Ameka — the AI computer and AI operating system.

Ameka is the world's first AI-native computer. The Ameka 01 is a custom AI workstation built on a new neural-fabric substrate: a 3 nm SoC with 128 tensor tiles, 2 terabytes of unified HBM3e memory, 8 TB/s bandwidth, and 4 PFLOPS of FP8 compute — all wrapped in a silent, fanless, monolithic heatsink chassis. Ameka ships with Ameka OS, the first AI operating system designed end-to-end by AI: models are first-class processes, memory is content-addressed, every inference is signed by a silicon root of trust, and tool-use is a kernel primitive. Ameka is built for local LLM inference, on-device reasoning, agent runtimes, multi-agent systems, private AI development, and frontier research. Reserve the Ameka 01 online with a refundable deposit from $100. Ships from Pasadena, California, beginning Q3 2027.

Reservations open · Ameka 01

A new
computer.

Ameka is the first computer designed for a mind that isn't human. New silicon. New memory. A new operating system, conceived end-to-end by AI — for the people building the future with it.

See the machine
LIVE · BENCH 04 AMEKA 01 / PROTOTYPE
34°N · 118°W NEURAL FABRIC · v0.9.2
01 — Why

For forty years we built
computers for people.
Then we asked them
to host another mind.
It deserves its own.

A general-purpose PC was never the right shape for what comes next. We pried a thinking machine into a chassis built for a typewriter, a browser, and a calculator. Then we sold the latency back to you.

Ameka starts from the model down. The substrate, the memory lattice, the cooling, the kernel, the OS — re-imagined as one continuous surface for inference, reasoning, and agency. Not retrofitted. Re-thought.

02 — The Machine

One artifact. Eight ways of seeing it.

01 · Form
The chassis is the heatsink.
A single milled volume of nickel-coated copper, vapor-chambered end to end. No fans. Silent under sustained 800-watt load.
02 · Memory
1 TB of unified memory.
HBM3e stacked directly above compute, with a coherent DDR5 spillover pool. A 70 B model fits and stays resident.
03 · Substrate
Neural fabric. Not GPUs.
A bespoke 3 nm SoC with 128 tensor tiles wired by an on-die optical interconnect.
04 · Power
Quiet, by design.
Two-phase immersion option for studios. 240 V single-line draw. 96% efficient.
05 · I/O
Six lanes of light.
Three 800 GbE ports, three optical breakouts. Cluster two Amekas with a single fiber.
03 — Architecture

Four pillars. One continuous substrate.

P / 01 Compute

Neural fabric

A 3 nm tensor lattice with 128 cores, optical interconnect, and deterministic latency end to end. Inference, training, and reasoning collapse onto one substrate.

4.0 PFLOPS · FP8
P / 02 Memory

One address space

1 TB HBM3e fused above compute with a coherent DDR5 spillover pool. No PCIe wall, no host/device boundary, no quantization-of-convenience. A frontier model loads once and stays resident.

Up to 6 TB/s bandwidth
P / 03 Operating System

Ameka OS

An AI-native kernel and userland written end-to-end by reasoning models. Models are first-class processes. Memory is content-addressed. Privacy is enforced at the silicon root.

Open SDK
P / 04 Privacy

Local by default

Every inference runs in a hardware-attested enclave. No telemetry leaves the device unless you sign a request. The model on your desk forgets nothing — and tells no one.

Confidential compute
04 — Engineering targets

The numbers, honest.

Ameka 01 is in development. The figures below are Phase 1 engineering targets for the first production silicon — within reach of the same 3 nm process node that ships in 2026, validated against current bring-up benches and the published roadmaps of our silicon partners. Final units are tuned at tape-out. Reservations help us pay for it.

4.0PFLOPS
Peak AI compute
FP8 dense / FP4 sparse across 128 tensor tiles. Target on TSMC N3.
1TB
Unified memory
512 GB HBM3e die-stacked + 512 GB coherent DDR5 pool.
6TB/s
Peak memory bandwidth
Out of HBM. Sustained on 70 B-class workloads.
128tiles
Tensor cores
On-die NoC at 1.2 Tb/s per lane. 12 lanes.
<2s
Model load time
Llama-class 70 B from cold cache to first token.
800W
Sustained draw
Single 240 V line. 96%-efficient PSU. Phase-change cooled.
1Mctx
Context window
Streaming KV cache + paged attention in the OS.
<25dB
Acoustic floor
Whisper-quiet. Heatsink is the chassis. No moving parts in the base config.

On the numbers: Ameka 01 specifications represent Phase 1 engineering targets for production silicon expected to tape out in 2027 and ship from 2027–2028. Reservations are refundable. As we progress through the bench, we will publish a public delta against these targets in Bench Notes. Targets validated against TSMC N3 PDK, Micron HBM3e roadmaps, and the published performance envelopes of NVIDIA Blackwell, Tenstorrent Blackhole, and Cerebras WSE-3.

05 — Operating System

An operating system written by AI, for the people who command it.

Ameka OS was conceived, scaffolded, and stress-tested end-to-end by a fleet of frontier reasoning models. Then ported, audited, and hardened with us. Every primitive — process, file, network, identity — was redesigned for a world where agents are first-class citizens.

First-class
Models are processes. They schedule, log, and inherit identity like any other.
Content-addressed
Files are hashed. The OS never re-reads what it already knows.
Attested
Every model invocation is signed by the silicon root of trust.
Composable
Tool-use, MCP, and agent runtimes are kernel primitives — not bolted on.
06 — Reservations

Choose your Ameka.

A refundable deposit secures your place in the first production batch. Configure the rest closer to delivery. Founders receive a numbered unit, signed firmware, and standing invitations to the bench.

Ameka Core

For builders bringing a frontier model to a single workstation.

  • 2.0 PFLOPS · 1 TB unified memory
  • Single-tile neural fabric
  • Ameka OS · open SDK
  • Numbered founder unit · signed firmware
From $12,500 $100 deposit
Est. delivery Q4 2027
Ameka Pro

Studios and small teams running production agents around the clock.

  • 4.0 PFLOPS · 2 TB unified memory
  • Full neural fabric · 128 tiles
  • Ameka OS · multi-tenant runtime
  • Priority queue position · founder kit
  • White-glove install at your studio
From $28,000 $500 deposit
Est. delivery Q3 2027
Ameka Architect

Labs and research collectives pushing past the frontier.

  • Two-node cluster · 8 PFLOPS combined
  • 4 TB unified memory · optical fabric link
  • Two-phase immersion option
  • Direct line to the Ameka bench team
  • Roadmap voting · early kernel preview
From $48,000 $1,000 deposit
Est. delivery 2028

A new kind of mind
deserves a new kind of machine.

We're taking reservations for the first 1,000 units.
Pasadena, California · 2026