonshot AI released Kimi K3 on 16 July, a 2.8 trillion parameter (connection points) open-source model that, within 48 hours of launch, overwhelmed its own infrastructure and forced a temporary suspension of new subscriptions. It performs close to Anthropic’s most advanced publicly available model at a fraction of the cost. Alibaba followed days later with Qwen 3.8. In mid-July, six of the top 10 models on OpenRouter – the developer marketplace where models compete on usage – came from Chinese companies and all of the top five did. Online marketplace, Airbnb, is using Alibaba’s Qwen for customer service; fintech company, Coinbase, halved its AI spending by switching employees to Kimi and Z.ai’s GLM models; a San Francisco company that builds AI work assistants switched from Anthropic to DeepSeek and saved millions; tech company, Mozilla’s CTO, switched to Kimi K3 within days of its launch and the list goes on.
This matters for the investment case for the US hyperscalers, whose models rest on the assumption that enterprises will pay premium prices for closed models, and that the proprietary nature of OpenAI and Anthropic’s systems justifies the subscription cost and the switching cost. Chinese open-weight models, freely downloadable and adaptable without ongoing licensing fees, undermine that pricing power directly. If a company can get 90% of the capability at 10% of the cost by self-hosting a Chinese open model, the revenue justification for $760 billion in US AI infrastructure weakens. The AI label no longer automatically signals a premium product in a market with no alternatives. There are now alternatives. They are good, and they are free.
Inflation and interest rates put a spoke in the US-AI wheel
Against this backdrop, US Federal Reserve (Fed), South African Reserve Bank (SARB), European Central Bank (ECB) and the Bank of England (BoE) all recently kept rates on hold, while citing concerns over inflation. The Fed voted nine-to-three this week to hold the federal funds rate at 3.50% to 3.75%, with three regional presidents dissenting in favour of a hike. Warsh used the post-meeting statement to reiterate the 2% inflation target as absolute and to avoid any forward guidance on cuts. The market now prices in between one and two Fed rate hikes before year end.
The bond market has delivered its own verdict. The 30-year US Treasury yield hit 5.236% on 29 July, its highest level since July 2007. The 10-year moved to 4.7%. These are not technical moves. They are the bond market pricing in an environment where inflation stays above target, central banks cannot cut, and fiscal deficits continue to require new issuance that the market must absorb at a price. Higher long-end yields raise the discount rate applied to every future cash flow in the economy. For AI companies specifically – where the cash flows being priced in are years or decades away, and where the depreciation drag from current capex will only begin showing up on income statements over the next several years – the mathematics of valuation tighten directly and immediately.
The need for careful calculations
The AI trade is not over. The technology is real, the demand is real, and the competitive dynamics between US and Chinese models are arguably accelerating progress rather than stalling it. What is over is the period where owning anything with AI exposure was sufficient. The next phase of this trade requires knowing which companies will generate actual returns from what they have spent, and at a 30-year yield above 5%, the cost to investors of being wrong about that is considerably higher than it was when government bond rates were at zero.
ENDS
