SerpApi.Org
AMD and Microsoft on Sunday announced a sweeping expansion of their AI infrastructure partnership, with Azure committing to deploy the AMD Helios Rackscale Solution at scale for frontier model inference — the kind of comprehensive, production-scale commitment that elevates Helios from a promising competitor into a verified hyperscaler platform. The timing is deliberate: AMD's Advancing AI 2026 conference opens at San Francisco's Moscone Center tomorrow, with CEO Lisa Su delivering the keynote at 9:30 AM Pacific Time on Thursday. Microsoft's confirmation arrived first, as a declaration, not a preview, according to AMD's official July 20 announcement.
The strategic significance runs deeper than the customer list. Both AMD and Microsoft are founding members of the UALink Consortium — the industry group that wrote the open-standard interconnect specification Helios is built on — as established when the UALink Promoter Group formed in May 2024. Azure deploying Helios is not simply one hyperscaler selecting a GPU vendor. It is two co-authors of an open interconnect standard validating the architecture they designed together, in the same procurement window where Nvidia's proprietary NVLink fabric still controls the rack-scale AI market.
What Helios Is and Why the Standard Matters
Helios is AMD's first rack-scale AI system: 72 Instinct MI455X GPUs, 6th-generation EPYC Venice CPUs, Pensando networking, and the ROCm software stack, all housed in an Open Compute Project-compliant chassis with liquid cooling, as detailed in AMD's CES 2026 specifications. Each MI455X carries 432 gigabytes of HBM4 memory, giving a single Helios rack 31 terabytes of aggregate high-bandwidth memory at 1.4 petabytes per second of combined memory bandwidth — enough to hold a multi-hundred-billion-parameter model's weights and a substantial inference cache on a single rack without tensor parallelism across multiple systems, as covered by TechTimes' prior EPYC Venice launch coverage.
AMD rates the rack at 2.9 exaFLOPS of FP4 inference compute and 1.4 exaFLOPS at FP8 precision. Those are AMD's own figures, not yet confirmed by independent MLPerf benchmarks. Lisa Su's keynote Thursday is expected to address whether those numbers hold in production inference conditions.
The structural argument for Helios is not primarily about peak throughput. It is about what connects the GPUs. The Helios scale-up fabric runs on UALink-over-Ethernet — AMD's open-standard alternative to Nvidia's proprietary NVLink. In initial H2 2026 deployments, AMD uses a Broadcom-co-designed Ethernet switch fabric rather than native UALink switching silicon, because the purpose-built UALink switches from partners like Astera Labs are not production-ready until 2027. That implementation detail matters: the 260 terabytes per second of aggregate intra-rack bandwidth AMD claims is directly comparable to Nvidia's NVLink72 at 259 TB/s on paper, but the real-world routing performance for mixture-of-experts model architectures — where AI tokens are constantly dispatched between specialized sub-networks — remains unverified in published benchmarks, as analyzed in TechTimes' earlier Helios versus Vera Rubin comparison.
The longer-term structural advantage is what matters to Microsoft as a procurer at scale: NVLink switches must be purchased exclusively from Nvidia and are incompatible with any other vendor's fabric. UALink, built on open Ethernet PHY standards, can be sourced from multiple hardware partners. A Helios customer is still making a platform commitment — the MI455X's HBM4 memory layout uses a proprietary 2,048-bit routing interposer that is incompatible with Nvidia server chassis — but the interconnect layer itself is not a single-vendor item.
How Azure Deploys Helios Across Three New VM Families
Microsoft's commitment covers three workload categories, each assigned its own VM series.
The ND MI455X v7 virtual machine series is the inference platform — Helios-powered, targeting reasoning, search, and agentic workloads behind modern AI services. Pricing, per-VM GPU allocation, and regional availability dates have not been announced. AMD will begin shipping Helios systems to Microsoft in the second half of 2026; Azure VM availability will follow integration and validation by some additional margin, likely putting meaningful capacity availability in early 2027 for most regions, per the AMD partnership announcement.
The Azure HDv2 is co-designed for agentic AI and data pipeline work — the CPU-intensive orchestration, retrieval, and preprocessing that feeds GPU clusters. Each instance offers approximately 500 physical EPYC Venice cores, 4 terabytes of RAM, 32 terabytes of local NVMe storage, and 400 gigabits per second of Azure Boost networking. EPYC Venice is the first high-performance computing processor on TSMC's 2nm nanosheet process, offering improved performance per watt at the transistor level before microarchitecture improvements are counted, as detailed in TechTimes' EPYC Venice launch analysis.
The Azure HXv2 targets semiconductor design, EDA simulation, and scientific computing. It offers 176 EPYC Venice cores at clock speeds above 5 GHz, 50 percent more addressable cache per core than the prior generation, up to 4 terabytes of RAM, and 800 gigabits per second of InfiniBand connectivity, per the Microsoft Azure infrastructure announcement. Synopsys, whose customers use Azure for chip design cycles, is already integrated into the HXv2 roadmap, with Synopsys Chief Product Development Officer Shankar Krishnamoorthy quoted in the Microsoft blog about enabling EDA workloads at cloud scale.
Alongside these VM families, AMD's Pensando DPUs are being integrated more deeply with Azure Boost — Microsoft's hardware-and-software offload layer that handles networking, storage, and virtualization at the edge of the rack, as confirmed in AMD's announcement. The result is AMD silicon at every functional layer of Azure's AI compute stack: GPU accelerator, CPU orchestration, and DPU networking.
Why the Interconnect Architecture Determines Who Wins at Scale
The importance of Microsoft and AMD's shared UALink membership goes beyond cost and flexibility. When a hyperscaler's infrastructure team evaluates whether to commit hundreds of megawatts of data center capacity to a single vendor's interconnect standard, the question is not just price — it is whether the vendor retains unilateral pricing power over a component that is physically impossible to replace without a full platform teardown.
NVLink is that component in Nvidia's architecture. It is custom silicon, sold exclusively by Nvidia, required for any rack-scale deployment of Nvidia GPUs, as documented in TechTimes' Helios versus Vera Rubin analysis. UALink, ratified as a public specification in 2025 and updated in April 2026, supports any conforming hardware — the standard is now publicly available and covers multiple new technical areas since its original April 2025 ratification. Microsoft and AMD helped write that specification. Azure deploying a system built on it is the validation event the UALink Consortium was designed to produce.
For a reader evaluating AI cloud infrastructure: if you run large-scale inference on Azure ND-series VMs today, the ND MI455X v7 will eventually be a deployable option on the same platform. If your organization is evaluating agentic AI infrastructure, the HDv2's density of EPYC Venice cores addresses the orchestration bottleneck that GPU-heavy clusters often encounter. Neither option is available to book yet — but the procurement planning window has opened.
Where AMD Genuinely Leads and Where Nvidia Still Leads
The memory capacity advantage is real and consequential. Helios carries 31 terabytes of HBM4 per rack versus the Nvidia Vera Rubin NVL72's 20.7 terabytes — a roughly 50 percent advantage, as measured in TechTimes' head-to-head comparison. For frontier model inference on systems that must keep a trillion-parameter model's weights and KV cache on a single rack, that difference can determine whether a workload requires one rack or two.
On raw FP4 throughput at the rack level, Nvidia's Vera Rubin NVL72 currently leads: 3.6 exaFLOPS versus Helios's 2.9 exaFLOPS. For training workloads in FP8 precision, the gap widens — 2.5 exaFLOPS for Vera Rubin versus 1.4 exaFLOPS for Helios — making Helios an inference-optimized platform that does not challenge Nvidia directly in large-scale model training, per TechTimes' benchmark comparison.
Power consumption favors AMD. Helios draws approximately 140 kilowatts per rack. Vera Rubin NVL72, the current-generation Nvidia rack-scale system, draws approximately 190 to 230 kilowatts per rack, as reported by ModulEdge in TechTimes' Helios comparison coverage — a meaningful difference for data centers managing power density constraints.
What ROCm Ecosystem Maturity Means for Real Deployment
The hardware comparison above is a specification exercise. The practical question for a developer or infrastructure buyer is whether AMD's software stack can match Nvidia's 18-year CUDA head start closely enough for production workloads.
In 2026, the honest answer is: for standard LLM inference using PyTorch and vLLM, ROCm-based AMD GPUs reach roughly 90 to 95 percent of H100 throughput on comparable hardware, per analysis from Spheron comparing UALink vs. NVLink deployment scenarios. ROCm 7.2.4, released in May 2026, is officially supported as a first-class compute backend in PyTorch 2.7.0 — not experimental, not community-maintained, as confirmed in TechTimes' AMD software ecosystem coverage.
The remaining gaps matter for specific workloads. TensorRT-LLM, Nvidia's optimized inference engine, has no full ROCm equivalent. FlashAttention 3, which extracts significant throughput from Nvidia's Hopper tensor cores, is architecture-specific and unavailable on AMD. Custom CUDA kernels — the ones enterprises spend years tuning — require a porting effort that ranges from minimal to substantial depending on kernel complexity, as detailed in TechTimes' ROCm gap analysis. An analysis published in June 2026 estimated that Nvidia's software ecosystem advantage translates to a performance equivalent of 30 to 99 percent additional hardware capability in optimized production deployments.
The AMD thesis for enterprise buyers is that inference workloads are the fastest-growing share of AI compute spending, that ROCm is sufficient for most production inference in 2026, and that the 15 to 30 percent cost advantage AMD hardware commands versus comparable Nvidia hardware on current cloud markets more than compensates for the ecosystem friction for those workloads, per Spheron's AMD-Nvidia deployment cost comparison.
Market Reaction and the Path to August
AMD shares rose approximately 5 percent intraday on July 20, reaching a session high near $528, as tracked by Yahoo Finance. The announcement arrived alongside analyst upgrades: Rosenblatt raised its target to $665 from $490, calling AMD its top semiconductor long idea into the August 4 earnings call, per TipRanks' analyst coverage. UBS lifted its target to $700 from $670, citing AMD's CPU and GPU roadmap strength and data center checks, with UBS analyst Timothy Arcuri and KeyBanc's John Vinh also raising his target to $725, projecting AI GPU revenue across AMD's full lineup at $16.8 billion in 2026 and $48.5 billion in 2027. Goldman Sachs had raised its target to $640 in a prior session.
AMD's Q2 2026 earnings on August 4 will be the first opportunity to quantify the Helios revenue ramp. Q1 2026 data center revenue reached $5.8 billion — up 57 percent year-over-year — and AMD guided Q2 total revenue at approximately $11.2 billion, according to CNBC's coverage of the July 20 announcement.
The Helios revenue story is a 2026 tail and 2027 story, not a Q2 event. AMD expects to ship systems in the second half of 2026, but volume is constrained: all 2026 HBM4 production is reportedly allocated to hyperscalers, limiting general availability beyond engineering samples through the end of this year, per TechTimes' supply chain analysis.
Microsoft's Helios commitment adds to an already confirmed roster: Meta, which announced a commitment covering up to 6 gigawatts of AMD GPUs — with 1 gigawatt of Helios racks targeted for this year — OpenAI, Oracle, and Tata Consultancy Services, as reported by CNBC.
Lisa Su is expected to address Helios deployment timelines, ROCm ecosystem progress, and likely the first detailed disclosure of MI500 architectural specifications during Thursday's keynote. In-person registration at the Moscone Center is full; AMD will stream the keynote on its YouTube channel for remote attendees.
Frequently Asked Questions
What is AMD Helios, and how is it different from a regular GPU?
AMD Helios is a rack-scale AI system that integrates 72 Instinct MI455X GPU accelerators with EPYC Venice server CPUs, Pensando networking, and the ROCm software stack into a single liquid-cooled unit. Unlike purchasing individual GPUs, Helios delivers a complete, pre-integrated platform where the interconnect fabric, compute, and networking are co-designed for large-scale AI inference. The key differentiator from Nvidia's competing Vera Rubin NVL72 is that Helios uses UALink-over-Ethernet — an open standard co-developed by AMD, Microsoft, and others — rather than Nvidia's proprietary NVLink switching hardware, which can only be purchased from Nvidia, as covered in TechTimes' Helios versus Vera Rubin analysis.
When will AMD Helios be available on Azure, and what should developers do now?
AMD expects to begin shipping Helios hardware to Microsoft in the second half of 2026, but Azure general availability for the ND MI455X v7 VM series will follow integration and validation, most likely in early 2027 for most regions, per AMD's partnership announcement. The HDv2 and HXv2 EPYC Venice VM families are also planned for H2 2026 without firm dates. Developers who want to evaluate AMD infrastructure now can access the current ND MI300X v5 series on Azure. If you plan to migrate to Helios, the best preparation is ensuring your inference pipeline is portable across compute backends — specifically, that it does not depend on CUDA-specific libraries like TensorRT-LLM or FlashAttention 3 that currently have no ROCm equivalents.
Does AMD's open-standard interconnect actually break Nvidia's hold on AI infrastructure?
At the rack-scale level, the open interconnect claim is real in structure: UALink specifications are publicly available, and multiple vendors can build conforming switches — unlike NVLink, which is exclusively a Nvidia product. In practice, Helios's initial H2 2026 systems use UALink-over-Ethernet rather than native UALink silicon, because the purpose-built UALink switching chips from partners like Astera Labs are not production-ready until 2027. So the openness is structural and contractually significant for hyperscalers, even if the specific implementation in 2026 leans on Ethernet PHY hardware as a bridge. The deeper competitive question is the software ecosystem: CUDA's 18-year advantage means that benchmark-matching hardware alone does not instantly break Nvidia's developer-level hold. Microsoft's and AMD's shared co-authorship of the UALink standard suggests they built this architecture specifically to have that structural answer ready — even if it takes another product cycle or two to manifest fully in market share.
How does ROCm compare to CUDA for developers evaluating AMD Helios in 2026?
For standard large language model inference using PyTorch and vLLM, ROCm on AMD's current Instinct GPUs reaches approximately 90 to 95 percent of equivalent Nvidia H100 throughput. ROCm 7.2.4, released in May 2026, supports PyTorch as a first-class compute backend, as documented in TechTimes' AMD software ecosystem coverage. The remaining gaps affect specific workloads: TensorRT-LLM (Nvidia's optimized inference engine), FlashAttention 3 (Hopper tensor-core-specific), and any custom CUDA kernels your team has written. If your production inference stack runs on PyTorch and vLLM with standard model architectures, AMD infrastructure in 2026 is a viable option worth benchmarking. If your pipeline depends on TensorRT-LLM or custom CUDA kernels, budget for a porting effort before treating AMD hardware as a drop-in replacement.
Related Articles
