AMD MI400 Ships With 432GB HBM4 to Contest Nvidia Rubin in Data Center AI
AMD launched the Instinct MI400 series at its Advancing AI 2026 event on July 23 in Santa Clara. The chip ships with 432GB of HBM4 memory, the highest capacity in AMD’s data center GPU line, targeting the same enterprise AI and HPC workloads where Nvidia’s Rubin B300 generation is now competing.
The company previewed the MI400 family at CES 2026 in January, making the Santa Clara announcement the formal commercial launch. AMD described the MI400 as “the next generation of AMD data center GPUs engineered to redefine AI and HPC at every scale.”
The market AMD is entering
The competitive context is sharpening. AI inference provider DeepInfra opened its first non-US data center in Toronto on July 8, deploying more than 1,000 Nvidia Blackwell B300 GPUs across 1.7MW of capacity. The facility gives Canadian developers their first domestic access to a purpose-built Blackwell-generation inference cluster, addressing data-residency requirements that were previously pushing workloads to US facilities.
That Toronto deployment illustrates what AMD is competing for: the next wave of purpose-built AI inference infrastructure, where buyers are committing to multi-year silicon bets at the rack level. DeepInfra went Blackwell for its Canadian build. The question for subsequent deployments is whether MI400 economics alter that calculus.
Specs and positioning
The 432GB HBM4 figure is the headline differentiator for memory-bound workloads, particularly large-context inference and frontier model serving where VRAM is often the first bottleneck. AMD has not disclosed the MI400’s full compute throughput specifications at launch, but the HBM4 configuration signals a direct challenge to Nvidia’s memory subsystem in the Rubin architecture.
AMD’s Anthropic relationship adds a commercial dimension: the two companies announced a deal in July in which AMD committed up to $5B in investment and won a 2GW MI450 compute contract for future Claude infrastructure. The MI400 launch sits upstream of that, establishing the product line that will eventually underpin those facilities.
What this means for buyers
The AMD versus Nvidia data center GPU contest is no longer a roadmap debate. Both vendors have product in production. The MI400 is shipping, the B300 is shipping, and AI infrastructure operators are making vendor decisions that will define their stack for the next two to three years.
GPU hourly rental rates and total-cost-of-ownership comparisons will drive procurement for mid-tier operators who cannot match hyperscaler buying power. AMD’s HBM4 memory advantage needs to translate into per-inference cost efficiency to move buyers away from established Nvidia toolchains, particularly CUDA, which retains a significant software ecosystem lead.
The answer will arrive in benchmark comparisons and pricing sheets over the next two quarters.