Keep pulling the thread on Reiner Pope.
MatX has raised a Series B funding round led by Jane Street and Leopold Aschenbrenner's fund, Situational Awareness.
Major AI labs like OpenAI, Anthropic, and Google are purchasing multi-gigawatt compute clusters that cost tens of billions of dollars.
MatX's chip architecture combines HBM for inference data and SRAM for model weights, aiming to deliver both high throughput and low latency.
The primary bottleneck for new AI data center construction is the availability of power and grid infrastructure.
A significant part of NVIDIA's competitive advantage stems from its CUDA software ecosystem, which has been developed for over a decade.
OpenAI has initiated its own internal chip design program.
MatX's chip architecture features a large systolic array that can be split into smaller pieces to efficiently process workloads like the attention mechanism in transformers, which do not map well to a single large array.
NVIDIA's CUDA software moat is less effective in the LLM market because there are only about five major "Frontier Labs," and they have the economic incentive and capability to write custom software for new, more efficient hardware.
The public perception of Google's AI capabilities improved significantly following the release of its Gemini 3 model, which was noted for being both high-quality and fast.
The high performance speed of Google's Gemini 3 model is powered by the company's custom chip hardware.
The Transformer architecture, which is foundational to modern LLMs, originated from research conducted at Google.
Google's TPUs were designed specifically for neural networks from the outset, in contrast to NVIDIA's GPUs which originated from graphics applications.