AI stock selloff reaches profitable suppliers and software companies

From credit risk to earnings risk

The selloff first looked like a CoreWeave problem. The company has more than $100 billion in backlog, but also heavy debt, rising interest expense and unsecured bonds yielding above 16%. Investors could treat that as a risk specific to a highly leveraged infrastructure provider.

In the second quarter, CoreWeave’s revenue rose 112% to $2.58 billion. Net interest expense was $640 million, and the company lost $626 million, even as customers bought nearly all the capacity it could offer.

Contracts produce revenue, but capacity takes upfront spending: processors, electricity and data centers. Lenders get paid in cash; shareholders get what remains. CoreWeave’s capital structure makes that divide unusually visible.

Now the selling has reached companies without the same financing strain. Broadcom does not need junk debt to operate. TSMC and ASML are highly profitable. Micron recently reported gross margins more typical of a software company than a memory maker.

The risk, in Osman’s analysis, is not primarily on their balance sheets. It is in the customer decisions that underpin their profits. Micron’s margins reflect an exceptional shortage of advanced memory; TSMC benefits from demand for leading-edge manufacturing; ASML sells equipment to expand capacity. Strong results encourage customers to spend more, which can eventually ease the shortages behind those results.

Record earnings need not vanish next quarter for share prices to fall. Investors can price in a future decline before it appears in reported results. Financial strength may protect a company without protecting the valuation investors put on its earnings.

More capacity can change the math

Osman draws a parallel with the telecom buildout of the late 1990s. Internet traffic did grow, and the world did need more fiber. But many investments still failed. Too many companies raised money to build networks and buy equipment on the assumption that capacity would remain scarce. Once it arrived, prices weakened, and some companies could not survive the infrastructure they had built.

The same tension runs through today’s AI investment cycle. Alphabet, Amazon, Meta, Microsoft and Oracle are spending hundreds of billions of dollars. Nvidia and other chipmakers supply processors; TSMC makes them; ASML sells equipment to expand production. CoreWeave and other infrastructure companies borrow to install hardware, while utilities build power plants and networks to meet demand.

No irrational decision is required for this spending to create excess capacity. Microsoft does not want to lose ground to Amazon or Google. Meta ties its spending to a future advertising platform. Chipmakers expand because customers ask for more capacity than is available. The combined result can still overshoot demand.

The bullish case assumes AI capacity will remain scarce for years, across advanced chips, data centers, transformers and grid connections. Customers book capacity in advance and pay heavily for it. Osman’s concern is that scarcity may be less reassuring to investors than it looks.

Planned US AI infrastructure is valued at about $10 trillion. Yet at least 47 major data-center projects were delayed or canceled in 2026, and analysts consider only about 35% of the announced pipeline credible.

Delays can support chip shortages and prices in the short term. They also raise a more basic question: how much of the announced pipeline will become operating infrastructure? Microsoft reportedly declined to be the anchor tenant for Nscale’s data-center project in Essex County after learning that grid power might not arrive until sometime in the next decade.

Nscale may find another tenant, another site or another power source, and the project may still happen. But the announced contract is now a problem to solve, not revenue already secured. Data-center agreements can depend on timing, conditions and rights to exit. A customer commitment backed by a working facility is worth more than one attached to a site still waiting for electricity.

A delay at one data center can push chip demand out of an expected quarter. A shifted chip order can affect factory utilization, which may eventually change demand for manufacturing equipment. The first disappointment may appear far from the company whose results ultimately suffer.

The return shareholders need

Microsoft, Meta, Alphabet and Amazon look safer than CoreWeave because they can fund much of their investment from operating cash flow. They do not need a lender’s approval for each new data center. But if Meta puts another $100 billion into AI infrastructure, suppliers can recognize revenue when they deliver equipment, while Meta depreciates that investment over years. Shareholders still need the spending to produce enough extra advertising revenue, subscription income or cost savings.

During construction, many companies benefit. Nvidia sells processors, TSMC makes them, ASML supplies more equipment, contractors build facilities, utilities sell power and Meta can report improved models. The harder question comes later: how much additional cash flow did the investment generate?

The 10-year US Treasury yield is about 5%. Against that benchmark, an 8% project return looks less attractive than it did when safe assets yielded close to zero. An investor can earn more than 5% without taking on construction delays, technological obsolescence and uncertain customer demand.

Osman argues that upcoming hyperscaler results should be read through this trade-off. If capital spending rises again, that could support chip and infrastructure suppliers while putting more pressure on hyperscaler free cash flow. If spending falls, hyperscalers’ cash flow may improve, but companies dependent on their orders could suffer.

He also advises watching how stocks respond to good news before trying to call a bottom. TSMC and ASML reported results that might have supported their shares, but the market reaction was weak. When a stock fails to rise on strong results, investors may be worried about what has not yet appeared in the numbers.

I think the most useful test here is not whether AI revenue is growing, but how much cash it takes to produce that growth. CoreWeave can expand its backlog while interest payments consume the economics of the business. A hyperscaler can grow AI revenue while capital spending grows faster. A chipmaker can report record margins while helping finance the capacity that may later compress them.

Osman is not saying AI stocks are finished. Some may rebound sharply after a fast decline, and some may already offer real value. His concern is the assumption that moving from riskier companies to the strongest players removes the risk.

Microsoft is unlikely to face CoreWeave’s financing problems, but its shareholders could still lose if returns on new investment disappoint. Micron may remain highly profitable as margins fall; ASML may retain a near-monopoly while orders shift; Dell may sell more AI servers without earning the margins investors expect.

The selloff began with companies whose debt made their vulnerability easy to see. It has spread to profitable suppliers and software companies investors treated as shelter. Hyperscalers sit above much of the supply chain: their spending supports the rest of it.

My guess is that the next phase turns on whether investors trust that spending to earn a return, not whether demand for AI survives. If selling reaches the hyperscalers, the doubt will extend beyond the weakest links to the source of money sustaining the whole industry. AI demand need not disappear for the selloff to continue; the cost of serving it only has to rise faster than the income shareholders receive.