Cole Brennan: Cyberpunk. At 1440p. On an ARM laptop.
Malcolm Reeves: That's your lede.
Cole Brennan: That's my lede! A journalist actually ran it at the Computex showcase on an RTX Spark system. And I — look, I don't know why that detail hit me harder than the spec sheet, but it did. Because Jensen Huang said every Windows app ever made would run on this thing and I've heard that promise before.
Malcolm Reeves: Now. The announcement itself — June 1st, Taipei. NVIDIA unveils RTX Spark. The N1X superchip, 128 gigabytes of unified LPDDR5X memory, 300 gigabytes per second bandwidth. Microsoft on stage co-announcing Surface Laptop Ultra as an RTX Spark device. ASUS, Dell, HP, Lenovo, MSI all committed.
Cole Brennan: Whoa — that's basically every major OEM in one morning.
Malcolm Reeves: It is. And that's not a coincidence — that's a platform play. NVIDIA isn't selling a chip, they're trying to own the Windows-on-ARM moment. But the chip itself, the Blackwell GPU with 6,144 CUDA cores, the Grace CPU — it's been shipping in the DGX Spark already.
Cole Brennan: Right — and Jensen's calling it 1 petaflop of FP4 compute. Which sounds massive. But I want to actually pull on that number because I'm not sure it means what they want us to think it means.
Malcolm Reeves: And that's where you have to slow down. Because the petaflop number — that's FP4, peak theoretical throughput. It's the most generous possible framing of what the Tensor Cores can do. But the chip itself, the N1X, it's a 70-billion-transistor part on TSMC's 3nm process, co-developed by NVIDIA and MediaTek. That silicon has been shipping in the Lenovo ThinkStation PGX for roughly 18 months. At $3,999. Running Linux.
Cole Brennan: Wait — the ThinkStation PGX is the same chip?
Malcolm Reeves: Same GB10/N1 silicon. Exact same hardware. What changed on June 1st wasn't the die — it was the Windows packaging, the OEM commitment from six partners, and NVIDIA's CUDA stack repositioned around personal AI agents.
Cole Brennan: So — I mean, the ThinkStation's been out there for a year and a half and nobody noticed. Which is actually wild to me. Because if the silicon works, why didn't that move the needle at all? Is it just... the form factor? Or is this actually about what Windows unlocks?
Malcolm Reeves: That's the distinction that matters. It's a platform play. The chip didn't change — the ecosystem around it did.
Cole Brennan: Which means the thing NVIDIA's actually selling is... the CUDA developer bet. Not the hardware.
Malcolm Reeves: And that's the bad take I want to name. Because there are two of them floating around right now and they're both wrong. First one — gaming. The Cyberpunk 2077 at 1440p demo. People are reading that as proof RTX Spark is a gaming machine. It's not. That demo is compatibility insurance. HP's OmniBook Ultra 16, the Dell XPS with the N1X — those were pitched explicitly for local AI workloads. Not games.
Cole Brennan: But — I mean, a journalist actually ran it. At the show. That's not nothing.
Malcolm Reeves: Compatibility is not performance. Windows-on-ARM x86 translation has a mixed track record — Qualcomm helped mature it, yes, but running Cyberpunk at 1440p and running it *well* are different claims. NVIDIA never published a frame rate.
Cole Brennan: Okay, fair. So then what number actually matters? Because if 1 petaflop FP4 is — wait, @SebAaltonen was pointing out on X that Apple and AMD already have bandwidth edges and this is basically year-old-and-a-half architecture. So what do I tell someone who asks what the real metric is?
Malcolm Reeves: 128 gigabytes. The unified memory pool. That's it. That's the structurally significant number. Discrete GPUs — even fast ones with GDDR7 — top out well below that for VRAM. And 300 GB/s is lower bandwidth than dedicated GDDR7, yes, but the ceiling on *model size* is what's been killing local AI workloads.
Cole Brennan: So the actual use case is — it's Tuesday morning, a developer opens a laptop, loads a 70 billion parameter model locally, and just... doesn't hit the wall.
Cole Brennan: And we just... don't know yet. I mean, that's the honest answer. RTX Spark laptops aren't hitting shelves until Q4 2026, earliest. Some of the OEM systems are sliding into early 2027. There are no independent benchmarks. None. So everything we're talking about — the CUDA ecosystem bet, the unified memory ceiling, whether it actually beats Apple Silicon M5 in a real local AI workload — that's all still theoretical.
Malcolm Reeves: And that's precisely where I'd sharpen the question. Because the chip story — that's almost settled. The silicon is documented, the ThinkStation PGX proves it runs. The open question is structural. Can NVIDIA make Windows-on-ARM the environment where developers actually want to run local AI agents? Not can they run it — can they make developers *choose* it over what Apple's already shipping now.
Cole Brennan: Right — and AMD Strix Halo's already in market with its own bandwidth advantages. So when Q4 finally comes and those benchmarks land... is NVIDIA's CUDA ecosystem enough to close that gap, or does the repackaged silicon catch up with them?
