Sequoia-backed Nuvacore is designing a new CPU for AI workloads and data centers

The fight over AI infrastructure is spreading beyond GPUs: startup Nuvacore wants to prove that the next bottleneck in data centers is the CPU. The company, led by former Apple architect Gerard Williams, is betting on a processor designed from the start for sustained, heavy AI workloads rather than adapted to them after the fact. According to Bloomberg, Nuvacore has already secured backing from Sequoia Capital and is closing its seed round, after which it expects to move to a Series A in the coming months. The startup is also opening offices in Silicon Valley, Austin and Markham, Canada, all major hubs of semiconductor expertise.

The company has only just emerged from stealth, but it states its ambition bluntly: to build a stronger central processor for the AI era. Nuvacore argues that GPU dominance alone is no longer enough. GPUs remain the main compute engine for training and inference, but CPUs manage the data flow, run system software, and coordinate memory, networking and the accelerators themselves. If that layer becomes a bottleneck, the overall efficiency of an AI cluster drops.

Nuvacore describes its product as a general-purpose CPU core built from scratch for data centers and AI infrastructure. Its main goals are maximum performance and efficient use of die area. The company stresses that it is targeting not just ordinary server tasks but long-running, continuous and compute-intensive workloads, including advanced AI systems and agentic scenarios. That emphasis matters: the market increasingly looks past peak benchmarks to how hardware behaves under constant load, where speed is only part of the picture alongside energy efficiency, rack density and total cost of ownership.

The team behind the project carries real credibility. Gerard Williams is one of the best-known CPU architects of his generation: at Apple, his name is associated with the development of high-performance ARM cores from the A7 era to the line that eventually led to the M1 family of chips. Nuvacore was co-founded with him by John Bruno and Ram Srinivasan, veterans of Apple, Nuvia and Qualcomm. The same group previously launched Nuvia, a server-processor startup that Qualcomm bought in 2021 for about $1.4 billion. The new venture suggests the team again wants to play at the level of core compute architecture, not just slot into other companies' roadmaps.

Against this backdrop, a bet on the CPU no longer looks like a niche idea but part of a broader shift. The largest cloud players and hyperscalers have long built their own silicon for specific workloads, because general-purpose products do not always offer the best balance of performance, power consumption and cost. GPUs usually get the attention, while CPUs are seen as a mature and almost exhausted category. Nuvacore challenges exactly that view: if AI services become persistent, autonomous and mass-market, the central processor should also be designed for that mode from day one.

The company has not yet disclosed key technical details, including its chosen instruction set architecture, and has not finally decided whether it will sell finished chips or license its design. But the first focus is already clear: data centers, where mistakes are costly and even a small efficiency gain quickly turns into big money.

The takeaway is simple: Nuvacore is not just another "AI" startup in the broad sense, but an attempt to redefine the CPU's role in the AI stack. A finished product is still far off, and the risk of such a project is enormous. But its very emergence shows that the next race in AI infrastructure will be fought not only over GPUs and accelerators, but also over which processor runs the whole system.