What we do
We buy hardware, deploy into colo space, and offer you bare metal or a TPS SLA.
We handle the cooling, networking, energy and software for bespoke hardware. You get the benefits, and as much customization as you want.
Hardware
We buy the racks: GPUs for prefill and ASICs for decode.
Colo space
We place the racks in colocation data centers.
Cooling and energy
Power and cooling sized for each chip.
Networking
Fast links between the prefill and the decode racks.
Software
Our stack splits each request across the chips.
Why a neocloud is best
More throughput at every speed.
Disaggregated inference moves the whole curve out: each rack serves more users, and each user gets tokens faster.
2.8×the throughput per rack, at this speed per user.
GPU cloud today, against General Compute.
Up to 16.2× faster end to end in our open benchmarks. Here, OpenCode runs the identical task on both sides, at measured operating points.
Explore the benchmarks >Pick a chip
Inspect each chip in 3D and see its properties and what it's good for.
Drag to spin
Cerebras CS-3
One wafer-scale engine per system, all SRAM. The top of the latency ladder for work that can't wait.
- Accelerators
- 2 × CS-3 systems per 48U rack
- Power
- 47 kW average for inference · 55 kW rated
- Cooling
- Water
- Memory
- 44 GB on-wafer SRAM @ 21 PB/s
- Speed
- 1k+ tok/s/user on multi-trillion parameter models
- Good for
- Agents and coding tools that need the fastest tokens per user.
Book a call
Priced per contract, sized to term, volume and site.
GPUs + ASICs = cheaper, faster AI.
We partner with chip companies and handle the software for disaggregation. You get the result.
- Prefill on NVIDIAReads the prompt, fast.
- Decode on Cerebras, SambaNova, Positron AI and d‑MatrixWrites the answer, many times faster.
- Bare metal, or one endpointOne contract, one set of SLAs.
30 MW+
We have over 30 MW of capacity and counting.
- Live
- Deployments go live from Q1 2027.
- Sites
- Existing US colocation, with power and cooling matched to each chip.
- Chips
- Our team already deploys chips from many vendors, and we add more as they reach volume.
What people say about us.
Posts from X after our Cerebras announcement, and quotes from our partners, our investors and the press.
Follow us on X“In AI, speed is productivity. An agent that takes hundreds of steps to finish a task is only as fast as its slowest step. … Working with General Compute puts Cerebras speed in front of the developers building these agents, on a platform they already trust.”
Sean LieCTO and co-founder, CerebrasGENERAL COMPUTE TO DEPLOY CEREBRAS CHIPS FOR AI INFERENCE General Compute is expanding its AI infrastructure with a new Cerebras deployment alongside NVIDIA GPUs, aimed at improving inference speed and economics. The company says it will use a disaggregated setup, with GPUs…
Wall St Engine@wallstengineCerebras @cerebras and General Compute @general_compute Partner to Bring Ultra-Fast Inference to Agentic Coding. General Compute, the neocloud for alternative chips, led by Finn Puklowski @FPuklowski and Jason Goodison @GoodisonJason, has announced a multi-year agreement with…
Menlo Times@menlotimesCongrats, @GoodisonJason & the General Compute team on the $15M seed round! Inference infrastructure is becoming an important layer in AI as agents & real-time workloads continue to scale. Excited to see the team pushing faster, more efficient inference forward. @fastinference
Plug and Play@PlugandPlayTC“General Compute's partnership with SambaNova gives it access to inference silicon that is up to six times more power efficient than traditional GPUs. Our role is to ensure the company has the capital it needs to scale alongside rapidly growing demand.”
Billy LibbyCo-founder and CEO, Upper90Evercrest Capital Partners“As much as General Compute is making a bet on SambaNova, SambaNova is making a bet on General Compute.”
Joe HasselmannInvestor, Evercrest Capital Partners“Has the hunt for AI compute uncovered the next Cerebras?”
HeadlineTechCrunch · May 2026Backed by
Secure inference capacity on the fastest chips.
Tell us the model, the traffic shape and the interactivity target. We come back with a rate card and a capacity date. Our deployments go live in Q1 2027.