The AI Spending Reckoning: Why the World's Most Profitable Chip Makers Got Hit

On July 16, 2026, TSMC reported its fifth consecutive quarter of record profits. Net income surged 77% to $22 billion. Revenue reached $40 billion, beating estimates. The quarter was, by most measures, difficult to fault. The stock fell 7.3% the next day anyway.

The market was not questioning whether demand was real. It was questioning whether demand was wise.

TSMC, the world's largest semiconductor foundry, had simultaneously announced it would spend $60 to $64 billion on new capacity in 2026 — a sharp increase from earlier guidance — along with an additional $100 billion commitment to build factories in Arizona. Higher capital expenditure means lower near-term free cash flow, larger depreciation charges, and more capital that must earn its keep before it reaches shareholders. For a stock that had climbed more than 60% over the preceding year, the math looked less comfortable.

That reaction was not isolated to Taiwan. On July 17, the Nikkei 225 fell 4.5% in Tokyo and the Taiex plunged 6.5% in Taipei. South Korea's KOSPI, which had more than doubled in the first six months of 2026, lost roughly a third of its value from its June peak within weeks. Semiconductor names from Tokyo Electron to SK Hynix were hammered across the region. The sell-off extended through July and August, punctuated by fresh declines each time hyperscaler earnings calls confirmed one thing: the spending was only getting bigger.

The scale is what makes the argument worth following. Amazon, Microsoft, Alphabet and Meta plan to spend between $725 billion and $775 billion on capital expenditure this year. The vast majority of the outlay goes toward AI compute: graphics processors, custom silicon, data centres and the networking and power infrastructure that supports them. Amazon alone guided to $200 billion — more than the entire publicly traded U.S. energy sector invests annually.

These are not idle projections. The companies have signed power purchase agreements, placed equipment orders, broken ground on data centres and locked in supply contracts. Nvidia, which supplies roughly 90% of AI accelerator chips, reported quarterly revenue of $96 billion in the second quarter of fiscal 2027 — more than doubling from a year earlier. TSMC, which manufactures almost all of Nvidia's advanced chips, guided to 40% revenue growth for 2026. The build-out is happening. The machines are being ordered, shipped and installed.

The trouble is the returns.

AI services — the products and platforms that sit on top of all this infrastructure — generated roughly $25 to $35 billion in revenue over the past year. That is 3% to 5% of the annual hyperscaler spend on building them. OpenAI's annual recurring revenue stood at around $20 billion. Anthropic's run rate had surpassed $9 billion. Combined, the pure-play AI vendors account for less than a quarter of even the most conservative spending estimate. The gap between what is being invested and what is being earned back is not a rounding error. It is the entire point of contention.

To be sure, the comparison is not perfect. Infrastructure always precedes revenue in a technology build-out. Cloud computing followed the same pattern in the 2010s: data centres were built for years before cloud services reached scale. The hyperscalers also argue that AI is already generating indirect returns — Meta's AI-driven ad targeting, Microsoft's Azure growing at roughly 40%, Google Cloud accelerating at 82% in the latest quarter. Amazon's AI services within AWS reached $25 billion in annualised revenue. None of this is zero.

Yet the infrastructure is growing faster than any of those revenue streams, and faster than history suggests is comfortable. Alphabet's chief executive Sundar Pichai acknowledged "elements of irrationality" in the current spending pace. A PwC survey published in January found that 56% of chief executives reported no revenue increase or cost reduction from their AI investments. The question is not whether AI will eventually prove worthwhile. It is whether the capital deployed in 2026 can earn back its cost within a timeframe that justifies the valuation the market has assigned to the supply chain.

That valuation is the mechanism through which the sell-off works. When the market was confident that hyperscaler spending would compound indefinitely at accelerating rates, the beneficiaries — TSMC, Nvidia, Tokyo Electron, Samsung, SK Hynix — traded at multiples that assumed that confidence would endure. The sell-off repriced those assumptions downward. It did not stop demand. It questioned sustainability.

Two structural concerns feed that question. The first is customer concentration. Four companies now account for roughly 61% of Nvidia's revenue, up from 54% a quarter earlier. When growth depends on a handful of buyers, each buyer's spending decision matters disproportionately. If one of them pulls back, the ripple effects travel upstream.

The second concern is substitution. Those same four companies are designing their own AI chips — Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia, Meta's MTIA. These custom accelerators do not need to replace Nvidia entirely. They target inference workloads, where cost per query matters more than raw performance. Even modest shifts in inference share away from Nvidia's GPUs would flow through to reduced chip orders, smaller foundry allocations and weaker equipment demand. The suppliers benefit from volume today while watching their customers build tomorrow's alternatives.

This is where the distinction between TSMC and the rest of the supply chain becomes useful. TSMC is a foundry — a manufacturer that does not compete with its customers. Google, Amazon and Microsoft are buying wafers from TSMC, not asking them to stop producing. If anything, custom chips increase TSMC's workload, since they too must be manufactured at advanced nodes. TSMC's margin compression — operating margins are projected to fall to roughly 57% in the second half of 2026 as it ramps its most advanced two-nanometer process — is a production cost, not a competitive threat. The company generated nearly $26 billion in operating cash flow in the second quarter alone and holds $96 billion in cash on its balance sheet.

Nvidia faces the substitution risk directly. Its forward price-to-earnings multiple is roughly in line with, or below, the S&P 500 — a signal that the market is pricing in eventual deceleration rather than an infinite runway. The company has responded defensively, licensing inference technology from Groq and locking in large chip orders with Meta. These moves stabilise the near term but do not eliminate the structural tension between a dominant training business and a competitive inference landscape.

The broader Asian semiconductor names — Tokyo Electron in equipment, Samsung and SK Hynix in memory — sit between the two. They benefit from volume growth while facing both the substitution risk (fewer Nvidia chips means fewer memory modules) and the cyclical risk (accelerating capacity additions in a cyclical framework typically signal investors to fade the group rather than chase it). SK Hynix fell as much as 30% in a single session in late July, triggering Korean circuit-breaker mechanisms. The declines reflect the fact that these companies' fortunes are tied to Nvidia's design wins, not just to aggregate foundry output.

The sell-off, then, was not a repudiation of AI. It was a repricing of the timeline between investment and return. The machines are being built. The chips are being shipped. The question investors began asking in mid-2026 is whether the revenue on the other end will arrive soon enough to justify what they have paid for the privilege of supplying the tools.

Some evidence points to a resolution. Nvidia guided to $108 billion in revenue for its next quarter — a figure that, while missing the loftiest analyst estimates, implies AI spending remains robust through September. TSMC's raised capital budget signals that customer orders are flowing, not drying up. The cloud backlogs — Alphabet's $240 billion, Microsoft's $80 billion in unfulfilled Azure orders — suggest demand is genuinely supply-constrained, not a speculative exercise.

Other evidence points the opposite way. UBS projects hyperscaler capital expenditure growth will slow to 25% in 2027 and 6% in 2028. JPMorgan estimates AI-related investment could reach $5.5 trillion by 2030, but warns that the sharp increase in capital intensity is consuming the free cash flow that once funded the build-out — leaving the companies increasingly reliant on external financing to keep the machines ordered. When the funding stops compounding, the build-out slows with it.

For an investor considering exposure to the Asian supply chain, the sell-off established a clearer boundary than the rally did. The companies at the centre of this trade are not broken. TSMC earns $4.22 per share in the latest quarter, far above its $3.81 estimate. Nvidia's profit doubled to $60 billion in one quarter. The financial performance is extraordinary. The question is whether the market's price for that performance reflects durable compounding or a period of premium pricing that will normalise as the revenue gap narrows — or widens.

The distinction matters less for the business than for the portfolio. A company can be well-managed and still overpriced. The sell-off was the market's way of remembering that relationship.