Welcome Timothy Arcuri to The Compute 100!
Today’s guest is Timothy Arcuri, Managing Director at US. Timothy has been covering the semiconductor landscape for over 20 years. His expertise is unmatched. If you’re looking for a primer on how to understand the semis value chain today, give this episode a listen!
fal raised a $1.3B Series F at a $16B valuation, double the $8B it was valued at in April. Bessemer Venture Partners led the round, with Nvidia, Sound Ventures and T. Rowe Price also investing. The company announced it alongside a Series E, bringing the two rounds to $1.65B in total, and Sequoia led the earlier round. fal hosts and runs AI video, image and audio models, and Bloomberg reported its annualized revenue passed $800M earlier this year (TechCrunch)
Private Companies
Sambanovapartnership
Supermicro will build, sell and support inference systems that pair Nvidia GPUs with SambaNova's SN50 chips, with an initial plan for 10,000 SN50 chips. The GPUs handle prefill, the compute-heavy step that reads the prompt, and the SN50 handles decode, the memory-bound step that writes the answer one token at a time. Sales start with neoclouds and sovereign clouds before moving to enterprises, and SambaNova expects the deal to add substantially more inference capacity in 2027 (SambaNova)
Fireworkspartnership
Fireworks is adding European capacity on Nvidia's Vera Rubin and Grace Blackwell platforms, with deployments beginning in 2027. The company did not disclose locations or size. At its first Fireworks Forge conference it also put Ultrafast, a latency-focused inference tier on Blackwell, into public preview, and added multi-day prompt caching, which reuses large shared context across repeated model calls, plus higher rate limits for long-running agents. Nvidia CEO Jensen Huang said, "I'm delighted to see demand for Fireworks skyrocketing" (Fireworks press release)
Public Markets
AMD reported after the close that third-quarter revenue rose 45% to $13.4B, with Data Center up 97% to $8.5B, and guided the fourth quarter to about $17B. Shares rose more than 6% in extended trading (AMD press release)
Emerging
Decoding inside DRAM: A paper from Kevin Xu, Tatsuya Kubo, Daichi Tokuda, Shinya Takamaeda-Yamazaki and Issei Sato, Accelerating LLM Decoding with In-DRAM Parallelism, runs the matrix-vector math of LLM decoding inside DRAM chips, packing a later layer's projection with the current one so the 65,536 DRAM columns stay busy. It reports up to 1.41x end-to-end speedup over a baseline processing-in-DRAM design and up to 2.45x over modeled processor execution, both for Gemma 12B. The results come from a Ramulator 2.0 simulator and an analytical model, not fabricated hardware, and the geometric-mean gains are about 1.2x (arXiv)