deepsilicon
Summer 2024InactiveSoftware and hardware to run neural networks faster and cheaper
Software and hardware to run neural networks faster and cheaper
Deepsilicon was a San Francisco-based AI infrastructure startup that participated in Y Combinator's Summer 2024 batch. Founded by Abhinav Reddy (CEO) and Alexander Nanda (CTO), the company built software and hardware to run transformer-based neural networks faster and at lower cost, using a technique called ternary quantization — compressing model weights from 16 bits down to 2 bits to dramatically reduce memory usage and increase inference throughput.[1]
Deepsilicon failed primarily because it attempted to execute a full-stack hardware-software co-design strategy — one that typically requires tens of millions of dollars and multi-year timelines — with $500K in total funding and a two-person team. The software layer faced rapid commoditization from open-source tools, while the custom ASIC ambition was structurally incompatible with the capital raised.
YC lists the company as "Inactive."[2] No acquisition, asset sale, or public shutdown announcement has been recorded. Alexander Nanda subsequently joined Mercor, signaling a clean departure from the venture.[3] The company's entire public life — from YC acceptance to last known activity — spanned roughly six months in 2024.
Deepsilicon was founded in 2024 by two technically credentialed but early-career engineers who identified a genuine bottleneck in AI deployment: running large transformer models at the edge was prohibitively expensive and slow on existing hardware.[4]
Abhinav Reddy, the CEO, brought a background in computer science and electrical engineering, with a focus on building simple software interfaces for fast hardware.[5] Alexander Nanda, the CTO, was a physics and computer science dropout from Dartmouth College who described himself as passionate about deploying massive neural networks on edge devices.[6] Nanda's Crunchbase profile indicates his role at Deepsilicon began in May 2024, suggesting the company was formally constituted in the spring of that year.[7]
No public record exists of how Reddy and Nanda met, what prior projects they collaborated on, or whether they had industry experience at semiconductor or AI infrastructure companies before founding Deepsilicon. This is a meaningful gap: hardware-software co-design ventures typically benefit from founders with prior chip design, fab relationships, or enterprise sales experience — none of which is documented in the available record.
The founding insight was straightforward and technically grounded: transformer models store weights as 16-bit floating-point numbers, but research had shown that weights could be quantized to just three values (-1, 0, +1) — a ternary representation — with acceptable accuracy loss on many tasks. If weights occupy only 2 bits instead of 16, the same GPU can hold roughly 8x more model parameters in memory, and purpose-built hardware can process those weights far more efficiently than general-purpose silicon.[8]
The initial vision was explicitly full-stack: a pip-installable software layer that worked on existing Nvidia hardware immediately, paired with a longer-term custom ASIC chiplet architecture that would unlock further power and cost reductions for edge deployment in autonomous vehicles, robotics, and industrial IoT.[9] The go-to-market was developer-first: remove all friction from the software entry point, build a user base, and use that traction to justify the hardware investment.
No major pivot is documented. The company appears to have pursued its original thesis from founding through its last public activity in September 2024, without a recorded change in direction.
Read the complete post-mortem, the rebuild playbook, and the exact reasons deepsilicon is still worth studying now.