Lisa Su has made AMD a credible second source for AI compute by pairing Instinct accelerators with EPYC CPUs, networking, systems, and the ROCm software stack. The opportunity is large, but the competitive test is not a peak hardware specification. AMD must make real models easy to deploy, operate, and move across a software ecosystem that has long been optimized for Nvidia CUDA.

This analysis was checked against public sources on September 14, 2026. AMD performance and roadmap statements are labeled as company claims unless supported by a regulatory filing or reproducible third-party test.

Su’s current mandate

AMD’s 2025 Form 10-K identifies Lisa Su as chair, president, and chief executive officer. The filing also describes AMD’s reportable segments, acquisitions, competitive risks, and dependence on external manufacturing. The annual filing is the authoritative source for her title and AMD’s audited results.

Su’s AI strategy extends the playbook AMD used in CPUs: maintain a regular architecture cadence, build products for several parts of the data center, and give large customers an alternative supplier. AI accelerators add a harder software problem. A customer that can compile and run a model is not necessarily ready to debug it, tune kernels, monitor it, or hire operators at scale.

Hardware progress created an opening

AMD’s Instinct MI300 family established commercial demand, and the company moved to MI350 with MI400 on its roadmap. Its 2024 annual report said AMD had launched MI325X, enhanced ROCm, and shifted to an annual accelerator cadence. Those are company disclosures, useful for chronology but not proof of a universal performance advantage.

The platform spans more than GPUs. EPYC CPUs, Pensando networking technology, accelerator systems, and software can be sold as a coordinated design. That gives AMD more control over system performance and a larger share of the deployment. It also raises the bar: customers will compare rack-level throughput, power, memory, networking, availability, and service, not only the accelerator.

ROCm is the adoption constraint

ROCm is AMD’s open software platform for GPU computing. Its public repositories let developers inspect releases, supported components, issues, and contributions. The ROCm project on GitHub provides stronger evidence of software availability than a marketing claim, though repository activity does not prove a specific workload will operate without changes.

For buyers, software compatibility should be tested at the model and framework level. Check required operators, quantization, distributed training, inference engines, monitoring, debuggers, and container support. Measure engineering time as well as tokens per second. A cheaper accelerator can have a higher total cost if a team spends months adapting software or maintaining a separate toolchain.

Revenue claims need the right denominator

AMD reports Data Center segment revenue, but that segment includes EPYC server CPUs and other products in addition to Instinct accelerators. Company statements about annual data-center GPU revenue can be useful, yet they are not a complete product income statement. They also do not reveal customer concentration, discounts, deployment utilization, or the cost of associated development commitments.

AMD’s second-quarter 2025 release reported record company revenue and cited demand for MI350, EPYC, and Ryzen. The filing with the release is a primary source. Its forward-looking language should not be treated as a realized result.

Competition is more than market share

Claims that one supplier controls a precise percentage of the AI accelerator market depend on how the market is defined and are often repeated without a stable source. This profile does not assign a percentage. Nvidia’s established software, networking, and developer ecosystem is a meaningful competitive advantage. Cloud providers’ custom chips and new accelerator companies add other alternatives.

AMD can succeed without displacing the market leader everywhere. Large buyers value supply diversity, negotiating leverage, and architectures optimized for different workloads. The practical goal is to win defined training and inference deployments where price, availability, memory, or system design outweigh migration cost.

A buyer’s evaluation

Run the same production-shaped workload across candidate systems. Include model quality, throughput, time to first token, tail latency, power, failure recovery, compiler stability, engineering labor, and capacity terms. Keep the evaluation harness and results under customer control. NIST’s AI Risk Management Framework can supplement the technical comparison with governance and deployment-risk questions.

Su has given AMD a plausible hardware and software path into AI infrastructure. Public filings confirm growing data-center scale and continued investment. They do not establish that AMD has erased the software gap or that its accelerators offer the best economics for every model. The credible conclusion is narrower: AMD is now a supplier that serious infrastructure buyers should benchmark, not merely mention as leverage.