AINEWS
Nvidia's first custom server CPU in a decade arrives as the chip wars move from GPUs to the silicon that runs AI factories, and the Intel/AMD compatible (x86) software moat is the real fight.
Nvidia just detailed its first custom-designed server CPU in nearly a decade, and the framing is a direct challenge to the Intel and AMD duopoly that has defined the datacenter for twenty years. The chip, called Vera, is built on a custom core called Olympus, paired with a memory subsystem, LPDDR5X, that is unusual for servers.
For most of the AI era, Nvidia paired its GPUs with CPUs it licensed from Arm, then sold the pair as an integrated accelerator. Vera changes the contract: the company is no longer renting the CPU, it is designing it, and treating the chip as a battleground product in its own right.
Vera is the first Nvidia-designed datacenter core since the Denver and Carmel cores of the Tegra mobile era, and the company has been unusually direct about how different it is from anything x86 (the Intel and AMD instruction set that has dominated servers since the 2000s) has shipped.
The Olympus core is wide and aggressive. It uses a 10-wide decode front end, the part of the chip that breaks instructions into pieces the processor can run, and prefetches data based on patterns like the graph structures that show up in AI workloads. Eighty-eight Olympus cores sit on a single monolithic compute die, a chip built as one piece of silicon, and they run 176 threads through a partitioned scheme Nvidia calls Spatial Multithreading. Spatial Multithreading divides the core's execution resources into independent lanes rather than sharing them opportunistically as traditional Simultaneous Multithreading (SMT) does.
The memory subsystem is where Vera makes its most unconventional choice. LPDDR5X is a type of low-power memory typically used in laptops. Nvidia hardened it for datacenter use with ECC, error-correcting code that catches silent data corruption, and full telemetry, and claims the result delivers up to 1.2 TB/s of bandwidth, roughly 3x the memory bandwidth per core and about 5x the bandwidth per watt of conventional DDR-based server designs. The third-party benchmarking outlet ServeTheHome has begun running SPEC CPU 2026, the industry-standard server benchmark suite, against Vera; that independent lane is the only way to verify whether the marketing numbers hold outside Nvidia's own test conditions.
Nvidia is shipping Vera in two forms, and the larger one signals how seriously the company is taking the CPU. A dense liquid-cooled rack packs 256 CPUs and more than 22,000 cores. A conventional air-cooled 2U server, the standard two-unit rack chassis, holds two sockets. Dell has committed to multiple PowerEdge systems based on Vera.
The compute die is one piece of silicon connected by a second-generation scalable coherency fabric, the on-chip network that keeps multiple chips in sync. Chiplets, smaller silicon tiles that are usually stitched together inside a single processor package, are used only for memory controllers and I/O. Nvidia measures bisection bandwidth across the die, the worst-case throughput between any two halves of the chip, at roughly 3.4 terabytes per second.
The headline performance claims are aggressive: roughly 2x faster Olympus core performance, 3x the core-to-core bandwidth of chiplet-based competition, and 40% lower memory latency under load, per Nvidia. The CNBC framing is the standard market read: AMD and Intel now have a fully custom Nvidia core in the room, on a roadmap that sits beneath the Rubin GPU platform.
The harder question, raised by analyst Ryan Shrout's column for Signal65 syndicated through TechSpot, is whether custom cores alone decide datacenter outcomes. Two decades of x86 software, every Linux distribution, every Java Virtual Machine, every database engine tuned for Intel and AMD, does not move because a new chip arrives. Vera ships on the Rubin roadmap, so the question for buyers is not whether the silicon is fast, it is whether the software stack, including the agentic AI workloads Nvidia is positioning the chip for, runs well enough on a non-x86 architecture to justify swapping a known commodity for a faster unknown.
Vera is the first datapoint that Nvidia now owns its CPU the way it owns its GPU, as part of a designed-together AI factory stack. The CPU has stopped being a rental and started being a product.