AMD used its Advancing AI 2026 event in San Francisco this week to do something it has never quite managed before: stand in front of the whole AI industry and look like the other company that can build all of it.

The headline launch is Helios, a full rack-scale AI system AMD is aiming squarely at Nvidia. Around it came 6th Gen EPYC “Venice” CPUs, the Instinct MI400 GPU series, an AI-native software stack called ROCm.ai, agentic AI advances, personal-AI silicon and a physical-AI push. A parade of the biggest names in AI walked on stage to vouch for it: OpenAI, Meta, Anthropic, Cerebras, AT&T and Cisco.

Two numbers frame the day. AMD confirmed a strategic partnership with Anthropic that includes an equity investment of up to US$5 billion into the Claude maker, and between OpenAI, Meta and Anthropic it now has roughly 14 gigawatts of publicly committed customer compute lined up. For context, a gigawatt is about the output of a large nuclear reactor. AMD is talking about 14 of them, pointed at AI.

For iTWire readers, the local angle is closer than it looks, and I’ll get to Pawsey, sovereign AI and the Anthropic-Australia connection further down.

Here’s the must-watch keynote presentation in full, after which the article continues – please watch when you can, and read on!

The number that matters: AMD is now a shareholder in its own customers

Start with the money, because it’s the most interesting thing AMD did all week – well, beyond launching market leading processing and AI technologies, of course!

Ahead of the keynote, AMD and Anthropic announced Anthropic will deploy up to 2 gigawatts of AMD compute, starting with the first gigawatt of Instinct MI450 Series GPUs in the first half of 2027, running on Helios racks with Venice CPUs and Pensando networking. Tucked inside that release: AMD is making a strategic equity investment of up to US$5 billion in Anthropic.

That mirrors the 6-gigawatt deal AMD signed with OpenAI in October 2025, which came with a warrant handing OpenAI up to 160 million AMD shares at 1 cent each, a stake CNBC reported could reach about 10% of the company if the milestones land.

So AMD gives its biggest customers equity, the customers commit gigawatts, and both sides get pulled onto the same side of the table. It’s a clever piece of financial engineering, and it’s also a little circular, which is the criticism you’ll hear from the bears. Nvidia has been doing versions of this too. The difference is that AMD needs the anchor tenants more, so it’s paying for them in stock.

Worth keeping honest about: “up to” is carrying weight in every one of these sentences. Up to 2 gigawatts, up to US$5 billion, up to 160 million shares. These are ceilings tied to deployment milestones stretching into 2027 and beyond, not cheques already cashed.

All infographics in this article independently created by iTWire using Gemini Nano Banana

Helios: the rack AMD points straight at Nvidia

AMD Chair and CEO, Dr Lisa Su, spent the first half of the keynote on hardware, and she clearly enjoyed it, at one point calling herself “the Vanna White of chips” as she held up module after module.

Helios is the centrepiece. It’s a liquid-cooled rack that AMD designs and sells as one system: 72 Instinct MI455X GPUs across 18 compute trays, 6th Gen EPYC CPUs, Pensando Vulcano AI NICs, and scale-up networking over UALink and Ethernet rather than Nvidia’s proprietary NVLink. AMD quotes 2.9 exaflops of FP4, 31 terabytes of HBM4 memory and 1.7 petabytes per second of memory bandwidth per rack. On stage Su put it in blunter terms: more than 18,000 GPU compute units and over 4,600 Zen 6 CPU cores in a single rack.

The comparison AMD wants you to make is against Nvidia’s Vera Rubin NVL72, the rack Jensen Huang put into production at CES 2026. Both pack 72 GPUs, so it’s a genuine head-to-head. AMD claims Helios delivers 15% more FP4 compute, 50% more HBM capacity, 50% more scale-out bandwidth and up to 30% more tokens per dollar than the Nvidia rack. Those are AMD’s own numbers, measured its own way, and the usual rule applies: wait for independent testing before you bank them.

The tell is that customers are buying before that independent testing exists. Su said Helios is in full production, with shipments starting at the end of the third quarter and ramping through the fourth. Anthropic, OpenAI, Meta, Oracle, Cerebras and a list of neocloud names including Vultr and Tensorwave are all in the queue, built by OEMs including HPE, Lenovo, Supermicro and Bull.

The best Helios story came from Anthropic Co-Founder and Chief Compute Officer, Tom Brown, who described putting a single engineer onto an earlier MI355 rack and expecting a slog.

“They spun up, they connected it to Claude, asked it, ‘Hey, bring up this machine.’ Left it going over the weekend, and we ended up with a graph of the actual performance of our leading model on it, just going up and up over the weekend,” Brown said. His point, that “anyone, human or AI, can now build real models on your platform,” is the whole open-ecosystem pitch in one anecdote. Su said her team offered to help and Anthropic’s engineers told them they had it covered.

Brown’s official line in the partnership release was drier, and it’s the sentence that explains the gigawatts: “Access to compute is central to keeping Claude at the frontier and meeting demand from our customers,” he said.

Instinct MI400: an MI455X for frontier AI, an MI430X for sovereign AI, and a 2,000x promise

The GPU that drives Helios is the Instinct MI455X, the frontier-class part of the new MI400 Series. Su reeled off the specs: 320 billion transistors, TSMC 2-nanometre and 3-nanometre process technology, 12 compute and I/O chiplets and 432GB of memory tied together with 3D stacking. AMD says it delivers up to 34 times the inference throughput of the previous MI355X at high concurrency, and up to 18 times more tokens per dollar.

The quieter, and for Australia more relevant, launch is the MI430X. It’s the same family reworked for sovereign AI and high-performance computing, with native hardware FP64 rather than emulation, which is the double-precision maths scientific workloads live on. AMD quotes 288 teraflops of FP64, roughly 9 times the competition, and says the MI430X ships in the first half of 2027. AMD Senior Vice President of AI, Vamsi Boppana, framed the split plainly: “The next generation of AI will span frontier AI, sovereign AI and scientific computing.”

AMD also isn’t standing still on the roadmap. Su previewed MI500 and, behind it, MI600 on a new CDNA architecture, and made the kind of claim that either ages brilliantly or gets quoted back at her: MI500 will “deliver the largest generational leap in the history of Instinct,” she said, putting AMD “on track to deliver more than 2,000 times higher inference throughput in just 4 years.” Even by the standards of AI keynote maths, 2,000x is a number to file and check later.

Venice and the CPU comeback nobody priced in

Here’s the part of the day that surprised me most, and it isn’t a GPU.

For a year, the story has been that AI is a GPU game and the CPU is plumbing. AMD spent a chunk of this keynote arguing the opposite, and it had Meta standing next to it to help. The pitch: agentic AI runs code, calls tools and queries data outside the model, and all of that runs on CPUs. Su said AMD now sees the server-CPU market climbing from roughly US$25 billion today to more than US$200 billion by 2030, a sharp upgrade on the roughly US$60 billion it forecast a year ago – and yet another example of AMD being astoundingly massive and dominant. 

Enter 6th Gen EPYC “Venice”, built on new Zen 6 cores and TSMC’s 2-nanometre process, with up to 203 billion transistors and up to 1.8 times the performance of the current Turin generation. Venice comes as a whole family of chips: a high-frequency Venice-HF for feeding GPUs inside Helios, a 256-core, 512-thread monster tuned for “agent sandboxes,” and a 128-core enterprise part, with a low-power Verano and a 3D V-Cache Venice X still to come.

AMD’s benchmark theatre was aimed at Arm as much as at Intel. Against the best Arm server chip it claims 20% higher per-core performance and 2.2 times the socket performance, plus up to 3.3 times more performance per watt at the rack level. The argument that actually matters to enterprise buyers is duller and stronger: the world’s software runs on x86, so Venice runs all of it without a recompile.

Meta Head of Infrastructure, Santosh Janardhan, has watched his company move through four EPYC generations from Milan to Venice, and he’s stopped treating CPUs as an afterthought.

“CPUs are becoming at least as important if not more,” Janardhan said. “You have to start thinking about CPUs and GPUs as conjoined things.” He also gave the most quotable line about how long infrastructure really takes: “One of my favourite stories is Mark comes to me and says, hey, I want a gigawatt worth of data centres. You should have talked to me two years ago.”

Meta’s commitment is not small. AMD says the company has deployed millions of EPYC CPUs and plans up to 6 gigawatts of infrastructure based on AMD GPUs, moving from MI300X through MI350 to a custom MI450-based design, with Helios built on Meta’s own Open Rack Wide specification through the Open Compute Project. Both EPYC and Instinct are full production launches, with Venice rolling out through the fourth quarter.

The partners doing the vouching: OpenAI and Cerebras

OpenAI was first to gamble on AMD Instinct, and OpenAI Vice President of Compute Strategy, Sachin Katti, came to collect a little credit for it.

“We were first to bet on AMD, and we are really thrilled with how that bet is turning out for us,” Katti said. He confirmed OpenAI got hands on Helios racks about 3 months ago and is running frontier-class workloads on them, with deployment “at massive scale starting towards the end of this year and then accelerating throughout” 2027, on top of the 6-gigawatt agreement the two signed last October. Asked what he needs next, Katti gave the answer Su says he always gives: “I need more compute more quickly.”

The genuinely new technical announcement came from Cerebras Co-Founder and CEO, Andrew Feldman, whose company builds a chip the size of a dinner plate. AMD and Cerebras are pairing Helios with the Cerebras Wafer-Scale Engine in a disaggregated inference setup: Helios handles the compute-heavy prompt processing, the Cerebras part handles the memory-bandwidth-heavy token generation, and each gets optimised separately.

“Until recently, customers could have high throughput or they could have extraordinary speed,” Feldman said. “By bringing together the Helios rack with the Cerebras wafer-scale engine, we give you 5 times the throughput while continuing to deliver this extraordinary speed.” AMD quotes up to 5 times the tokens per second per watt versus a Cerebras-only setup, arriving through Cerebras Cloud in the second half of 2026.

ROCm.ai, and the moment AI started writing the kernels

Hardware was half the show. The other half was AMD’s long-standing weak spot, software, and this is where the keynote made its most quietly radical claim.

ROCm.ai is an AI-native layer over AMD’s open-source ROCm stack. It plugs AMD expertise into the coding agents developers already use, specifically Claude, Cursor and Codex, and adds an optimisation engine called HyperLoom that profiles a workload, writes and tests GPU kernels, and iterates toward a performance target on its own. AMD says ROCm releases now ship every 6 weeks instead of every 4 months, and quotes average gains of 3.3 times on inference and 2.4 times on training over the previous ROCm 7, on the same hardware.

Boppana said the part that made his own engineers double-take was watching AI-generated kernels come out “shockingly good. Better than what we expected. Sometimes better than the most manually tuned versions.” AMD Corporate Vice President of AI Software, Anush Elangovan, put the pitch as agents that “don’t just answer questions, but profile, debug and drive workloads toward peak performance.”

The credibility here came from OpenAI Member of Technical Staff and Triton creator, Philippe Tillet, who wrote one of the most important pieces of AI infrastructure software going. He was frank about where his own job is heading.

“Everyone in this room knows how much agents are taking over our jobs as kernel engineers,” Tillet said, adding that they’ve become “extremely good at generating high quality GPU kernels in a way that wasn’t possible before.” His argument for AMD is unsentimental and specific: because AMD’s compiler stack down to LLVM is open source, agents can be trained to write excellent low-level code for it, “and this has led to very, very significant performance gains that I don’t think we would have been able to achieve in a fully closed source stack.” That is a real strategic edge against Nvidia’s more closed CUDA world, if it holds.

Enterprise and the edge: an air-cooled MI350P, Ryzen AI Halo, Cisco and AT&T

AMD Senior Vice President and General Manager of Compute and Enterprise AI, Dan McNamara, took the enterprise slot, and the smartest launch in it is the Instinct MI350P: an air-cooled GPU designed to drop into existing enterprise servers without a facilities upgrade. A single MI350P handles models up to 260 billion parameters, and AMD claims up to 4.2 times more tokens per second per dollar than the competition. The point is friction: most enterprises can’t re-plumb a data centre for liquid cooling, and this lets them add serious inference without trying.

McNamara also showed AMD eating its own cooking. AMD’s IT team runs open-weight models on EPYC and MI350P behind an intelligent router, sending sensitive or heavy work to local hardware and the rest to frontier models. The result, he said, was a 43% cut in token cost and up to 3 times faster responses on the local workloads.

AMD Senior Vice President and General Manager of the Computing and Graphics Group, Jack Huynh, handled personal and physical AI. The Ryzen AI Halo developer platform runs models up to 200 billion parameters locally with 128GB of unified memory, its successor pushes that to 192GB and 300 billion parameters, and every unit ships with a free year of Hugging Face Pro. “AI is moving from a cloud-only experience to something deeply personal, local and always available,” Huynh said.

The Cisco partnership puts a management layer around all those local agents. Cisco President and Chief Product Officer, Jeetu Patel, delivered the line of the day on why this is suddenly urgent: “Humans click, but agents swarm.” His worry list for CIOs, network bandwidth, run-away token costs, and monitoring agent behaviour in real time, is exactly the anxiety Cisco Cloud Control is built to answer, with the ability to quarantine an individual desk-side box if its agent starts consuming too much. Patel said the Halo hardware is on sale now, with the management stack reaching general availability in the US in early autumn (northern), which is our spring.

AT&T grounded it all in production reality. AT&T Chief Technology Officer, Jeremy Legg, said the carrier is running more than 1 trillion tokens a month across over 100 GenAI models, from customer care to fraud to working out where to place a cell tower, and transcribing more than 300,000 calls a day. AT&T used the event to launch OTel 2.0, an open-source telco model it trained on AMD, following 18 million downloads of the first version.

“Data sovereignty for us means not being tied to a specific chipset,” Legg said, which is about the most quietly damaging thing anyone can say about Nvidia’s pricing power, delivered with a smile.

Physical AI: taking the fight to Nvidia’s Jetson

The last frontier Huynh walked through was physical AI, meaning robots. AMD introduced the Kria AI System-on-Module, powered by a new Ryzen AI Embedded X100 processor, pulling CPU, GPU, NPU and unified memory into one module so a robot can run perception, reasoning and real-time control at once.

The comparison AMD picked is Nvidia’s Jetson Thor, the default brain for a lot of current robotics. In third-party testing AMD cited, the Kria module delivered 3.4 times better real-time results, 2.3 times more concurrent agents and 1.6 times more CPU capacity. AMD paired it with a turnkey robotics developer platform built on ROCm and ROS 2, and pitched the breadth only it can assemble: Ryzen for the brain, Versal for the spine, the former Xilinx parts for the joints, Spartan for the sensors. AMD Embedded Corporate Vice President, Kirk Saban, called it “an exciting new chapter for physical AI.”

Is it enough to dislodge Jetson, which owns developer mind-share the way CUDA owns the data centre? Not on one keynote. But it’s the first time in a while AMD has had a credible robotics story, and the numbers are aimed at exactly the right target.

What it means for Australia

The through-line for iTWire readers is that AMD is already inside Australia’s most serious computing, and this launch aims at the parts of the market Australia is spending on right now.

Start with the hardware we already run. Australia’s flagship research supercomputer, Setonix at the Pawsey Supercomputing Research Centre in Perth, is built on AMD EPYC CPUs and Instinct GPUs, more than 200,000 EPYC cores across roughly 50 petaflops, and remains one of the fastest machines in the Southern Hemisphere. Every FP64 and sovereign-AI point AMD made about the MI430X is a pitch aimed straight at the next Setonix, and at the research and defence buyers who need double-precision maths and data that never leaves the country.

That lands as Australia builds AI infrastructure at a pace it’s never managed before. NEXTDC has committed to a A$4.6 billion AI campus tied to an “OpenAI for Australia” push, Firmus raised US$505 million for its Project Southgate sites, and sovereign specialists like ResetData are selling exactly the “your data never leaves the fence” model AMD’s sovereign silicon is built for. Most of that buildout runs on Nvidia today. AMD’s entire keynote was an argument for why the next tranche should not be single-vendor, and it now has the racks, the CPUs and, with ROCm.ai, a software answer to the CUDA objection.

Then there’s the partner Australia knows best. AMD’s marquee customer this week, Anthropic, has been unusually close to Canberra: the Australian government signed an MOU with Anthropic on AI safety and research, and Anthropic’s own Economic Index found Australians over-index on Claude adoption relative to the country’s size. Follow the chain: a large share of the Claude usage Australians rely on is about to run on up to 2 gigawatts of AMD Helios. Australian AI habits are quietly becoming AMD demand.

The telco read-across is just as direct. AT&T’s open-model, cost-controlled, sovereignty-first playbook is precisely the argument Telstra, Optus and TPG are working through as their own token bills climb. An Australian carrier watching AT&T post-train a model on AMD and cut AI costs by as much as 80% through smart routing is watching a template it can copy.

The AMD stock angle

None of this is happening in a vacuum on the market, and iTWire’s investor-minded readers have been paying attention.

AMD shares have had a remarkable 2026, up around 145% on the back of data-centre momentum, changing hands around US$553 heading into Advancing AI, with some analysts now running US$700 targets and Goldman Sachs resetting its own. The bull case is the roughly 14 gigawatts of OpenAI, Meta and Anthropic commitments, plus the CPU re-rating, turning into data-centre revenue that has already been growing at a rate that surprised the street.

The bear case is the one I flagged up top. A lot of this backlog is “up to,” priced in stock AMD is handing to the same customers doing the buying, and dependent on Helios shipping in volume and performing in the wild the way the slides promise. It’s a strong hand. Whether it converts into shipped racks and revenue is the story of the next 18 months – but all signs point to AMD actively monumental deals on all fronts for the foreseeable future. 

The verdict

AMD walked in needing to prove it can sell a system, not just a chip, and it did. Helios is a coherent answer to Nvidia’s rack, EPYC has quietly become the reason the CPU line item is growing again, and ROCm.ai is the first time AMD’s software story has sounded like a strength rather than an apology. The Anthropic, OpenAI, Meta, Cerebras, AT&T and Cisco line-up is the sound of the biggest buyers in AI validating the roadmap in public.

The honest caveats stay exactly where they were. The performance numbers are AMD’s own until reviewers and hyperscalers confirm them – but AMD isn’t putting on its biggest and most consequential keynote presentation yet on questionable numbers or claims. AMD means business, after all, and it’s not mucking around. 

The gigawatts and the dollars are ceilings tied to milestones out past 2027. And the customer-equity model that anchors the whole thing is either brilliant alignment or uncomfortable circularity, depending on the week and the share price.

For Australia, the takeaway is simpler and more useful. AMD already powers the country’s biggest research machine, its sovereign-AI and FP64 pitch fits what Australian research, defence and government actually need, and its highest-profile customer is the same AI lab Canberra has chosen to get close to. When the next Australian AI factory goes out to tender, there’s now a real second name on the shortlist. Competition is good for buyers, and Australian buyers just got some.

People and links

  • AMD Chair and CEO, Dr Lisa Su, is on LinkedIn.
  • Anthropic Co-Founder and Chief Compute Officer, Tom Brown, is on LinkedIn.
  • OpenAI Vice President of Compute Strategy, Sachin Katti, is on LinkedIn. His title was recently elevated from Head of Compute Infrastructure.
  • Meta Head of Infrastructure, Santosh Janardhan, is on LinkedIn.
  • Cerebras Co-Founder and CEO, Andrew Feldman, is on LinkedIn.
  • AMD Senior Vice President of AI, Vamsi Boppana, is on LinkedIn.
  • OpenAI Member of Technical Staff and Triton creator, Philippe Tillet, is on LinkedIn.
  • AMD Senior Vice President and General Manager of Compute and Enterprise AI, Dan McNamara, is on LinkedIn.
  • AMD Senior Vice President and General Manager of the Computing and Graphics Group, Jack Huynh, is on LinkedIn.
  • AT&T Chief Technology Officer, Jeremy Legg, is on LinkedIn.
  • Cisco President and Chief Product Officer, Jeetu Patel, is on LinkedIn.
  • AMD Corporate Vice President of AI Software, Anush Elangovan, and AMD Embedded Corporate Vice President, Kirk Saban, spoke on the ROCm.ai and physical-AI segments; no personal LinkedIn profiles could be independently confirmed at time of writing. AMD is on LinkedIn.

Sources