Code and Silicon: Unconventional AI's [un]CFO on Cutting AI Power a Thousand-Fold
Key points
Key takeaways from Source interview with Unconventional AI [un]CFO Ali Esfahani (October., 2026):
1000× power target. Unconventional AI is engineering the next substrate for intelligence, aiming to cut AI power consumption around a thousand-fold by running neural networks directly on semiconductor physics rather than simulating them on digital hardware.
Science answered in months. In five months the team went from no chip team to a tape-out, sending the design to TSMC on June., and bringing back real power numbers, which Esfahani says are orders of magnitude below traditional systems; an open-source model released in mid-June demonstrated computation with dynamics. The company now says the science phase is over and is on the engineering journey, aiming for product in less than three years rather than the five-year original timeline.
Sparsity is the scaling key. The key discovery: cutting the chip's connectivity from near-100% to roughly two-to-three percent actually improves output quality and learning — and makes a physically realisable chip possible, since a trillion wired connections cannot be routed in two dimensions.
Brain-inspired, physics-based design. There are no floating-point operations and no instruction set; inference traverses learned neural pathways, and memory inheres in the evolving state of the system rather than in dedicated SRAM or HBM. Reasoning and retrieval still need separate memory, which the company is solving concurrently.
A CFO who codes. Esfahani says he has written around 400,000 lines of code over the past nine months, building three internal applications, and that the company uses AI everywhere it can; he joined inJanuary,. as employee number 14,, and headcount iso closer to 60 now.
A full-stack team in a hot market. Hiring is challenging because the company spans everything from semiconductors to AI models;;"one of the only full stack teams," Esfahani says, co-designing hardware and software together;;the cap table includes the $475M seed raised before he joined and roughly $70–80M more raised since, led by CEO Naveen Rao.
Cheaper compute flips an energy paradox. At a thousandth of the power, a mini AI node could sit ina container by a cell tower and tap neighborhood power;;the team sees demand for intelligence as infinite,, citing Jevons' paradox — cutting the cost of intelligence drives consumption up, not down;;the ultimate goal is human-level intelligence in one rack for $20 an hour.
Ready for the robotics wave. The company is deliberately not chasing robotics yet, judging the market still early, but low-power silicon positions it for the coming explosion in edge devices, robotics, wearables and on-device AI;;"we need the data centers and the power infrastructure" — the fix is efficiency, not halting construction.
Read more: YouTube · SiliconANGLE
