An invite-only AMD fireside chat for industry analysts last month had a refreshing vibe, signaling a company that no longer thinks in product silos. It framed AMD as a three-pillar business spanning data center, end-user products, and embedded.
That matters because AI is no longer a single market; it is a stack and a deployment pattern. The most interesting line in the discussion may have been the simplest framing of where AI lives across that stack: cloud is for scale, edge is for immediacy, devices for intimacy.
That is a creative way to describe how AMD has evolved over the past several years, and certainly in 2025, into a company that can credibly sell compute from training clusters to inference at the edge to on-device AI PCs.
In the late 1990s and much of the 2000s, AMD often looked poised to invent great products, only to struggle with execution at the pace and consistency the market demanded. There were persistent execution gaps, missed windows, and uneven platform follow-through — just ask Compaq or Dell during that era. The market’s skepticism was palpable, as I can attest firsthand, having worked at Compaq and Dell during AMD’s dark days.
Today, the tone is different. Not because the competition eased, but because AMD’s culture since the arrival of CEO Lisa Su has been far more disciplined in roadmaps, partnerships, and sustained delivery.
One reason the conversation landed is that AMD is openly describing the “how,” not just the “what.”
AMD SVP and GM of Adaptive and Embedded Computing, Salil Raje, put it plainly: AMD “was not traditionally a software company,” but has “turned our attention to software” with ROCm and an explicit push to be “more open source friendly” so the developer community can work with it at scale.
ROCm and the Platform Question
That represents a subtle but massive shift from the old AMD. The old playbook leaned on a sound chip and hoped the ecosystem would follow. AI has taught the entire industry that the ecosystem must be engineered, funded, and obsessively supported. AMD is now talking like a company that understands that rule.
During the call, Raje said AMD was trying to make ROCm feel less like a data center-only toolkit and more like a unifying layer across its portfolio.
In one section on local AI workflows, the speaker describes ROCm as a “unified stack” that is “primarily composed of open-source software,” spanning “low-level kernels to high-level end user applications,” and intended to accelerate workloads “seamlessly between our cloud infrastructure, data center, and personal devices.”
That is the right ambition. It is also where AMD still has something to prove. Nvidia has years of developer muscle memory behind CUDA. AMD has real momentum, but the market will ultimately judge ROCm on the last-mile details: installation friction, framework parity, model zoo maturity — the breadth, quality, and production readiness of supported models — performance stability, and the speed of fixes when something breaks.
AMD highlights recent ROCm performance gains and broader platform support, underscoring progress while leaving last-mile execution as the real test.
Still, AMD’s 2025 announcements show a company moving from a component vendor to an infrastructure participant, which is what AI data centers demand.
It’s All About the AI Data Center
2025 was a transformative year for AMD, particularly in its strategic alliances. Significant partnerships were formed that could shape the future of AI infrastructure, most notably the company’s pact with OpenAI and HPE on the Helios rack-scale architecture.
The OpenAI deal is especially notable. AMD and OpenAI entered a multiyear agreement under which OpenAI will deploy up to 6 gigawatts of AMD Instinct GPUs, beginning with 1 gigawatt of Instinct MI450 GPUs in the second half of 2026, and will include joint technical work to align roadmaps. On the commercial side, AMD also issued OpenAI a warrant tied to deployment and scaling milestones, which signals both commitment and accountability.
Then there is the cloud scaling narrative, which increasingly reads like “rack scale, not just GPU.” For example, Oracle plans to launch a publicly available AI supercluster powered by 50,000 Instinct MI450 Series GPUs in Q3 2026, using the Helios rack architecture with next-gen EPYC Venice CPUs and Pensando Vulcano networking.
Vultr, which provides affordable, high-performance virtual private servers (VPS) for developers and businesses, plans to build a 50 MW AI supercluster in Ohio, powered by 24,000 Instinct MI355X GPUs, with plans to integrate MI450 and Helios.
Furthermore, AMD expanded its collaboration with HPE around Helios, promising high-bandwidth, low-latency rack-scale connectivity. This is the fundamental shift from a legacy bread-and-butter processor company to an AI data center contender. It is not a single GPU. It is CPU, GPU, networking, systems design, and software.
AMD is also using the embedded story to remind us that AI’s next growth wave will not remain confined to hyperscale data centers.
Raje described AMD’s ability to go “from cloud to edge to endpoints” and paired it with a major market claim: an AI infrastructure TAM of “about a trillion dollars in 2030,” with expectations to grow that market at an 80% CAGR over the next three to five years. He then pivoted to embedded and physical AI, arguing that the next inflection is “when the AI moves from the cloud to the physical world,” with billions of devices and a $200 billion TAM by 2035.
The embedded group’s design-win cadence is presented as evidence of customer trust and long-term visibility, including approximately $16 billion in design wins in 2025.
Whether you accept every number, the strategic point remains sound: the company is trying to be relevant in every layer of AI deployment, not just the glamorous training clusters.
AMD’s Lukewarm 2026 CES Announcements
So where do AMD’s CES plans fit in this bigger AI progress story?
Though this may come across as less than enthusiastic, AMD’s CES announcements appeared to be a continuation of its current playbook rather than a surprise pivot.
ROCm 7.2 arrives this month for Radeon and Ryzen on Windows and Linux, with new support for Ryzen AI 400 processors and integrations like Comfy UI.
AMD’s ROCm 7.2 update, announced at CES 2026, reflects continued software execution across Ryzen and Radeon platforms rather than a disruptive platform shift.
In gaming, AMD says it is expanding the Ryzen 9000 X3D portfolio and pushing FSR Redstone across its RX 9000 Series cards, with claims such as “up to 4.7x” performance in a ray tracing example. That will keep AMD highly visible in gaming and reinforce a strength that has been in AMD’s DNA for years.
But frankly, these announcements feel largely incremental, reflecting execution against an established roadmap rather than a step-change moment.
Realistically, companies that want that magical “CES moment” that resets the whole industry’s expectations typically need either a brand-new platform or a new ecosystem announcement.
What AMD announced at CES is undeniably important and will indeed sell real units. It just does not appear to change the competitive physics overnight. The bigger AMD AI story, the one that could change physics, is happening in the data center partnerships and rack-scale systems that start to land in 2026.
AMD’s Positioning in 2026
That leads to the balanced outlook I see for 2026.
On the upside, AMD has positioned itself to participate in the next phase of AI infrastructure buildout in a way it could not have a few years ago. The OpenAI deployment begins in the second half of 2026. Oracle targets Q3 2026 for its MI450-powered supercluster. The Humain joint venture targets a first phase planned for 2026 as part of a longer 1 GW roadmap. These are not small experiments. They are system-scale bets.
On the risk side, 2026 is when execution becomes visible, and visibility cuts both ways. The real tests are straightforward:
- Can AMD deliver MI450-class platforms at the volumes and reliability customers expect?
- Can it keep ROCm improving fast enough that developers feel confidence, not curiosity?
- Can it make Helios- and Pensando-based networking feel like a coherent “AI factory” alternative rather than a collection of parts?
- Can AMD do all of that while Nvidia continues to move the goalposts with relentless cadence, as it will certainly not stand still?
Still, I’m confronted with this nagging question after taking in what AMD has accomplished in 2025 and as it turns its sights on 2026: Is AMD trying to win AI by building the best single chip, or by creating the most adoptable platform?
Discussions with AMD executives, up and down the company’s command chain, suggest the latter. It talks about “AI factories,” rack-scale solutions like Helios, and software that spans cloud to personal devices.
In my view, that is precisely the right lesson from AMD’s own history. In the rougher years, AMD often needed a breakout product to compensate for weaker platform gravity. Under Lisa Su’s leadership, the company has increasingly built durable systems around its silicon. Its 2025 partnership list reads like a continuation of that strategy.
If 2026 delivers on the first wave of these deployments, AMD will look less like an alternative supplier and more like a structural part of the AI data center market. If it stumbles — which I do not expect — the market will not be patient, because AI customers do not buy on hope. They buy predictable outcomes.
AMD’s stellar performance in 2025 suggests it understands that reality. More than any single benchmark chart, that understanding is the clearest sign of how far the company has come — and how decisively it has moved beyond the execution woes of the late 1990s and early 2000s.
The images featured in this article are credited to AMD.
Loading ...
