AMD Helios

AMD Helios is a rack-scale artificial intelligence system from AMD that packages 72 Instinct MI455X GPUs and 18 sixth-generation EPYC "Venice" CPUs into a single double-wide cabinet, wired so that the whole rack behaves like one large accelerator rather than a row of separate servers. AMD launched it on 2026-07-23 at Advancing AI 2026 in San Francisco, saying the platform was "now in production to be deployed by leading AI companies at gigawatt scale." ^[1][2][3] A single rack, on AMD's published figures, delivers 2.9 exaFLOPS of peak dense FP4 compute, 1.4 exaFLOPS of peak FP8, 31 TB of HBM4 memory, 1.7 PB/s of aggregate memory bandwidth, 260 TB/s of scale-up bandwidth among the GPUs, and 43 TB/s of Ethernet scale-out bandwidth. ^[1][2][4] Helios is built on Open Rack Wide (ORW), the double-wide standard Meta contributed to the Open Compute Project (OCP), and it is AMD's direct answer to NVIDIA's rack-scale machines, the GB200 NVL72, the GB300 NVL72, and the Vera Rubin NVL72. ^[14][15][6]

AMD first previewed Helios at its Advancing AI event on 2025-06-12, showed a physical reference design at the OCP Global Summit on 2025-10-14, put working hardware on public display at CES 2026 in January, and launched the finished platform about thirteen months after that first preview. ^{[24][14][18][19]} The design connects its accelerators using open industry standards rather than a single vendor's proprietary fabric, and it marks the point where AMD stopped selling AI mostly as chips and boards and started selling it as a full rack you wheel onto a data center floor.

From chips to rack-scale AI factories

For most of the last decade, an AI accelerator was a card or a board. You bought GPUs, put eight of them in a server, and wired many servers together with whatever network you had. That model started to strain as frontier models grew. Training and serving the largest models now spreads a single job across dozens or hundreds of GPUs that have to act like one tightly coupled system, and the limit is often the links between chips rather than the chips themselves. ^[22][6]

The industry's response is the rack-scale system, sometimes marketed as an "AI factory." Instead of treating the server as the unit of sale, vendors treat the whole rack as the product. Every GPU in the rack shares a fast internal fabric, the CPUs and network cards are co-designed with the accelerators, and power and liquid cooling are engineered for the rack as a whole. NVIDIA moved first here with its GB200 NVL72 and the later GB300 NVL72, which link 72 GPUs over NVLink so they look like a single accelerator to software. Helios is AMD's version of that idea. The scale-up domain grows from 8 GPUs on the previous MI355X generation to 72 on Helios, which is the single biggest architectural change in the product. ^[7][20]

This shift changes what buyers compare. A faster GPU helps, but a model trainer cares about how much memory and compute a rack can pool, how fast the GPUs talk to each other inside the rack, and how cleanly many racks scale out into a cluster. Helios is built to be judged on those terms.

What is AMD Helios?

Helios is AMD's first rack-scale AI platform: a liquid-cooled, double-wide rack that combines Instinct MI400 series GPUs, sixth-generation EPYC "Venice" CPUs, and Pensando networking silicon into one coherent system, unified by AMD's open ROCm software stack. ^[1][2] AMD describes the platform as extending its open hardware philosophy "from silicon to system to rack." ^[14]

One point of vocabulary matters, because AMD's own materials are inconsistent about it. The GPU inside Helios is the Instinct MI455X, a specific part number; the MI450 Series is the family name AMD uses in commercial announcements, and several customer press releases say "MI450 Series" where the product pages say MI455X. ^[3][8][12] The MI400 Series AMD launched on 2026-07-23 has two members: MI455X for frontier AI and AI factory deployments, and the MI430X, an FP64-capable sibling rated at up to 288 TFLOPS of hardware FP64 for HPC and sovereign AI. A third part, the Instinct MI440X for on-premises enterprise deployments, was announced as an MI400 Series member at CES 2026 but was not part of the July launch. None of those siblings ships in a Helios rack, and neither does the enterprise-oriented MI350P. ^[3][18][21]

It also matters that Helios is not, strictly speaking, a product AMD sells. AMD's own FAQ describes it as "a reference design, not a product for sale," a blueprint that OEM and ODM partners build into their own branded systems on open ORW standards. ^[4] The systems customers actually buy come from Bull, HPE, Lenovo, Supermicro and others. ^[3]

Inside the rack

The compute tray

Helios arranges its GPUs in two groups of nine compute trays, each 1 OU high, for 18 trays and 72 GPUs. ^[2][7] Each tray holds four MI455X accelerators, one sixth-generation EPYC "Venice" 9006 CPU in the SP7 socket, one AMD Pensando "Salina" DPU at 400 Gb/s for front-end traffic (management, API and storage), and up to twelve Pensando "Vulcano" 800 Gb/s AI NICs. ^[6][7] In the reference configuration reported at launch the host CPU is a 96-core Venice part boosting to about 5 GHz, paired with sixteen 64 GB DIMMs for 1 TB of system memory and five E1.S SSD slots; OEMs and hyperscalers can specify Venice CPUs with up to 256 cores if they want more host compute. ^[6][7] AMD Infinity Fabric carries CPU-to-GPU traffic inside the tray. ^[4]

The MI455X itself is a 24-chiplet package built with TSMC's CoWoS-L packaging. Eight accelerator complex dies on TSMC 2 nm sit atop two 3 nm fabric-and-cache dies, with two further 3 nm dies handling chip-to-chip I/O. ^[5][6][7] Each fabric-and-cache die carries 96 MB of L2 (192 MB per package) and hosts the memory controllers for six of the chip's twelve 36 GB HBM4 stacks. Notably, the two cache domains are not coherent with each other: a work group processor attached to one die cannot read the other die's L2, which makes the part behave more like a chiplet-based EPYC CPU than like previous Instinct GPUs. ^[7] AMD rates the chip at 320 billion transistors, 256 work group processors, a 2.4 GHz peak engine clock, 432 GB of HBM4 and 23.3 TB/s of peak memory bandwidth. ^[5][7]

The scale-up fabric

The fabric that makes 72 GPUs behave as one is UALink over Ethernet (UALoE), which tunnels the UALink protocol across standard Ethernet transport. Each MI455X exposes 36 UALoE interfaces of 400 Gb/s (two 200G lanes each) for 3.6 TB/s of peak bidirectional scale-up bandwidth per GPU. ^[5][7] Six switch trays carry twelve switch ASICs between them; AMD's specification describes each as a 512-lane 200G part with 216 active UALoE links, delivering 21.6 TB/s bidirectional, and twelve of those give the 260 TB/s rack figure. ^[4][7] Those switches are Broadcom Tomahawk 6 silicon rated at 102.4 Tb/s, which is the practical payoff of AMD's open-fabric choice: because the transport is Ethernet, AMD does not need to build a bespoke switch and system integrators can buy merchant Broadcom parts. ^[6] The topology is single-hop all-to-all, so any GPU reaches any other without traversing an intermediate switch. ^[7]

Scale-out and front-end networking

Beyond the rack, Helios uses standards-based Ethernet aligned with the Ultra Ethernet Consortium. Scale-out runs over AMD Pensando Vulcano 800 Gb/s AI NICs, mounted on two boards per tray that can be populated with four or six NICs depending on how much backend bandwidth the customer wants. ^[2][7] Fully populated, each MI455X gets three Vulcano NICs for 2.4 Tb/s of scale-out per accelerator, which works out to the rack-level 43 TB/s AMD quotes. For comparison, The Register noted that NVIDIA's Vera Rubin platform pairs each GPU with a single 1.6 Tb/s ConnectX-9 NIC. ^[6]

Power, cooling and the rack itself

The chassis follows the Open Rack Wide specification, a double-wide cabinet about 1.2 m wide, 1.3 m deep and 44 OU tall, roughly twice the footprint of NVIDIA's NVL72. ^[6][7] The unit there is the OpenU, which is 48 mm rather than the 44.45 mm of a conventional EIA rack unit, so a 44 OU rack stands about 47 ordinary rack units high; writing "44 rack units" understates it. ^[26] Meta introduced ORW and contributed it to the Open Compute Project at the OCP Global Summit in October 2025, for the power, cooling and serviceability needs of next-generation AI systems. AMD did not contribute the rack standard; it built Helios on Meta's open design and aligned it with other open compute standards including OCP DC-MHS, UALink and Ultra Ethernet. ^[14][15] A centralized power shelf distributes power over a vertical busbar, and a cooling manifold feeds coolant to compute and switch trays through quick-disconnect fittings so a sled can be swapped without recabling. ^[4]

AMD has published no official rack power figure. Its Helios blog states plainly that "actual power consumption, thermal design, and system configuration may vary by deployment and are available to qualified customers and partners under a mutual non-disclosure agreement." ^[1] The most specific public number comes from The Register, which reported after AMD's launch briefing that the rack's 50 V liquid-cooled DC busbar can deliver between 225 and 245 kW under load, and estimated (as its own guess, not an NVIDIA disclosure) that Vera Rubin racks would peak somewhere between 240 and 250 kW. ^[6] Separately, Schneider Electric and AMD published a co-engineered data center reference design on 2026-07-23 that supports Helios racks up to 246 kW and modular clusters up to 10.4 MW of IT load, using liquid and hybrid cooling capable of removing up to 84 percent of rack heat and targeting a power usage effectiveness as low as about 1.12 at full load. ^[16] Those two figures are consistent with each other and are the best public evidence for where Helios sits, but neither is an AMD specification.

A moving spec: 19.6 versus 23.3 TB/s

Per-GPU memory bandwidth is the one Helios number that has genuinely changed, and AMD's own website carries both values at once. The MI455X product page, the formal specification tables and the headline spec block on the Helios page all say 23.3 TB/s. ^[4][5] The compute-tray description further down the same Helios page, and the platform FAQ beneath it, both still say "up to 19.6 TB/s," a figure that dates from the pre-launch materials. ^[4] Oracle's October 2025 announcement, written when the part was still called MI450 Series, said "20 TB/s." ^[12]

The 23.3 TB/s figure is the one that reconciles with everything else AMD published at launch. Twelve HBM4 stacks on 2,048-bit interfaces at roughly 7.6 GT/s per pin produce about 23.3 TB/s, and 72 GPUs at that rate give the 1.7 PB/s rack aggregate AMD quotes in both its launch blog and its newsroom release. ^[1][2][7] The older 1.4 PB/s rack figure, which AMD's October 2025 Helios blog still carries and which coverage derived from it repeats, is simply 72 multiplied by 19.6. ^[15] Independent teardown analysis by Chips and Cheese, working from AMD's Advancing AI disclosures, arrives at 23.3 TB/s per GPU and 1.7 PB/s per rack. ^[7] This article uses the launch-day figures and flags the older ones where they appear.

Why does Helios bet on open standards?

The sharpest difference between Helios and NVIDIA's racks is not the silicon. It is the wiring philosophy. NVIDIA's NVL72 systems use NVLink and NVSwitch, interconnects NVIDIA controls end to end. That gives NVIDIA a tightly integrated product, and it gives NVIDIA leverage, because a customer buying into NVLink is buying into one supplier for the fabric. ^[22][20]

AMD took the opposite route across two different jobs. For scale-up, the fast links among GPUs inside a rack, AMD backs UALink, developed by a consortium that includes AMD, Broadcom and others as an open alternative to NVLink; in this first Helios generation that traffic runs as UALink over Ethernet while native UALink switch silicon matures. ^[4][6] For scale-out, the network that joins many racks into a cluster, AMD uses Ultra Ethernet, an effort to tune standard Ethernet for AI traffic, with the Pensando Vulcano NICs as the on-ramp. ^[1][2] Using Ethernet at both layers lets operators reuse familiar tooling and a broad supplier base instead of a proprietary network, and it is why merchant Tomahawk 6 switches can serve as the rack's scale-up fabric. ^[6]

The rack itself is the most pointed part of the strategy. As AMD data center chief Forrest Norrod put it at the OCP announcement, "Open collaboration is key to scaling AI efficiently. With 'Helios,' we're turning open standards into real, deployable systems, combining AMD Instinct GPUs, EPYC CPUs, and open fabrics to give the industry a flexible, high-performance platform built for the next generation of AI workloads." ^[14] Helios also ships defense-in-depth security features including a hardware root of trust, device identity, continuous attestation, hardware-enforced isolation, and encrypted memory and interconnects, which AMD positions as the basis for multi-tenant AI hosting. ^[2][4]

How does Helios compare to NVIDIA's rack systems?

AMD framed Helios against NVIDIA's Vera Rubin NVL72, which arrives in a similar window. Every comparison below is AMD's own, produced by AMD Performance Labs, and each carries a numbered endnote with a test configuration. Against a Vera Rubin NVL72 rack, AMD claims 15 percent more peak FP4 performance, 50 percent more HBM capacity, 6 percent more HBM bandwidth, 50 percent more scale-out bandwidth, and up to 30 percent more tokens per dollar. ^[1][2][3] At the GPU level AMD's published comparison puts MI455X at 432 GB and 23.3 TB/s against Vera Rubin's 288 GB and 22.0 TB/s, and 40 PFLOPS versus 35 PFLOPS of peak FP4. ^[5]

Two qualifications matter. First, the widely quoted "up to 30 percent more tokens per dollar" is endnote MI400-025, and it is partly a pricing model rather than a benchmark: AMD calculated it on the Kimi K2 Thinking workload at 32K input and 8K output, using estimated aggregate throughput across low, medium and high interactivity operating points and projected hourly GPU pricing based on market conditions. ^[1][3] Change the assumed cloud price and the figure moves. Second, The Register reported that Vera Rubin's adaptive compression is expected to give NVIDIA roughly a 25 percent lead over Helios at FP4 for inference workloads that can use it, with AMD's 15 percent peak FP4 advantage applying to those that cannot. ^[6]

No independent benchmark results for Helios existed as of 2026-07-27. AMD's MLPerf Inference 6.0 submission was built on MI355X GPUs, and AMD's own write-up describes those multinode results as "a stepping stone" toward Helios rather than a measurement of it. ^[17] Neither AMD nor NVIDIA had submitted rack-scale results for their 2026 platforms. The Register also cautioned, correctly, that peak theoretical FLOPS are a mathematical ceiling derived from operations per clock at a fixed frequency, while real accelerators follow boost curves tied to power and thermals and real workloads rarely reach maximum achievable matrix throughput. ^[6] The honest summary is that AMD claimed parity or better on paper and has yet to prove it in deployment.

Vendor performance claims and their test configurations

Scroll sideways for more →

At the OCP Global Summit in October 2025, before final specifications were set, AMD's Helios blog also claimed the rack would deliver "up to 36x higher performance compared to previous generations." Its footnote (MI350-047A) sets a 72-GPU MI450 Series rack against an eight-GPU MI355X platform, so most of the multiple is the nine-fold jump in GPU count rather than a per-chip gain. The figure predates the launch numbers above and should not be read as a shipping measurement. ^[15]

Specifications

Unless noted otherwise, every figure below is vendor-reported by AMD, drawn from its launch blog, its newsroom release, its Helios product page and the MI455X specification page, with system-level detail from The Register's launch briefing coverage and the Chips and Cheese architectural analysis. ^{[1][2][4][5][6][7]}

Scroll sideways for more →

Customers, partners and availability

AMD named a long list of adopters at launch: OpenAI, Anthropic, Meta, Microsoft, Oracle, HUMAIN, TensorWave, Vultr and Cirrascale. Systems will be built by OEMs including Bull (Eviden), HPE, Lenovo and Supermicro, with infrastructure partners Sanmina and Wiwynn. AMD's Helios page carries a wider partner wall that adds Dell, Aligned, OneQode, River and Amp to that list, and links a separate Celestica collaboration announcement. ^[3][4][25] The timelines differ meaningfully by customer and should not be collapsed into a single date.

Scroll sideways for more →

The Anthropic agreement, announced 2026-07-22, is the newest and the most unusual. Alongside the hardware commitment, AMD said it "has committed to make a strategic equity investment of up to $5 billion in Anthropic in the future," and the two companies launched a multi-year engineering collaboration to use Claude to optimize workloads for Instinct GPUs and accelerate ROCm development, with AMD adopting Claude across its engineering teams. ^[8] AMD's release did not describe the investment's structure; CNBC and Reuters reported that the payments are tied to undisclosed deployment milestones rather than paid up front. ^[9] Anthropic already runs AMD's current-generation MI355X GPUs. ^[8] Separately, Cerebras and AMD announced a combined offering pairing Cerebras ultra-low-latency inference with Helios high-throughput rack-scale capacity. ^[3]

On availability, three different statements are in circulation and they mean different things. AMD's launch release says Helios is "now in production." ^[3] AMD's Helios FAQ, which was not updated for launch, still says the reference design "is being shared with partners now, with volume deployments expected in 2H 2026." ^[4] And the Q4 2026 date that circulated widely after the launch is specifically OpenAI's expectation for when it will bring Helios online, not an AMD ship date. ^[3] AMD had held to a 2H 2026 window since at least February 2026, so the platform arrived roughly on the schedule it set. ^[23] As of 2026-07-27 no customer had reported a Helios rack running production workloads.

Why does Helios matter for AMD?

Helios is the centerpiece of AMD's attempt to win real share in data center AI, a market NVIDIA has dominated. Selling racks rather than chips changes AMD's position in two ways. It raises the value of each deal, since a rack bundles GPUs, CPUs, NICs and integration work. It also makes AMD a credible single-vendor option for an operator that wants a turnkey AI cluster, which until now usually meant going to NVIDIA. ^[22][20] The commercial commitments are large and they come from buyers who also lean heavily on NVIDIA, which suggests those buyers want a second credible supplier. The open-standards approach reinforces that, because UALink, Ultra Ethernet and a Meta-derived OCP rack lower the cost of running AMD gear next to everything else in the building.

AMD used the launch to lay out a multi-year cadence rather than a single product. Next-generation Instinct MI500 Series GPUs are due in 2027, paired with a Helios 500 rack using EPYC "Verano" CPUs and Pensando "Como" and "Monza" networking; a Helios 600 follows with MI600 Series GPUs, EPYC "Ferrara" CPUs and Pensando "Palma" and "Levanzo" parts. ^[3] Annual GPU generations tied to matching rack, CPU and networking refreshes is the same rhythm NVIDIA established, and holding to it is arguably a bigger test for AMD than any single specification.

Limits and open questions

The headline numbers are AMD's own projections and calculations for a platform with no shipping deployments and no third-party measurements. Several specific risks remain open as of 2026-07-27.

Software is the largest. ROCm has historically trailed NVIDIA's CUDA in maturity and breadth, and rack-scale performance depends as much on the software stack as on the silicon; AMD's decision to spend engineering effort (and up to $5 billion) on an Anthropic partnership aimed partly at accelerating ROCm development is a tacit acknowledgement of the gap. ^[8][9] The open interconnects are also still early: Helios runs UALink tunnelled over Ethernet rather than native UALink hardware, so the interoperability promise remains partly aspirational until several vendors ship compatible parts. ^[6] AMD's own materials carry inconsistent memory-bandwidth figures, which is a small thing but a reminder that specifications were still settling at launch. ^[4] And a 225 kW to 245 kW rack is a facilities problem before it is a compute problem, which is why the Schneider Electric reference design exists at all; the number of sites that can take delivery of one is smaller than the number that would like to. ^[6][16]

Helios is a clear statement of direction for AMD's AI data center strategy, and its standing against NVIDIA will be settled by 2027 deployments and independent measurements rather than by launch-day slides.

See also

• AMD Instinct MI455X
• AMD Instinct MI400
• AMD Instinct MI430X
• AMD Instinct MI350P
• AMD Instinct MI355X
• AMD EPYC Venice
• AMD Pensando
• AMD Advancing AI 2026
• NVIDIA GB200 NVL72
• NVIDIA Vera Rubin
• Broadcom Tomahawk 6
• UALink
• Ultra Ethernet
• ROCm

References

1. ^Jerry Zhang and Chris Raymond. "AMD Launches Helios: The Highest Performing Rackscale AI Infrastructure Solution." AMD blog, 2026-07-23. amd.com/...ng-rackscale-ai-infrastructure-solution
2. ^AMD. "AAI 2026: AMD Launches AMD Helios Rackscale Solution for Frontier AI." AMD Newsroom, 2026-07-23. newsroom.amd.com/...aai-2026-helios-update
3. ^AMD. "AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era." AMD Investor Relations, 2026-07-23. ir.amd.com/...stack-compute-for-the-agentic-ai-era
4. ^AMD. "AMD Helios Rackscale Solution: Powering Frontier AI." AMD product page, accessed 2026-07-27. amd.com/...helios
5. ^AMD. "AMD Instinct MI455X GPUs." AMD product page, accessed 2026-07-27. amd.com/...mi455x
6. ^Tobias Mann. "AMD attacks the rack with Helios systems that rival Nvidia's." The Register, 2026-07-23. theregister.com/...5277246
7. ^Chips and Cheese. "AMD's Instinct MI455X: Aiming for the Sun." 2026-07-23, edited 2026-07-25. chipsandcheese.com/...stinct-mi455x-aiming-for-the
8. ^AMD. "AMD and Anthropic Announce Strategic Partnership to Deploy Up to 2 Gigawatts of AMD Instinct MI450 Series GPUs." AMD Investor Relations, 2026-07-22. ir.amd.com/...ts-of-amd-instinct-mi450-series-gpus
9. ^CNBC. "AMD to invest up to $5 billion in Anthropic as part of computing power deal." 2026-07-22. cnbc.com/...amd-anthropic-ai-chip-investment
10. ^AMD. "Microsoft to Deploy Next-Gen AMD Instinct and AMD EPYC Processors as the Companies Expand Their Long-Term Strategic Partnership." AMD Investor Relations, 2026-07-20. ir.amd.com/...heir-long-term-strategic-partnership
11. ^AMD. "AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs." AMD newsroom, 2025-10-06. amd.com/...nai-announce-strategic-partnership-to-d
12. ^Oracle. "Oracle and AMD Expand Partnership to Help Customers Achieve Next-Generation AI Scale." Oracle news, 2025-10-14. oracle.com/...-next-generation-ai-scale-2025-10-14
13. ^AMD. "AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs." AMD newsroom, 2026-02-24. amd.com/...a-announce-expanded-strategic-partnersh
14. ^AMD. "AMD Showcases 'Helios' Rack-Scale Platform Built on the Open Compute Project Open Rack for AI, Introduced by Meta." AMD newsroom, 2025-10-14. amd.com/...ases-helios-rack-scale-platform-built-o
15. ^AMD. "AMD 'Helios': Advancing Openness in AI Infrastructure Built on Meta's 2025 OCP Open Rack for AI Design." AMD blog, 2025-10-14. amd.com/...-ai-rack-built-on-metas-2025-ocp-design
16. ^Schneider Electric and AMD. "Schneider Electric and AMD release first Helios platform reference design to accelerate AI Factory deployment." GlobeNewswire, 2026-07-23. globenewswire.com/...elerate-AI-Factory-deployment
17. ^AMD. "AMD Delivers Breakthrough MLPerf Inference 6.0 Results." AMD blog, 2026-04-01. amd.com/...eakthrough-mlperf-inference-6-0-results
18. ^AMD. "AMD and its Partners Share their Vision for 'AI Everywhere, for Everyone' at CES 2026." AMD newsroom, 2026-01-05. amd.com/...s-partners-share-their-vision-for-ai-ev
19. ^ServeTheHome. "AMD's EPYC Venice, Instinct MI455X, & Helios Hardware On Display for First Time at CES 2026." January 2026. servethehome.com/...lay-for-first-time-at-ces-2026
20. ^Anton Shilov. "AMD debuts Helios rack-scale AI hardware platform at OCP Global Summit 2025, promises easier serviceability and 50% more memory than Nvidia's Vera Rubin." Tom's Hardware, 2025-10-14. tomshardware.com/...memory-than-nvidias-vera-rubin
21. ^AMD. "AAI 2026: AMD Launches AMD Instinct MI400 Series GPUs for Frontier AI, HPC." AMD Newsroom, 2026-07-23. newsroom.amd.com/...aai-2026-mi400-instinct-update
22. ^Timothy Prickett Morgan. "Oracle First In Line For AMD 'Altair' MI450 GPUs, 'Helios' Racks." The Next Platform, 2025-10-14. nextplatform.com/...altair-mi450-gpus-helios-racks
23. ^Timothy Prickett Morgan. "AMD Says 'Helios' Racks And MI400 Series GPUs On Track For 2H 2026." The Next Platform, 2026-02-23. nextplatform.com/...ries-gpus-on-track-for-2h-2026
24. ^AMD. "AMD Unveils Vision for an Open AI Ecosystem, Detailing New Silicon, Software and Systems to Power the Next Era of AI." AMD newsroom, 2025-06-12. amd.com/...s-vision-for-an-open-ai-ecosystem-detai
25. ^Anton Shilov. "HPE adopts AMD's Helios rack architecture for 2026 AI systems." Tom's Hardware, 2025-12-02. tomshardware.com/...chitecture-for-2026-ai-systems
26. ^Open Compute Project. "Open Rack Wide (ORW) Meta Design Specification, v1.0.1." OCP Rack and Power Project, 2025. opencompute.org/...-specification-v1-0-1-final-pdf