Zettascale Computing

Market position

Zettascale positions itself as an alternative to general-purpose AI accelerators, arguing that language-data scaling is exhausted and that hardware optimized for experience generation and recursive self-improvement will determine the pace of progress. It cites backing from Y Combinator, Soma Capital and Olive Tree Capital.

Rather than scaling conventional accelerators, the company builds an architecture intended to reduce energy spent on data movement and to unify dense and sparse computation with the precisions used in production, extending to clusters that present as a single chip.

Technology

The XPU architecture is built around minimizing data movement, which the company identifies as the dominant energy cost in AI workloads, on the premise that datacenter capacity is limited by power. The design targets dense math and sparse models with the numeric precisions used in practice within a single architecture. The first XPU, Grasshopper, is prototyped on an FPGA; Monolith connects multiple XPUs so that the cluster behaves as one larger chip, supporting agents, experience generation and training. The company also maintains a public GitHub presence for open-source work.

Go-to-market

Direct engagement through the company website for partnerships, early access and investment inquiries, supported by published thesis essays and announcements; hiring is conducted via open application with a request for demos of prior work rather than credentials.

Operators of AI training and inference infrastructure, including organizations running agentic workloads, experience generation and large-scale model training, and applications in scientific, materials and medical discovery.

Geography

Headquartered in San Francisco, California, United States, with in-person work required for engineering roles.

History

The company adopted the name Zettascale in an announcement dated 2025-09-04, published together with a progress update. It subsequently published an essay, 'The Final Frontier of Intelligence,' dated 2026-06-01, setting out its view that LLMs are plateauing and that the next phase of AI requires different hardware.