AI Stocks and the Semiconductor Sell-Off: When the Bill Comes Due

The screen turns red, and the questions change

A sharp fall in a favourite stock has a way of making yesterday’s confident explanation sound incomplete. The AI opportunity has not disappeared overnight. People still use the tools. Companies still need computing power. Yet the shares that helped carry the story can suddenly fall together, leaving readers wondering whether they missed a warning.

The October 22 sell-off brought that tension into focus. The Nasdaq Composite fell about 2.6%, the S&P 500 lost about 1.2%, and NVIDIA declined 4.4%, according to Associated Press. Many US shares nevertheless advanced. That distinction matters: a concentrated technology decline can make the headline index look weaker than the typical stock. [1]

This article examines that session and the business questions behind it. It is analysis dated October 23, 2026, rather than a live price feed or a prediction of the next trading day.

The spending story meets the cash-flow statement

Meta’s October 21 results offered a useful example. Third-quarter revenue rose 26% to $64.60 billion, while capital expenditure, including finance-lease principal payments, reached $34.91 billion. Operating cash flow was $32.79 billion; its reported free cash flow was negative $2.12 billion. The company narrowed its full-year capital-expenditure range to $135–145 billion from $130–145 billion. The ceiling did not rise in that update. [2]

Meta also reported a $9.89 billion litigation charge. Treating every change in reported profit as an AI expense would therefore misread the accounts. Management said AI was helping its advertising business while expecting newer AI initiatives to take longer to produce returns. Those are management’s assessments, not independent proof of future profitability. [2]

The wider issue is the timing gap. Buildings, servers, networking and power commitments can require cash before the applications using them generate an adequate return. A supplier may book revenue today while its customer waits much longer for the economic payoff. My interpretation is that this gap helps explain why investors can become uneasy even while reported demand remains strong.

Elon Musk, Tesla and the smaller spending plan

Tesla added another uncomfortable question to the AI story: what happens when a company plans to spend less? Reuters reported on October 22 that Elon Musk’s company lowered its 2026 capital-spending forecast to about $20 billion from more than $25 billion, following lower-than-expected third-quarter spending on AI and robotics. The report said Tesla did not identify which areas would receive cuts. [4]

The regulatory filing provides an important qualification. Tesla’s third-quarter Form 10-Q says it expects 2026 capital expenditure to exceed $20 billion. That is a company-wide forecast covering compute infrastructure, data centres, manufacturing, research facilities and other assets. It is not a separately disclosed AI-only budget, and it does not establish a corresponding reduction at xAI. [5]

A lower spending plan can improve near-term cash flow. It can also leave investors asking whether expansion is taking longer, projects have been reprioritised, or existing resources are being used more efficiently. Those interpretations have different consequences. A budget change alone cannot tell us which one will dominate.

For semiconductor investors, the connection is straightforward but conditional: if a major buyer delays computing investment, some supplier revenue could arrive later. That is an analytical possibility, not confirmation of a cancelled chip order. The next useful evidence is the timing and composition of actual purchases, followed by what suppliers report about orders and cash collection.

A strong business can still be an expensive stock

A share price reflects more than the latest sales figure. It also reflects the growth, margins and durability investors expect in future years. If those expectations become unusually demanding, a company can deliver impressive growth and still disappoint its shareholders.

Consider a deliberately simplified example, unrelated to any company’s actual valuation. A business earning $10 a share is valued at 40 times earnings, giving a $400 share price. Earnings then rise 20% to $12, but investors become willing to pay only 28 times earnings. The resulting price is $336—16% lower despite higher profits. The example ignores dividends and other valuation inputs, but it shows how changing expectations can outweigh improving earnings.

That is why a technology story and an investment result can travel in different directions. To evaluate a sell-off, ask whether expected cash flows deteriorated, whether investors reduced the price paid for those cash flows, or whether both changed. A price chart alone cannot answer that.

The evidence against declaring the AI boom finished

There is substantial operating activity behind the theme. In its August 26 release, NVIDIA reported second-quarter fiscal 2027 revenue of $96.2 billion, up 106% year over year, including $89.0 billion of Data Center revenue. It guided the following quarter to $108 billion, plus or minus 2%. That guidance was a forecast when issued, not a completed result. [3]

These figures do not settle whether a particular share price is attractive. They do make a blanket claim that the industry has no revenue difficult to sustain. The more useful debate concerns the durability and profitability of the spending, its financing, and how much future success a valuation already assumes.

Both sides of the debate can contain something true. AI can create useful products while some projects earn disappointing returns. A chip designer can sell more hardware while competition eventually pressures pricing. A customer can have a successful core business and still allocate too much capital to a new initiative. The outcome need not be uniform across the sector.

‘AI stocks’ describes several very different businesses

The label hides important differences. A chip designer develops processors and related systems; a foundry manufactures chips; a memory producer supplies storage and working memory; an equipment company sells machinery used to manufacture semiconductors. Cloud operators purchase infrastructure and rent computing capacity. Software businesses try to turn that capacity into useful, paying applications.

Those positions create different sensitivities. A designer may be exposed to architecture changes and competitive pricing. A manufacturer faces large capital commitments and factory-utilisation risk. A memory supplier must manage supply, product mix and pricing cycles. An infrastructure renter may be especially sensitive to financing costs and the gap between committed spending and customer payments.

NVIDIA, AMD, TSMC, Micron and semiconductor-equipment manufacturers belong in different parts of that discussion. Mentioning them together does not establish that their balance sheets, valuations or prospects are equivalent. Their latest filings need to be assessed individually before drawing a company-specific conclusion.

Why the selling can spread beyond the original trigger

One earnings release can change how investors think about an entire spending chain. If a large infrastructure customer faces pressure to justify its budget, investors may revisit assumptions for the suppliers expecting to benefit from that budget. The questions spread faster than a new set of company results can arrive.

Portfolio mechanics can amplify such moves. Funds may own overlapping holdings; investors may cut several correlated positions at once; leveraged exposures may need to shrink when volatility rises. These are possible transmission mechanisms, not verified explanations for every trade on October 22. Establishing the role of forced selling requires evidence beyond falling prices.

The same care applies to interest rates. Higher required returns can reduce the present value of distant profits, but that does not mean bond yields rose during every technology sell-off. AP reported falling Treasury yields on October 22 as investors sought safety. A generic ‘rates went up’ explanation would not fit that session. [1]

The chip cycle still matters

Semiconductors connect a fast-moving product cycle to expensive physical capacity. Building supply takes time. Customers make forecasts, suppliers commit resources, and demand can change before all the equipment and factories are ready. AI introduces new demand, but it does not remove the possibility of excess capacity or product transitions.

It also matters which chip is in short supply. Strong demand for a particular accelerator or specialised memory product does not automatically mean every semiconductor category has the same pricing power. Consumer devices, industrial applications and data centres can follow different paths.

Watch inventories alongside sales. Rising inventory is not automatically a warning: a company may be preparing for a launch. It becomes more concerning when inventory, receivables and production commitments rise faster than cash collection and credible customer demand. The explanation and the supporting numbers need to agree.

A lower price needs a new valuation, not an old target

A stock that falls 25% is cheaper than its own previous price. Whether it offers better value depends on what happened to the business and the forecast. If expected earnings also fell sharply, the apparent discount may be smaller than the chart suggests.

Revisit the assumptions underneath the valuation: revenue growth, gross margin, operating spending, dilution, capital expenditure and the return demanded by shareholders. Use comparable periods and distinguish reported earnings from adjustments. An attractive-looking multiple based on a temporary earnings peak can be misleading.

It helps to write down what would justify the valuation before choosing a conclusion. Does the model require exceptional growth for several years? Does it assume unusually stable margins? Does it allow customers to spend less? This makes the reasoning easier to challenge than a statement that a share ‘must recover’ because it once traded higher.

What the next disclosures need to clarify

The next useful evidence is likely to come from operating disclosures and financing terms, rather than a confident social-media forecast. The following questions provide a way to organise that evidence without assuming the answer in advance.

• Customer spending: are budgets supported by operating cash flow, borrowing, equity issuance or contractual commitments?
• Cash conversion: are growing sales turning into cash, and are receivables or payment terms changing?
• Economics: do gross margins and utilisation support an adequate return after power, depreciation, financing and operating costs?
• Demand quality: how much revenue depends on a small group of customers, and what flexibility do those customers have to delay purchases?
• Competition: are alternative processors, customer-designed chips or more efficient software changing the economics?
• Execution: are delivery schedules, power availability and product transitions consistent with management’s previous expectations?

Three paths worth keeping open

In a more constructive scenario, AI usage and monetisation expand, customers fund their commitments comfortably, and suppliers continue converting demand into cash. Valuations could still remain below earlier peaks if investors demand a larger margin for uncertainty.

In a middle scenario, demand grows but spending becomes more selective. Some capacity earns attractive returns while other projects are delayed. Share-price performance then depends more on company execution and starting valuation than on the AI label.

In a weaker scenario, customer financing tightens, utilisation disappoints, orders are postponed, or pricing and margins deteriorate. Earnings forecasts and valuation multiples could fall together. These are analytical scenarios, not assigned probabilities or promises about what happens next.

The question to carry beyond the sell-off

A red screen encourages a quick verdict. A useful research process asks for a little more patience: what changed, which assumption was wrong, and what evidence would support a different conclusion? That discipline applies equally to a confident bull and a convinced sceptic.

The October sell-off does not, on its own, establish that AI has failed or that chip shares have reached a bargain price. It brings the financing and return on enormous investment commitments into sharper focus. The lasting question is how much durable cash each part of the AI economy can generate for the capital it consumes.

Keep the accounts, the valuation and the market reaction in the same conversation. They tell different parts of the story, and the most expensive mistakes often begin when one of them is ignored.