Borrowers Down, Payers Up

Update: one week later (as of Sep 15, 2026, 11:01 AM ET)

The original piece looked at a single premarket session. Here is where the same names stand on 1-week and 1-month % change, using TradingView data as of Sep 15, 2026, 11:01 AM ET.

Horizontal bar chart of 1-week % change as of Sep 15, 2026, 11:01 AM ET. Spenders: META +2.68%, GOOGL +1.75%, MSFT +0.74%, AMZN -3.06%. Suppliers: NVDA -2.54%, AVGO -5.41%, ORCL -4.91%, AMD -1.63%, CRWV -7.37%, NBIS -5.51%, GLW -13.05%. Benchmarks: QQQ -1.28%, SPY -0.92%. — One week later: spenders vs suppliers. 1-week % change as of Sep 15, 2026, 11:01 AM ET. Data: TradingView.

On the spender side, META was up 2.68% over the week and 11.46% over the month. GOOGL was up 1.75% on the week and down 0.69% on the month. MSFT was up 0.74% on the week and 0.59% on the month. AMZN, which wasn't on the original screen, was down 3.06% on the week and 6.03% on the month.

Every supplier name was down on the week: NVDA -2.54%, AVGO -5.41%, ORCL -4.91%, AMD -1.63%, CRWV -7.37%, NBIS -5.51% and GLW -13.05%. Over the month the picture is similar, with one exception. AMD was up 4.13%, while NVDA was down 6.18%, AVGO 16.88%, ORCL 8.31%, CRWV 22.41%, NBIS 17.79% and GLW 10.13%.

For comparison, QQQ was down 1.28% on the week and 3.77% on the month, and SPY was down 0.92% and 2.73%.

So the split has continued, with AMZN sitting on the red side of it rather than with the other spenders. One week is a short window, and these numbers don't say why any single stock moved. The original article follows below.

Monday's screen

Here's the screen I woke up to. All figures below are premarket, Sept 14, around 08:25 ET, versus Friday's close (TradingView).

The names that pay for AI were mostly green. META was up 2.2%, GOOGL up 1.6%, MSFT up 0.7%, and AAPL was flat.

The names that build AI, or sell to the people building it, were red. NVDA was down 2.4%, AVGO down 3.9%, ORCL down 4.8% and AMD down 5.5%. CRWV, NBIS and GLW were each down roughly 8%. QQQ was down 1.6%.

Bar chart of premarket % change on Sept 14, 2026 versus Friday's close: META +2.2%, GOOGL +1.6%, MSFT +0.7%, AAPL 0.0%, NVDA -2.4%, AVGO -3.9%, ORCL -4.8%, AMD -5.5%, CRWV -8.1%, NBIS -7.7%, GLW -8.2%. — The payers were green, the builders were red. Premarket change vs Friday close, Sept 14, 2026 (~08:25 ET; CRWV, NBIS, GLW ~07:50 ET). Data: TradingView.

Tech mornings like this usually move together. This one split: the customers up, the suppliers down. So I went through the numbers to see why that split might make sense, and why it might not.

Oracle: the balance sheet everyone's watching

I'll start with Oracle, because it's the cleanest example of building ahead of revenue with borrowed money.

At the end of fiscal 2026 (May 31), Oracle had $129.5 billion in notes payable. Free cash flow for the year was negative $23.7 billion. It brought in $32.0 billion of operating cash and spent $55.7 billion on capex.

The first quarter of fiscal 2027, reported Sept 11, kept going the same way. Capex was $28.5 billion against about $23 billion of operating cash flow, and free cash flow came in at negative $5.4 billion. Capex guidance for the full year is still $90 to $95 billion. Interest expense rose 55% from a year earlier to $1.4 billion. Oracle also completed a $20 billion at-the-market stock sale during the quarter.

Bar chart of Oracle operating cash flow versus capex: fiscal 2026 $32.0B versus $55.7B (free cash flow -$23.7B); Q1 fiscal 2027 about $23B versus $28.5B (free cash flow -$5.4B). — Oracle is spending well ahead of the cash coming in. US$ billions. Sources: Oracle Q4/FY2026 results; ERP Today, Sept 11, 2026.

The credit market noticed before this weekend. In July, S&P cut Oracle from BBB to BBB-, one notch above junk, with a stable outlook. S&P projects free operating cash flow of roughly negative $42 billion in fiscal 2027. Later that month, Oracle's five-year credit default swaps traded near 203 basis points, reported as the highest in nearly 18 years.

Then there's the backlog. Remaining performance obligations hit $664 billion last quarter, up $209 billion in a year. That looks great until you ask who it's from. In July, S&P said roughly half of Oracle's then-$638 billion backlog was tied to OpenAI, and called OpenAI a "central credit risk."

S&P made another point I keep coming back to. Oracle's long-term data center commitments can't easily be terminated or moved to other customers on comparable terms. Oracle builds first and collects over many years. If the anchor tenant slows down, Oracle still owns the buildings and the debt.

To be fair to Oracle, management says most of the $30 billion-plus in new AI contracts signed last quarter are prepaid or bring-your-own-hardware, so they don't need extra Oracle capital. That's a real mitigant. It just doesn't change the contracts already on the books.

One more data point from this morning: Larry Ellison cancelled a trading plan that would have let him sell up to $7.5 billion of Oracle stock. The stock was still down almost 5%.

Neoclouds: more leverage, fewer customers

If Oracle is leveraged, the neoclouds are leveraged with less cushion.

CoreWeave's 2025 annual report showed $5.1 billion of revenue, and about 67% of it came from one customer, Microsoft. In Q2 2026 revenue more than doubled to $2.58 billion, and the backlog grew to $104.2 billion. Those are big numbers. At June 30, though, CoreWeave also carried about $35 billion of total debt. Net interest expense for the quarter was $640 million, up from $267 million a year earlier. Capex guidance for 2026 is $35 to $39 billion.

So net interest alone is running above $2.5 billion a year.

Nebius is earlier in the curve. Q2 revenue was $582.3 million and it had $8.04 billion of cash at June 30. The company leans on convertible debt, with $4.34 billion raised in the first half. On Aug 24 it closed another $5.75 billion: $3.45 billion at 0.50% due 2030 and $2.3 billion at 4.50% due 2034. Its big anchor is a Meta agreement worth up to about $27 billion over five years, with deliveries starting in early 2027.

Table of interest costs and new debt: Oracle Q1 FY27 interest expense $1.4B, up 55%; CoreWeave Q2 2026 net interest expense $640M vs $267M a year earlier; CoreWeave total debt about $35B at June 30, 2026; Nebius convertible notes $4.34B in H1 2026 plus $5.75B closed Aug 24 ($3.45B at 0.50% due 2030, $2.3B at 4.50% due 2034). — Interest costs and new debt. Sources: ERP Today; MarketDaily; Nebius 6-K (SEC); Yahoo Finance.

Both run the same model. Borrow, build, then rent the capacity out for years to a short list of very large customers.

"Pace the frontier," and why builders would feel it first

The headline everyone pointed to was a weekend of statements from the top of the industry. On Sept 12 one lab chief published an essay arguing the frontier needs to be paced, with a plan built around independent evaluators, industry coordination and international agreements. By Sept 13, OpenAI's Sam Altman, xAI's Elon Musk and Microsoft's Satya Nadella had publicly backed it.

What wasn't said matters more to me. No lab announced a compute cut. No capex number changed. The Register went through the statements on Sept 14 and found no concrete link between "pacing" and less spending. None of the statements suggested progress would actually stall.

My read is that the market isn't pricing a cut. At most it's pricing a slightly higher chance of one. If that's right, builders should feel it more than spenders, because of the order of operations.

A builder like Oracle or CoreWeave spends the money up front, often borrowed, and gets paid back over years of rental. Lab demand, and OpenAI demand especially, sits behind a big share of those backlogs. If frontier training slows even a bit, the odds rise that contracts get stretched, renegotiated or never fully used. Lenders get paid before shareholders, so the equity takes the first hit.

The chip and optical suppliers sit one step further back. NVDA, AMD, AVGO and GLW all sell into that same capex. Corning's Q2 Optical Communications sales rose 32% to $2.07 billion, and Enterprise Networks grew 65%, growth that tracks data center build volumes pretty closely. Broadcom has already shown how fragile the financing chain can be. In May it reportedly said it would fund the first $18 billion phase of OpenAI's custom chip program only if Microsoft agreed to buy roughly 40% of the chips.

I didn't find Corning-specific news from the last few days. Its move this morning looks like it's riding with the group.

Rates and oil are doing their part

The rate backdrop got worse over the same weekend.

Early Monday the 10-year Treasury yield was 4.977%, the highest since 2007, and the 30-year was at 5.367%. Goldman Sachs switched to expecting a 25 basis point hike at the Sept 15-16 FOMC meeting, and said market pricing drove the change more than its own outlook. JPMorgan expects hikes in both September and December, citing hotter-than-expected August consumer and producer prices, oil above $100 and rising yields.

Oil is its own problem. Brent settled at $104.61 on Friday, up 8.7% on the week. Early Monday it was above $107 after attacks on Gulf oil infrastructure, including a Saudi pipeline shutdown.

Higher rates hit the red names harder in three ways. Refinancing gets more expensive for a BBB- issuer like Oracle, and for neocloud convertibles and secured loans. Contract revenue collected years from now is worth less today when the discount rate goes up. And interest bills are already climbing, with Oracle's up 55% and CoreWeave's more than doubled, right when these companies need tens of billions in new funding.

A company funding capex out of ad revenue feels that far less.

The case against my own framing

I don't want to oversell a neat "borrowers down, payers up" story, because there's a decent argument against it.

First, the "cash-rich spenders" aren't as cash-rich as they were. Meta's Q2 capex was $31.08 billion and free cash flow fell to $784 million, from $8.55 billion a year earlier. It guides to $130 to $145 billion of capex for 2026 and carries $83.66 billion of long-term debt. Alphabet's Q2 capex was $44.9 billion and free cash flow was negative $5.9 billion, its first negative quarter since going public. It raised 2026 capex guidance to $195 to $205 billion, and its debt went from about $16 billion to about $100 billion in a year. The spenders' own buildout has already eaten their free cash flow.

Second, nobody has actually cut anything. The lab statements are about process, and none of them touched a budget. President Trump pushed back with "whoever wins AI wins," and Wedbush's Dan Ives said the debate is unlikely to change the "trillions" of expected AI capex.

Third, today might just be a rates-and-oil morning with a convenient headline on top. Nasdaq futures were already down about 1.5% as of 4 a.m. ET, and the selling hit memory and chip names broadly, not only AI builders. And I can't rule out stock-specific news behind any single move, META's included.

One premarket session isn't a trend.

What I'm watching next

These are things I'll be looking at, not calls on how they turn out.

Oracle credit. Its CDS spread and bond prices. If the equity keeps falling but spreads hold near July's levels, that tells me something different than if spreads push past 203 basis points.

The Fed on Sept 16. Whether it actually hikes, what the new projections say about December, and where the 10-year goes after. For leveraged builders, the path of yields probably matters more than any single meeting.

Anything concrete from the labs. A real number on compute or spending, versus more talk about evaluators and coordination. So far there's no number.

Capex guidance and neocloud funding. The next quarterly updates from Meta, Alphabet and Microsoft, since that money is revenue for everyone on the red list. And the terms on any new CoreWeave or Nebius converts and loans, which will show how lenders feel faster than the stock price will.

Oil. Whether Brent stays above $100. That feeds inflation, which feeds the Fed, which feeds everyone's cost of debt.

That's the picture through the numbers as of Monday morning. I'll revisit once the Fed has spoken.