Startup Irreversible Unveils Analog AI Chip with 1000x Power Reduction
Canadian company Irreversible has announced a "physical" analog compute-in-memory architecture for AI at the edge. The technology promises a 1000x reduction in power consumption compared to digital processors and targets ultra-efficient devices such as drones and wearables.
The Analog Revenge: Why Irreversible's Chip Is Not Just About Energy Efficiency, but Dismantling Digital Orthodoxy
The Core: What's Really Happening
Canadian startup Irreversible has done what the academic community has dismissed for the last 40 years—created a working analog chip for neural network inference that doesn't just compete with digital architectures but redefines the very paradigm of computing. The claimed 1000x reduction in power consumption is not marketing hype but a physical limit achievable only by moving from discrete logic to continuous physical processes. The chip performs matrix multiplication directly in memory, using Ohm's and Kirchhoff's laws instead of clock signals, registers, and data buses. This is not an improvement of existing architecture—it's a replacement.
But the real news runs deeper. Irreversible didn't just make an analog accelerator—they solved the problem that had sunk analog computing for decades: noise, parameter drift, and lack of exact reproducibility of results. Their "physical" architecture uses noise as a functional element rather than fighting it. This is a philosophical break from the digital paradigm, where the main idol was accuracy, portable across time and space.
Timeline and Context
The roots go back to 2014, when Professor Rahul Sarpathwar from the University of Illinois published a paper on computational memory based on resistive elements. The problem was fundamental: analog computing offered tremendous speed and low power consumption, but a 1-2% error per operation made multi-layer networks useless. In 2018, DARPA launched the LUMOS program with a $78 million budget, trying to find a compromise between accuracy and efficiency. By 2022, Mythic AI had raised $70 million for analog matrix processors, but the product never made it beyond test samples—noise killed accuracy above 4 bits.
Irreversible's breakthrough, which occurred in May 2026, was the result of three converging technologies: (1) hafnium oxide memristors with record state stability from Stanley Williams' group at HP Labs, (2) a stochastic resonance method where controlled noise increases rather than decreases computational accuracy, and (3) hardware-in-the-loop training, where the model is initially trained for a specific hardware instance with its noise profile.
CEO Niv Jain, who previously headed the neuromorphic systems division at Intel Labs until its closure in 2024, assembled a team of 14 people, including three former Loihi engineers. The company is registered in Toronto, but the main lab is in Sherbrooke, home to one of the leading centers for photonics and quantum electronics. The startup raised $22 million from SOSV and Canadian venture fund BDC Capital at an $85 million valuation in its seed round.
Who Wins and Who Loses
Winners include drone and wearable device manufacturers—a segment suffocating from thermal constraints. A drone with an Irreversible chip could perform full YOLO inference onboard, consuming 8 mW instead of 8 W, radically changing flight time and thermal profile. Qualcomm, whose Snapdragon chips dominate drones through the Flight RB5 series, loses differentiation.
Losers include manufacturers of digital neuromorphic chips like BrainChip and SynSense. Their advantage of "event-driven architecture" now looks like a half-measure compared to a direct analog approach achieving even lower power without complex spike representations. BrainChip's AKD2500 project, announced last week, loses some investment appeal.
An unexpected loser: battery manufacturers for IoT. If a sensor device consumes 10 µW instead of 10 mW, energy harvesting becomes more realistic. This shifts added value from chemical power sources to the silicon substrate.
The biggest winner: the US defense sector through DARPA. The LUMOS program gets a working commercial chip that can be immediately classified and used in passive surveillance systems. The Irreversible chip, consuming microwatts, is ideal for acoustic sensors operating for years without maintenance.
What the Media Isn't Saying
The first non-obvious insight: Irreversible's architecture is fundamentally incompatible with cloud computing. An analog chip physically cannot be virtualized or emulated on a GPU cluster while preserving its properties. This means infrastructure giants like AWS, Azure, and GCP cannot offer this technology as a service. All value is concentrated in the physical device at the edge. This creates an economic model opposite to the current trend of AI centralization and calls into question the business cases of companies betting on cloud monetization of inference.
The second insight concerns lithography. The Irreversible chip is manufactured on an outdated 45nm process at Tower Semiconductor's fab in Israel. This is not a bug but a feature: analog circuits do not benefit from process shrinking—noise only increases, and dynamic range decreases. Physics works in reverse: the larger the transistor, the lower the thermal noise and the more stable the analog state. Thus, Irreversible is independent of TSMC, sanctions, queues for 3nm wafers, and the entire lithographic scaling race. This completely bypasses the geopolitical semiconductor crisis.
The third insight: training models for a specific chip instance creates a situation where software and hardware become inseparable. A model trained for chip number 1764 will not work optimally on chip number 1765 due to the unique noise profile of each piece of silicon. This gives rise to a new paradigm for licensing and distributing AI models, radically different from the current one—models become physically tied to hardware.
Forecast: Next 30 Days and 90 Days
Next 30 days (until mid-June 2026). I expect DARPA to announce an expansion of the LUMOS program with additional funding of at least $40 million, with Irreversible as the prime contractor. Qualcomm and AMD will begin closed-door talks about acquiring or licensing the technology; the most likely buyer is Qualcomm with its drone and XR device platform, with a deal size of $300–400 million. Chinese investors via proxies in Southeast Asia will also try to enter the cap table, but the Canadian government will block the deal under the Investment Canada Act, citing national security.
Next 90 days (until mid-August 2026). At least one major reorganization in the industry will occur: either Intel will close the remaining part of its neuromorphic division, or BrainChip will begin merger proceedings with a larger player lacking an analog portfolio. I expect the first commercial product with the Irreversible chip to be announced not in North America but in Japan—Sony has exclusive access to the design through its partnership with Tower Semiconductor and could embed the chip in the next Sony Airpeak for industrial inspections. The startup will raise its next Series A round of $80–100 million at a valuation of around $400 million.
But the most important shift will happen in regulation. The FCC and European regulators will face a challenge: a chip with no digital debug interface and operating on physical principles cannot undergo standard software certification. How do you verify that an analog neural network in a drone has not been trained on prohibited patterns if the weights are physically fused into memristors and do not exist in digital form at any stage? This question will open a new chapter in AI regulation, and Irreversible will become a precedent-setting case—a company that accidentally, through physics, bypassed the emerging system of algorithmic oversight. That is the true scale of what is happening: analog computing does not just save energy—it returns computation to the physical world, where human laws no longer apply with the same inevitability as in the world of digital copies.
— Editorial Team
No comments yet.