AMD

Overview

AMD sells CPUs, GPUs, adaptive silicon and, increasingly, complete rack-scale AI systems. Its Instinct accelerator line is the product that matters for inference, and it occupies a specific commercial position: the alternative that is architecturally similar enough to Nvidia to be a realistic substitute, rather than a bet on a different computing model.

That similarity is deliberate. AMD's pitch to large buyers is not that GPUs are wrong for inference — it is that they should not have to buy every GPU from one company. Concentration risk, pricing power and allocation queues are commercial problems that hyperscalers and large AI labs have strong incentives to solve, and AMD is the most immediately available solution to them.

The strategy has begun to produce results at scale. Data centre revenue reached $5.8 billion in the first quarter of 2026, up 57% year over year, and $6.7 billion in the second quarter, up 107%. Across the first half of 2026 the segment produced $12.5 billion, an 81% increase, driven by EPYC server processors and Instinct MI350 series GPUs.

Chip and product families

The ROCm.ai performance figures are AMD's own published claims measured on its own hardware. They are reported here as vendor claims and have not been independently verified by this publication.

Role in inference

AMD's structural advantage in inference has consistently been memory. Because decode is usually bound by memory bandwidth and capacity rather than arithmetic, an accelerator with more HBM per device can hold larger models — and larger KV caches for long-context workloads — without splitting the model across more devices. Fewer devices per model instance means less communication overhead and, often, better cost per token.

The persistent weakness has been software. ROCm is genuinely open, which matters to a certain class of buyer, but the ecosystem of hand-tuned kernels, serving frameworks and quantisation tooling that grew up around CUDA has no exact equivalent. AMD has spent heavily to close this and the gap is narrower than it was, particularly for mainstream open-weight models served through common frameworks. It is not closed for unusual model architectures or for teams doing deep kernel-level optimisation.

The practical result is a workload-dependent picture. For a team serving a widely-used open-weight model through vLLM or SGLang, AMD is a straightforward option with real cost advantages. For a team writing custom kernels against a novel architecture, the migration cost is still substantial.

Strengths and key questions

Leadership

Dr Lisa Su has been chief executive since 2014 and chair since 2022, and is generally credited with the turnaround that took AMD from marginal to competitive across CPUs and then accelerators. Forrest Norrod leads the Data Center Solutions Business Group, which owns both EPYC and Instinct. Vamsi Boppana leads the AI group, including ROCm and the software effort that most analysts regard as the deciding factor in AMD's inference position. Jean Hu is chief financial officer.

AMD's leadership has been publicly consistent about the diagnosis — that software, not silicon, is the constraint — which is a useful signal when reading the company's investment priorities.

Funding and financials

AMD is a public company reporting on a calendar fiscal year.

AMD does not break out Instinct accelerator revenue separately from EPYC within the data centre segment, so accelerator-specific revenue is not publicly disclosed. Segment figures above include both.

Customers and partners

The most consequential disclosed relationship is with OpenAI. In October 2025 the two companies announced a multi-year, multi-generation partnership under which OpenAI will deploy 6 gigawatts of AMD Instinct GPUs, beginning with the MI450 series and rack-scale systems. AMD issued OpenAI a warrant for up to 160 million shares, vesting in tranches as deployment milestones are reached; the first tranche is tied to an initial 1 gigawatt deployment scheduled to begin in the second half of 2026.

Beyond that, AMD Instinct parts are available through major cloud providers, and Oracle Cloud Infrastructure has been among the more visible large-scale MI355X adopters. AT&T has been cited by AMD as an MI355X customer. Reporting has also described a Meta agreement structured similarly to the OpenAI deal; that arrangement is less thoroughly documented in primary sources and is noted here as reported rather than confirmed.

Technical positioning

AMD sits immediately adjacent to Nvidia on the architecture spectrum — same general computing model, same programming abstractions, similar system topology. This is the source of both its advantage and its ceiling. The advantage is that migration is a software-porting exercise rather than a rethink of how inference is deployed. The ceiling is that AMD cannot claim an order-of-magnitude architectural win the way a specialist such as Etched or d-Matrix can; it competes on price, memory, supply and openness.

The MI400 generation and Helios represent a shift in how AMD frames that competition. Selling a rack rather than a card puts AMD into the same conversation as Nvidia's GB200/GB300 systems, and moves the comparison from per-accelerator specifications to delivered tokens per second per rack per watt — a comparison AMD would not have invited two generations ago.

What to watch

• First OpenAI gigawatt. Whether the initial 1 GW MI450-series deployment lands on schedule in the second half of 2026, and what it reveals about Helios in production.
• Independent inference benchmarks. MI355X and MI455X results measured by third parties, at realistic batch sizes and context lengths, rather than vendor-supplied numbers.
• ROCm release cadence. Time-to-support for newly released frontier model architectures is the cleanest proxy for whether the software gap is closing.
• Cloud instance breadth. How many providers offer Instinct instances on demand, not just through committed contracts.
• Sovereign programmes. MI430X is explicitly aimed at sovereign AI and HPC, a segment where openness and export terms can outweigh raw performance.

Sources

1. AMD — Form 10-Q for the quarter ended 27 June 2026 (SEC EDGAR). Q2 and first-half data centre segment revenue.
2. AMD — Form 10-Q for the quarter ended 28 March 2026 (SEC EDGAR). Q1 data centre segment revenue.
3. AMD — Instinct MI400 Series launch, Advancing AI 2026. MI455X and MI430X positioning, HBM4 and ROCm.ai claims.
4. AMD Investor Relations — AMD and OpenAI strategic partnership. 6 GW commitment, MI450 series, warrant terms.
5. OpenAI — AMD strategic partnership announcement. Counterparty statement of the same agreement.
6. AMD — Fourth quarter and full year 2025 financial results. Full-year 2025 data centre revenue.

Where AMD publishes performance comparisons against its own prior products or competing hardware, those are identified above as vendor claims. See the editorial methodology for the source hierarchy used on this site.

Related pages