Nvidia Builds Its Own Brain: Inside the 88-Core Vera CPU and the Olympus Cores

For most of its history, Nvidia has been synonymous with one thing above all else, and that is the graphics processor. Yet the company has spent recent years quietly expanding its ambitions, and its latest move takes it into territory long dominated by others, namely the central processor at the heart of the data center.

The reveal came at Hot Chips 2026, one of the industry's most closely watched technical conferences. There, Nvidia laid out the full details of a new chip called Vera, a data-center processor that signals just how serious the company has become about controlling every single layer of the modern computing stack.

A first for Nvidia

What makes Vera particularly notable is not only its specifications but also its very origin. According to the technical disclosures, Vera marks the first time that Nvidia has designed the actual processor core itself, rather than relying on the existing licensed designs it had used with its earlier Grace chip.

This is a genuinely meaningful shift in the company's strategy. By building the core from the ground up, Nvidia gains far greater control over both performance and efficiency. The processor's instruction set remains compatible with the widely used Arm architecture, keeping it familiar to developers while letting the firm tailor the silicon to its own precise needs.

The numbers behind Vera

At the very heart of the chip sit 88 custom cores, which Nvidia has chosen to name Olympus. Each of these individual cores is capable of executing two hardware threads at once, which brings the grand total to 176 threads. It is a configuration clearly built for handling a great many demanding tasks in parallel.

The physical design of the chip is equally deliberate. Rather than stitching together several smaller chiplets, as has become common practice today, Nvidia opted instead for a single, monolithic compute die housing all 88 of the cores. This approach can simplify communication between the cores and reduce certain kinds of latency inside the processor.

Speed and memory

Moving data quickly is just as important as raw processing power, and here Vera makes some bold claims. According to the reports, the 88 cores are linked together by a second generation of the company's Scalable Coherency Fabric, which reportedly reaches an internal bidirectional bandwidth of up to 3.4 terabytes per second.

The memory setup is rather unusual as well. Vera carries a unified level-three cache of 164 megabytes shared across the entire chip, with each core holding a further two megabytes of its own. For main memory, Nvidia replaces conventional modules with new, replaceable units known as SOCAMM2 that use LPDDR5X, with up to eight of them per processor.

Built for agentic workloads

All of this careful engineering is aimed at one very specific target. According to Nvidia, Vera is designed to serve as the host processor for so-called agentic systems, the increasingly autonomous software agents that carry out complex tasks with only limited human input along the way.

The chip is also central to the company's much broader plans. Reports indicate that Vera is intended to pair with the upcoming Vera-Rubin platform, in which the new processor works alongside Nvidia's next generation of accelerators to form the very backbone of large and highly demanding computing installations.

The performance claims

Nvidia is certainly not shy about the numbers it expects the chip to deliver. According to the company, the Olympus cores provide a 1.5 times increase in the instructions processed per clock cycle, which translates into performance gains of around 50 percent over the standard processors of the kind used widely today.

At a much larger scale, the ambitions grow further still. According to the reporting, a full rack built around the Vera processor packs some 256 liquid-cooled chips and is claimed to deliver up to a sixfold gain in processing throughput, a figure that hints at the sheer density of modern computing facilities.

A challenge to Intel and AMD

Taken together, Vera represents a fairly direct challenge to the companies that have long ruled the server processor market, chiefly Intel and AMD. By designing both its own core and its own memory approach, Nvidia is positioning itself to compete head-on in a segment where it was once merely a paying customer.

Whether Vera ultimately lives up to its lofty promises will only become truly clear once it reaches real data centers. What is already evident, however, is the sheer scale of Nvidia's ambition. The company that first made its name on graphics now wants to shape every single part of the machines that power the next era of computing.