NewIntroducing XPU Grasshopper. Pre-orders are open.
NewIntroducing XPU Grasshopper, an AI chip co-designed with AI for the era of recursive self-improvement

AI silicon
built for discovery

XPU chips run inference and training on a fraction of the energy, built for the era of recursive self-improvement and superintelligence.

Explore XPU GrasshopperJoin the team
Backed byY CombinatorOlive Tree CapitalSoma Capital

Products

XPU Grasshopper (VU47P), today

End-to-end inference live on FPGA (VU47P). Serving API coming soon.

Status as of Sep 2026

XPU Grasshopper

Our XPU, running end to end on FPGA today and heading to silicon. Nearly all the energy in AI goes to moving data, so the XPU keeps it close.

FPGA prototype, in-house. Not generally available.

Explore XPU GrasshopperPre-order

Early access, with the cluster

XPU Frontends

Keep the framework you already use. One import puts your model on the XPU and it runs fast from the first token.

Request early access
 
 
 
import torch
import xpu.torch
 
model = Model()
model.to("xpu")

Soon

Monolith

A cluster of XPUs that behaves as one chip, hosted. Rolling out in batches, soon. Early access ticket holders are the first batch.

Explore Monolith
Language
 
weights = Buffer(WEIGHTS).at(g0.dram[0])
tokens  = Buffer(256, base=0).at(g0.spm[0])
next    = Buffer(4, base=256).at(g0.spm[0])
 
@kernel
def gpt2(x):
    # one hardware loop
    for _ in range(12):
        heads = (
            attention.at(0, 0)
            | attention.at(1, 0)
            | attention.at(0, 1)
            | attention.at(1, 1)
        )
        x = (heads >> mlp.at(0, 0))(
            x, weights
        )
    return argmax(x, next)
 
graph = gpt2.graph(embed(tokens))
kernel = graph.assemble(g0)

XPU Kernel Development

libxpu is a C ABI over the whole machine. You control every buffer and every byte it moves from any language that can call C, and build whatever abstractions you want on top. It is planned to be fully open source.

Request early access

Our thesis

The substrate for all future knowledge

AI is moving from reading the internet to discovering new science, new materials, and new medicine. That is the road to superintelligence, and we build the hardware that takes it there.

Language is spent

Models have read nearly everything humanity has written, and each new round of scale buys less than the last. More text will not produce the next leap.

Models now learn from experience

Models improve by acting, being scored, and learning from the result. Most of that compute goes to generating experience, and each round bakes what was learned into the model itself. That is recursive self-improvement.

Silicon sets the ceiling

Moving data costs far more energy than computing on it, and datacenters are capped by power. So the machine that moves the least data gets the most discovery out of every watt.

Read the full essay

01 / Light

Bend Light

Every chip is written with light, a pattern focused onto a wafer a few atoms at a time. What you draw decides what the machine is good at. On every chip today the arithmetic costs almost nothing, and moving the data costs nearly all of the energy.

02 / Sand

Teaching Sand To Think

Silicon is sand, refined into a single flawless crystal and written with billions of switches. Everything AI has ever done ran on this. It has read nearly everything humanity has written. What it learns next it has to make for itself, as experience, and the price of experience is compute.

03 / Chip

Polymorphic Computing

GPUs were built for graphics and TPUs for tensor math. XPUs are built for frontier AI, with reconfigurable silicon that changes with the work so the hardware keeps up with the models it runs.

End-to-end inference live on FPGA (VU47P). Serving API coming soon.

Explore XPU Grasshopper

04 / Monolith

Monolith

Connecting XPUs creates a bigger XPU. Monolith is a cluster that behaves as one chip and runs the long, stateful work that dominates modern AI, from agents to experience generation to training, on one machine.

Explore Monolith

05 / Factory

AI Discovery Factory

When labor and intelligence are nearly free, the only scarcity left is new knowledge. A discovery factory manufactures it. The AI proposes, a simulator checks, and the result trains the next proposal. Its output is useful trajectories per second, and every one makes the next one better.

Read the essay

06 / Dyson Swarm

Dyson Swarm

Compute is bounded by energy, and the Sun is the most abundant source of energy we know of. A discovery factory in orbit around it turns that light into trajectories, experience, and knowledge that did not exist before. The chips we are building today are the first step on that path.

Join the team

News & Announcements

Careers

Build the future of computing

We're hiring founding engineers across the stack to bring silicon back to Silicon Valley. In person in San Francisco, with a direct hand in the research and the products.

Degrees and titles don't matter here. Send a demo of something you're proud of and we'll talk.

See all open roles
Open roles01
01Founding Silicon EngineerEngineering / HardwareSan Francisco$200K - $300K

Don't see your role? Send a demo of something you've built to careers@zscc.ai

Contact

Get in touch

Partnerships, serving access, investment, or anything else. Headquartered in San Francisco, CA, USA.

General inquiriescontact@zscc.aiCareerscareers@zscc.aiFollow alongXOpen sourceGitHub