Has the AI Bubble Burst? Why Nvidia, Sandisk, Micron, CoreWeave and Other Stocks Are Falling

The AI trade is already in a deep correction. Across 49 AI and AI-infrastructure stocks in the US, Taiwan and South Korea, the median company is 40.7% below its all-time high, while 13 have fallen by at least 50%.

Yet AI demand is still surging. Microsoft’s Azure grew 43%, and Nvidia’s Data Center revenue jumped 92% in their latest quarters. The market is not questioning AI demand. It is questioning whether record spending on chips, data centres and power will generate enough profit and cash flow to justify it.

Let’s break down what fell, why memory and leveraged AI builders were hit hardest, which parts of the AI trade remain financially sound, and the cash-flow test that matters more than another bubble headline.

AI Stock Crash: How Nvidia, Microsoft, Micron, CoreWeave and Other AI Stocks Have Fallen

The broad U.S. market has not crashed. On July 29, the S&P 500 was about 2% below its record, while the Nasdaq was nearing a 10% correction and the Philadelphia Semiconductor Index was already in a technical bear market. Beneath those indices, the damage was much worse.

Source: Google Finance, CompaniesMarketCap

The cleaner measure of the latest reversal is even more striking. 42 of the 49 stocks made their all-time high within the preceding 90 days. Their median decline was 38.2% and their median peak was only 41 days old. Nine of those recent peakers had already halved.

This is a 3-speed correction. Cash-rich platforms and scarce foundry capacity held up better. Memory, optics and connectivity fell faster. Leveraged capacity builders such as CoreWeave were punished most. The market is repricing the quality of each company’s demand and funding, not rejecting every business with “AI” in the story.

Why Are AI Stocks Falling? Demand, Cash Flow, Funding and Valuation

Instead of asking whether there is one AI bubble, test four circuits. A stock can pass the first and still fail the other three.

The demand circuit is still on. The cash circuit is under strain. Funding quality varies sharply. The expectations circuit is where the correction began.

Think of an expensive new metro line. It may be crowded and useful from day one, but it is not automatically a good investment. Fares and cost savings still need to cover construction, interest, electricity and maintenance. AI can transform the economy while some AI stocks deliver poor returns because too much was paid, too much capacity was built, or cash arrived too late.

Memory stocks show how this works. Sandisk, SK hynix and Micron reported extraordinary demand and pricing, yet fell 41% to 57% from recent highs. Investors are not disputing today’s shortage. They are discounting tomorrow’s supply. A low price-to-earnings (P/E) ratio near peak profits can be a trap if new capacity later pushes memory prices and earnings down.

Circular financing adds another layer. When a chip supplier invests in a cloud customer, or a customer prepays for capacity, both companies gain visibility but their risks become linked. Nvidia, for example, made a new $2 billion equity investment in CoreWeave in January. This does not make the revenue fake. It does mean one stressed customer can hit the supplier, financier and data-centre owner together. CoreWeave’s two largest customers supplied 65% of Q1 revenue, while the company carried $25.1 billion of debt principal at quarter-end, according to its SEC filing.

AI Demand Is Strong, but Microsoft, Alphabet, Amazon and Meta Face a Cash-Flow Test

Recent earnings do not support the claim that enterprise and cloud AI demand has collapsed. They do support a more uncomfortable claim: revenue growth and shareholder cash flow are separating.

Sources: Microsoft, Alphabet, Amazon, Meta, Nvidia, Oracle, CoreWeave, TSMC, Micron, Sandisk Investor Relations

The strongest balance sheets can survive a slow payback. Microsoft still generated $19.6 billion of quarterly free cash flow, Alphabet held $242.5 billion of cash and marketable securities, and Nvidia earns supplier economics without financing hyperscale data centres itself.

The fault line is farther down the funding ladder. Oracle spent $55.7 billion on capex in FY26 against $32 billion of operating cash flow and plans to use a mix of debt and equity to fund expansion. CoreWeave combines a visible backlog with high leverage, heavy interest expense and customer concentration. If AI demand merely arrives later than planned, a self-funded platform can wait. A leveraged builder may need fresh capital.

Wall Street’s disagreement is therefore not mainly about whether AI exists. It is about how long pricing, growth and margins can persist.

Sources: Goldman Sachs, Morgan Stanley, Barron’s, Reuters, Benzinga

Is the AI Stock Crash a Dot-Com Bubble 2.0?

The comparison is useful, but the copy-paste version is wrong. The Nasdaq traded near 70 times forward earnings in March 2000 and later lost about 75%. Today it is near 30 times. Current leaders such as Microsoft, Alphabet, Meta and Nvidia are highly profitable, not pre-revenue websites. The S&P 500 was also only 2% below its record at the cutoff. This is not yet a broad technology collapse, and it does not resemble the bank-centred systemic failure of 2008.

The modern risk is concentration and capital intensity. Semiconductor companies now represent a record 19% of the S&P 500, more than twice their weight in 2000. Meanwhile, the Bank for International Settlements estimates that strategic competition could push AI investment to about 1.5 times the efficient level in its conservative baseline. That is a model, not an observed fact, but it shows how rational companies can collectively overbuild when none wants to fall behind.

Headline valuations are mixed rather than uniformly extreme. Nvidia traded near 19 times 12-month forward earnings, its lowest multiple in more than a decade, while Micron had touched 5.4 times in May. AMD, Intel and Marvell still traded above their longer-run averages. A low multiple can signal value, but for a cyclical chip company it can also signal that investors expect peak earnings to fall.

The strongest counterargument to the bubble case is visible in the earnings table: cloud growth, backlogs, margins and multi-year supply agreements are real. The strongest answer is that real demand does not settle the return question. Railways, telecom networks and fibre booms all created lasting infrastructure. Investors who financed too much capacity at the wrong price still lost money.

Are AI Stocks a Buy After the Crash? The $700 Billion Capex Payback Test

Microsoft signals about $175 billion of calendar-2026 capex, Alphabet $195 billion to $205 billion, Amazon about $200 billion, and Meta $130 billion to $145 billion. Combined, that is roughly $700 billion to $725 billion of 2026 capex. The figures are not perfectly comparable, and not all of the spending is AI. Microsoft’s reported total changes with lease classification, while Amazon’s plan also includes chips, robotics and satellites. To avoid false precision, take $700 billion as a low-end starting point and test what the investment must earn.

The model applies a 12% annual hurdle rate, zero residual value and different assumptions for AI’s share of capex, asset life and post-tax cash contribution margin.

The base case says $490 billion of assumed AI investment must produce about $136 billion of annual post-tax cash contribution for five years. At a 50% cash contribution margin, that means roughly $272 billion of annual incremental revenue or equivalent cost savings. At the same assumptions, a one-year delay raises the requirement to about $304 billion.

This is not a price target or a claim that Big Tech will miss. It is a burden-of-proof calculator. Shorter GPU lives, low utilisation, power costs and delays push the hurdle up. Longer-lived buildings, ad-conversion gains and internal productivity push it down. It also warns against double counting: hyperscaler capex, Nvidia revenue and CoreWeave contract value are different sides of overlapping transactions, not three independent returns.

After a 50% fall, a stock is cheaper than its peak, not automatically cheap. The right post-correction question depends on the business model.

AI Stock Outlook: Has the AI Bubble Burst or Is This a Valuation Reset?

The evidence does not show that AI demand has burst. It shows that the market has stopped treating demand as automatic proof of shareholder returns. That is a healthy correction in some stocks and a financing warning in others.

Bull case: Azure, Google Cloud, AWS and Nvidia Data Center are still growing at 28% to 92%. Backlogs are large, memory suppliers have multi-year visibility, and cash-rich platforms can absorb a slow payback. If cloud revenue and cost savings begin to outrun capex and depreciation, the recent sell-off may prove to have overshot.

Bear case: Peak memory margins fade, frontier-model customers slow spending, cheaper models or custom chips reduce GPU economics, and external funding becomes more expensive. That would hit leveraged builders first, then suppliers whose earnings were priced as a permanent shortage.

For investors, the answer is not to dump every AI stock for “traditional” sectors simply because the chart broke, or to buy every 50% fall as a bargain. Diversification should be by business model and funding quality, not just sector label. Re-underwrite the cash-rich platforms, wait for margin and order proof in cyclicals, and demand a much wider safety cushion from companies that need capital markets to stay open.

The key debate for the next two quarters is simple: does AI-generated cash flow start catching up with AI capex? If it does, this was a violent valuation reset inside an intact boom. If it does not, July 2026 may be remembered as the moment the market finally noticed that a great technology and a great investment are not the same thing.