This Pullback Isn't Over: The Market Is Flashing Liquidity Warnings

Disclaimer: This article is for informational and educational purposes only and does not constitute financial, legal, or investment advice. The views expressed are based on public information available on the date of publication and may change without notice.

The FOMC decision came out yesterday, and two of the hyperscalers reported earnings. Nasdaq futures (NQ) swung sharply after the close, then started to slide in the overnight session.

Longtime readers should remember what I said last weekend: this bounce would be weak, and it would get sold. On Monday, Tuesday and Wednesday the Nasdaq rally still looked strong. Now it is starting to break down.

So I’m writing this update right away. Lots to discuss!

How the Rally Was Built

A quick recap of the past three sessions:

• Nasdaq 100: +0.8% Monday, +1.3% Tuesday, +0.3% Wednesday, about +2.4% over three days. Monday and Tuesday were its first back-to-back gains of more than 1% in three months.
• S&P 500: from 7,623.68 on Friday to about 7,744 on Wednesday, +1.6%.
• Philadelphia Semiconductor Index: about +6.5% over three days. Memory names up by about +7.5%.
• SOXX closed above 500 on both Tuesday and Wednesday.
• QQQ closed at 721 on Tuesday and about 723 on Wednesday, right at the top of my key 700–725 zone.

All three days had one thing in common: breadth was negative. The equal-weight S&P underperformed the cap-weighted index on Monday, Tuesday and Wednesday.

Decliners outnumbered advancers in the S&P. The gains came from AI-linked mega caps, and everything else fell. Semiconductor ETFs took in a record $11 billion last week, while gold, utilities, financials and small caps saw outflows. Money rotated out of defensives and back into AI.

What drove the rally?

• First, the financing overhang cleared. On Monday, Meta raised $95 billion in equity and SpaceX raised $40 billion in debt. On Tuesday, Broadcom’s $63 billion SPV closed. Anthropic filed its S-1, and its IPO timeline is unchanged.
• Second, selling pressure exhausted. Goldman Sachs data show hedge fund net exposure at its lowest level of the year, and the past two weeks saw the largest institutional tech liquidation on record. JPMorgan says there is little room left for concentrated selling. When selling exhausts, prices can rise without much buying.
• Third, dealer hedging changed direction. The S&P flipped back into positive gamma on Monday, so dips got bought. QQQ is still in negative gamma, but the large open interest at the 700 strike held, and heavy call open interest near 725 pulled price toward that level.

What Happened During Wednesday’s Session

The Fed: The Fed held rates at 3.75–4.00%. Two of the 12 voting members dissented: Cook and Hammack favored a 25 bp hike. The statement upgraded its description of economic activity from “solid” to “strong.” It dropped a line from the September statement, “a rate increase would help return inflation to 2% in a more timely manner,” and replaced it with “gathering more information will help determine the appropriate policy stance.” At the press conference, Warsh said financial conditions had already tightened and gave no forward guidance. Fed funds futures price a 90% probability of a December hike.

Bonds: Long-end yields rose 5–6 bp while the front end was little changed, so the curve steepened.

Oil: WTI rose 2.9%, back above $100. Oil above $100 pushes up inflation expectations.

Why did long-end yields rise when the Fed did nothing?

Last Thursday’s Treasury rally was flight-to-safety buying: money left AI stocks and went into Treasuries. This week AI stocks bounced, that safe-haven bid unwound, and Treasuries were sold.

Supply mattered too. Treasury auctioned $183 billion of 2-, 5- and 7-year notes this week, and both the 2-year and 5-year auctions tailed. A tail means the auction cleared at a higher yield than the market yield just before the auction, which shows weak demand. In the same week, Meta issued $95 billion in equity, Broadcom’s SPV issued $63 billion in debt, and SpaceX issued $40 billion in debt. All of them competed for funding in the same capital markets.

AI borrowing is absorbing capital at an extreme pace, and borrowing costs are rising for everyone, including the U.S. government. That is what Warsh meant by tighter financial conditions:

The market is doing the tightening, so the Fed has to do less.

That is also why gold fell instead of rallying.

Four After-Hours Reports: How the Market Graded Them

What Free Cash Flow is, and why it is the only number that matters now. Free cash flow = operating cash flow − capital expenditures. It measures how much cash a company has left after maintaining and expanding its business, which it can use to pay down debt, buy back stock or pay dividends. If it is negative, it shows how much the company has to raise from outside.

In Q3, Google generated $42.3 billion in operating cash flow and spent $51.4 billion building data centers, a shortfall of $9.1 billion. That $9.1 billion has to come from somewhere: existing cash, new debt or new equity. Google already issued $84.8 billion of stock in June. Two straight quarters of negative free cash flow, plus 2027 capex guidance above $300 billion (another 50% increase over 2026), mean the gap will be larger in 2027 and Google will need to borrow more money.

You may see analysts highlight 103% cloud growth. But the cash that business generates is already included in the $42.3 billion of operating cash flow. In other words, after counting all of its revenue growth, the company is still $9.1 billion short.

A company deserves a high valuation only if today’s investment will generate more cash in the future than it costs.

Google’s problem is not a lack of growth. Its problem is that it has not said when capex will start to come down and when cash will start to show a net inflow. Financing conditions are tough, and credit is deteriorating. If a company cannot tell the market how long it will keep borrowing in this environment, the market prices in the worst case.

Oracle is the opposite. Its cash flow is also negative, but at its Investor Day it told the market that capex will peak in FY27–28, free cash flow will turn positive in the second half of FY28, and FY29 free cash flow will exceed $30 billion. That tells the market Oracle needs to borrow only about $20 billion more over the next four years. The commitment puts an exact number on how much Oracle has to borrow, so the market sees a borrower with limited financing needs and a clear end date.

CoreWeave rose more than 10% the same day. Oracle is one of the largest customers of CoreWeave and the broader neocloud group. When Oracle says its returns pencil out, it is indirectly saying that the customers of the compute rental business can afford to pay.

Microsoft’s small after-hours gain follows the same logic. Capex came in at $45.2 billion, below its own guidance, and the shortfall becomes extra free cash flow. From here we can connect the dots.

Part of Microsoft’s capex goes to companies like CoreWeave and Nebius. If those data centers are not energized on schedule, Microsoft does not have to pay for them this quarter, so capex comes in below plan. For Microsoft this is good news. It spent less, and Azure still grew 48%, which shows it is not short of compute for now.

For CoreWeave and Nebius, the same fact means their data centers were not delivered on time. Revenue recognition slips to a later quarter, while interest on the debt they took on to build those data centers is still due. That is why CoreWeave and Nebius each fell 5% after hours. Microsoft is CoreWeave’s largest customer, so every dollar Microsoft doesn’t spend is a dollar CoreWeave doesn’t collect.

I’m teaching you to connect the dots. Learn it.

Last, ServiceNow. ServiceNow sells software. Its subscription revenue used to be very stable, with gross margins around 83%. AI has changed the nature of its costs. Every AI model call consumes compute, and compute is billed by token usage or by GPU-hour. The more customers use AI, the more ServiceNow pays its compute providers.

Costs are rising, but the company cannot raise prices. Its competitors, including Salesforce, Workday and Microsoft itself, are all using AI features to win the same enterprise customers, and whoever raises prices first loses customers. So software companies are using their own gross margin to subsidize customers’ AI usage.

They are waiting for one of two things: lower unit costs for compute (Jevons Paradox), or customers who depend on AI features enough that prices can go up later.

ServiceNow’s falling gross margin shows that software companies cannot pass compute costs on to customers. Microsoft’s gross margin beat shows that cloud providers can. ServiceNow’s gross margin has now fallen for four straight quarters. The after-hours drop broke through the entire positive-gamma range, and dealer hedging switched from buying the dip to selling into the decline, which made the drop bigger.

Why Volatility Is Just Starting to Worsen