semiconductorsGraveyard archive2018–2025

Untether AI

Untether AI bet seven years and ~$152M on a radical at-memory chip that shattered energy-efficiency benchmarks — then watched the AI world pivot to large language models its hardware could never adequately serve. In June 2025 AMD acqui-hired its engineering team for a reported sum under $100M; the speedAI accelerators and imAIgine SDK were discontinued, and Untether filed for bankruptcy that October.

Years active

2018–2025

Total raised

$152M

Peak headcount

~151

The bet

Untether AI emerged from the University of Toronto's chip-research community in 2018, founded by Martin Snelgrove, Darrick Wiebe, and Raymond Chik. The central thesis was architectural: conventional AI accelerators waste most of their power and time shuttling data between memory and compute. Untether's "at-memory compute" design embedded processing elements directly alongside SRAM cells, attacking data movement at its root.

The startup came out of stealth in April 2019 with a $13M Series A led by Intel Capital — early validation from the world's largest chipmaker — with Toronto AI specialist Radical Ventures also participating. It set up in Toronto's growing AI district and began a multi-year silicon program.

The build

In late 2019 Untether named technology veteran Arun Iyengar CEO, professionalizing the leadership as chip development ramped. The company recruited aggressively across chip design, compilers, and systems software, eventually peaking near 151 employees.

In July 2021 it closed a $125M Series B — one of the largest Canadian deep-tech rounds of the year — led by Tracker Capital and co-led by Intel Capital, with CPPIB and Radical Ventures joining. The round funded the tape-out of its flagship silicon and the imAIgine SDK, the software layer meant to give developers an automated path from trained networks to Untether hardware. Lifetime capital reached roughly $152M.

The high-water mark

Untether shipped the speedAI240 Preview and then the speedAI240 Slim — a 75W PCIe card it billed as the world's most energy-efficient inference accelerator. The MLPerf numbers backed the claim: more than 3x the energy efficiency of rivals in the Datacenter Closed Power category, ~6x in Edge, and record-low single-stream latency. In raw throughput the Preview posted the highest single-PCIe-card result in its class.

In September 2023 Untether recruited Intel corporate VP Chris Walker as president; he became CEO in January 2024, signalling a commercial push toward generative-AI and enterprise LLM use cases. But the optimism papered over a deepening structural mismatch.

The wall

Untether's architecture was purpose-built for the inference that defined AI before late 2022: convolutional networks, image classification, object detection — workloads where weights fit on-chip and memory bandwidth is the bottleneck. The at-memory design was superb for that. Large language models are a different beast: they demand enormous off-chip capacity, high-bandwidth memory, and strong chip-to-chip interconnects for multi-chip scaling. The speedAI family had none of those at scale.

When ChatGPT reoriented the whole industry toward generative AI, Untether held a chip that excelled at benchmarks customers had stopped prioritizing. Computer-vision inference — autonomous vehicles, edge cameras, industrial inspection — remained real, but not the fast-moving market drawing hyperscaler budgets. The company tried to re-narrate toward LLMs, but the silicon could not follow. Nvidia's CUDA moat deepened and AMD's HBM-dense MI300X took the LLM-inference market Untether aspired to. A capital raise in early 2025 failed.

The acqui-hire and aftermath

In early June 2025 AMD announced it was bringing on Untether's engineering team — compiler engineers, kernel developers, SoC designers, verification specialists — into its own AI-hardware organization. It was an acqui-hire: AMD took the people, not the products. The speedAI line and imAIgine SDK were immediately discontinued, no longer supplied or supported. The reported price was under $100M — a fraction of the ~$152M raised.

On October 15, 2025 Untether AI made a formal assignment in bankruptcy, with PwC as trustee; filings showed roughly $25M in assets against $128M in liabilities. "Lack of working capital/funding" was the cited cause. The at-memory dream didn't entirely die, though: co-founders Snelgrove and Chik — who had left before the deal over differences in market vision — had already founded Hepzibah AI in Toronto, pursuing a leaner, software-first take on edge AI. The idea just changed address.

What worked, what broke

What worked

Raised ~$152M including a standout $125M Series B backed by Intel Capital, CPPIB, and Tracker Capital — among Canada's largest deep-tech rounds at the time
speedAI240 Slim set MLPerf records for energy efficiency (>3x Datacenter, ~6x Edge) and lowest single-stream latency in its submission class
Built and shipped production silicon (speedAI240 Preview and Slim) plus the imAIgine SDK — clearing the hardest hurdle in the chip-startup playbook
Validated the at-memory-compute thesis well enough to draw AMD's engineering-talent deal and inspire a successor company (Hepzibah AI)
Grew from University of Toronto research roots to a ~151-person commercial hardware organization in seven years

What broke

Architecture optimized for vision/CNN inference was structurally ill-suited to LLM inference, which needs HBM, large off-chip memory, and chip-to-chip interconnects the speedAI family lacked
The late-2022 ChatGPT moment reoriented the AI-chip market to generative AI just as Untether prepared to commercialize — too late to change the silicon
Failed to raise additional capital in early 2025, ending the company; ~$128M in liabilities against ~$25M in assets at bankruptcy
Nvidia's CUDA ecosystem and AMD's HBM-dense MI300X dominated the LLM-inference market Untether tried to pivot into, leaving no viable lane
The acqui-hire returned under $100M to investors who had put in ~$152M — a loss in a capital-intensive sector with no second chances

Sources

Obituary authored Jun 4, 2026 via Sonnet 4.5 + web_search.