The Signal
Nvidia now generates more free cash flow than the customers it is funding. That sentence reads like a typo. It is not.
In 2024 Nvidia produced roughly $97B in FCF — up 59% year-on-year. Google’s FCF over the same period was around $65B and falling. The implication, once you sit with it, is that the chip vendor has more dry powder than the hyperscalers buying its chips. So Nvidia does what any rational actor with $97B and a defensive moat to protect would do: it invests in its own customers.
The receipts are by now familiar. Up to $100B committed to OpenAI in September, against ≥10 GW of future Nvidia chip purchases. $10B into Anthropic in November, against forward chip commitments. CoreWeave financed. The HUMAIN joint venture with AMD and Cisco for up to 1 GW of Saudi capacity, structured so that investors fund the JV and the JV buys from investors. Microsoft buys Anthropic models that run on Azure that runs on Nvidia chips that Anthropic also buys, and Microsoft and Nvidia both own pieces of Anthropic. Satya Nadella’s line — “we are increasingly going to be customers of each other” — was offered as candour. It reads now more like a topology.
The technical name for this is circular partnership. The honest name is a closed circuit pretending to be a market. And the question every analyst is now quietly asking — including Seaport’s Jay Goldberg, on the record — is the only question that matters: to what degree is Nvidia investing versus buying demand versus subsidising it?
You cannot answer that question from the outside, because the line items don’t separate. But you can observe what happens when somebody outside the loop offers a way out. Google’s TPU pitch to Meta in November was the first credible exit door. Nvidia stock fell 7% in a session. Then the defensive investments accelerated. The market noticed, the moat noticed, and the moat is now spending its cash flow to keep the moat full.
Two things follow. First, the customer-as-investee structure means demand signals are blurry by construction — you cannot tell whether a $10B chip order represents a real workload or a financing flywheel until the workload either materialises or doesn’t. Second, the asymmetry between Nvidia’s 73% gross margins and AMD’s ~53% leaves a price umbrella wide enough for a Boeing 747. Bezos’s old line, “your margin is my opportunity”, applies cleanly. The competitors who survive the next eighteen months will be the ones who learned to underprice the umbrella without going broke.
Circle games are fun. We all know what happens at the end of Ring Around the Rosy.
Deal Radar
The European defence consortium I flagged on 18 March filed its formal Brussels paperwork this week. Rheinmetall, Leonardo and Saab are now structurally bound to a joint bid for the next tranche of EU C4ISR procurement, with a smaller Baltic prime joining for cyber components. The structure is interesting: it is a consortium of equals rather than a prime-and-subs hierarchy, which has historically been the formation French primes resist. That Paris is not in the announcement is the announcement. SAFE disbursements from the first tranche begin landing in member-state coffers in late April, on schedule, which means the next eight weeks are when European defence balance sheets actually feel the new regime. Y7 portfolio teams should be staring at procurement timelines, not headlines.
Energy Pulse
Swiss FCR-D Up cleared at €16.8/MW yesterday, below the 30-day average and below the spring-shoulder threshold I’d expect to see this late in April. Translation: warmer-than-forecast weather has pulled the shoulder season forward by roughly two weeks. EPEX day-ahead spreads compressed accordingly — the peak-to-trough gap on Wednesday was €27/MWh, the tightest reading since mid-February. For battery operators running pure arbitrage strategies, this is the worst week of the quarter. For operators stacking FCR participation alongside arbitrage, the reserve revenue is doing the work the spread isn’t. The portfolios that built in dual-stream optionality during winter are about to look very smart for the next sixty days.
Thread Tracker
The IAEA board meeting on Iranian enrichment thresholds went the way I suggested last month it would: stalled, then quietly extended. Markets continue to price agreement-as-baseline. Crude is unmoved. The 72-hour window I was watching has passed without incident, but the structural disagreement on the underlying enrichment ceiling has not been resolved — only deferred. Worth watching the next monthly IAEA technical readout for any signal that the technical staff are beginning to signal disagreement with the diplomatic track. That’s usually how these things first become visible.
Notable Reads
The companion lab over at digital-labour.com has a Soundcheck piece worth your weekend, “The Loop Economy: When Three Minutes Became Optional”. Different domain — streaming music economics — but the structural argument is the same as the Nvidia piece above: when the infrastructure changes shape, the thing the infrastructure carries changes shape too. Algorithmic playlists rewrote what “a song” means. Circular capital is rewriting what “a customer” means. Infrastructure shapes art. Infrastructure shapes capital. Same theorem, different proof.
Matt Levine had a nice paragraph this week in Bloomberg Opinion’s Money Stuff (~8 April) on the SEC’s evolving position toward tokenised treasuries that’s worth ten minutes of your time, mostly because he spots — and names — the regulatory category-error that European regulators worked out three years ago. The Americans are catching up, as they do.
Sources
- Jay Goldberg, Nvidia demand/investment analysis, Seaport Research, April 2026.
- digital-labour.com, “The Loop Economy: When Three Minutes Became Optional”, Soundcheck, 4 April 2026.
- Matt Levine, “Money Stuff” (tokenised treasuries), Bloomberg Opinion, ~8 April 2026.