Operating✦ Frontier trackUntether AI

Untether AI puts memory and compute on the same chip for faster AI inference

Toronto startup building at-memory AI inference chips — memory and compute elements combined in one chip to kill the data-movement bottleneck of CPUs and GPUs, sold as runAI processors and accelerator cards for cars, machines, and edge devices.

شركة كندية من تورونتو تصمّم رقائق ذكاء اصطناعي تجمع الذاكرة والحوسبة في شريحة واحدة لتسريع الاستدلال وخفض استهلاك الطاقة في السيارات والآلات والأجهزة الطرفية.

Next →
  1. The Bottleneck

    What was broken?

    In traditional servers compute and memory are separate, so running AI models means shuttling data back and forth — a movement bottleneck that costs energy, latency, and performance on every inference.

  2. The Move

    Why it worked

    At-memory computation fuses memory and compute elements in the same chip, cutting the electricity needed for data transfers and delivering high-throughput, low-latency inference without sacrificing accuracy — a bet that efficiency, not raw training scale, wins edge and physical-AI workloads.

  3. The Trap

    Battle scars

    Selling chips against entrenched GPU incumbents; the capital intensity of silicon development across multiple generations; and converting pilot deployments into volume orders — the company ultimately ended its independent run when AMD acqui-hired its engineering team in June 2025.

The problem

Every AI query on a traditional server pays a movement tax: models live in memory, compute happens in the processor, and the ones and zeros travel back and forth constantly. That shuttling burns energy and adds latency — the exact bottleneck that makes fast, efficient inference hard, especially outside the data center.

How it works

Untether AI bet on at-memory computation: fuse the memory and computing elements into the same chip so neural-network weights are processed where they sit. Its runAI processors and accelerator cards target inference (running trained models on live data), promising high throughput at a fraction of the data-movement energy — packaged for data centers, cars, and edge machines.

Pain points

Cloud operators face soaring inference power bills; automakers and industrial buyers need on-device AI that works within tight power and latency budgets; and the market wants credible alternatives to incumbent accelerators on performance-per-watt.

Business model

Fabless chip company: design the silicon in Toronto, sell processors and accelerator cards to enterprise, automotive, and industrial customers, and support them with inference software and collaborations (including an Arm Automotive Enhanced tie-up for vehicle workloads).

Challenges

Silicon is brutally capital-intensive across generations; unseating incumbents requires not just better benchmarks but volume supply chains and software ecosystems; and despite real product shipments, Untether AI could not sustain independence — its engineering team joined AMD in 2025.

Funding

  • Raised: over $150M in venture capital total, including an oversubscribed $125M round in July 2021 co-led by Intel Capital and Tracker Capital, with CPP Investments and Radical Ventures participating.
  • Valuation: MISSING.

Latest — July 2021

Untether AI announced the oversubscribed $125M round to deploy its high-performance inference-acceleration chips, built around the runAI200 processor and a four-chip server card claimed at about 2 quadrillion operations per second.

Latest — October 2024

The company released an AI chip aimed at physical-AI applications in machines — including cars and agricultural devices — pushing its efficiency-first inference story from the data center to the edge.

Latest — June 2025

Semiconductor giant AMD acqui-hired the team behind Untether AI, as originally reported by CRN; the terms of the deal weren’t disclosed, ending Untether AI’s run as an independent chipmaker.

Similar sparks

⚡️ SERIES DVal: $2B pre-money (Dec 2024)

Canada · 2026-09-28

Tenstorrent sells open-source AI chips and RISC-V IP to challenge Nvidia

A Toronto-founded chip company selling AI processors built on Tensix cores with open-source software, plus licensed AI and RISC-V silicon IP for custom chip designs.

بالعربي

شركة رقائق أسسها مهندسون في تورونتو تبيع معالجات ذكاء اصطناعي مفتوحة المصدر وتراخيص سيليكون لتحدي هيمنة إنفيديا.

  • semiconductors
  • #canada
⚡️ LATE-STAGE-PUBLIC

China · 2026-08-14

SMIC is China's foundry answer to the chip blockade

China's largest contract chipmaker: a Shanghai foundry fabricating logic and AI-related chips at mature and 7nm-class nodes, expanding capacity under US export controls.

بالعربي

أكبر مسبك للرقائق في الصين: تصنيع الرقائق المنطقية ورقائق الذكاء الاصطناعي بتقنيات متقدمة مع توسيع الطاقة رغم قيود التصدير الأمريكية.

  • semiconductors
  • #china
🌱 Idea

China · 2026-07-05

Dongfang Suanxin bets 3D chip stacking can route around US curbs

A stealth-mode Shanghai AI-chip startup led by a Tsinghua professor that emerged in July 2026 with software-defined chips and 3D stacked near-memory computing on a fully domestic supply chain.

بالعربي

شركة صينية ناشئة لرقائق الذكاء الاصطناعي في شنغهاي بقيادة أستاذ من تسينغهوا خرجت من السرية عام 2026 مراهِنة على الرقائق المعرّفة بالبرمجيات والتكديس ثلاثي الأبعاد بسلسلة إمداد محلية بالكامل.

  • semiconductors
  • #china
⚡️Spark playbook copied! Go build.