Eliza Ward: Hey. Good to be back.
Brian Reed: Yeah, good to be back. You saw the NVIDIA thing this morning?
Eliza Ward: I did. Okay — here's what actually happened. August 10th and 11th, NVIDIA announced strategic partnerships with six Wall Street firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — targeting five hundred billion dollars in third-party capital to finance AI compute infrastructure.
Brian Reed: Five hundred billion.
Eliza Ward: Five hundred billion. And the headline lands huge. But wait — none of these firms have committed actual capital. They signed MOUs. Memorandums of understanding. Which is — I mean, it's basically six banks agreeing in writing that they'd like to build a toll road together. Nobody's broken ground. Nobody's wired money.
Brian Reed: So the toll road exists on paper, and the paper is what got announced.
Eliza Ward: That's the confirmed part, yeah. No specific financial terms, no individual firm commitment amounts, no deployment timetable disclosed. That's directly from NVIDIA. And Jensen Huang simultaneously posted on X framing AI compute as — his phrase — an 'investable asset class.' NVIDIA has moved, in his words, from selling chips to creating 'AI factories.'
Brian Reed: And David Solomon was physically onstage with Huang for this. Goldman Sachs wanted to be seen at the announcement even with no committed terms.
Eliza Ward: Right — but the part that actually breaks the clean version is Reuters. Because Reuters reported that Huang indicated NVIDIA could backstop up to a hundred and twenty-five billion of potential deals. Which is roughly twenty-five percent of the five hundred billion target. That's not in the press release. NVIDIA hasn't confirmed it applies to every transaction.
Brian Reed: Hold on. If that's real — if NVIDIA is absorbing a quarter of the default risk — that's not a financing innovation. That's NVIDIA not believing third-party capital will actually show up at market rates.
Eliza Ward: That's the inference, yeah. But we don't have confirmation. So hold that — reported, not confirmed.
Brian Reed: Okay, so KKR's co-CEOs, Joe Bae and Scott Nuttall, put out a joint statement on day one. 'Delivery, not ambition, is the hard part.' That's not a victory lap. That's a hedge. On the same day they signed the MOU.
Eliza Ward: That line is doing a lot of work.
Brian Reed: It is. And then Apollo's Jim Zelter calls compute 'a scarce, mission-critical asset class.' Which is — I mean, that's the bull case in four words. So you've got two partners in the same deal, same week, pointing in genuinely opposite directions. Zelter's saying this is the new utilities play. Bae and Nuttall are saying wait, we've seen infrastructure promises before. Neither of them is wrong to say what they said. That's actually the problem.
Eliza Ward: And the press release — NVIDIA's own language — is 'dedicated pools of capital at significant scale at attractive rates.' Which tells you the pitch. Countries, governments, enterprises, startups. But attractive to whom at what yield? Analysts are putting required investor returns at eleven to seventeen percent. That's not utility pricing. Utilities clear at four to six.
Brian Reed: So the actual new thing here — not the five hundred billion number, that's the headline — it's that NVIDIA is attempting to get capital markets to treat AI compute the way they price infrastructure. Toll roads. And the market is saying, we heard you, but we're pricing it like junk. The MOU is NVIDIA getting six logos on a stage. The yield spread is the market's actual answer.
Eliza Ward: And that spread — eleven to seventeen percent versus four to six — that's actually where the toll-road analogy collapses. Because if Apollo and BlackRock and the rest genuinely believed chips behave like toll roads, they'd price them like toll roads.
Brian Reed: No, I don't buy that they believe it.
Eliza Ward: Right — so that's the take circulating that I think is wrong, or at least unproven: that this is straightforwardly infrastructure financing. Jensen Huang's framing it that way. 'AI factories.' Utilities. Toll roads. But a toll road doesn't become functionally obsolete when a faster road gets built. A GPU does.
Brian Reed: Hang on, though. Couldn't the yield premium just be pricing power and regulatory risk? Not hardware depreciation specifically?
Eliza Ward: Maybe — actually, that's NVIDIA's answer, sort of. They're pointing at CUDA. The argument is that software updates, the whole CUDA ecosystem, keep chips productive and valuable indefinitely. So the asset doesn't depreciate, it appreciates. But that's NVIDIA's claim about NVIDIA's own platform. Eleven to seventeen percent says sophisticated lenders aren't buying it.
Brian Reed: So Brookfield is pricing in obsolescence that Jensen Huang says can't happen.
Eliza Ward: That's the contradiction sitting right there in the deal structure. And nobody's named which assumption is actually driving the mechanics — whether it's the software-appreciation story or the junk-bond risk story. Think of a pension fund CIO in Toronto, Wednesday afternoon, 2027, and her board is asking: does this go in the infrastructure sleeve or the high-yield sleeve? The deal's architects haven't answered that.
Brian Reed: And the part I keep thinking about — we haven't even touched the thing that makes all of this worse if the collateral starts moving.
Eliza Ward: China. Yeah. That's coming.
Brian Reed: Right — China. And the thing is, nobody in this deal is saying it out loud. Not NVIDIA, not Apollo, not Brookfield, none of the six. If low-cost Chinese compute floods the market, the whole 'scarce, mission-critical asset class' framing that Jim Zelter put his name on — that evaporates. The loans are backed by commodities, not infrastructure.
Eliza Ward: NVIDIA made no public statement on China competitive risk. That's a confirmed gap.
Brian Reed: And nobody — not one of the six partners — has put a number on how likely or how near-term that price shock actually is. So picture a loan officer at, I don't know, a sovereign wealth fund, Friday afternoon, collateral review, and she's asking: what's my floor on GPU value if Beijing decides to subsidize compute exports? There's no answer in the deal documents.
Eliza Ward: Wait — and that's the sourcing problem too. The China risk comes from analysts flagging it. Not from, like, a disclosed stress test. We genuinely don't know the timeline.
Brian Reed: Right. So we should say that plainly — we don't know how near-term this is. Nobody in the deal has said.
Eliza Ward: But then — okay, even setting China aside — the five hundred billion target includes power and cooling systems. Energy infrastructure is baked into the structure. And the grid can't keep up. That's not a capital markets problem. Analysts are flagging grid and power bottlenecks as a binding constraint on AI data-center expansion that exists whether or not the financing closes.
Brian Reed: And that's — yeah, that's the part that actually surprises me. You can finance chips. You cannot finance your way around grid permitting. Capital doesn't fast-track transmission buildout.
Eliza Ward: It's like — actually, the analogy is almost too clean — it's like financing a toll road when the state hasn't approved the land. KKR said delivery is the hard part. The grid is delivery.
Brian Reed: So the two things to watch — and I mean watch specifically, not just 'monitor the situation' — are whether any Chinese compute price move compresses what these loans are actually backed by, and whether power permitting timelines slip relative to deployment schedules. Either one of those moves, and the collateral question gets very real, very fast.
Eliza Ward: And that's — I mean, that's where I actually land. The whole structure rests on one thing NVIDIA hasn't proved yet. Not in a press release. In actual collateral performance, across a full hardware cycle. If CUDA and the software updates genuinely preserve chip value the way Jensen Huang says — not just in benchmarks, permanently, in practice — then this reshape how capital markets price AI compute. Permanently. That's not nothing. But if the first generation of loans hits a hardware refresh and collateral values move, the default cascade runs faster than any toll road or power plant would, because chips aren't power plants. They're electronics with a refresh cycle.
Brian Reed: And we genuinely don't know yet. August 10th, 2026 is day one of — I don't know, a five-year experiment? We don't have a hardware cycle of data. Nobody does.
Eliza Ward: No confirmed stress test. No disclosed timetable. The MOUs don't answer it. The Reuters backstop claim — unconfirmed. So the question that would actually settle this isn't one more announcement. It's whether Blackstone or Brookfield disclose real collateral terms on an actual funded deal. That's the signal.
Brian Reed: Yeah. And until then — wait, actually, that's the only honest place to stop, isn't it. We don't know.
