The most crowded trade of the past two years was buying AI hardware. Nvidia GPUs, Broadcom custom chips, TSMC advanced process capacity, optical modules, servers, power and cooling equipment formed the clearest beneficiary chain of the investment wave.
By November 2026 the logic is changing. Investors who worried about being unable to buy enough chips now worry about buying too many. Order growth and capacity expansion have given way to capital returns, customer financing capacity, utilisation rates and cash flow.
From buying hardware to cutting exposure
On 5 November US semiconductor stocks were sold hard. The Philadelphia semiconductor index fell 12.2 per cent in a single session, its largest drop since April 2025, while Nasdaq fell 5.5 per cent. Qualcomm’s disappointing results lit the fuse, but Reuters reported that hedge funds had been actively selling semiconductor and new-cloud names for weeks, with some funds still short or reducing leverage.
The move crossed into Asia. On 16 November China’s 5G communications index fell 7.5 per cent and the STAR 50 index 5.2 per cent, with Zhongji Innolight down 12.3 per cent. In Japan, Tokyo Electron fell 7.1 per cent and Advantest 11.1 per cent.

Reuters reported on 16 November that directly AI-exposed assets among major global companies amount to at least USD 25.3 trillion. For large institutions, AI is no longer a standalone theme but a shared risk factor embedded in equities, bonds and other assets. PGIM has begun screening nearly USD 1 trillion of fixed-income assets for business-model exposure and credit stress from slowing AI capex.
This is not a uniform exit. Saxo Bank’s analysis of client trading showed buying in Nvidia, AMD, Micron and San Disk after the 5 November drop, with Nvidia’s average daily buying activity rising. The market is choosing different prices and different names, not abandoning the sector.
The forecast that changed the frame
On 13 November SemiAnalysis published a revised accelerator model that cut forecasts for capital spending, chip shipments, HBM demand and power demand. Global AI accelerator capex for 2027 fell from USD 770 billion to USD 627 billion, a 19 per cent cut. The 2028 figure dropped 28 per cent, from USD 1.054 trillion to USD 758 billion. Estimates for 2029 and 2030 were cut 41 and 40 per cent.
By 2030 the forecast is USD 923 billion rather than USD 1.531 trillion, a reduction of USD 608 billion. Spending still rises from roughly USD 392 billion in 2026, but the slope matters for companies priced for the steeper line.
HBM assumptions softened too. SemiAnalysis cut 2028 global HBM shipments from 9.0 EB to 6.4 EB, a 30 per cent reduction, and expects supply and demand to improve enough to lower future prices. Volume growth alone no longer guarantees average selling prices or margins.
Nvidia took the largest cut. Its compute-revenue-related shipment forecasts for 2027 through 2030 were lowered 22, 39, 56 and 66 per cent, with lower average selling prices on some next-generation products. Google’s TPU, Amazon’s Trainium and custom ASICs developed with Broadcom are giving large buyers alternatives, and any of them offering better performance per dollar, lower energy use or more controllable supply can shift procurement.
Broadcom held up better. Its cumulative 2026 to 2030 ASIC revenue forecast fell 11 per cent, largely on HBM cost and pricing assumptions rather than a volume collapse. The supply chain can no longer be treated as one asset.
Who pays for the build-out
JPMorgan Private Bank reported that capital expenditure at three large cloud companies came in below expectations in the third quarter of 2026, and that upgrades to the next twelve months’ capex forecasts were at their lowest since the fourth quarter of 2023. A Federal Reserve Bank of St Louis study found the cash-to-asset ratio at large cloud companies fell from 35 per cent in 2018 to 16 per cent in 2026, while investment assets nearly doubled.
Financing pressure is unevenly distributed. Between the fourth quarter of 2024 and the second quarter of 2026, average quarterly net debt financing at data-centre builders rose from USD 3.23 billion to USD 30.17 billion, while semiconductor companies moved from net borrowing to net repayment.
Chinese hardware suppliers are exposed to the same cycle. Zhongji Innolight, Eoptolink and TFC Optical occupy different positions in the chain, but all are priced against global data-centre construction. The question investors now ask is simpler than before: does the customer have the money, when is delivery, can the equipment be used, and does revenue cover depreciation, interest and operations?
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1102576590_355012.