Co-Founder & CEO at Zettascale
Energy efficient chips for AI
Building energy-efficient reconfigurable dataflow chips for AI training and inference. The XPU architecture can be optimized per model, delivering up to 27.6x more efficiency than NVIDIA H100 GPUs — saving data centers hundreds of millions in annual energy costs. Self-taught engineer from Gothenburg, Sweden who started coding at age nine. Studied Computer Science and Computer Engineering at Chalmers University of Technology before dropping out at 21 to build Zettascale (formerly Exa Laboratories). Zettascale (YC S24) is tackling one of AI's biggest bottlenecks: the growing energy demand of training and running large models.
Zettascale (formerly Exa Laboratories) is building energy-efficient reconfigurable chips called XPUs for AI training and inference. The architecture can be optimized per AI model, and early simulations show up to 27.6x better efficiency than NVIDIA H100 GPUs.
The chips are built on Learnable Function Units (LFUs), which are specialized components that approximate any univariate function and run asynchronously without traditional instruction fetching. The architecture reconfigures itself per model using three techniques: localization (keeping data close to compute), instruction fusion (combining operations), and layer fusion (merging model layers).
Modern GPUs consume 600-1000W each and power draw keeps going up. Global data center energy demand is expected to grow from 60 GW to over 170 GW by 2030. Zettascale's reconfigurable chips aim to deliver dramatically higher performance per watt, with early simulations showing up to 2.3 TFLOPS/W at 400W.
Zettascale (YC S24) raised a $500K pre-seed and a seed round of several million dollars in September 2024. Investors include Climate Capital, Failup Ventures, Geek Ventures, Multimodal Ventures, Olive Tree Capital, and Y Combinator.
Elias Almqvist (CEO) is a self-taught engineer from Gothenburg, Sweden who started coding at 9. He studied Computer Science and Computer Engineering at Chalmers University of Technology before dropping out at 21 to start the company. Prithvi Raj (CTO) holds an MEng from Cambridge's Computational Statistics & Machine Learning Lab and previously interned at Microsoft, specializing in scientific machine learning and electrical engineering.
The team felt that exascale computing was no longer ambitious enough given their internal breakthroughs. The new name reflects their goal of reaching zettascale computing, the next order of magnitude beyond exascale.
