On-device AI hardware
A USB stick or embedded chip that runs state-of-the-art AI directly on your device. No cloud, no subscription, no data leaving your machine.
Figures are design targets from simulation. No silicon exists yet. Here's where we actually are.
Turn any device into a local AI powerhouse.
Scroll to connect
Simple, blazingly fast & private.
No more waiting
Code suggestions appear instantly. Inference runs locally on-device at 10,000+ tokens/s with zero network latency.
Fully offline
Works on planes, in secure environments, or anywhere without internet.
Full data privacy
Your code never leaves your machine. No telemetry, no training on your IP.
Why not just a bigger GPU
It is not a software problem. It is architecture. As long as weights live in DRAM and get fetched on every token, there is a hard limit on how fast and how cheap inference can get.
Generating each token means reading the entire set of weights from memory. GPUs are built for parallel throughput, not sequential reads, so even an H100 or B200 tops out around 100-150 tokens per second per user. More FLOPs do not fix a bottleneck that is about moving data, not crunching numbers.
Every token depends on the one before it. Agentic workflows like coding assistants also chain dozens of sequential tasks together, so the wait compounds. You cannot parallelize your way out of a dependency chain: the next word/task literally cannot be computed until the last one is known.
Quantization, mixture-of-experts routing, speculative decoding, and multi-token prediction each shave real latency, and stacked together they can multiply throughput. But they optimize around the same DRAM bottleneck rather than removing it, so gains will plateau.
Wafer-scale and SRAM-resident chips like Cerebras or Groq push past GPUs by keeping weights on-chip instead of in DRAM. That is directionally the right idea, but each machine costs millions, draws kilowatts, and lives in a data center. It is faster cloud, not on-device.
How it's possible
This is our design, validated in simulation, not yet in silicon. The weights live in the chip itself. No RAM, no data movement, no cloud. Just computation at the speed of physics.
Each chip is fabricated for a specific model. There is no general-purpose instruction set, no runtime scheduler, no OS overhead. The architecture of the chip mirrors the architecture of the model, layer by layer, operation by operation.
General purpose inference hardware loads model weights from DRAM on every forward pass, a process that bottlenecks throughput and dominates power draw. Our chips will have no RAM. The weights are embedded directly into the analog compute fabric during manufacturing. There is nothing to fetch, nothing to transfer, and nothing to cache.
Cloud inference adds 50-300 ms of round-trip latency on every call, plus serialization, TLS, load balancers, and queue contention. Running on-device eliminates every layer of that stack. On our chip tokens are generated the moment you make the call. No waiting, no throttling, no quota.
This results in a chip the size of a USB dongle that draws under 5 W, requires no RAM, and outputs over 10,000 tokens per second. Those numbers come from simulation today — proving them in silicon is the work ahead of us.
Why on-device
Cloud inference turned AI into a rented utility. We make it something you own.
The right column reflects design targets, not shipping hardware. And the trade-off is real: weights are baked in at fabrication, so one chip runs one model — a major model update means new silicon. We think that's a fair price for speed, privacy, and ownership.
Sustainability
Every cloud inference call burns lots of energy in a data center on the other side of the world. Let's build a more sustainable alternative.
100x
Less energy
A fraction of the energy used by a general purpose GPU server.
<5W
Idle power
Runs continuously on the same power budget as a USB hard drive.
0
Cloud servers involved
No data center cooling, no long-haul network transmission, no hyperscaler overhead per query.
Hardware Options
Whether you're building the next product or tinkering in your garage, we'll offer a format that fits your workflow.
Bring private, local AI to the laptop or smartphone you already own with a plug-and-play USB stick. It also connects to Raspberry Pi, Arduino, ESP32, and other edge devices for prototyping, robotics, and embedded projects.
Order bare chips in small or large quantities. Drop them into your own PCB designs or prototype boards and start running on-device inference in your hardware.
Access the full chip design files and integrate our inference architecture directly into your own hardware products. Ideal for OEMs and teams building purpose-built AI devices at scale.
Need something else? Get in touch and we'll find a path together.
Transparency and honesty.
Most hardware startups borrow credibility they haven't earned. We'd rather build ours by being straight with you.
Roadmap
No dates. Real chips take years and tens of millions of dollars, and anyone at our stage who gives you an exact timeline is guessing. Here is the honest path from idea to hardware.
The architecture exists and every figure on this site comes from simulation. This website exists to find out whether you want it built.
The design running on programmable hardware. The first real benchmarks measured outside a simulator, published openly.
A test chip from a foundry. This is the expensive step: masks and fabrication runs cost millions. It is what serious capital is for.
USB sticks and maker shields in your hands. We ship nothing before the silicon works. no pre-orders, no exceptions.
Waitlist
Your signup helps us show investors there is real demand for private, fast, on-device AI. Help us prove this should be built.
No deposits. No spam. One email when we launch, nothing more. Your data will never be shared. By joining, you consent to the storage of your email address as described in our Privacy Policy.