Zettascale

Zettascale builds AI silicon and XPU chips for discovery-oriented workloads, with lower energy use, mixed dense and irregular compute, and prototype-to-cluster scale support.

AI Developer Tools Other

Visit Website

Overview

Zettascale is building AI silicon, including XPU chips, for discovery-focused workloads. The company positions its hardware as infrastructure for AI systems that do more than generate text: they propose ideas, simulate them, test candidates against reality, and use the results to improve the next cycle.

The site presents Grasshopper as the first XPU, currently prototyped on an FPGA, and Monolith as a cluster in development that connects XPUs into a larger machine. Across the public pages, the product is framed around energy-efficient reconfigurable dataflow chips for AI training and inference, with emphasis on minimizing data movement, supporting dense and irregular compute, and spanning a wide precision range.

Core capabilities

Energy-efficient AI hardware

The company says its XPU chips are built to run every AI workload on a fraction of the energy, with a focus on discovery-oriented compute rather than only text generation.

Data-movement reduction

Grasshopper is introduced as an FPGA prototype of the first XPU, designed to minimize data movement, which the site identifies as a major energy cost in AI.

Mixed workload support

The architecture is described as supporting dense and irregular workloads in one machine, so it can handle both matrix-heavy compute and sparse, conditional execution.

Broad precision range

The thesis page says the hardware should cover a precision range from fp8 to fp64, allowing low-precision training as well as higher-precision scientific computation.

Cluster-scale discovery workflow

Monolith is presented as a cluster that behaves as one chip, intended to support training, simulation, and verification together in a discovery loop.

Reconfigurable dataflow design

The public pages frame the product around reconfigurable dataflow chips, indicating an architecture that can be adapted across AI training and inference workloads.

Where it fits

Discovery-oriented AI research
Use when a team is building AI systems that need to propose candidates, simulate outcomes, and verify results in one discovery loop rather than only serving text responses.
Mixed-compute AI workloads
Use for workloads that mix dense linear algebra with sparse or irregular execution, where a single fixed accelerator may be too narrow and a general GPU may waste capacity.
Energy-conscious training and inference
Use in projects where energy use matters and data movement is a significant cost, especially if the workload benefits from a dataflow-oriented design.
Scientific and simulation-heavy work
Use when a research group needs higher-precision computation in parts of the workflow, such as scientific or simulation-heavy steps that need more than low-precision tensor math.
Clustered discovery pipelines
Use for team-scale systems that need to connect multiple XPUs into one logical machine for training, simulation, and verification.

Pros and Cons

Pros

Targets both dense math and irregular execution instead of optimizing for only one workload pattern.
Supports a stated precision range from fp8 to fp64, which is relevant for both training and scientific workloads.
Frames the system around reduced data movement, a key energy cost in AI hardware.
Extends from a single XPU concept to cluster-scale operation with Monolith behaving as one chip.

Cons

The public site does not provide detailed setup documentation, integrations, or platform compatibility information.
Pricing, plan structure, and purchasing terms are not published on the visible pages.
The hardware is described in development terms, so some product capabilities appear to be roadmap rather than shipping features.

FAQ

What does Zettascale do?

Zettascale is building energy-efficient reconfigurable dataflow chips for AI training and inference. The site positions the hardware for discovery-oriented AI workloads rather than general-purpose consumer apps.

What is Grasshopper?

The site describes Grasshopper as the first XPU, prototyped on an FPGA. It is presented as a way to minimize data movement, which the company says is a dominant energy cost in AI.

What is Monolith used for?

Monolith is described as a cluster in development that connects multiple XPUs so they behave as one chip. The intended workflow is the discovery loop: training, simulation, and verification in a single machine.

How is Zettascale priced?

The public pages do not show product pricing, plan tiers, or a self-serve checkout flow. The pricing URL appears to function as a company page rather than a product pricing page.

What setup or integrations are supported?

The source emphasizes hardware for dense math, a full precision range from fp8 to fp64, and irregular execution. It does not document software integrations, setup steps, or platform compatibility details.