AI Spending Vs. Nasdaq Sell-Off: Why Wall Street May Be Misreading the Boom

The Sell-Off vs. the Prints

• SpaceX: AI is now its biggest business, and the compute plan keeps growing
• Meta: capex up 80%, cash flow below zero, early returns in the core
• Credit: why costlier money could make this cycle last longer

Last week, the Nasdaq 100 fell 5.5%, its steepest weekly drop since the April 2025 tariff shock, and semiconductor stocks posted their longest losing streak since 2008.

The fear is financing. With the 10-year Treasury yield above 5%, investors are asking whether the AI buildout slows, and whether everyone building it can still pay for it.

Then the companies reported. TSMC grew revenue 41% and Lam Research 54%, and ASML said its planned capacity won’t fully cover demand through 2028. And this week’s two biggest prints, SpaceX and Meta, both raised their spending plans.

Stocks are pricing a slowdown; the prints describe the opposite. Here’s how we square the two.

 1. SpaceX: The Newest Hyperscaler Is Speeding Up

Two quarters after its IPO, AI is already SpaceX’s largest business. Revenue rose 161% to $14.6 billion, and the AI segment brought in $7.6 billion against $5.2 billion for Starlink, a lead that holds even without Cursor’s $608 million acquisition.

The bigger surprise was the build plan:

• September: 1.7 gigawatts (GW) of compute, up from 0.6 GW a year earlier.
• Year-end: 2.5–3 GW, up from roughly 2 GW guided in August.
•  End of 2027: at least 13 GW, per Elon Musk, over four times this year’s target.

Spending follows. Capex was $19.6 billion in Q3, $16.4 billion of it on AI, with $35–40 billion guided for Q4. Musk also floated $500 billion to $1 trillion of recurring revenue by late 2027.

Why build this fast into a sell-off? The returns. Management puts the payback on new compute at less than 12 months at today’s prices, and Musk expects rising rental rates to shorten it. The limit he named was power and cooling, with turbine parts 12–18 months behind, not chips or customers.

Our read: the company adding capacity fastest is telling us demand isn’t the bottleneck. Funding is the bigger question. At this quarter’s cost per gigawatt, adding 10 GW next year implies more than $200 billion of capex, over twice the $90 billion of cash and securities SpaceX holds. Its investment-grade bonds yield less than half of what weaker AI borrowers pay, yet even its credit spreads have doubled since July. We dive deeper into this in section 3 below.

2. Meta: Spending More, and Starting to Show Why

Meta was the week’s test of whether the biggest spenders would pull back. It didn’t.

Capex reached $34.9 billion, up 80% year over year, and Meta lifted the bottom of its 2026 range to $135–145 billion. That implies about $54 billion in Q4 alone, more than Meta spent in all of 2024. For 2027, the message was to keep investing aggressively, with numbers due on the Q4 call.

With this level of capital spending, Meta’s free cash flow went negative, at –$2.1 billion versus +$10.6 billion a year ago, for the first time in at least a decade. The balance sheet is now helping: long-term debt is up $22 billion this year, and Meta names capital raising as one use for a new non-voting share class.

Where are the returns? So far, mostly in the core business:

• Pricing: revenue grew 26%, with the average price per ad up 14%, and Meta credits its AI investments with adding ad revenue.
• Costs: running content analysis on its own models cut processing costs by more than 80% in offline tests.
• New products: partnership ads nearly doubled to a run rate above $15 billion.

And the highlight this quarter has been Muse, Meta’s AI agent. Meta drew over 10 million sign-ups in roughly six weeks, and its wider rollout is waiting on compute.

Our read: this is what a spender that can afford to keep going looks like: a core business growing 26%, AA-range credit and borrowing costs not far above Treasuries. Negative cash flow says Meta is early in the investment. For suppliers, the bellwether spender just raised its floor.

3. Credit: Wider Spreads, a More Durable Cycle

For background: a credit default swap (CDS) is insurance against a company defaulting on its debt. It’s priced as a spread: 100 basis points means paying $10,000 a year to protect $1 million of bonds.

CDS is the best real time measure of credit quality of a debt issuer.

Given the unprecedented spending on AI CDS spreads continue to raise. But they are affecting the companies on the higher end of risk curve significantly more. Five-year CDS spreads, in basis points:

CoreWeave: about 1,000, after its debt roughly tripled in a year

• Oracle: near 300, about six times its level a year ago
• SpaceX: about 240
• Meta: about 115
• Microsoft, Amazon and Nvidia: below 100

The market isn’t pricing the end of the buildout. It’s pricing who pays for it. Spreads widened most where borrowing funds spending that cash flow can’t cover, and least where a profitable core business does the funding.

We expect the spending to continue. Street estimates have the five largest U.S. cloud builders going from about $770 billion of capex this year to nearly $1.2 trillion in 2027, before counting SpaceX. What costlier money stops is the marginal project: the speculative campus, the landlord financing GPUs at 15%. Fewer of those keeps capacity tight and pricing firm. Rental rates for NVIDIA’s B200 rose about 28% between January and August, which is how SpaceX can talk about paybacks within a year.

Our read: for the spenders, returns will increasingly depend on credit quality, since borrowing costs now differ by several percentage points from one company to the next.

Across the chain, we think returns are likely to rise: less exuberant spending means less overbuilding, and a more durable cycle. As a result we expect GPU rental rates to continue to higher.

Among suppliers, we prefer the products that benefit from mix shift rather than broad data center spending.

Bottom Line

• The sell-off and the prints are measuring different things. Stocks are pricing the cost of money; earnings are measuring demand.
• Demand: SpaceX and Meta both raised spending; power, not customers, sets the pace.
• Funding: credit markets are separating strong balance sheets from stretched ones, not shutting the buildout.
• Returns: fewer marginal projects should mean tighter capacity and better economics for those that get built.
• Expensive money is sorting the builders. It isn’t stopping the build.

What We’re Watching

Microsoft (October 28) and Amazon (October 29), for whether the other hyperscalers raise spending as Meta did. Oracle’s investor day next week, for how the most credit-stretched of them plans to fund its backlog. The Fed’s decision that week, with a December hike largely priced in. And CoreWeave’s results, a read on the highest-cost borrowers.