The AI Selloff Is a Capex Audit, Not an AI Collapse

What actually sold off

The July 17-20 move was not a clean story of AI demand disappearing. It was a global repricing of AI-linked infrastructure stocks during a broader macro shock. AP reported that on July 17 the S&P 500 fell 1%, the Nasdaq lost 1.4%, and the Dow fell 0.8%, with Nvidia down 2.2%, Applied Materials down 5.6%, and TSMC down 7.3%. Asian exposure was hit harder: Taipei fell 6.5%, Tokyo 4%, and Shanghai 3% in that session, according to AP.

The selloff continued unevenly on July 20. South Korea's Kospi dropped 4.5%, with Samsung Electronics and SK Hynix both down more than 4%, while Taiwan's Taiex slipped only 0.5% and TSMC rebounded 1.3%, AP reported in its follow-up. That pattern matters. This was not a simple China crash or a one-day panic. It was a reassessment of the supply chain behind AI compute: chips, memory, tools, foundries, cloud platforms, and financing.

The better frame: utilization and returns

The easy headline is that the AI bubble burst. The facts do not support that as the main conclusion. Cloud demand is still growing fast, but the spending required to serve it is now large enough to test balance sheets and free cash flow.

Microsoft said in its April 2026 quarter that AI business ARR surpassed $37 billion, Azure and other cloud services revenue grew 40%, and quarterly capex was $31.9 billion. Meta reported Q1 capex of $19.84 billion and raised 2026 capex guidance to $125 billion to $145 billion. Amazon said AWS sales rose 28% to $37.6 billion, while cash capex rose to $43.2 billion and trailing 12-month free cash flow fell to $1.2 billion, mainly because of AI-related property and equipment spending.

That is the practical issue. Investors are not only asking whether AI is useful. They are asking whether hundreds of billions of dollars in servers, chips, data centers, power equipment, and networking gear will be used heavily enough, priced high enough, and depreciated slowly enough to justify the buildout. S&P Global Ratings has estimated that five large cloud providers could spend about $750 billion on capex in 2026, or 38% of revenue. That is infrastructure-industry scale, not a software feature rollout.

Cheaper models complicate the chip story

The China-linked trigger was model competition. AP reported that Moonshot AI's Kimi K3 revived a DeepSeek-style shock by challenging assumptions about how much compute is needed to produce competitive models. But cheaper architectures do not automatically mean lower total chip demand. They may reduce the need for maximum frontier-training clusters while increasing inference use if AI becomes cheaper to deploy.

DeepSeek's own technical report showed a 671 billion-parameter mixture-of-experts model with 37 billion active parameters per token and 2.788 million H800 GPU hours for full training. That is real evidence of efficiency pressure. It is not proof of full company economics, enterprise reliability, or long-run monetization. The market is reacting to a credible challenge to scarcity pricing, not to proof that compute demand has vanished.

Energy and geopolitics are part of the valuation

Oil is not a separate background story. AP reported Brent crude near $88 after U.S.-Iran attacks raised Strait of Hormuz supply concerns. Higher energy costs matter because data centers buy power, and marginal electricity supply often depends on gas generation, grid capacity, transformers, and local permitting.

The IEA estimates data-center electricity use was 415 TWh in 2024 and could more than double to about 945 TWh by 2030. That makes power availability a direct constraint on AI economics. So are export controls, TSMC capacity, HBM supply from SK Hynix, Samsung, and Micron, and chip-tool capacity from firms such as Tokyo Electron and Applied Materials.

The useful reading of this selloff is not AI collapse. It is a market correction forcing a harder test: utilization, margins, customer ROI, power access, and supply-chain resilience. Those questions should have been central before valuations assumed the buildout would pay for itself.