Lila Soto: Hugo, hey — rough week to be an Intel shareholder, I think.
Hugo Vance: Mm. Or an AMD one, for that matter.
Lila Soto: So Nvidia dropped full Vera CPU specs on July twenty-first — and Jensen Huang at GTC 2026 called it, I'm quoting directly, 'the world's first processor purpose-built for the age of agentic AI.' That's the framing. And what's underneath that framing is a twenty-billion-dollar revenue projection for this year alone.
Hugo Vance: Twenty billion is the number I'd want to interrogate carefully. That's not a roadmap — analysts are saying it as though the orders are placed.
Lila Soto: And some of them kind of are? OpenAI, Anthropic, SpaceX physically received Vera chips in June — before the spec sheet was even public.
Hugo Vance: Yes. Which is the part that gives me pause. The full spec data — eighty-eight Olympus cores, one-point-two terabytes per second memory bandwidth — that dropped July twenty-first. The chips were already deployed a month earlier. We're evaluating a product that's already running in the wild on the thinnest independent validation.
Lila Soto: Phoronix put out the first public benchmarks — but only on agentic workloads, which is exactly what Nvidia designed for. So, I mean, what does a test on your own terrain actually tell you?
Hugo Vance: Well, that's the crux of it, isn't it. Think of it this way — imagine a restaurant kitchen where the prep station, the stove, and the pass-through are all within arm's reach. That's Vera's monolithic memory design. AMD's EPYC, Intel's Xeon — chiplet architectures — you're sprinting between three separate rooms every time you need an ingredient. The latency is the sprint. Nvidia's claim is: we abolished the sprint.
Lila Soto: Okay, that actually lands. So the one-point-two terabytes per second — that's not a speed boost on top of the old layout. That's a different layout entirely.
Hugo Vance: Correct. And Nvidia's own numbers say forty times lower peak latency versus the EPYC Turin 9755 at ninety percent memory utilization. That's a striking figure. But — and I want to be precise here — that test was run at high utilization on agentic workloads. What nobody has yet published is what happens when a supply-chain analyst runs a warehouse report. A database query on a Tuesday. Something Xeon handles ten thousand times a day without drama.
Lila Soto: Has anyone even tried that?
Hugo Vance: Phoronix went independent — yes — but their results were still scoped to agentic AI workloads. Which is, you see, Nvidia's chosen terrain. The benchmarks that would expose weakness simply haven't been published. And Nvidia controls which tests get designed first.
Lila Soto: Mm — and the fifty percent higher instructions per cycle versus Grace, the one-point-eight times faster task completion versus x86... those are all measured on agentic and reinforcement learning tasks. So what's actually new is kind of narrow, even if it's genuinely new inside that lane.
Hugo Vance: Exactly the right framing. The architectural argument is real — eighty-eight Olympus cores, ARM v9.2, Spatial Multithreading across a hundred and seventy-six threads — that is not marketing. But the proof is scoped to Nvidia's own category. And — I'll say this now because we'll need to come back to it — the twenty-billion-dollar projection only makes sense if Vera isn't really competing against AMD and Intel on open merits. There's something else underneath that number.
Lila Soto: Okay, but that's actually the wrong take circulating out there — the one I keep seeing. 'Nvidia just disrupted Intel and AMD.' Like, that's the headline. And I don't think it's even the right frame, because — who bought Vera first? OpenAI. Anthropic. SpaceX. Those aren't switchers. They've never run a meaningful x86 workload in their lives.
Hugo Vance: No, they have not. And that's the point precisely. Those three organizations are GPU-native infrastructure shops. Vera speaks ARM v9.2 — not x86. For them, there's no migration friction whatsoever. They're already inside the Nvidia ecosystem. This is vertical deepening, not market conquest.
Lila Soto: Which means the twenty-billion number — I mean, does it only make sense if Vera is basically mandatory inside every Rubin GPU rack sale? Like, not winning open RFPs, just... riding along as a line item?
Hugo Vance: That is the more honest read, yes. And I'll concede it freely — if Vera is bundled into Nvidia's AI factory stack alongside Rubin, the twenty billion doesn't require displacing a single Intel Xeon account. It just requires that every Rubin buyer takes the whole rack.
Lila Soto: Which neither Intel nor AMD has an answer to right now.
Hugo Vance: Neither does. And here's where I'd be cautious about the disruption framing — it also assumes broader enterprise never needs to follow. AWS Graviton, Google Axion, they're ARM too, yes. But Nvidia is simultaneously supplying those hyperscalers and competing with them on custom silicon. That tension hasn't resolved. The moment a traditional manufacturer tries to retrofit Vera into an x86 workflow, you run directly into forty years of software certifications, OEM relationships — institutional friction that is not a technical problem.
Lila Soto: So the disruption story is real — just not the one being told. It's not Vera beating AMD EPYC on open merits. It's Vera making the question of open merits kind of... irrelevant, inside the bundle.
Hugo Vance: And that's the question I'm left holding, honestly. Jensen Huang framed Vera at GTC 2026 as a direct challenge to AMD and Intel — his words, not mine. But if that twenty-billion projection unwinds... I mean, if enterprises eventually demand modular sourcing, separate CPU and GPU contracts, the way procurement departments actually work — does the bundle hold? Or does it start to look like the kind of lock-in Intel built, and eventually got punished for, just wearing a different name?
Lila Soto: Yeah. And I don't think Nvidia knows the answer to that yet either.
Hugo Vance: No. I rather think they don't. Good thinking through it with you.
