Authors: Leo Gergs, Research Director & Larbi Belkhit, Principal Analyst

AMD’s Advancing AI 2026 event in San Francisco last week provided a concrete demonstration of how far the company has progressed over the past 12 months. Said progress was visible not only in the much larger scale of the event compared with last year, but also in the growing maturity of AMD’s overall strategy, spanning both technology and commercialization.

Throughout the event, AMD projected a level of confidence that reflected how far it believes it has come as a company from a technology perspective. The message was clear—AMD believes that it is now a serious leader in Artificial Intelligence (AI) infrastructure, rather than simply a thorn in NVIDIA’s side.

Enter Venice

While Helios and the MI400 series Graphics Processing Units (GPUs) were central to AMD’s messaging at the event, the company also placed notable emphasis on the strategic role of the Central Processing Unit (CPU) in the emerging Agentic AI era. AMD’s argument is that data center CPUs will increasingly segment into three functional tiers: agent sandboxes used to run and coordinate thousands of agents, AI head nodes responsible for feeding and orchestrating GPU racks, and general-purpose CPUs that continue to support traditional enterprise Information Technology (IT) applications and infrastructure.

Against that backdrop, AMD positioned Venice—its first x86 CPU built on a 2 Nanometer (nm) process—as a product family intended to address all three tiers. The company said Venice delivers 1.8X the performance of its Turin CPUs, is already in production, and is scheduled to roll out in 4Q 2026.

The competitive context is also important. Shortly before Advancing AI, NVIDIA published its Vera CPU whitepaper, showing Vera ahead of AMD’s EPYC 9755 by roughly 3% on the SPECrate Integer. AMD used the Venice launch to respond directly, citing the same benchmark and claiming 20% higher per-core performance and 2.2X higher per-socket throughput for Venice relative to Vera. ABI Research anticipates that the move by both AMD and NVIDIA to showcase their CPU performance on industry-standard benchmarks is a welcome and helpful move to advance the industry forward for Agentic AI.

MI400 Series

AMD’s Instinct MI400 Series is the new data center GPU family unveiled at Advancing AI 2026 and the first product line built on the CDNA 5 architecture, which AMD confirmed is manufactured on a 2 nm process node. The lineup splits into two distinct products aimed at different customer groups: the MI455X, targeted at frontier AI training, fine-tuning, and high-volume inference; and the MI430X, aimed at sovereign AI and scientific High-Performance Computing (HPC) workloads, with hardware-based FP64 acceleration alongside standard AI precision support.

Each MI455X carries 432 Gigabytes (GB) of HBM4 memory, roughly a 50% increase over the prior-generation MI355X’s HBM3E capacity, with AMD citing per-GPU memory bandwidth of up to 19.6 Terabytes per Second (TB/s). Meanwhile, the MI430X delivers up to 288 Tera Floating-Point Operations per Second (TFLOPS) of hardware-based FP64 performance for scientific computing. AMD framed this split as a purpose-built strategy rather than stretching a single general-purpose part across all use cases, arguing that frontier labs, sovereign AI programs, and research institutions have sufficiently different infrastructure requirements to justify separate silicon lines rather than one-size-fits-all accelerators.

AMD is positioning the MI430X specifically to target sovereign AI and national HPC buyers, a customer set with different procurement priorities from the hyperscalers and frontier labs that the MI455X is pursuing. Both parts support UALink for scale-up connectivity, giving AMD a consistent interconnect story across the family rather than fragmenting it by product line. Whether this two-part split proves to be a genuine advantage or simply added product complexity will depend on how clearly customers segment between the two. Still, the broader signal from AMD is a wider AI portfolio: purpose-built silicon for at least two distinct classes of AI infrastructure buyer, with the MI455X also serving as the anchor for the Helios rack-scale platform.

Helios

Bringing these component announcements together, AMD announced the production of Helios, its rack-scale reference architecture, at Advancing AI 2026, moving it from roadmap slide to something other companies are now willing to build businesses around. AMD confirmed engineering samples in the second half of this year, with volume production ramping up through 2027, alongside commercial commitments larger than anything the company has previously disclosed for a single platform launch.

AMD confirmed a strategic partnership with Anthropic that includes an equity investment of up to US$5 billion, and between OpenAI, Meta, and Anthropic, it now has roughly 14 Gigawatts (GW) of publicly-committed customer compute. Microsoft Azure and Oracle were named as early Helios customers, and Oracle committed to deploying 50,000 MI450 GPUs on Oracle Cloud Infrastructure (OCI) beginning in 3Q 2026. Separately, Schneider Electric and AMD released their first jointly engineered reference design for deploying Helios, covering racks up to 246 Kilowatts (kW) and clusters up to 10.4 Megawatts (MW) of IT capacity, giving operators a validated blueprint rather than just a specification sheet to plan around.

Moreover, AMD and Cerebras formalized a disaggregated inference partnership leveraging Helios and Cerebras’ Wafer-Scale Engine to split prefill and decode between the 2 platforms, respectively. This follows a wider trend ABI Research has observed in 2026—inference is disaggregated across different compute architectures, not simply different GPU clusters. Cerebras has made similar deals already this year with AWS, but AMD retains its flexibility in observing the next major accelerator architecture to add to its portfolio.

Reaction on the show floor was largely positive, although the rack's double-wide Open Rack Wide (ORW) form factor came up more than once as a point worth watching, because it's a different physical footprint than what most cloud and colocation operators have standardized around. The bigger question sits a level above the rack: how long will centralized, mega-scale systems like Helios stay the default shape of AI infrastructure.

There's a reasonable case that the market fragments over time. Inference increasingly needs to sit closer to where data are generated, pulling some capacity toward the edge rather than into a handful of giant campuses. Grid constraints push in the same direction, particularly in Europe, where power availability and interconnection timelines are already shaping where operators can realistically site multi-hundred-megawatt clusters. None of that undermines the case for Helios today, but it's worth watching as AMD, other semiconductor manufacturers, and their infrastructure partners keep designing around the assumption that bigger, more centralized racks are the durable answer.

ROCm Strengthens

On the software side, one of the more notable announcements from the event was ROCm.ai, which is intended to simplify how developers build, deploy, troubleshoot, and optimize AI workloads across AMD hardware. ROCm.ai combines a Command-Line Interface (CLI), agent skills, and Hyperloom into a more unified software experience. The agent skills extend into coding assistants such as Claude Code, Cursor, and Codex, while Hyperloom provides an optimization layer aimed at end-to-end inference performance through automated profiling, bottleneck identification, targeted tuning, and validation.

From an industry perspective, the significance of ROCm.ai lies less in the individual tools themselves and more in AMD’s attempt to reduce the operational and workflow friction that has historically limited broader adoption of alternative AI software stacks. By embedding AMD-specific guidance into increasingly agent-driven development environments, the company is directly addressing one of CUDA’s long-standing advantages: the perceived cost and complexity of porting, tuning, and maintaining workloads on a non-NVIDIA platform. If AMD can make that transition meaningfully easier, ROCm.ai could become an important enabler of broader ecosystem adoption. That said, the extent of its impact will depend on execution. AMD will need to demonstrate that these tools are reliable in production settings and support them with clear documentation, repeatable deployment examples, and customer evidence that the platform delivers measurable gains in real-world environments.

Cloud Partner Ecosystem Focus

The broader partner ecosystem that AMD assembled was one of the clearest signals from this year’s event. Advancing AI 2026 was visibly larger than the 2025 edition, both in the number of named partners AMD brought on stage and in the breadth of categories they represented, from frontier labs and hyperscalers to specialist neoclouds and Agentic AI-focused providers. That expansion matters as much as any individual deal: a year ago, the ecosystem around AMD’s AI push was thin; this year, AMD had enough committed partners across enough different business models to begin demonstrating a genuine ecosystem rather than a launch-day customer list.

Within the neocloud layer specifically, several distinct positioning strategies emerged. TensorWave has built its identity around being an AMD-exclusive cloud, serving customers developing frontier models and large-scale inference services, with Helios now anchoring that offering. It is a narrower bet than most neoclouds make, but one that gives the company a clearer story to sell than simply offering AMD capacity alongside everything else.

Vultr, by contrast, is incorporating AMD into a broader multi-vendor portfolio rather than building its identity around a single silicon partner, with Helios-based compute opening for pre-order in the fourth quarter as one strong option among several. Crusoe is positioning more narrowly around specific workload categories, with Helios earmarked for Agentic AI use cases.

Taken together, these partners illustrate three different ways of building a business around the same underlying infrastructure shift: exclusive identity, portfolio expansion, and workload specialization, all launched commercially at the same event. That variety is itself a useful signal, suggesting that operators are still experimenting with how best to package and sell this new class of capacity rather than converging on a single obvious model.

Conclusion

The announcements at the show reflected a more confident AMD than in previous years. The combination of product and credible customer announcements reflect how AMD is building a full AI infrastructure strategy spanning compute, software, systems, and commercialization.

The larger question now is how sustained the customer momentum will be for said products, especially around the Helios system and the Turin/Venice CPUs. Proving utilization, efficiency, and the accelerated ROCm update cadence will be a top priority and something that ABI Research will be observing closely. 

Related Research:

Artificial Intelligence and Machine Learning Market Data Update: 2Q 2026 

AI Inference Providers: Market Developments

AI Cloud Workloads Market Data Overview: 2Q 2026