AMD rewrites its AI roadmap around agentic orchestration and challenges Nvidia's reign in the data center

By Zendoric · July 24, 2026.

At its Advancing AI 2026 conference, Advanced Micro Devices (AMD) executives gave the most detailed roadmap yet of their bet on agentic artificial intelligence (AI) —AI systems capable of chaining together several steps and tools to complete a task without constant supervision—. CEO Lisa Su said the company aims to capture more than 50% of the market for server central processing units (CPUs), relying precisely on the fact that this type of AI demands greater orchestration capacity, the coordination of tasks from start to finish, which falls on the CPU and not only on the graphics processing unit (GPU).

The most concrete figure came from Dan McNamara, head of AMD's server business: in the "outer years" —his way of referring to a horizon several years out, without committing to a date—, agentic AI applications could represent "probably half" of the entire server CPU opportunity, according to his own internal estimate based on conversations with customers. Su went further and suggested that the ratio of GPUs per CPU could be inverted: if today some configurations use up to four GPUs for every CPU, agentic AI could bring that ratio to less than one GPU per CPU, even two CPUs for every GPU in certain scenarios. These are AMD's own projections, not figures audited by a third party, and it is worth treating them as such.

In parallel, AMD confirmed that its Helios systems —full compute racks designed for large-scale AI, not just standalone chips— will begin shipping in September, with a production ramp extending into the first half of 2027. According to Su, the main customers in this phase are Anthropic, OpenAI and Meta. For Anthropic specifically, the first shipments of the "first gigawatt" of committed capacity will start in the first half of 2027, as part of a deal that already envisaged up to 2 gigawatts tied to the MI450 accelerator and that MarketBeat has previously placed at around 5 billion dollars.

The most revealing part is not the chip itself, but the type of relationship AMD says it is building. Su insisted that no major customer chooses an accelerator for a single generation, given the engineering effort required to adapt software to each architecture; that is why AMD is already discussing with Anthropic its plans beyond the MI450, including the future MI500 and MI600. Vamsi Kompella, who leads AMD's AI business, added an important nuance: the company is working directly with Anthropic to tune and optimize the Claude model itself on AMD hardware, and maintains a similar collaboration with OpenAI around Codex. In other words, AI labs are no longer just buying chips: they are co-designing how their software runs on them.

This matters because it shifts the playing field. Over the past few years, Nvidia's advantage lay not only in the performance of its GPUs, but in CUDA, its software ecosystem that tied developers to its hardware. AMD is now betting on ROCm.ai —which Kompella called the company's "biggest software leap" since the start of its AI strategy— and on a more open approach to instruction sets, compilers and tools, precisely to reduce that friction of switching providers. If it works, it would be good news for the industry in general: less dependence on a single compute provider usually translates into more competitive prices and greater room to maneuver for those building with AI, something aligned with the thesis that an abundance of compute resources ends up benefiting more players, not just the giants who arrived first.

AMD was also explicit about the bottlenecks that do not depend on it: electric power, its partners' manufacturing capacity and its own customers' capital. Su said the company already has "between 12 and 18 months" of visibility on energy planning, and that its capacity is sized for "significant" growth in 2027 and 2028; for 2029 and 2030, she admitted, the pace will depend on the entire ecosystem —memory manufacturers, energy providers, integrators like Sanmina and Wiwynn— growing in step. It is a useful reminder: the limit on AI deployment increasingly looks less like an algorithm problem and more like a problem of physical infrastructure, of watts and factories.

Our reading is that this announcement, more than a quarterly-results story, is a signal of where the competition for AI compute is headed. The dominant narrative of the past two years has been "who has the fastest GPU"; AMD instead argues that agentic AI shifts the question toward "who best orchestrates a complete system" —CPU, GPU, networking, memory and software— and that it can compete there without needing to beat Nvidia chip by chip. It is a reasonable bet, but not yet proven: much of what was said at the conference consists of goals and ongoing conversations (MI500, MI600, the relationship with Anthropic "beyond" the MI450), not capacity already delivered. The real thermometer will come in 2027, when it becomes clear whether the Helios shipments and the first gigawatt for Anthropic are delivered on the promised timelines, and whether the promise of more than 50% share in server CPUs survives contact with real energy and factory demand.