Chip stocks just had their worst week in months, and Wall Street is asking whether the artificial-intelligence bubble is finally bursting.

A Bank of America analyst says the question comes too late. On valuations, the burst has already happened.

Vivek Arya told clients in a note dated Friday that Nvidia Corp. (NVDA), Broadcom Inc. (AVGO) and Micron Technology Inc. (MU) trade at about 13 times forward earnings on average.

That sits inside the 10-to-20 times range at which Cisco Systems Inc. (CSCO) traded after the dot-com bubble burst.

"Investors are paying post-crash multiples for a cycle that hasn't crashed," Arya said.

The Chip Sector Is Already In A Bear Market

The Philadelphia Semiconductor Index (SOX), which tracks the 30 largest U.S.-listed chip stocks, was down about 9% for the week by Friday morning and on track for an eighth straight daily decline.

The index, closely tracked by the Invesco PHLX Semiconductor ETF (SOXQ), now sits about 22% below its June 22 record close. A 20% drop from a peak is the usual definition of a bear market.

It is the second one this year. Between June 22 and July 29, the index lost 28.6% before rallying back.

Memory stocks took the heaviest hit. Micron fell 13.1% this week to around $894, after closing at a record $1,088 on Oct. 7.

Broadcom lost 6.5% and Nvidia 5.7%, while the S&P 500 slipped just 0.8%.

The selling started with memory.

Early in the week, Chinese memory maker CXMT announced plans for a server-memory product built on a new architecture that could pack in more storage, raising concerns that it could take market share from Samsung, SK Hynix, and Micron.

Then inflation joined in.

A hotter-than-expected inflation reading on Wednesday revived worries about further Federal Reserve rate hikes and a higher cost of capital for the AI trade.

Collectively, 20 of the world's most valuable AI and semiconductor stocks shed more than $1 trillion in market value through Thursday.

This Is Not 2000, According To Bank Of America

The dot-com bust left miles of fiber-optic cable unused, known as "dark fiber." Arya says AI hardware is the opposite. It runs near 100% utilization, and rental prices for older Nvidia chips launched between 2020 and 2024 are up 12% to 37% this year.

Rising prices for old equipment point to scarcity, not a glut.

Demand keeps climbing. Weekly usage on OpenRouter, a platform that routes requests to AI models, hit about 170 trillion tokens in early October, up from roughly 3 trillion a year ago. A token is a small chunk of text an AI model reads or writes.

Arya sees agentic AI, software that completes multi-step tasks on its own, as the next tenfold jump in computing demand.

Nvidia Is Bank Of America's Gigawatt Winner

With power and land now scarcer than chips, Arya expects spending to favor the most productive hardware. He argues that Nvidia captures more than 70% of AI system spending, backed by over $100 billion in announced buybacks.

Cloud and neocloud capital spending is expected to reach about $1.01 trillion in 2026, up 98%, and $1.39 trillion in 2027, up 37%.

Arya named three risks: a credit event at a neocloud, higher interest rates, and delays in power and permitting.

Neoclouds, such as CoreWeave Inc. (CRWV) and Nebius Group N.V. (NBIS), rent out AI computing capacity, often funded with debt.

The cash math is stretched.

Big cloud companies are expected to spend about 106.5% of their operating cash flow on capital projects in 2026.

The bank still sees that as sustainable, since the top five hyperscalers have raised about $382 billion this year, mostly in long-dated debt and equity.

Cisco shows how long a reset can last. Its stock peaked at $82 in March 2000, fell to $8.12 by October 2002, and didn't reach that peak again until March 2026.

Bank of America bets that chip earnings, unlike Cisco's, will keep growing.