First up is Brium, a stealthy outfit fiddling with inference compilers. According to AMD, “Brium brings advanced software capabilities that strengthen our ability to deliver highly optimised AI solutions across the entire stack.” In other words, it's trying to get more performance out of its gear without brute-forcing it with more silicon.
Enosemi, a silicon-photonics specialist, was swept into AMD's claws. AMD senior VP Brian Amick reckons, “Now as part of AMD, the team will help us immediately scale our ability to support and develop a variety of photonics and co-packaged optics solutions across next-gen AI systems.”
Meanwhile, Untether AI’s engineering squad has been absorbed not for their product, but for their know-how. AMD confirmed, “The transaction brings a world-class team of engineers to AMD, focusing on advancing the company’s AI compiler and kernel development capabilities as well as enhancing our digital and SoC design capabilities.” Translation: we needed their brains to build leaner, meaner inference chips.
ZT Systems joins the fray to give AMD a direct line into rack-scale deployment, meaning AMD can now sell more than just chips. Add in previous buys like Nod.ai, Silo AI, Pensando and Mipsology, and suddenly AMD doesn’t just have a stack it has a cunning plan.
This is all part of what AMD hopes will be a “full-stack AI challenger” to Nvidia’s polished CUDA-and-CuDNN regime. It’s banking on compiler magic from Brium, speedy optics from Enosemi, and hands-on deployment from ZT to bridge the massive gulf between MI300 silicon and actual real-world AI dominance.
The company’s Instinct MI350 and MI400 chips are already out of the gate, along with Helios-branded racks. Big names like OpenAI, Meta and Microsoft are supposedly lining up to use them.
AMD says its latest silicon offers 40 per cent more performance per dollar than Nvidia’s B200, but without proper benchmarks and independent tests, that number could be as meaningful as one of Nvidia’s keynote charts.
Still, the acquisitions make it clear that AMD is done playing second fiddle. The only question is whether this Franken-stack can be stitched together in time to matter.