AI Wiki

AMD Advancing AI 2026

AI CompaniesAI HardwareAI Infrastructure
21 min read
Updated
Suggest editHistoryTalk
RawGraph

Last edited

Fact-checked

Sources

31 citations

Revision

v2 · 4,145 words

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Advancing AI 2026 (styled AAI 2026 in the company's own materials) was an annual product and strategy conference held by AMD in San Francisco on July 22 and 23, 2026.[23] It was the fourth event in AMD's "Advancing AI" series and the first to run across two days. Chair and CEO Lisa Su delivered the main keynote on July 23, launching the Helios rackscale AI system, the Instinct MI455X GPU, and 6th Gen EPYC "Venice" server processors, alongside embedded and robotics silicon and an AI-assisted software toolchain called ROCm.ai.[1]

The event's framing device was the "agentic AI era": AMD's argument that inference driven by autonomous agents shifts data center economics toward memory capacity, tokens per second per dollar, and tightly coupled CPU-plus-GPU systems rather than raw training FLOPS. The corresponding pitch was that AMD could now supply a whole rack, CPU, GPU, DPU and software, as an integrated product rather than as components.[1][2]

Every performance figure AMD presented was generated or estimated by AMD Performance Labs. As of 2026-07-25 no independent third-party benchmarking of Helios, the MI455X, or 6th Gen EPYC had been published. Analysts covering the event were broadly positive on AMD's strategic position while stating that the numbers remained unverified.[2][3]

Background: the Advancing AI series

AMD has used "Advancing AI" as the umbrella brand for its flagship AI hardware launches since 2023, on a roughly annual cadence alternating between San Jose and San Francisco.

EventDateLocationHeadline launches
Advancing AI 20232023-12-06San Jose, CAInstinct MI300X, MI300A, ROCm 6
Advancing AI 20242024-10-10San Francisco, CAInstinct MI325X, 5th Gen EPYC 9005 "Turin", networking and AI PC updates
Advancing AI 20252025-06-12San Jose, CAInstinct MI355X and the MI350 series, first public disclosure of the Helios rack concept and the MI400 series
Advancing AI 20262026-07-22 to 2026-07-23San Francisco, CAHelios in production, Instinct MI455X, 6th Gen EPYC "Venice", Ryzen AI Embedded X100, Kria AI robotics platform, ROCm.ai

Sources: AMD press releases and event pages for each year.[1][4][5][6]

The 2026 event was the delivery date for commitments made at the 2025 event. Helios and the MI400 series were announced as future products in June 2025 with a 2026 target; Advancing AI 2026 was where AMD had to show they existed.[6]

What was announced

AMD's press release mixes products that are shipping, in production, sampling, and several years out on a roadmap. The table separates them. Availability language is AMD's own except where noted.

CategoryProductStatus as stated at the event
Rackscale systemAMD Helios (72x MI455X plus 18x 6th Gen EPYC "Venice")"Now in production to be deployed by leading AI companies"; OEM systems from Bull, HPE, Lenovo and Supermicro; infrastructure partners Sanmina and Wiwynn. Customer deployments begin Q4 2026.[1]
Data center GPUInstinct MI455XIn production; the compute element of Helios. Part of the MI400/MI450 series.[1]
Data center GPUInstinct MI430XAnnounced for HPC and sovereign AI, with up to 288 TFLOPS of hardware FP64. AMD's own MI400 series FAQ commits only to "expected to be available in 2027" and names no half or quarter.[1][26]
Data center GPUInstinct MI350PPCIe card, 144 GB HBM3E, CDNA 4. This part was introduced on 2026-05-07, eleven weeks before the event, and had already appeared at Computex, Dell Technologies World and HPE Discover. It was re-promoted at AAI 2026 with new AMD benchmark claims rather than debuted there.[1][8][9]
Server CPU6th Gen EPYC "Venice" (EPYC 9006 series)In production, but announced rather than shipping. SP7 systems are not available until Q4 2026, and AMD shared no SKU tables at the embargoed briefings, which Phoronix called "more of a soft launch". Flagship EPYC 9996 with 256 cores and 512 threads across eight 32-core CCDs, up to 1 GB of L3, 16 memory channels at 12.8 GT/s, PCIe Gen 6, TSMC 2 nm. Those 256 cores are the dense Zen 6c variant rather than the high-frequency Zen 6 core, per AMD's own briefings; AMD's published 9006 product pages carry no core-variant label at all, and the one AMD page that names an architecture alongside the core count, the Helios page, says "Zen 6" with up to 256 cores.[1][10][11][31]
NetworkingAMD Pensando "Vulcano" 800 Gbps NICs inside HeliosShipping as part of Helios. Helios uses UALink over Ethernet for scale-up rather than a proprietary fabric.[1][7]
Edge and roboticsRyzen AI Embedded X100 SeriesAnnounced. Sampling since June 2026; mass production scheduled for Q4 2026, with supply committed through 2037.[12][13]
Edge and roboticsKria AI SOMs and Kria AI Robotics Developer PlatformSOMs "expected to be available from ODM partners starting in the fourth quarter of 2026"; the developer platform is "sampling with early access customers" with general availability expected in Q4 2026.[12]
SoftwareROCm.ai, ROCm CLI, AMD Skills, and the open-source Hyperloom optimization agentIntroduced at the event; Hyperloom published as open source.[14][15]

Helios

Helios was the centerpiece of the event, and it is worth being precise about what AMD is actually selling. AMD's own Helios FAQ says the rack "is a reference design, not a product for sale," a blueprint that OEM and ODM partners build their own branded systems against.[25] Each rack holds 18 liquid-cooled compute trays, each pairing four MI455X GPUs with one 6th Gen EPYC "Venice" CPU, for 72 GPUs and 18 CPUs per rack. It uses the Open Compute Project Open Rack Wide form factor, the double-wide standard Meta contributed to OCP, measuring roughly 1.2 m wide and 44 OpenU tall (about 47 conventional rack units).[7][25]

AMD's launch figures for a full rack are 31 TB of HBM4, 2.9 exaFLOPS of peak OCP MXFP4, 1.4 exaFLOPS of peak OCP MXFP8, 1.7 PB/s of aggregate HBM bandwidth, 260 TB/s of scale-up bandwidth and 43 TB/s of scale-out bandwidth.[24] A figure of 1.4 PB/s for rack memory bandwidth is in wide circulation and is wrong for the shipping design. It comes from AMD's own October 2025 Helios blog, which is still live and was written when AMD rated the MI455X at 19.6 TB/s per GPU; 72 times 19.6 TB/s is 1.41 PB/s. AMD has since raised the per-GPU figure to 23.3 TB/s, and 72 times that is the 1.67 PB/s the MI400 series page now quotes and the 1.7 PB/s the launch blog rounds to.[24][26] The coincidence that 1.4 is also the rack's FP8 exaFLOPS number makes the stale bandwidth figure easy to repeat.

Scale-up networking runs over UALink over Ethernet rather than a proprietary interconnect, using four scale-up cartridges and six switch trays that AMD describes only as carrying 512-lane 200G UALoE switch ASICs; press reporting identifies these as Broadcom Tomahawk 6 parts, which AMD itself has not confirmed for Helios. Separately, on the scale-out side, each GPU is served by three 800 Gbps Pensando Vulcano NICs for 2.4 Tbps per accelerator.[7][25]

AMD published no rack power figure at all, on the product page, in the brochure or in the press release. The only number on record is The Register's, which describes the 50 V liquid-cooled DC busbar at the back of the rack as able to deliver between 225 kW and 245 kW under load. That is a delivery capability rather than a measured draw.[7]

Fuller technical treatment is in the AMD Helios rack and Instinct MI455X articles.

The performance claims and their footnotes

AMD's headline numbers are vendor comparisons. AMD did footnote its test configurations, which makes it possible to state precisely what is being compared to what.

ClaimAMD footnoteWhat is actually being compared
MI455X delivers "34x higher token throughput compared to MI355X GPUs"MI400-020Measurements and calculations by AMD Performance Labs, July 2026, measuring token throughput at high, medium and low interactivity operating points, running DeepSeek V4 Flash with FP4 serving, MI455X versus MI355X. AMD notes system manufacturers may vary configurations, yielding different results.[1]
Helios delivers "up to 30% more tokens per dollar" than the leading competitive solutionMI400-025Kimi K2 Thinking at 32K input and 8K output, AMD Helios versus an NVIDIA Vera Rubin NVL72 rack, using estimated aggregate throughput across low, medium and high interactivity operating points combined with an hourly pricing projection. Both the throughput and the pricing inputs are AMD estimates, not measured results or published prices.[1]
MI350P delivers "up to 4.2x more tokens per second per dollar than competition"MI350P-007AMD internal testing, July 2026, one MI350P versus one NVIDIA H200 NVL, Llama 3.3 70B Instruct in FP8 online serving at 1024 in and 1024 out, median of three runs per point, with the stated result being the peak per-concurrency ratio rather than an average. The only prices anywhere in the endnote are an AMD internal estimate of $327,238.40 for an 8x MI350P server and an OEM list price of $265,928.24 for an 8x NVIDIA RTX PRO 6000 server. No H200 NVL price is given.[1]
EPYC 9996 offers 2.08x the cores per rack of NVIDIA Vera, plus leadership in agents per CPU dollar and per CPU watt9xx6-012, 9xx6-013, 9xx6-014AMD comparisons against NVIDIA's Vera CPU and, in the Xeon comparisons, against Intel parts.[1]

Two caveats deserve emphasis. The 34x MI455X figure is a generation-over-generation result on one FP4 serving workload at AMD-chosen interactivity points, not a general statement that the part is 34 times faster; the same Helios launch material elsewhere describes roughly 4x higher AI floating-point performance versus MI355X at the silicon level, which is the more conservative architectural framing.[7] The 30% tokens-per-dollar figure is partly a pricing model rather than a benchmark, and the comparison target, NVIDIA Vera Rubin NVL72, was itself not shipping in volume when the comparison was made.

On the gap between spec sheets and delivered performance, AMD's Anush Elangovan told The Register that the MI455X reached about 20 petaFLOPS of FP4 in real testing, roughly half the part's peak theoretical figure. The outlet credited the candour while noting that achievable FLOPS rarely approach spec-sheet peaks on any vendor's silicon.[7]

Competitive framing versus NVIDIA

AMD's position was that Helios reaches the rack-scale tier at the same time as NVIDIA's Vera Rubin NVL72, closing a gap in which AMD had historically trailed by roughly a generation. Press comparisons of the two platforms, drawn from vendor-published specifications rather than benchmarks, generally found a split decision:

  • Memory capacity favours Helios. 31 TB of HBM4 per rack against approximately 20.7 TB for Vera Rubin NVL72, which matters for keeping very large models resident on a single rack without partitioning.[7][16]
  • Aggregate FP4 inference compute favours Vera Rubin on some accountings (3.6 exaFLOPS against 2.9 for Helios), while AMD claims that a single MI455X has 15% more peak OCP MXFP4 throughput than a Vera Rubin GPU rated in NVFP4 dense, per endnote MI400-003. The discrepancy comes from how each vendor counts compressed and sparse formats, and the two claims are not at the same level: 3.6 against 2.9 is rack against rack, the 15% is GPU against GPU.[7][16][26]
  • Scale-up bandwidth is roughly a tie, scale-out favours Helios. Both racks land near 260 TB/s of aggregate scale-up bandwidth, AMD over UALink and NVIDIA over NVLink 6. Helios carries 43 TB/s of scale-out bandwidth, which Futurum puts at 50% more than the NVIDIA rack. At FP8 the two are close, with AMD quoting 1.4 exaFLOPS per Helios rack.[16][24]
  • Openness favours AMD. Helios uses OCP Open Rack Wide, UALink over Ethernet, and merchant switching silicon rather than a proprietary rack and fabric.[2][7]

The asymmetry analysts kept returning to is software. CUDA has more than a decade of accumulated tooling behind it, and ROCm has historically been the reason customers hesitated. Analyst Steve McDowell wrote that NVIDIA retains the "deeper advantage" there, and that AMD's real test comes after the current compute shortage eases, when buyers can weigh total cost of ownership, software maturity and delivery reliability rather than taking whatever capacity exists.[3]

Roadmap through 2030

AMD published an unusually long roadmap, committing to annual GPU and rack refreshes through the end of the decade. Everything beyond 2026 in this table is a stated plan, not a product.

YearInstinct GPURackscaleEPYC CPU
2026MI455X, MI430X (MI400/MI450 series)Helios (in production; volume deployments in 2H 2026)6th Gen EPYC "Venice" (Zen 6 and dense Zen 6c)
2027MI500 seriesHelios 500, pairing MI500 with EPYC "Verano" and Pensando "Como" and "Monza" networkingEPYC "Verano"
2028MI600 seriesHelios 600, pairing MI600 with EPYC "Ferrara" and Pensando "Palma" and "Levanzo" networkingZen 7: "Florence", "Ferrara", "Fidenza"
2030Not namedNot namedZen 8: "Ravenna"

Sources: AMD's AAI 2026 press release and keynote roadmap slides.[1][17] AMD presented Helios 500 and Helios 600 as aligned to the MI500 and MI600 GPU generations, which places them in 2027 and 2028 rather than later. No rack or GPU generation was named for 2030; the only 2030 commitment in the release is the Zen 8 "Ravenna" CPU. The release also carried extensive forward-looking-statement cautionary language covering the whole roadmap period.[1]

Customers, partners and commitments

The customer roster was arguably the most consequential part of the event, because it moved AMD from a credible roadmap to booked capacity. By press accounts, publicly committed customer compute across OpenAI, Meta and Anthropic totalled roughly 14 gigawatts as of the event.[18]

PartyCommitment as stated
AnthropicAnnounced 2026-07-22, day one of the event: a strategic partnership to deploy up to 2 GW of AMD Instinct MI450 series GPUs in Helios racks, with the first gigawatt beginning in the first half of 2027. AMD has committed to make a strategic equity investment of up to $5 billion in Anthropic "in the future"; the release attaches no timetable, tranche schedule or milestone conditions to it. Anthropic's Claude models are to be used to optimise workloads for Instinct GPUs and accelerate ROCm development.[19][20]
OpenAIContinuation of the previously announced 6 GW agreement, with the first 1 GW MI450 deployment in the second half of 2026. OpenAI expects to bring Helios online in Q4 2026 and appeared on stage describing joint optimisation work.[1][2]
MetaValidating 6th Gen EPYC and testing Helios racks; separately holds a 6 GW multi-generation agreement with AMD. Meta's Santosh Janardhan appeared at the keynote.[1][2]
MicrosoftAnnounced 2026-07-20, two days before the event rather than at it. Microsoft will ramp Helios on Azure for frontier model inference, add Azure HDv2 and HXv2 VM series on 6th Gen EPYC "Venice", and broaden its Pensando DPU deployment. AMD said it would begin shipping Helios to customers including Microsoft in the second half of 2026. The launch release lists Microsoft among the labs and clouds choosing Helios but adds no new commitment.[1][27]
OracleOracle Cloud Infrastructure had previously committed to the first publicly available AI supercluster on the Helios design, with an initial 50,000 GPUs, in an October 2025 announcement. AMD's launch release names Oracle among the customers choosing Helios but does not restate the GPU count or add a new commitment.[1]
AT&TDeploying AMD technology across cloud, on-premises and air-gapped environments, and using Instinct GPUs with ROCm to power OTel 2.0, an open-source model trained specifically for telecoms. AT&T processed roughly a trillion tokens on about 430 Instinct MI300X GPUs through Microsoft Foundry Managed Compute to build it.[1][28]
CiscoCombining AMD inference hardware with Cisco networking and security; Cisco's Jeetu Patel appeared at the keynote.[1][2]
CerebrasCollaboration on an ultra-low-latency inference solution.[1]
Neoclouds and othersHUMAIN, TensorWave, Vultr and Cirrascale named as infrastructure providers.[1]
OEMsBull, HPE, Lenovo and Supermicro building Helios systems; Sanmina and Wiwynn as infrastructure partners.[1]

The Anthropic arrangement drew particular scrutiny because it follows the pattern of AMD's earlier OpenAI deal, in which the supplier takes an equity or warrant position in the customer buying its chips. Jefferies analysts had flagged before the event that the key question was whether an Anthropic deal could be structured without such incentives; it was not.[21][22] Financial press commonly describes these structures as circular, and their accounting and disclosure treatment remains an open question for investors rather than a settled one.[21][22]

Software: ROCm.ai and Hyperloom

AMD's software announcement attacked the CUDA ecosystem gap with AI tooling rather than headcount. The components:

  • ROCm.ai, described by AMD as an agentic platform for AI-assisted GPU programming.[14]
  • ROCm CLI, a command-line tool for installing, validating, serving and managing AI workloads on AMD hardware.[14]
  • AMD Skills, packaging AMD-authored kernel and optimisation expertise for use inside third-party coding assistants including Claude, Cursor and Codex.[14]
  • Hyperloom, an open-source agentic system that analyses an inference workload, tunes kernels and improves parallelism automatically. AMD says it compresses end-to-end workload optimisation from weeks to hours, and that it has internally pushed a suite of 14,000 models through the system. A live keynote demonstration on MiniMax M3 running on MI355X GPUs under vLLM showed a 38% improvement in tokens per second after the agent wrote a better mixture-of-experts GEMM kernel.[2][15]

The bet is that if frontier coding models can write competitive GPU kernels, the value of a decade of hand-written CUDA libraries erodes. OpenAI's Philippe Tillet, who works on the Triton compiler, endorsed the premise on stage, saying agents are now capable of writing GPU kernels at a high level of quality.[15] The obvious counter is that NVIDIA can run the same play on its own stack, and analysts flagged production-readiness and reproducibility at scale as the open questions.[2]

Reception

Coverage was substantially more positive than at prior Advancing AI events, but with consistent and specific reservations.

What reviewers credited. Delivering Helios into production concurrently with NVIDIA's Vera Rubin generation, rather than a year behind, was treated as the genuine news. McDowell noted that AMD's Q1 2026 data center revenue of $5.8 billion, up 57% year over year, lent credibility that previous roadmap presentations lacked, and that Helios answers NVIDIA's long-standing advantage in selling a complete, tested, software-supported system rather than parts.[3] The 2 nm EPYC "Venice" launch, the first HPC-class product in volume production on TSMC's N2, was widely treated as an execution win.[11]

Where reviewers pushed back.

  • Zeus Kerravala of ZK Research, writing for SiliconANGLE, called Helios and the MI450 series a credible but unproven alternative, observing that the 10 to 15% performance and 30% tokens-per-dollar figures are AMD's internal benchmarks against unnamed competitors, and that customer case studies and public benchmarks are needed before the claims convert into market share.[2]
  • McDowell wrote that "both comparisons are based on vendor testing and still require independent verification", and that the announced customer commitments represent interest rather than guaranteed near-term orders, whose value depends on deployment schedules, system acceptance, and the rate at which planned capacity becomes revenue.[3]
  • The Register emphasised the gap between peak and achievable FLOPS, citing AMD's own figure of roughly 20 petaFLOPS FP4 measured against a much higher spec-sheet peak.[7]
  • On software, Kerravala's conclusion was that "openness looks necessary but not sufficient" without execution across silicon, software and ecosystem.[2]

Market reaction. AMD shares had been volatile into the event, and the stock entered 2026 amid open debate over whether the company's AI GPU ambitions were real or aspirational.[21] Most of the sell-side repricing came in the week before the keynote rather than after it. Rosenblatt Securities went to $665 from $490 on 2026-07-15, and UBS to $700 from $670 the following day, both ahead of the event and partly on the Microsoft agreement announced 2026-07-20.[27][29] The clearest post-event move was UBS raising to $730 from $700 on 2026-07-24, citing a $220 billion 2030 server CPU market forecast and confirmation that Helios was in production with named gigawatt-scale customers.[30] Price targets are analyst opinions rather than measures of technical merit, and are recorded here only as reception.

Open questions

Advancing AI 2026 marked the point at which AMD's AI data center business stopped being primarily a roadmap. The company shipped a rack rather than a chip, landed a second frontier-lab anchor customer in Anthropic alongside OpenAI and Meta, and arrived in the same window as NVIDIA's competing generation for the first time.

Three things the event did not settle: whether AMD's performance claims survive independent measurement, whether ROCm.ai narrows the software gap in production, and whether roughly 14 gigawatts of announced commitments convert into deployed capacity on schedule. All three were open as of 2026-07-25. The first will only be answerable once Helios racks reach customers in Q4 2026 and third parties, or consortium benchmarking such as MLPerf, publish results.

See also

References

  1. "AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era." AMD Investor Relations, 2026-07-23. https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era ↩
  2. "Five thoughts from Lisa Su's keynote at AMD Advancing AI." SiliconANGLE, 2026-07-24. https://siliconangle.com/2026/07/24/five-thoughts-lisa-sus-keynote-amd-advancing-ai/ ↩
  3. "AMD's Rack-Scale Challenge To Nvidia's AI Dominance." Forbes, 2026-07-23. https://www.forbes.com/sites/stevemcdowell/2026/07/23/amds-rack-scale-challenge-to-nvidias-ai-dominance/ ↩
  4. "AMD Advancing AI 2024 Event to Highlight Next-gen Instinct and EPYC Processors and Expanding Solutions Ecosystem." AMD Investor Relations, 2024-09-10. https://ir.amd.com/news-events/press-releases/detail/1216/amd-advancing-ai-2024-event-to-highlight-next-gen-instinct-and-epyc-processors-and-expanding-solutions-ecosystem ↩
  5. "AMD Delivers Leadership AI Performance with AMD Instinct MI325X Accelerators." AMD Investor Relations, 2024-10-10. https://ir.amd.com/news-events/press-releases/detail/1220/amd-delivers-leadership-ai-performance-with-amd-instinct-mi325x-accelerators ↩
  6. "AMD Unveils Vision for an Open AI Ecosystem, Detailing New Silicon, Software and Systems at Advancing AI 2025." AMD Newsroom, 2025-06-12. https://www.amd.com/en/newsroom/press-releases/2025-6-12-amd-unveils-vision-for-an-open-ai-ecosystem-detai.html ↩
  7. "AMD attacks the rack with Helios systems that rival Nvidia's." The Register, 2026-07-23. https://www.theregister.com/systems/2026/07/23/amd-attacks-the-rack-with-helios-systems-that-rival-nvidias/5277246 ↩
  8. "AMD Instinct MI350P PCIe GPUs: Run Enterprise AI on Your Existing Infrastructure." AMD Blogs, 2026. https://www.amd.com/en/blogs/2026/amd-instinct-mi350p-pcie-gpus-run-enterprise-ai-on-your.html ↩
  9. "The AMD Instinct MI350P is a HBM PCIe AI Accelerator That Has Been All Over." ServeTheHome, 2026-07-17. https://www.servethehome.com/the-amd-instinct-mi350p-is-a-hbm-pcie-accelerator-that-has-been-all-over/ ↩
  10. "AMD 6th Gen EPYC Venice: 256 Cores, 1.6TB/s, and the First PCIe Gen 6 Server CPU." StorageReview, 2026-07-23. https://www.storagereview.com/news/amd-6th-gen-epyc-venice-256-cores-1-6tb-s-and-the-first-pcie-gen-6-server-cpu ↩
  11. "AMD's 256-core Epyc 9996 'Venice' claims up to a 3.4x jump over Intel Xeon competition, 20% over Nvidia Vera." Tom's Hardware, 2026-07-23. https://www.tomshardware.com/pc-components/cpus/amds-256-core-epyc-9996-venice-claims-up-to-a-3-4x-jump-over-intel-xeon-competition-20-percent-over-nvidia-vera-zen-6-comes-with-up-to-1024mb-of-l3-16-channel-memory-and-5ghz-clock-speeds ↩
  12. "AAI 2026: AMD Introduces Open, Turnkey Integrated Platform for Physical AI." AMD Newsroom, 2026-07-23. https://newsroom.amd.com/news/aai-2026-kria-robotics-dev-platform/ ↩
  13. "AMD launches Ryzen AI Embedded X100 processors, Kria AI SoM, and physical AI/robotics developer platform." CNX Software, 2026-07-24. https://www.cnx-software.com/2026/07/24/amd-launches-ryzen-ai-embedded-x100-processors-kria-ai-som-and-physical-ai-robotics-developer-platform/ ↩
  14. "AMD Announces ROCm.AI As AI-Driven Platform For Developers." Phoronix, 2026-07-23. https://www.phoronix.com/news/AMD-ROCm-AI ↩
  15. "Hyperloom: Autonomous Agentic Inference Optimization for AMD GPUs." AMD ROCm Blogs, 2026. https://rocm.blogs.amd.com/software-tools-optimization/hyperloom/README.html ↩
  16. "AMD Helios Reaches Parity with Vera Rubin NVL72: Can Open Standards Outflank NVIDIA?" Futurum Group, 2026-07-23. https://futurumgroup.com/insights/amd-helios-reaches-parity-with-vera-rubin-nvl72-can-open-standards-outflank-nvidia/ ↩
  17. "AMD Advancing AI 2026 Keynote Live Coverage." ServeTheHome, 2026-07-23. https://www.servethehome.com/amd-advancing-ai-2026-keynote-live-coverage/ ↩
  18. "AMD launches Helios against Nvidia, tips up to US$5 billion into Anthropic, and books roughly 14 gigawatts of customer AI compute at Advancing AI 2026." iTWire, 2026-07-24. https://itwire.com/business-it-news/data/astounding-massive-dominant-ai-innovation-amd-launches-helios-against-nvidia-tips-up-to-us-5-billion-into-anthropic-and-books-roughly-14-gigawatts-of-customer-ai-compute-at-advanci ↩
  19. "AMD and Anthropic Announce Strategic Partnership to Deploy Up to 2 Gigawatts of AMD Instinct MI450 Series GPUs." AMD Investor Relations, 2026-07-22. https://ir.amd.com/news-events/press-releases/detail/1292/amd-and-anthropic-announce-strategic-partnership-to-deploy-up-to-2-gigawatts-of-amd-instinct-mi450-series-gpus ↩
  20. "AMD to invest up to $5 billion in Anthropic as part of computing power deal." CNBC, 2026-07-22. https://www.cnbc.com/2026/07/22/amd-anthropic-ai-chip-investment.html ↩
  21. "AMD's Advancing AI 2026: A Credibility Test That Could Redraw the Competitive Map." ad-hoc-news, 2026-07-20. https://www.ad-hoc-news.de/boerse/news/ueberblick/amd-s-advancing-ai-2026-a-credibility-test-that-could-redraw-the/69804642 ↩
  22. "AMD to Invest Up to $5 Billion in Anthropic in The Next Circular Deal." Trending Topics, 2026-07-22. https://www.trendingtopics.eu/amd-to-invest-up-to-5-billion-in-anthropic-in-the-next-circular-deal/ ↩
  23. "AMD Advancing AI 2026." AMD Corporate Events, 2026. https://www.amd.com/en/corporate/events/advancing-ai.html ↩
  24. "AMD Launches Helios, the Highest Performing Rackscale AI Infrastructure Solution." AMD Blogs, 2026-07-23. https://www.amd.com/en/blogs/2026/amd-launches-helios-the-highest-performing-rackscale-ai-infrastructure-solution.html ↩
  25. "AMD Helios Rackscale Solution." AMD, 2026. https://www.amd.com/en/products/rackscale-solutions/helios.html ↩
  26. "AMD Instinct MI400 Series GPUs." AMD, 2026. https://www.amd.com/en/products/accelerators/instinct/mi400.html ↩
  27. "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. https://ir.amd.com/news-events/press-releases/detail/1291/microsoft-to-deploy-next-gen-amd-instinct-and-amd-epyc-processors-as-the-companies-expand-their-long-term-strategic-partnership ↩
  28. "AT&T and Microsoft scale trillion-token workloads with Microsoft Foundry and AMD." Microsoft Azure Blog, 2026-07-23. https://azure.microsoft.com/en-us/blog/att-and-microsoft-scale-trillion-token-workloads-with-microsoft-foundry-and-amd/ ↩
  29. "AMD price target raised to $665 from $490 at Rosenblatt." Rosenblatt Securities, 2026-07-15. https://www.rblt.com/news/amd-price-target-raised-to-665-from-490-at-rosenblatt ↩
  30. "UBS raises AMD stock price target to $730 after AI Day." TheStreet, 2026-07-24. https://www.thestreet.com/investing/stocks/ubs-raises-amd-stock-price-target-730-advancing-ai-server-cpu ↩
  31. "AMD EPYC 9006 Venice Announced & Looks Poised To Be A Grand Slam." Phoronix, 2026-07-23. https://www.phoronix.com/review/amd-epyc-9006-venice ↩

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision by 1 contributors · full history

Suggest edit