News vs. Noise: AI Cash Flow Separates the Winners

Long-term rates jumped to an 18-year high, and AI earnings split sharply between spending and results.

The Fed held its benchmark rate steady for a fifth straight meeting on Wednesday, July 29, signaling that it’s stepping back from further hikes for now, resulting in the market doing more of the tightening on its own. The 30-year Treasury yield — the rate the government pays to borrow for three decades, and a benchmark that ripples through the cost of every long-term loan in the economy — climbed to roughly 5.24% that day, its highest level since 2007, and the Dow fell more than 1,150 points, its worst session since April 2025 [1].

For the past two years, simply being connected to artificial intelligence — regardless of whether that exposure was actually making a company money — was enough to get a stock rewarded. That’s over. Investors are now separating companies that have to prove their AI spending converts into cash flow from those that profit from the buildout itself, regardless of whether any single customer’s bet pays off. Last week’s earnings showed the split clearly: Microsoft’s cloud and AI business beat estimates even as free cash flow — the cash left over after a company pays for its operations and capital spending, not the same thing as profit — kept falling, and its shares rose about 8%; Meta’s revenue grew strongly but free cash flow collapsed 91% year-over-year against a sharply higher spending outlook, and its shares fell nearly 9%; Alphabet reported its first-ever negative free-cash-flow quarter while raising its own capex forecast further; and Amazon raised its 2026 capital spending outlook to $220 billion, citing rising memory costs [1, 2].

We’re treating the level of long-term yields as more important for portfolio positioning than any single day’s headline, and we’re watching whether the market keeps drawing a hard line between AI spending and AI cash flow in the earnings still to come, rather than assuming this quarter’s split holds for every company going forward.

NVIDIA already finances parts of the AI buildout it supplies — and the newest deal on the table would dwarf everything it’s done before.

NVIDIA is reportedly discussing a roughly $250 billion guarantee — a promise to cover a partner’s payments if that partner can’t — so OpenAI can lease a 10-gigawatt data-center campus that SoftBank’s SB Energy is developing in Ohio, and may separately finance up to $350 billion of the chips that would fill it. Separately, NVIDIA was identified as the previously unnamed tenant behind a $50 billion, 15-year data-center lease with Hut 8 in Texas. Neither NVIDIA nor OpenAI has confirmed the terms of the OpenAI deal. This kind of vendor financing isn’t new territory for NVIDIA: its most recent quarterly filing discloses $3.5 billion of maximum exposure across the partner lease guarantees it already carries on its books — a deal the size being discussed for OpenAI would dwarf that existing book many times over [3, 4, 5].

The distinction that matters here isn’t the dollar figures. It’s who is taking on the credit risk. Customer-funded demand is when a buyer pays for a product using its own cash flow or financing it raised independently. Vendor financing is when the seller invests in, finances, or guarantees the buyer so the buyer can afford the seller’s own product. Both can be legitimate — Alphabet reportedly does something similar, backstopping Anthropic-linked data-center leases — but they don’t deserve the same valuation multiple, the price investors are willing to pay for each dollar of a company’s earnings or cash flow. The market appears to be noticing: the cost of insuring NVIDIA’s own debt against default jumped by the most on record the same week, with the annual cost of that insurance on its five-year debt rising as much as 0.14 percentage points, to about 0.82 percentage points of the amount insured. The closest historical parallel is Lucent Technologies, once the world’s largest telecom equipment maker, and its vendor-financing binge during the 1999–2000 telecom buildout, which preceded a stock decline of more than 95% once the customers it had financed started failing — not a prediction of the outcome here, but a reminder of the mechanism at play [3, 6].

We’re watching how much of the AI infrastructure buildout is being funded by suppliers backstopping their own customers, versus genuine independent demand, as a read on how much credit risk is quietly building up in a sector still priced like a pure growth story.

Growth is cooling, inflation is easing — and the continued conflict in the Middle East is adding a fresh, independent source of inflation risk.

The U.S. economy grew at just a 1.5% annualized rate in the second quarter, according to the Bureau of Economic Analysis’s advance estimate released Thursday morning, July 30, below the roughly 1.8% consensus estimate and a step down from the first quarter’s 2.1% pace, reflecting weaker business investment and exports. Released the same morning, the Fed’s preferred inflation gauge — the personal consumption expenditures (PCE) price index — actually cooled: headline PCE inflation eased to 3.7% year-over-year from 4.1%, and core PCE, which excludes food and energy, held at 3.3%. Both readings remain well above the Fed’s 2% target, but the direction was cooling, not heating up [7, 8].
That combination of slower growth and moderating-but-elevated inflation was quickly overshadowed by a separate shock: Iran launched a missile attack on U.S. forces in Jordan late Tuesday, July 28, and the U.S. responded the following night with a heavy wave of strikes on Iranian military targets. Oil, which had fallen roughly 8% earlier in the week on hopes of de-escalation, reversed and rose more than 3% on the exchange of strikes — and as of Monday morning, oil was falling again, a reminder of how quickly this kind of risk can move in either direction [9, 10].
We’re watching the growth-and-inflation data as the more durable signal for the rate path, and treating the Middle East situation as a real but separate risk to monitor for its effect on energy prices and inflation expectations, rather than a reason to make portfolio changes on the news itself.

Last week’s most violent swings — in momentum stocks and Korean tech — were mostly forced buying and selling unwinding, not a verdict on AI.

Goldman Sachs’ U.S. High Beta Momentum basket — an index of the stocks that move hardest and fastest, in both directions, relative to the broader market — was down more than 37% month-to-date through last week, on track for its worst month since the series began in 1999, with the bulk of that decline concentrated in Wednesday’s Fed-driven selloff and the days around it. South Korea’s KOSPI, the benchmark index for stocks on the Korea Exchange and that country’s rough equivalent of the S&P 500, showed the same dynamic at greater scale, round-tripping nearly 40% down and then sharply back up within the same week [1, 11].

Some of the down-move had an identifiable, one-time cause: Situational Awareness, a hedge fund run by former OpenAI researcher Leopold Aschenbrenner, sold the bulk of its public stock portfolio to Citadel, Ken Griffin’s Chicago-based hedge fund and one of the largest in the world, after its prime brokers, the banks that lend a hedge fund the cash and securities it trades with, issued margin calls, demands for more cash or collateral, on leveraged AI-infrastructure and semiconductor positions, with its holdings falling from a peak near $45 billion earlier in the month to roughly $10 billion. Removing a large forced seller helped fuel the rebound — but a forced-liquidation unwind doesn’t explain Microsoft’s or Amazon’s earnings, which is why the rally wasn’t purely mechanical either. Korea’s record one-day rebound was itself partly a short squeeze — investors who had bet against the market being forced to buy back shares — layered on top of a genuine sovereign-fund policy announcement [12, 2].

We’re not reading the size of last week’s swings, in either direction, as a signal about the underlying AI investment case. Forced selling and forced buying both distort how violent a move looks without changing what the companies involved are actually worth.

A record-setting IPO pop and a surprising earnings-day stock drop were both about expectations and available shares, not fundamentals.

CXMT, a Chinese memory maker that listed on the Shanghai exchange, closed its first day up 466%, briefly making it one of the most valuable companies listed in mainland China — but only about 7% of its shares were actually available to trade, so a relatively modest amount of buying demand moved the price dramatically. Meanwhile, SK Hynix, the South Korean memory-chip maker and one of the world’s largest suppliers of the memory chips used in AI servers, posted a record quarter — revenue up 257% and operating profit up 557% to a 76% margin — and its shares still fell 9.6% [13, 1, 2].

Both moves were about positioning, not results. SK Hynix’s drop reflected how much perfection had already been priced into the stock, and analysts attributed the shortfall to HBM4 chip shipments, the newest generation of high-bandwidth memory that feeds data to AI processors, coming in below expectations, not slower than the company’s own plan. It was also one of the names hit hardest in Situational Awareness’s forced liquidation described above — two unrelated sources of selling pressure landing in the same week as a genuinely strong quarter. CXMT’s pop reflected enthusiasm colliding with an artificially thin float, not a considered valuation of the company [1, 12, 13, 14].

We’re not treating either move as new information about the health of the memory sector. A stock can fall on a great quarter, and a stock can nearly quintuple on a thin float — neither tells you much on its own about the business underneath.