Every hyperscaler capex model built since January 2023

TO: Every hyperscaler capex model built since January 2023
FROM: The discount rate you kept assuming would behave
RE: Brent at $100, FOMC on deck, and the arithmetic you've been ignoring

You built beautiful models. Terminal growth rates that assumed AI demand compounds forever, WACCs pinned near 8% because rates were "normalizing," and free cash flow curves that only ever bent upward. Nobody stress-tested the denominator. That was the mistake, and this week it's collecting.

Brent crude closed above $100 a barrel on Thursday for the first time since May, the sixth-plus consecutive day of strikes on Iran doing what Hormuz disruptions always eventually do to a barrel of oil — and CME's FedWatch tool now prices an 82% chance the Fed hikes in September, up from under 53% a week earlier. Read that again. Not "holds longer." Hikes. Fed funds futures assign a 38% probability to a quarter-point move at the very next meeting, versus 12% seven days ago. Jobless claims fell to 187,000, the lowest since 1969, which under any other circumstance would be a labor-market victory lap. Instead it's ammunition for Kevin Warsh, who has spent his tenure so far telling anyone who'll listen that prices are still too high and that the Fed's easing bias is gone from the statement — because now he has a jobs report that lets him prioritize the inflation fight without pretending he's risking a recession to do it.

This is the scenario your model didn't run. Nine of eighteen FOMC participants already flagged the case for at least one hike before year-end back in June, when oil was still behaving. Now core PCE is tracking near 3.4%, headline near 4.1%, and gasoline just crossed $4 a gallon nationally for the first time in over a month. The inflation isn't demand-side, and everyone knows it — it's a fifth of the world's oil supply getting bottlenecked through a war zone — but the Fed doesn't have a "supply shock, please disregard" line item on its mandate. It has a number, and the number is uncomfortable, and Warsh has made abundantly clear he intends to act on the number rather than the narrative.

So here's the mechanism, spelled out slowly for the models that skipped this part: AI infrastructure — the fabs, the power contracts, the GPU clusters depreciating over five years while promising returns over fifteen — gets financed against long-run discount rates. Raise the policy rate, and the entire curve used to price those distant cash flows repricing higher moves with it. A hyperscaler's $500 billion data center buildout doesn't get cheaper to justify because Jensen Huang says demand is insatiable. It gets more expensive to justify, mechanically, every time the ten-year or the terminal Fed funds assumption ticks up. Net present value compression isn't a vibe. It's a formula, and the formula doesn't care that Nvidia beat estimates last quarter.

The market has started pricing this correctly, even if the sell-side notes haven't caught up. The Roundhill Magnificent Seven ETF shed more than 5% this week. Semiconductor names that carried the entire 2025 rally are the ones getting sold hardest into strength elsewhere — energy and commodities led last week while growth and duration got dumped, which is exactly the rotation you'd expect when the market starts pricing a hawkish surprise instead of a dovish drift. Tesla's Q2 operating profit missed by $1.3 billion and the stock dropped 18% in a single session, a reminder that "growth story" valuations don't survive contact with an earnings print once the discount rate stops being your friend. And DeepSeek — the one lab that was supposed to prove frontier AI could be built cheap — quietly told backers it's pausing its second fundraising round. Make of that what you will; capital is getting more selective everywhere the story requires patience.

None of this means the hike happens. FactSet's consensus among economists still calls for no 2026 hikes at all, with modest cuts penciled into 2027, and the Fed is close to unanimous that it holds steady at 3.50%–3.75% at this week's meeting. But consensus was also under 53% on a September hike seven days ago, and now it's 82%, and that kind of repricing in that short a window tells you more about market psychology than any single dot plot. The Fed doesn't need to actually hike to do damage here. It just needs to keep the odds elevated long enough for every capex model in Silicon Valley to quietly haircut its terminal value assumptions, and for every CFO on an August earnings call to get asked, on the record, what happens to the payback period if the tenth year discount rate is 200 basis points higher than the deck assumed.

You built your models assuming rates were the boring variable. Iran, the IEA's record-setting 400-million-barrel reserve release, and a Fed chair who's made hawkishness a personality trait have just reminded you that the denominator was never boring. It was just quiet. It isn't anymore.

— filed ahead of the July 29 FOMC decision, before anyone's spreadsheet catches up