"Black Myth" Didn't Arrive; NVIDIA Wants to Tell a New Story with the Debut Sale of the Vera CPU

Whenever I see hardware manufacturers talk about the “future,” what catches my attention first is not the specs, but who they actually want to cast as the protagonist. This time, NVIDIA unveiled the Vera CPU at GTC a few days ago, and my initial impression was very clear: it is not here to pick a direct fight with traditional CPUs, but rather to fill a role that has long been overlooked in the AI era.

In terms of positioning, Vera’s role is actually quite distinct. NVIDIA itself said this is the first CPU product it has ever sold externally, and the goal is not to “kill the CPU.” Jensen Huang also made it clear after the event that NVIDIA is not anti-CPU. I can actually understand that way of putting it—it feels like someone in a relationship finally admitting: you are not here to replace anyone, but to rewrite what collaboration looks like.

This CPU is defined very directly—as something built to serve AI agents and reinforcement learning. According to Jensen Huang, it is twice as efficient as traditional CPUs and 50% faster. At the same time, NVIDIA explained why it decided to build its own: over the years, it has been buying the very best, even the most expensive CPUs, because once the CPU falls behind, it drags down GPU systems worth millions of dollars. Put plainly, the CPU used to be like that old teammate in the squad—important, but always stalling at critical moments, leaving everyone both anxious and frustrated.

What impressed me is that Vera’s design philosophy is not exactly the same as the traditional path. NVIDIA believes that simply pushing CPUs forward by piling on more transistors is close to reaching its limits. So Vera does not focus solely on “the more cores, the better,” but places greater emphasis on single-thread and IO performance. That is actually crucial, because in AI training and agent workloads, what often hurts is not a lack of theoretical compute power on paper, but poor coordination across the entire system.

In terms of specs, Vera has 88 cores and supports hyper-threading, LPDDR5X memory bandwidth reaches 1.2TB/s, and NVLink-C2C bandwidth reaches 1.8TB/s, with the latter described as 7 times that of PCIe 6.0. These numbers are certainly impressive, and enough to show that this is not a CPU aimed at the mainstream consumer market, but one targeted squarely at high-end AI infrastructure.

But what pulls me out of the narrative is that the single-thread performance—the factor that most truly determines whether this “role” can hold up—still has no credible data behind it. It is like a character trailer that looks thrilling and delivers great lines, but once the full feature begins, whether she can really carry the scene still depends on actual performance. Right now, Vera’s setup is in place and the emotional buildup is there, but the most critical step has yet to land.

If viewed from a broader trend, NVIDIA’s move looks more like an effort to complete its control over the AI server narrative. In the past, CPUs were largely the home turf of AMD and Intel. Now NVIDIA is bringing CPU, GPU, and interconnect together in its own hands, and its influence over future full-system platforms will only grow stronger. For ordinary users, there is no need to rush into asking “what can I buy” in the short term, because this is not a product for gaming PCs. But if you care about where the industry is heading, this is something worth remembering: it could shape how future AI compute platforms are built, who gets to lead them, and how traditional CPU makers respond next.

My own judgment is that Vera feels like a new character who enters the stage already knowing exactly how to speak—clearly positioned, emotionally compelling, and indeed striking at one of NVIDIA’s real pain points. But whether she can truly develop that arc will depend not on how lively the launch event was, but on whether the follow-up performance—especially the single-thread data—can actually carry the story forward.