Nvidia gained 2.78% and AMD jumped 11.04% on Friday, recovering only part of the losses suffered a day earlier as an AI-related market rout spread from Wall Street to Asia.

The rebound did little to resolve the larger question facing investors: whether the selloff represents an overdue correction or the early stages of an AI bubble bursting. Concerns over data-center restrictions, the enormous cost of AI infrastructure and the sustainability of current valuations continue to weigh on the sector.

Nvidia rose 2.78% to $186.64 after plunging 9.3% on Thursday. AMD climbed to $449.04 after collapsing 17.3% the previous day. Both shares remained well below their October peaks, underscoring how little a single relief rally had repaired the damage.

The turmoil extended beyond U.S. markets. South Korea’s KOSPI fell 6% on Friday, while Taiwan’s benchmark declined more than 8% over the week, its worst performance in months. Both economies host some of the world’s largest chipmakers, whose shares carry heavy weightings in the local benchmark indexes. Taiwan Semiconductor Manufacturing Co. (TSMC) reached its daily limit-down threshold during the selloff, the maximum one-day decline allowed under the exchange’s price-limit rules. Reuters reported that Taiwanese officials asked state-owned banks to help support market prices.

The U.S. Tech 100 cash-market proxy, tied to the tech-heavy Nasdaq 100, also recovered, rising 1.85% to 27,007.3. It was still down more than 5.5% for the week. Its daily price action showed the rebound following a much larger decline from October levels rather than signaling a return to the previous trend.

Why the AI trade is unraveling

The U.S. midterm elections intensified concerns that Democratic gains could bring tougher scrutiny of data-center construction. According to Ballotpedia, voters favored limits on development in 40 of 41 local ballot contests held on November 3, including by rejecting measures that supported construction. Some of the votes were advisory, but the results conveyed a clear political message.

The issue is significant for chipmakers because data centers are central to their growth plans. Each facility requires substantial spending on processors, servers, electricity and cooling. New restrictions could delay orders or force companies to build in other locations.

However, doubts about the AI investment cycle began before the election. AI companies and their customers have been spending aggressively to expand capacity, often before the resulting returns are clear. Higher borrowing costs make those investments more difficult to sustain. Reuters quoted market strategist Steve Sosnick as warning that heavily owned stocks were already “priced to perfection.” The phrase describes valuations that leave little room for disappointment.

Correction or bubble burst?

A correction generally refers to a market decline of at least 10% from a recent peak. It does not necessarily indicate that the underlying business is shrinking. For example, a chipmaker whose shares trade at 60 times annual earnings could lose one-third of its market value if investors become willing to pay only 40 times those same earnings, even if the company’s profits remain unchanged.

A bubble bursting is more severe. It occurs when expectations prove unrealistic and the money supporting a boom begins to disappear.

The dot-com crash provides a historical comparison. The Nasdaq lost roughly 77% between March 2000 and October 2002. The internet was not a fad, but investors had significantly overestimated how quickly companies could turn the technology into profits, as Goldman Sachs’ history of the 2000 dot-com bubble describes.

Telecom operators spent heavily on fiber-optic networks for customers who had not yet arrived. Global Crossing, which built an international network, filed for bankruptcy in 2002. The technology survived, but many companies and shareholders did not.

AI could face a similar mismatch between its long-term promise and near-term returns, and some investors are already drawing that comparison. The RAM shortage is also creating a business opportunity that companies are attempting to exploit. Former cryptocurrency miners are renting out their hardware, while cloud providers and inference startups are emerging across the market.

Consider a hypothetical company that borrows billions of dollars to build a data center filled with Nvidia chips, expecting to rent the computing capacity at premium prices for years. If customers spend less than expected, borrowing becomes more expensive or construction is delayed, the debt still must be repaid.

The operator could cancel chip orders, reducing work for contractors. Utilities that planned additional capacity could lose expected customers. Lenders and investors holding the company’s debt could also suffer losses. A stock-market selloff could therefore develop into a broader economic problem.

There is an important difference from 2000, however. Microsoft, Alphabet, Amazon and Meta are profitable companies with established customers and substantial cash flows, unlike many speculative dot-com startups. They are better positioned to withstand setbacks than smaller AI-cloud providers that depend on repeated borrowing to expand.

AI could continue transforming industries even as numerous individual AI companies fail, and investors in successful companies could still have paid too much for their shares.

Financial warning signs are also appearing in the riskiest parts of the market. Reuters reported on Friday that the cost of insuring against a default by AI-cloud provider CoreWeave had risen above 1,000 basis points—the equivalent of more than 10 percentage points of the insured amount per year. Such protection functions somewhat like insurance on a loan: As its price rises, investors are signaling greater concern about repayment.

Broader bond markets, however, were still operating relatively normally. That suggests increasing strain in the riskiest areas of AI financing, rather than providing evidence of a system-wide crisis.

The coming week will offer more concrete tests than another session of rising or falling prices. Results from CoreWeave, Applied Materials, which supplies chipmaking equipment, and Cisco, a major supplier of data-center networking hardware, could show whether customers are continuing to order capacity and equipment. Inflation data will also influence expectations for borrowing costs.

Stable orders and manageable debt would support the view that the market is undergoing a correction. Repeated project cancellations, failed refinancing efforts and declining sales would suggest that the boom’s financial foundations are beginning to crack.

The central question is not whether AI has a future, but whether every company investing billions in that future can afford to reach it.