AMD is gaining strong ground in the AI compute race, announcing a series of major partnerships and product innovations that underscore its ambition to challenge NVIDIA’s market dominance. Just days after revealing a 6 GW compute deal with OpenAI, the Lisa Su-led semiconductor giant has secured another major win — a commitment from Oracle to deploy 50,000 of its upcoming MI450 GPUs starting in Q3 2026.
In tandem, AMD introduced Helios, a rack-scale system powered by 72 MI450 GPUs built on TSMC’s advanced 2-nanometer process technology. The system represents a major leap in AI infrastructure design, combining high-performance compute with next-generation efficiency to position AMD as a formidable player in large-scale AI workloads.
A central pillar of AMD’s resurgence is its revamped software ecosystem, marked by the release of ROCm 7 — the company’s most comprehensive software upgrade to date. Anush E., Vice President of AI Software at AMD, explained, “You can now do prefill, decode, and disaggregation — which means that you could actually scale to a massive size deployment with a few nodes.” He further highlighted that fine-tuning massive models like Llama 405B can now be achieved with a single AMD GPU boasting 288 GB of High Bandwidth Memory. “We want to have the largest software ecosystem on AI, we want to double down on all parts of it and not just documentation,” Anush emphasized.
The software improvements are already reshaping perceptions among developers and analysts. SemiAnalysis, a firm previously critical of AMD’s software reliability, observed that “the quality of AMD software is totally different from that of last year,” noting that bug occurrences are now “orders of magnitude lower.” Its benchmarking reports show AMD’s MI300X and MI355X GPUs approaching performance-per-dollar parity with NVIDIA’s H100 and B200 models.
As Anush E. summed up, “ROCm means open. ROCm means accessible. ROCm means performant.” With high-profile deals, improved developer trust, and a maturing software ecosystem converging, AMD is signaling that it’s not just catching up — it’s positioning itself to lead the next chapter in AI compute innovation.
