UNTETHER AI

Untether AI was founded in Toronto in 2018 by a team of semiconductor engineers and researchers with deep backgrounds in processor architecture and machine learning hardware. The founding insight came from a frustrating reality in AI chip design: most chips spend 90% or more of their time moving data between memory and processing units.

The actual computation is fast. The data movement is slow, expensive, and power-hungry.

The founders asked what would happen if you stopped treating memory and compute as separate problems and designed them as one system. The answer became Untether's core architecture — at-memory computing — and the company has been building it out ever since.

Untether AI sells custom AI inference processors to companies that need to run AI models at the edge or in data centers without relying on GPU-class power budgets. Their target customers are enterprises, OEMs, and cloud operators who need high-throughput inference at dramatically lower power consumption than what GPUs offer.

The revenue model is chip sales combined with software licensing for the toolchain that maps AI models onto their hardware. They are not a consumer product — they are deep in the B2B semiconductor supply chain.

Untether's flagship product family uses at-memory computing architecture — placing processing units directly adjacent to memory arrays to eliminate the bottleneck of data movement. Their chips are designed for AI inference workloads, particularly computer vision and natural language processing applications at the edge and in data centers.

The company has also developed a software compilation stack that handles the mapping of standard AI models (trained in PyTorch or TensorFlow) onto their hardware, which is essential for enterprise adoption. They have demonstrated significant improvements in inference throughput per watt compared to GPU-based alternatives.

Untether has grown by being technically credible in a market where technical credibility is everything. Intel Capital investing in the Series B was not just money — it was a signal to the industry that the architecture is worth taking seriously.

Their strategy has been to target the inference market specifically, rather than competing with Nvidia on training. Training is Nvidia's fortress.

Inference is still being contested. Edge AI inference, in particular, is a market where power efficiency matters more than raw compute, which is exactly where Untether's at-memory approach has an advantage.

The AI chip market is extraordinarily difficult to compete in. Nvidia's CUDA ecosystem is 15 years deep and almost every AI model in production is optimized for GPU.

Switching to a new chip architecture requires companies to retool their software stack, retrain their engineers, and accept some performance uncertainty. Untether has to convince customers that the power and cost savings are worth that switching cost.

They also have to survive long enough for a generation of engineers to get comfortable with their toolchain. That takes time and capital.

And the hyperscalers — Google, Amazon, Microsoft — are all building their own inference chips, which reduces the addressable market.

Untether AI's Series B in 2021 was led by Intel Capital and Tracker Capital Management, alongside other institutional investors. Intel Capital's participation is notable — Intel has its own chip ambitions and invests in companies that complement or extend the chip ecosystem rather than directly compete with their core products.

Earlier rounds brought in seed and Series A investors from the Canadian and international venture community. The company has also benefited from support from Canadian government innovation programs, which have historically backed deep-tech semiconductor companies in the Toronto region.