AMD Gets 58.2% of Revenue From Data Centers • The Weighted Average

AMD’s Data Center segment generated $6.718 billion of its $11.536 billion second-quarter revenue, making the segment 58.2% of the company. That majority earns AMD a place in the next accelerator RFP, but absent public Helios prices and independent production benchmarks, it does not justify a fleet-wide switch.

Three of every five AMD dollars now come from the data center

The transformation is visible across comparable quarters. AMD reported Q2 Data Center revenue of $2.834 billion in 2024, then $3.240 billion in 2025, and now $6.718 billion in Q2 2026. The latest figure is up 107% year over year and 16% sequentially; dividing it by total company revenue produces the derived 58.2% share.

The longer view is stronger than the one-year headline. Data Center added $3.884 billion of quarterly revenue in two years. Calculating (6.718 ÷ 2.834)^(1/2) − 1 gives 54.0% annualized growth, stitched from the 2024 and 2026 releases. That trajectory makes AMD material enough to improve a buyer’s negotiating leverage even before it becomes the default accelerator supplier.

Profit followed. AMD’s Q2 earnings slides put Data Center operating income at $2.103 billion, implying a 31.3% segment operating margin when divided by $6.718 billion. Do not compare that mechanically with the prior year’s $155 million loss: Q2 2025 included an $800 million charge tied to U.S. export restrictions on MI308 inventory. Excluding that charge, the apparent margin leap narrows sharply.

Demand is still accelerating at the company level. AMD guided Q3 revenue to roughly $13.0 billion, plus or minus $300 million, implying about 13% sequential growth at the midpoint. That forecast does not isolate Data Center and should not be treated as an accelerator-sales promise, but it gives procurement teams a reason to secure evaluation hardware before the next ramp rather than wait for a mature market that may allocate supply elsewhere.

This quarter also marks a product shift. AMD no longer pitches only an accelerator; it pitches Helios, a rack-scale system joining MI455X GPUs, sixth-generation EPYC CPUs, Pensando networking and ROCm software. The earnings release names Anthropic, Meta, Microsoft, OpenAI, Oracle and others, but named use is not the same as disclosed production volume or independent performance.

The strongest forward commitment comes with an important asterisk. AMD and Anthropic announced up to 2 gigawatts of MI450 deployments, with the first gigawatt beginning in the first half of 2027. AMD also committed up to $5 billion of future equity investment in Anthropic. The capacity signal is real, but the financing link means it is not pristine evidence of arm’s-length demand.

For buyers, the new choice sits beside—not beneath—SpaceX’s capital-heavy neocloud offer. More viable stacks can reduce dependency on one accelerator, one cloud and one software roadmap. Our earlier examination of AMD’s Core Scientific capacity strategy made the same point: supply diversification matters only when workloads survive the move.

Put AMD in the RFP, not across the whole fleet

The best near-term candidate is bounded inference on existing PCIe infrastructure. Tom’s Hardware reports AMD’s MI350P specifications list 144GB of HBM3E, 4 TB/s of memory bandwidth and up to 600 watts of board power, configurable to 450 watts. A team can compare that device against an incumbent inside a familiar server topology without waiting for a full Helios rack. What it cannot do is infer savings: AMD publishes no list price, matched cloud rate or independently verified cost per token.

Software remains the gating cost. The ROCm 7.2.3 compatibility matrix is specific to GPU, operating system and framework version, and includes exceptions such as unsupported TensorFlow combinations on MI350-series hardware. Procurement therefore needs an application-level test: custom kernels, quantization, collective communication, orchestration, observability, recovery and image maintenance. A nominal benchmark win can disappear beneath porting labor and lower utilization.

Who should move now? Large inference operators with meaningful Nvidia spend should put AMD into their next RFP and use a credible alternative to negotiate hardware, networking and support. Enterprises with portable inference and available platform engineers should pilot MI350P. Frontier labs planning 2027 campuses should reserve dual-source capacity. CUDA-extension-heavy training teams and small shops without accelerator specialists should wait or restrict the experiment to one workload.

Budget duplicated pipelines, rebuilt kernels, SRE training, validation and an incumbent fallback pool. Also budget power at the system level: the published 600-watt board ceiling excludes host CPUs, networking and cooling. “Up to 2 gigawatts” describes a commitment boundary, not installed, utilized or revenue-producing capacity.

The thesis breaks if Helios slips, ROCm regressions erase hardware economics, independent tests show lower useful utilization, or Nvidia answers with pricing that makes migration irrational. Export rules can also recur; last year’s $800 million charge proves that risk has already reached the income statement. Vendor-financed demand deserves monitoring until unaffiliated customers produce repeat orders.

Evidence for a broad switch should include public matched prices, reproducible production-model benchmarks, fleet uptime, on-time first-gigawatt delivery and at least two quarters of MI450 or Helios revenue and margin contribution. Until then, the verdict is disciplined: pilot AMD to create leverage, but migrate only the workloads whose measured total cost beats the incumbent after software labor and fallback capacity.