You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
The AI deals continue. I guess TSMC will have no trouble filling their new FABS till 2030 and beyond:
AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs February 24, 2026 7:00 am
News Highlights
Meta is partnering with AMD to rapidly scale AI infrastructure and accelerate the development and deployment of cutting-edge AI models
AMD and Meta agree to a definitive multi-year, multi-generation partnership to deploy up to 6 gigawatts of AMD Instinct™ GPUs
Built on the AMD Helios rack-scale architecture announced at the 2025 Open Compute Project Global Summit, shipments to support the first gigawatt deployment are expected to begin in 2H 2026, powered by a custom AMD Instinct GPU based on the MI450 architecture and optimized for Meta’s workloads
AMD and Meta are deepening their collaboration to align their GPU and CPU silicon, systems and software roadmaps
The big deal is AMD potentially giving away even more of the company:
" As part of the agreement, to further align strategic interests, AMD has issued Meta a performance-based warrant for up to 160 million shares of AMD common stock, structured to vest as specific milestones associated with Instinct GPU shipments are achieved. The first tranche vests with the initial 1-gigawatt of shipments, with additional tranches vesting as Meta’s purchases scale to 6 gigawatts. Vesting is further tied to AMD achieving certain stock price thresholds and exercise is tied to Meta achieving key technical and commercial milestones."
Google says AMD has ~ 1.63B common shares outstanding, so this is roughly 10%.
The "pay your customer who pays you" loop continues..
This is very true. Given TSMC owns the 3nm and 2nm nodes the next 5 years is going to be a busy time. Building fabs is the easy part, hiring people to do the actual fab work will be a challenge! Exciting times in the semiconductor industry, absolutely!
In my opinion, power affects every day people far more than wafer constraints or packaging constraints or architectural constraints, and it is what industry leaders — Jensen, Altman, Nadella, et al — are all saying the is bottleneck. So that makes sense to be what the media talks about.
But as a “deal target”, is setting a power consumption target not completely antithetical to efficiency goals? I get there’s probably some more complex provisions in the deal, but wouldn’t you want to target FLOPS? GPU count? Both of those would structure the deal to have everyone involved optimize for deploying what matters (compute capacity) and not just stuffing bad products in a data center to hit a power consumption goal.
What a weird time we’re in. You can bet that these type of deals would never, ever fly in a normally operating market environment…
. 