AMD put a price on its pending purchase of Taalas: about $2.9B in cash and stock, including up to about $1.4B in contingent consideration, according to its quarterly filing. AMD agreed in August to buy the Toronto inference-chip startup without disclosing terms. Taalas builds chips that hardwire a specific AI model into the silicon instead of running it on a general-purpose processor. The same filing values AMD's separate all-stock deal for World Labs at about $8.2B. Both acquisitions are expected to close by the end of 2026, subject to regulatory approvals. (AMD 10-Q)
Private Companies
Tenspressofundraise
Tenspresso, a chip startup about three months old, is in talks to raise at a valuation of roughly $500M, Bloomberg reported. Altimeter Capital has approached the company about an additional investment, following a $20M round from Striker Venture Partners. Coverage of the report puts the new round at $50M or more, and terms are not final. The company has 15 employees and has not completed a chip design. (Bloomberg)
Public Markets
Sandisk fell 17% to $861.54, its lowest level since April, as memory and chip stocks sold off after Democrats won control of Congress and investors weighed what it means for data center buildouts. Micron fell about 13%. Data center was $4.8B of the $11.03B in fiscal first-quarter revenue, up 61% from the prior quarter, and Sandisk guided the next quarter to $12.2B–$12.6B (Benzinga)
AMD fell 13.1% to $424.85 as the Nasdaq dropped as much as 5.7%. Qualcomm fell more than 10%, its worst day since March 2020, and Nvidia and Broadcom each fell about 9%. AMD had reported third-quarter revenue up 45% to $13.4B on Nov. 3, with Data Center up 97% to $8.5B and fourth-quarter guidance near $17B (CNN)
Emerging
Optical switches for decode: A paper from Maximilian Thürmer and eight co-authors, Workload-Static Optical Circuit Switching For Scalable LLM Inference Decode, observes that decode traffic does not change for a fixed model and parallelism plan, so optical circuit switches can be configured once per deployment instead of using packet switches. In a 576-accelerator domain serving 5T-parameter mixture-of-experts models with up to 5M tokens of context, all-optical fabrics improve interactivity by up to 65% and energy efficiency by up to 119% over 4:1 oversubscribed packet-switched fabrics. The results come from an emulator and network models calibrated to hardware measurements, not a deployed system (arXiv)
Flash for trillion-parameter models: A paper from Minho Ha, Myunghyun Rhee, Euiseok Kim and Hoshik Kim, HyPAIC: Personal AI Computer with High Bandwidth Flash for High Performance Trillion-Parameter LLM Inference, keeps read-only model weights in High Bandwidth Flash, which the authors say offers over 10x the capacity of HBM at comparable bandwidth, and keeps changing data in LPDDR memory. In simulation of a 1T-parameter mixture-of-experts model, one HyPAIC matched the throughput of eight baseline systems without flash (up to 1.04x) and delivered up to 3.59x the TDP-normalized throughput. All results are simulated (arXiv)