Before diving into today’s newsletter… check out our latest episode of The Compute 100!
Our guest this week is Austin Lyons, Analyst and Founder of Chipstrat. For anyone at the start of their compute learning journey, we often point to Austin’s writing as the place to start. He’s worked across the ecosystem, from time at Intel to working on startups, to now going all in on Chipstrat. We covered Austin’s takes on Intel, startup opportunities in chip design, and what being an independent analyst in this market looks like. Thank you Austin for joining! And be sure to subscribe to Chipstrat.
Lambda is raising up to $4B at a $14.5B pre-money valuation, led by Coatue and Blackstone, in what it is positioning as its last private round before a 2027 IPO. Its backlog grew from $15B in June to $50B in September, driven largely by a $35B Anthropic commitment signed in late August for Nvidia capacity hosted at a Texas facility operated by Hut8. A pre-IPO round priced on a backlog this concentrated in one customer ties the valuation to Anthropic's ability to fund its leases. (TechCrunch)
Private Companies
Coherecustomer
Cohere committed about $525.6M over roughly five years to Boost Run for dedicated Nvidia GB300 GPU cloud capacity, with part of the payment due as a prepayment. Rack acceptance is expected to begin in the second quarter of 2027, and the agreement lifts Boost Run's total contract value above $2.6B. A model lab is paying ahead for GPUs that will not be delivered for two quarters, which shows how far in advance capacity is being reserved. (Boost Run 8-K)
Public Markets
Lisa Su said AMD will substantially increase CPU and GPU supply in 2027 and needs more advanced wafer capacity, after meetings with TSMC and Foxconn in Taiwan and ahead of a stop in South Korea for memory supply. (Reuters)
SpaceX is seeking $40B to buy Nvidia chips, about $10B in bank loans and $30B in investment-grade debt led by Apollo, with Pimco among the lenders in talks and closing expected in 2027, the Financial Times reported. (Reuters)
Emerging
Meta NIC: A paper from Prankur Gupta, Alexander Duyck, Jakub Kicinski and colleagues at Meta, From ASIC to Fleet: Lessons from Building and Operating a Hyperscaler NIC, describes running fbnic, Meta's custom multi-host network interface controller, across hundreds of thousands of production hosts. Physical isolation, hardware-in-the-loop testing and a firmware stack with fewer than ten message types produced a 12x reduction in unplanned unavailability, 37% lower mean time to repair and 2.3x fewer hardware swaps. Fleet operability is now a stated reason hyperscalers build their own NICs, which raises the bar for merchant NIC vendors competing for those sockets. (arXiv)
TSMC Texas: TSMC is evaluating a potential investment in Texas on top of the $265B it has committed to Arizona, according to two people familiar with the matter. Arizona is planned to hold 12 fabs plus advanced packaging and R&D, and CEO C.C. Wei has said four or more additional fabs are likely there. TSMC declined to comment, and the plans are not final. A second U.S. state would widen where advanced-node capacity can be added as customers press for more wafers. (Reuters)