1The AI capex question got answered this quarter — and the answer says buy the suppliers, not the spenders
For two years the bear case on AI capex was never that the numbers were big. It was that nobody could tell you what they earned. That objection died on the Q2 calls, and it died on a specific number: Amazon's Andy Jassy told the market that the company's short-life data-centre assets — the chips and networking gear, the part everyone was actually worried about — break even in two to three years against a five-to-six year useful life. That converts an unquantifiable spend into a datable return, and the market re-rated the supply chain within days.
The backlogs behind it are committed demand, not pipeline. AWS is carrying $496bn, Azure $678bn, Google Cloud $514bn — over $1.5trn between them. Growth is accelerating, not decaying: AWS to 37%, Azure guided from 43% to 45%, Google Cloud at 82%, all at roughly 50% incremental operating margins.
Here is the part that cuts against instinct. Two T. Rowe Price portfolio managers spoke on separate podcasts within 24 hours, and both landed on the same trade — and it is not to own the companies whose returns were just proven. Sam Ruiz, on Equity Mates, put it as "solve the easier problem": who wins the compute race is genuinely unknowable, so he is neutral on the hyperscalers and constructive on the hardware that gets paid whoever wins — cooling, optical, memory, CPUs, racks. Dom Rizzo, on Excess Returns, reached the same place from the semiconductor side, arguing the recent 30% drawdown is a 1998-style correction rather than a top. His evidence is fundamental, not chart-shaped: in 1998 semiconductor revenue fell 8% into that drawdown; in 2026 it is up 64%. Same price action, opposite earnings backdrop.
Both now model 2027 hyperscaler capex at $1.2–1.6trn against roughly $800bn this year.
Two caveats to hold onto. First, these two are the same firm. They reached compatible conclusions from different seats, but this is one house's view appearing twice — not independent corroboration. Second, they openly disagree on where inside the trade to sit: Rizzo wants logic over memory into 2027, on the view that memory's 350% growth rates have to decelerate; Ruiz still likes memory, while flagging that expectations are full. Same firm, same week, unresolved.
2The market is pricing a Fed hike into a jobs report that lost 23,000 jobs
July payrolls came in at −23,000 against expectations of +80,000, with May and June both revised down. That is a dovish event by any normal reading. Polymarket moved one point on it.
September now prices no change at 58.5%, a 25bp hike at 40.5%, and a 25bp cut at 1.45% — on $25.7m of volume and $2.77m of liquidity, the deepest contract on the board. The market is not debating whether to ease. It is debating whether to tighten.
The reconciliation is the participation rate. Unemployment actually fell to 4.1%, a two-year low, but only because the labour force shrank — participation dropped to 61.4%, the lowest since February 2021. If workers are leaving rather than being laid off, the economy can shed jobs without the labour market loosening. Supply contracted alongside demand. That reading is inflationary, and it is the one the market has picked.
Inflation pricing corroborates it. July CPI resolves tonight, and over the last 24 hours the distribution drifted upward — probability left 3.3% and moved into 3.4% (now modal at 39%) and 3.5%. Core CPI month-on-month consolidated on 0.2%, with mass leaving both tails. Benign, but not soft, and not a case for easing.
Hold the disagreement in view, because it is real. Katherine Anne Edwards, the labour economist on Prof G Markets, reads the same print as straightforwardly bad: jobs lost and people leaving is the worst pairing, the falling headline rate is an artefact, and the freeze is in hiring rather than firing — which is why wage growth is flat, since wage gains come from job-switching. So the economist and the market are looking at one number and disagreeing about which side of the labour market broke. If she is right, September is mispriced. If the market is right, the weakness does not buy any easing anyway.
3Where a headline and a prediction market disagreed, back the market
Google News carried "Pakistan says U.S. and Iran close to deal," and a TradingKey piece credited the prospect with igniting risk appetite. Polymarket prices a final US–Iran nuclear deal by 31 August at 1.7%, on $14.2m of volume. By 30 September, 5.5%. It does not clear 25% until year-end.
What is holding is the ceasefire — 95.5% through 14 August, and it firmed a point. President Trump's own Monday post responded to Iranian compensation demands by counter-demanding compensation from Iran. That is a hardening posture, not a converging one.
The distinction matters for energy specifically. A ceasefire stops the escalation premium building. Only a deal returns sanctioned barrels to the market. With a deal priced near zero inside three weeks, the structural bid under crude stays. Retail gasoline printed $4.006/gal on 10 August, down from $4.079 the week before, still historically elevated.
What the sources recommended
Only eight single-name calls came out of four videos, and two of the four buys carry a caveat from the person making them.
The buys. (Vertiv) for rack cooling, with Sam Ruiz on Equity Mates flagging the weak last print as execution rather than demand. (Siemens) for European datacentre infrastructure. (SK Hynix) over Samsung on cash-flow durability — immediately followed by "don't back up the truck at three times earnings," his point being that cyclicals look cheapest at the top, so treat it as conditional. , from Rich Gelfond's CEO interview on The Compound, though note this is a reiterate of Michael Batnick's May-2024 $17 pitch at roughly 3x ago, not a fresh initiation — and the interview had no bear case put to it.
The sells. One thesis from Dom Rizzo on Excess Returns covers three of the four: (Salesforce), (ServiceNow) and (Workday) become "dumb data pipes" into the frontier labs as the interface shifts from humans to agents, and he says he has seen no compelling AI-native product from any incumbent app-software vendor. Separately — Ruiz is steering clear, not because the AI-compute case is weak but because the rocket business makes the whole thing unsolvable.
The thesis gets marked to market in 24 hours
reports after the close tonight and tomorrow. Both are optical components — the exact sub-sector Ruiz named as his preferred hardware exposure. follows tomorrow after the close, tonight, this morning. July CPI resolves tonight. If the AI-hardware read from the first section is right, these are where it shows up first.
One move worth its own line: fell 4% on a $15bn common-stock offering to fund AI, and fell 3% on a $500bn Wall-Street infrastructure financing package. Both are financing events, not demand events — idiosyncratic to the issuer, not a read-through to the sector.
