AMD MI355X Benchmarks and Infrastructure Contracts, September 2026

AMD submitted a 512-GPU MI355X cluster in MLPerf 6.1 as NVIDIA made its first Vera Rubin submission, and the company secured a $14 billion Core Scientific lease and a roughly $450 million LUMI-AI contract in Europe — but NVIDIA still added twice AMD's data-center revenue in a single quarter.

The September 2026 round of MLPerf 6.1 offered a compressed picture of AMD's competitive position in AI compute. AMD submitted results for a 512-accelerator MI355X cluster — its current data-center GPU generation — in the same cycle where NVIDIA filed both Blackwell-class results and, for the first time, a submission from the Vera Rubin architecture. The specifics of AMD's performance figures were not published in available reporting, making a direct quantitative comparison between the two vendors' submissions impossible to state here. The round's structure is nevertheless informative: NVIDIA is benchmarking its next-generation platform at the same moment AMD is establishing its current one at scale. AMD's answer — the MI455X, announced in August as the first 2-nanometre data-center GPU paired with the Helios rack platform — remains a forthcoming product rather than a benchmarked result, and AMD's own materials describe Helios as capable of reducing AI data-center operating costs by thirty percent, a figure that awaits independent corroboration.

The commercial record offers the more immediate story. AMD assembled significant infrastructure commitments, spanning three separate deployments on three continents. The company secured a fifteen-year, $14 billion contract with Core Scientific for 529 megawatts of AI data-center capacity across five campuses — a transaction that, combined with earlier agreements, brings Core Scientific's total contracted AI portfolio to 1.1 gigawatts. In Europe, AMD was awarded the contract to supply accelerators for the LUMI-AI supercomputer in Finland, with the deal reported at approximately $450 million and Bull appointed as systems integrator; production operations are targeted for the second half of 2027. In Saudi Arabia, AMD Instinct MI355X systems went live as part of HUMAIN's national AI infrastructure alongside Cisco networking and Together AI software. Three large-scale deployments committing to AMD hardware within the same reporting window marks a qualitative change in the company's position outside its traditional North American hyperscale relationships.

The infrastructure wins carry more weight alongside the software work AMD has conducted in parallel. ROCm 10 and the general availability of ROCm.AI, both announced in August, represent a decade of iteration on an open-source alternative to CUDA whose earlier versions were widely criticised for incomplete kernel coverage. The release does not undo CUDA's incumbency, but it provides a stable, production-grade baseline for enterprise planners to commit against. Complementary moves reinforce the picture: AMD and Intel's AI compute extensions have been merged into GCC, enabling compiler-level AI acceleration for x86 without proprietary toolchains; SiFive is collaborating with AMD to optimise ROCm on RISC-V data-center servers, opening a CPU-agnostic deployment path; and Alphabet disclosed it is funding both NVIDIA's NVLink Fusion and AMD's UALink interconnect standard simultaneously. That last point is a hedge rather than an endorsement — Alphabet is protecting itself against vendor lock-in, not predicting a winner — but it signals that AMD has crossed a threshold: hyperscalers now treat it as a credible second vendor at the interconnect layer, not merely in the accelerator slot.

The financial picture measures the distance still to close. AMD's FY2025 capital expenditure reached $974 million, up 53.1 percent from $636 million in FY2024, against revenue of $34.64 billion — a capex intensity of roughly 3 percent, placing AMD thirteenth of fifteen chip-sector peers by that measure in Slicast's compilation. The Core Scientific lease, spread across fifteen years, does not register in that PP&E line in the conventional sense. AMD's CFO raised the company's estimate of the total AI chip addressable market to $3 trillion during the month; the stock gained 5.9 percent in the September 8 session, closing at $505.74, and has since firmed to $512.50. The market's direction is not irrational, but it sits alongside a data point from August: one analyst's review of quarterly results found that NVIDIA added twice AMD's entire data-center revenue in a single quarter — a gap that the Core Scientific and LUMI-AI contracts will require years of chip deliveries to meaningfully narrow.

Three signals are worth tracking as the next product cycle unfolds. First, whether the HUMAIN and LUMI-AI deployments expand into MI455X hardware as they scale — a yes would confirm that today's MI355X design wins lead to next-generation procurement rather than ending there. Second, whether AMD enters larger cluster configurations in future MLPerf rounds, enabling a quantified comparison with Blackwell at matched scale; the 512-unit submission in MLPerf 6.1 does not provide enough context to assess the performance gap. Third, whether AMD's data-center revenue growth rate in coming quarters reflects the scale of the infrastructure commitments described above; a step-change would indicate contracts converting into revenue at the pace the valuations imply. Running alongside these opportunities is a softer risk: multiple outlets reported in September that AMD's consumer RDNA 5 line and NVIDIA's consumer Rubin GPU have been pushed to 2028, though neither company has confirmed this and the reports are confined to the consumer segment rather than the data-center accelerators central to the AI buildout. AMD's market position has improved materially over the past year; translating that position into commensurate revenue share against an NVIDIA that is itself accelerating remains the open question.