A semiconductor selloff that wiped out more than $1.3 trillion in market value over the past three weeks is forcing investors to make distinctions they refused to make when every AI bet looked like a winner.
The week ending June 26 was ugly by any measure. SoftBank plunged more than 11% on Friday as concerns over the rising cost of AI infrastructure bled from Wall Street into Asian markets. Micron, which had surged on blockbuster third-quarter earnings just a day earlier, fell more than 5% Friday to cap what CNBC described as a "wild week of big swings." Intel shed 3%, Arm lost nearly 4%, and Sandisk dropped 10%. The Nasdaq has now logged five consecutive losing sessions. This is not a blip.
The chain of events traces back to early June, when Broadcom reported fiscal Q2 earnings and delivered AI networking revenue of $4.1 billion, roughly 14% short of the $4.8 billion analysts expected. The company also declined to raise its full-year AI chip forecast, which had been priced at $17.2 billion by the Street. Broadcom stock fell 14% in a single session, and the Philadelphia Semiconductor Index cratered more than 6%. Over the following weeks, the selling accelerated. Bloomberg reported on June 24 that traders are now warning that investors have "maxed out" the AI trade, with leveraged semiconductor ETFs amplifying volatility on both sides. Goldman Sachs, according to TradingKey, has since advised clients to reduce chip sector exposure and rotate into cloud service providers like Amazon and Microsoft instead.
What changed isn't the underlying demand for AI infrastructure. Micron's revenue of $41.46 billion in its most recent quarter, up 346% year over year, tells you the build-out is real. What changed is that the market spent most of 2025 pricing chipmakers as though flawless execution and unbroken guidance upgrades were guaranteed. Some semiconductor stocks were trading at price-to-earnings multiples above 100x on forward earnings assumptions. Broadcom's modest miss didn't break the AI thesis. It broke the thesis that the AI trade required no selectivity at all.
Investors who are trimming megacap semiconductor exposure aren't necessarily leaving the AI trade. They're moving down the supply chain to companies that were overlooked during the euphoric phase. Tokyo Electron, Japan's largest semiconductor equipment manufacturer, surged as much as 14% in a single session earlier this month. Advantest, which makes chip testing equipment, rose more than 5.5%. Japan's Nikkei is up nearly 33% in 2026, and much of that gain reflects portfolio managers finding AI-adjacent exposure at more reasonable multiples than anything trading on the Nasdaq. South Korea's semiconductor exports hit an estimated $110.4 billion in the first four months of the year, following record annual exports of $173.4 billion in 2025. The money rotating there isn't defensive, it's selective.
The infrastructure side of the trade, power, cooling, networking, has a similar logic. These suppliers didn't benefit from the same speculative compression that megacap chip stocks did, so they don't carry the same unwinding risk. Whether that rotation proves durable depends on whether spending from hyperscalers actually converts into revenue across the supply chain, or whether it stalls as reported concerns about AI infrastructure costs grow. OpenAI is reportedly considering delaying its IPO to next year, citing both SpaceX's disappointing post-debut performance and AI sector volatility. That's a significant signal about how even the best-known AI companies are reading the public market's appetite right now.
What founders and VCs should take from this
The private market hasn't cracked yet. Crunchbase reported that global venture investment hit $300 billion in Q1 2026 alone, an all-time high, and seed-stage AI companies are still commanding a roughly 42% valuation premium over non-AI startups. Capital is still flowing. But the public market repricing of AI expectations almost always travels to private valuations with a lag, and that lag has historically been measured in quarters, not years.
The specific risk for founders raising at late-stage multiples is that those multiples were often calibrated against a public market that priced AI exposure at 80 to 100 times forward earnings. If public market AI multiples compress toward 40 to 50 times as selectivity sets in, the comparables that justified aggressive private valuations start to look stale. A company that raised its Series C at a 60x ARR multiple because Broadcom was at 90x P/E has a problem the next time it opens a term sheet. Investors will notice.
The fundraising narrative that worked in 2025, "we are an AI company therefore we command a premium," is losing its explanatory power. What replaces it is specificity: which part of the AI stack you serve, what your actual revenue trajectory looks like, and whether you have a credible path to margin. The Goldman Sachs rotation toward cloud providers and away from pure semiconductor plays is essentially the same argument applied to public markets. Investors want to know which AI spending they can point to in your numbers, not just in your pitch.
Frankly, that's a healthier environment than the one it replaces. The last 18 months rewarded founders who could tell a plausible AI story regardless of execution. The next 12 months will reward founders who can show one. The distinction between those two things is exactly what the public market just started making.
Also read: Regulators are finally building the AI they need to police the markets they oversee • Trump's 100% tariff threat over digital services taxes reshapes the cost calculus for every US tech company operating in Europe • Framework Ventures raises $400 million to chase AI and robotics beyond its crypto roots