Hope Sterling: August 1st, 2026. Tobias Mann at The Register drops this full technical deep dive on something that — honestly, if you blinked back in May, you probably missed it entirely.
Hope Sterling: Nvidia's new CPU. A real one — not Grace, not the licensed Arm Neoverse thing from before. Something genuinely new called Vera, running on a core called Olympus that Nvidia designed themselves.
Hope Sterling: First time they've ever done that.
Hope Sterling: It was unveiled at GTC Taipei, May 31st — Nvidia called it "the CPU for agents" — and then it kind of… just quietly disappeared from the main feed while everyone was still processing everything else, and then July 21st more architectural detail surfaces, ServeTheHome covers the SPEC CPU 2026 benchmarks, and THEN The Register piece lands and suddenly the full picture is actually there.
Hope Sterling: Eighty-eight Olympus cores. One monolithic die. TSMC 3nm. That's the chip.
Hope Sterling: And Olympus is Arm v9.2-A compatible — so software compatibility isn't the problem — but the core itself is not licensed Arm IP the way Grace was. Nvidia built it.
Hope Sterling: And then the day BEFORE The Register piece — July 31st — Los Alamos National Laboratory goes on record saying their Mission and Vision supercomputers run on Vera, co-designed with Nvidia, claiming up to 7x performance gains.
Hope Sterling: LANL. National lab. Supercomputers. Seven times faster.
Hope Sterling: I need to talk about all of this.
Hope Sterling: Because there's a version of this story that's just a chip announcement — and there's a version where Nvidia is quietly redrawing a line the whole industry thought was fixed.
Hope Sterling: The die is what matters here.
Hope Sterling: All 88 Olympus cores, on one single reticle-sized piece of silicon. One die. TSMC 3nm. The memory and I/O are disaggregated off to the side via CoWoS-R packaging — but the compute itself? Monolithic.
Hope Sterling: That is the OPPOSITE of what AMD EPYC does. EPYC is chiplets — you spread cores across multiple smaller dies, you get better yield, fewer wasted wafers. It's the efficient, sensible, modern thing to do.
Hope Sterling: Nvidia looked at that and went — no.
Hope Sterling: Because with chiplets you get latency variability. One core talks to another core on a different die and there's a tiny, real penalty for that trip. For most workloads, honestly, fine. But agentic AI — and this is where the whole architecture starts to make sense — agentic AI is not most workloads.
Hope Sterling: I mean, we're not talking about just generating a response. Agentic AI is a system that autonomously executes multi-step tasks — code execution, tool use, sandboxing, orchestration — it's making decisions, spinning up environments, running things, checking results. A lot of small things fast, in sequence, and the latency of each handoff matters.
Hope Sterling: That's why Nvidia called Vera 'the CPU for agents.' That's not a vibe. That's literally the architectural brief.
Hope Sterling: And Spatial Multithreading is part of how they serve that — each of Vera's two threads per core gets dedicated, partitioned resources. Not time-sliced. Not shared. One thread runs high-throughput, the other handles lighter work, and they're not fighting over the same resources. Which — for an agent spinning up a sandbox while something else is orchestrating? That matters.
Hope Sterling: Add 1.2 terabytes per second of aggregate memory bandwidth, up to 1.5 TB capacity per socket via LPDDR5X SOCAMM2 — and you have a chip that's genuinely built around a specific vision of what compute looks like when AI is running the show.
Hope Sterling: Now. The benchmark question.
Hope Sterling: Nvidia claims up to 80% faster sandbox environment performance. Eighty percent. That's a number that makes you sit up.
Hope Sterling: But that is a vendor-stated figure. No independent third-party validation. And then SPEC CPU 2026 — which ServeTheHome covered — shows Vera beating dual-socket EPYC by about 3% on SPECrate. Three.
Hope Sterling: So which number is real? Honestly — both, probably? SPEC is a general compute benchmark. Sandboxing is the specific thing Vera is supposedly built for. Those are different tests measuring different things, and Nvidia gets to pick which demo shows their chip well. That's not a scandal, it's just… worth holding.
Hope Sterling: The thing that does make me go — okay, actually — is that Nvidia is deploying Vera in its own EDA workflows. To design future chips. They're eating their own cooking, like, for real.
Hope Sterling: That's not marketing. That's a bet.
Hope Sterling: Look — if the agentic AI thesis is right, if that IS the defining workload of the next era of compute, then the monolithic die philosophy isn't a yield problem to manage around. It's the WHOLE point. If it's wrong… then Nvidia just made a very expensive, very large piece of silicon for a market that didn't materialize quite that way.
Hope Sterling: But none of that is settled yet. Not even close.
Hope Sterling: Fall 2026. That's the first gate. Vera is announced, it's not purchasable — you can't go to a system builder or a cloud partner and get one right now. That window is supposedly coming, but no pricing has been disclosed. Nothing. Which, for a chip that's supposed to reshape what data center compute looks like — that's a pretty significant thing to not know yet.
Hope Sterling: I get it, enterprise pricing is always murky at announcement. But still.
Hope Sterling: The first real-world proof point we have is LANL — Los Alamos National Laboratory — and their Mission and Vision supercomputers, announced July 31st as built on Vera CPUs. Seven times performance gains, co-designed with Nvidia. That's the headline.
Hope Sterling: Co-designed. With Nvidia.
Hope Sterling: Like — do you hear what I'm saying? LANL isn't an arms-length customer who bought a chip and ran their own test. They built the thing TOGETHER with Nvidia. So the 7x figure, as dramatic as it is, is coming out of a partnership, not an independent evaluation. That doesn't make it wrong. But it means we don't yet know what Vera looks like when someone who wasn't in the room during the co-design process just… uses it.
Hope Sterling: That's the validation gap.
Hope Sterling: What we need — and this is the watch — is independent benchmark validation. Not Nvidia's developer blog, not ServeTheHome covering Nvidia's own data drops, not a national lab that co-designed the chip. Someone with no stake in the outcome running Vera against real agentic workloads and publishing the results. The 80% faster sandbox claim especially. That number lives on a vendor slide right now.
Hope Sterling: The EDA thing is interesting — Nvidia running Vera internally for their own chip design workflows, genuinely eating their own cooking — but that's still a controlled environment. It's Nvidia, testing Nvidia's chip, on Nvidia's workloads. Not nothing. But also not the open-market scrutiny that actually answers the question.
Hope Sterling: The story isn't done. It's just — we're getting to the part where the claims actually meet reality. Fall 2026 shipments, Mission and Vision going live, and whoever gets a review unit first and doesn't work for Nvidia… that's the moment. That's when we find out if Vera is the CPU for agents, or just the most confident branding in the data center.
Hope Sterling: But here's what I can't stop sitting with — Nvidia isn't waiting to find out. They're already using Vera. Internally. In their EDA workflows. To design whatever comes AFTER Vera.
Hope Sterling: The chip is designing its own successor. Before it has commercially shipped. Before a single system builder has priced it out. Before anyone outside Nvidia has run an independent benchmark on a real agentic workload. The next Vera is already being built on this Vera, and the rest of us are still waiting for fall 2026 to even find out what it costs.
Hope Sterling: That's either the most confident move in the data center space in a decade — or it's the most expensive assumption anyone has made about where AI is going. Because if the agentic AI thesis scales the way Nvidia is betting it will, Vera becomes the CPU that every infrastructure team has to actively justify NOT buying. That's the lock-in logic. Not "buy this because it's good" — it's "explain to your CTO why you didn't." And if the thesis is wrong? Then Nvidia has been running its own chip design workflows on a very beautiful, very specific answer to a question the market didn't quite ask.
Hope Sterling: The chip is already self-referential. It's just waiting to find out if the world catches up.
