AMD announced this week the largest deal in its history: supplying up to 6 gigawatts of custom Instinct MI450 GPUs to Meta in a multi-year agreement valued at $60 billion. AMD shares jumped 14% to $349.54 after a DA Davidson upgrade.
The most interesting part of the deal is not the size. It is the structure: a warrant that gives Meta about 10% of AMD equity if both sides deliver. Meaning Meta now has skin in the game to make this work. It is not a customer — it is almost a co-investor.
Why it matters
For the first time in years, the largest hyperscaler in the Western world is making a SERIOUS bet outside Nvidia. Not for experimental workloads — for production, at gigawatt scale.
Technical translation: Meta will train Llama 5, Llama 6, and the next Muse Spark models on AMD hardware. That means the CUDA ecosystem will need to share mindshare with ROCm (AMD) through 2027-2028.
What the MI400 series brings
- 432 GB of HBM4 per GPU (vs 192 GB on Blackwell B200)
- Performance within 10-30% of Nvidia on enterprise workloads
- Pricing 40-50% less per unit (analyst estimate)
Cumulative pressure
This is the third recent big bet on Nvidia alternatives:
- Google invested $40B in Anthropic with 8.5 GW of dedicated TPU
- Amazon invested $25B in Anthropic with Trainium
- Meta now $60B in AMD MI450 with skin in the game
Add it up: ~$125B in non-Nvidia compute committed in the last 90 days. The monopoly is not over, but it has started to crack.
Sources
- Techi (April 2026): AMD Meta $60B AI Chip Deal
- LinkedIn News: AMD, Nvidia unveil new AI chips
- CES 2026: AMD reveals new AI PC chips, details next-gen data center chips