Summary
Y Combinator’s Paper Club asks a provocative question: what if AI stopped relying on GPUs? This session, hosted as the Alternative Compute Club, brings together researchers and founders exploring hardware and algorithms that depart from the dominant GPU-plus-transformer-plus-backpropagation stack.
The opening talk traces how hardware and model architecture have co-evolved since AlexNet in 2012. Early convolutional networks favored raw floating-point efficiency, but the rise of transformers after GPT-2 and GPT-3 shifted Nvidia’s priorities toward memory capacity and bandwidth, in part because attention scales quadratically. The speaker argues that gains in gigaflops per joule have largely stalled in the last two years, and contrasts today’s power-hungry systems with the human brain’s roughly 20-watt budget. They also question whether backpropagation will remain the training method of choice, pointing to the brain’s largely feed-forward structure and the weight transport problem.
Later talks cover optical and photonic computing, including how wavelength multiplexing can encode weights across hundreds of wavelengths, why feature sizes in optics are larger than in electronics, and the challenges of nonlinearity, analog-to-digital conversion, and optical storage. The audience Q&A digs into fixed-weight, application-specific designs and the practical barriers that have kept electronics dominant. Anyone following AI chips, inference hardware, and post-GPU architectures will find a wide-ranging look at what could come next.
📺 Source: Y Combinator · Published October 02, 2026
🏷️ Format: Course Lesson
