How to Trade This Pullback, and What to Watch Next Week

The S&P 500 is down 2% over two weeks. The Philadelphia Semiconductor Index is down 23%, eight straight losing sessions, the longest streak since 2006 . Over the same two weeks, CoreWeave’s cost of borrowing went from 5% to 15% .

Those are the same story. This piece explains why.

Disclaimer: This article is for informational and educational purposes only. It does not constitute financial, legal, or investment advice. Views are based on public information as of the publication date and are subject to change.

What Happened in the Last Two Weeks

• October 6: the S&P and Nasdaq close at record highs.
• October 7: the 10-year Treasury yield touches 5.37% intraday, the highest since 2002, before a strong afternoon auction pulls it back. Rate pressure peaks .
• October 8: the FT reports OpenAI’s annualized revenue is about $50 billion, not the $70 billion that had been circulating. The gap is later confirmed to be a difference in accounting definitions. The Nasdaq falls 1.25%. The semiconductor losing streak starts here.
• October 12: news of a memory breakthrough at China’s CXMT, and Broadcom’s CEO says capital, not chips, is the constraint on AI. Memory stocks lead the decline; Korea’s KOSPI drops 5.5%.
• October 13: September CPI runs hot. CoreWeave’s bond yields cross 15%.
• October 14: PPI prints +1.3% month over month, the hottest since 2022. The FT reports Anthropic’s Q3 revenue at $16.8 billion, up 45% quarter over quarter, down sharply from 147% the quarter before . Semis sell off hard.
• October 15: CoreWeave’s CDS crosses 1,000 basis points, implying better than 60% odds of default within five years. The KOSPI falls 7.6% and trips its circuit breaker. Reuters reports Anthropic’s IPO valuation discussions have slipped from $2 trillion to $1.5 trillion .
• October 16: Brent settles at $109.86, the highest since May. Of 45 software stocks, 41 close down. The VIX touches 22 intraday. Rumors of a liquidation at the AI hedge fund VARA circulate online and remain unconfirmed .

Eight days in one sentence: the market questioned AI demand, then AI debt, and then oil pushed the rate question back onto the table.

What the Market Is Pricing

There are three kinds of participants in the AI trade. The ones selling chips, Nvidia above all, are collecting cash. The ones borrowing to buy chips and build data centers are Microsoft, Google, Amazon, Meta, Oracle, and newer names like CoreWeave. The ones lending to the second group are bond investors.

Only the first group has actually been paid. The other two are earning expectations of future profit.

For the past year the market watched demand: how fast AI companies grew revenue, how large the cloud backlogs were. In the past two weeks it switched questions: can the borrowers pay it back?

Three data points forced the switch:

1. OpenAI’s revenue came in below the rumor.
2. Anthropic’s growth is decelerating.
3. CoreWeave’s borrowing cost went from 5% to 15%.

All three point at one question: if AI companies grow more slowly than expected, how do the companies that borrowed tens of billions to build data centers service the debt?

Five hyperscalers have guided to $825 billion of capex this year. The AI revenue you can actually see, OpenAI, Anthropic, AWS’s AI business and custom silicon, the AI portion of Google Cloud, adds up to $210 to $250 billion . Twenty-five to thirty squares out of a hundred.

The lower chart is a second ledger. Four hyperscalers spent $434 billion over the twelve months to March; their income statements have only recognized $149 billion of depreciation. The remaining $285 billion lands in reported earnings quarter by quarter through 2027 and 2028, whether or not AI revenue keeps pace . The money went out as cash. It comes back through the P&L as depreciation, two to three years later.

Demand for compute can grow exponentially; the capital to finance it cannot.

What Happened to Valuations (Educational)

This section resolves an apparent contradiction: S&P Q3 earnings are growing 29%, and the index is falling .

A stock price is earnings times a multiple. The multiple is the price-to-earnings ratio. A P/E of 19 means you pay $19 for $1 of annual profit.

If 19 feels abstract, flip it. One divided by 19 is 5.26%. Put $100 into the index and the companies behind it earn $5.26 a year on your behalf. That number is the earnings yield.

It works exactly like rental yield. A $1 million house renting for $52,600 a year yields 5.26%. The higher the price, the lower the yield; when prices fall, the yield rises. P/E and earnings yield are house price and rental yield.

Now compare that 5.26% to the 10-year Treasury at 5.12%. Same $100, the Treasury pays $5.12 a year, returns your principal at maturity, and carries no risk.

Stocks 5.26, Treasuries 5.12. The first requires you to absorb every swing in the equity market, and pays you 14 cents a year more than the second. Those 0.14 percentage points have a name: the equity risk premium. It is the extra return you earn for taking equity risk.

Since 1950 that premium has averaged about one point. It is now close to zero. Does that mean stocks have to fall? History says not necessarily. From 1996 to 2002, with the 10-year between 5% and 6.5%, the premium was negative for six years . Investors believed earnings would grow fast enough that they needed little compensation. A low premium can persist as long as earnings keep delivering.

A low premium does carry a cost: there is no cushion. With a two-point premium, bad news gets partly absorbed and the price does not have to fall as far. With a zero premium, bad news transmits straight into the price.

That is why the news flow of the past two weeks looked less bad than the price action. OpenAI was an accounting definition. Anthropic is still growing fast and still highly profitable. Stocks fell anyway, and faster each day.

Put all of the above into one formula:

Multiple = 1 ÷ (Treasury yield + risk premium)

There are only two variables. Move each one and you get a map of where the S&P goes. See the chart: