AI inference beyond NVIDIA
Make any AI chip dominant.
We write the software that runs today's biggest open AI models fast on chips other than NVIDIA's. We started with Intel Gaudi2.
The problem
Good chips. Software that can't keep up.
New AI models arrive every few months. Chip makers' software often falls behind.
On Gaudi2, Intel's own runtime couldn't run new models like DeepSeek V4.1 or GLM‑5.3 at all. Powerful hardware, sitting idle.
What we did
So we rebuilt the stack from the silicon up.
We wrote new code for every engine on the chip, one model at a time.
As it got faster, we replaced the vendor's runtime piece by piece. Today none of it is left.
The result
Faster, with the exact same answers.
Every change we make is checked word by word against a reference. If a single word differs, the change doesn't ship.
Unlocked
Unlimited inference for every agent you run.
Behind the gate is our runtime. Open models run fast and exact on hardware you choose, with room for as many agents as your work needs.
Two ways to use it
Use our cloud, or run it on your own servers.
Call open models we host
Send requests to models running on our own Gaudi2 servers.
Our stack, your machines
We install, tune and support it. Your data never leaves the building.
Hardware
Gaudi2 today. More chips next.
The method carries over to any accelerator: build a reference, write the kernels, prove they match.
- Intel Gaudi2 Running
- Intel Gaudi 3 Next
- Tenstorrent Next
- AMD Instinct Planned
- Qualcomm Cloud AI Planned
- Arm Planned
Talk to us
Run frontier AI on the chips you choose.
Tell us what hardware you have, or plan to buy. We'll tell you what it can do.
hello@dominantinference.com