# Five thoughts from Lisa Su’s keynote at AMD Advancing AI > Source: > Published: 2026-07-24 18:36:25+00:00 ### Five thoughts from Lisa Su’s keynote at AMD Advancing AI [Advanced Micro Devices Inc.’s Advancing AI 2026](https://www.amd.com/en/corporate/events/advancing-ai.html) keynote this week was Chief Executive Lisa Su’s bid to redefine the company from a “graphics processing unit alternative” to a full-stack artificial intelligence infrastructure vendor and to make this the year the central processing unit officially rebounds as a first-class AI platform. The message from Su (pictured) was ambitious and often compelling, but it also sharpened the competitive contrast with [Nvidia Corp.](https://www.nvidia.com/en-us/) and [Intel Corp.](https://www.intel.com/), raising as many questions as it answered. Though the obvious theme of the event was that AI is moving from pilots to production, that has been the theme of every event I have attended this year. Beyond that, there were several other sub-themes. Here are the five most notable: ### 1. Helios and MI450: Finally a credible rack-scale alternative — with a catch Su opened by turning Instinct MI450 and the Helios rack into a single, rack-scale product story aimed squarely at Nvidia’s system-level dominance. In AMD’s benchmarks, Helios delivers “an average of 10% to 15% more performance than the competition” at fixed rack power on “the highest throughput workloads” and “leading inference modes,” and she translated that into “up to 30% more tokens per dollar than the competition.” Though Nvidia has set the standard for the systems approach, if AMD can deliver on the savings it claims at comparable performance, it can use that to position itself as a credible alternative. The partner lineup was the strongest evidence that these claims are real. Su said demand for Helios is “extremely strong… from the largest AI labs to hyperscalers,” and she highlighted [OpenAI](https://openai.com/) as “one of our deepest and earliest partners deploying Helios,” noting that joint engineering teams are already running GPT-class workloads. OpenAI’s infrastructure lead described AMD and OpenAI engineers “working side by side to optimize the software stack” and said they expect to deploy Helios “at massive scale, starting towards the end of this year, and then accelerating toward 2027.” The catch is that “average of 10% to 15% more performance” and “30% more tokens per dollar” remain AMD-run numbers against unnamed “competition” and unspecified model mixes. Until cloud instance specs, public benchmarks and customer case studies show comparable gains in the wild, Helios is only a strong narrative and a promising design, not yet a proven market-share shift. Nvidia still owns software mindshare and the incumbent installed base; AMD must convert a handful of flagship design wins into a durable ecosystem. ### 2. CPU rebounds: Venice turns agentic AI into a three-tier compute story One of the more interesting and underappreciated parts of the keynote was AMD’s aggressive effort to reset the CPU narrative in AI. For the past few years, CPUs have been cast as glorified I/O controllers in GPU boxes. Su pushed back hard on that, arguing that AI infrastructure is splitting into three CPU roles: - GPU servers where “the CPU’s job is basically to drive the GPUs.” - Dense “agent servers or what we call agent sandboxes” where “the priority is actually density and the highest-performing cores per watt to run thousands of agents at once.” - Traditional genera-purpose servers where it’s “all about efficiency” for databases, data services and enterprise apps. Venice, AMD’s new Epyc family on Zen 6 and TSMC ‘s two-nanometer process, is the company’s attempt to own all three tiers. Su called it “one of the largest generational gains in the history of Epyc,” claiming “up to 1.8 times more performance than Turin” and up to 512 threads per socket. She then broke Venice into a family: Venice HF for GPU host nodes at up to 5 GHz; a 256-core Venice with “the highest compute density in the industry” for agent sandboxes; and a 128-core part tuned for enterprise performance per dollar, plus variants like Venice X and Verano for HPC and AI host interconnect. Her punchiest line of the keynote, “Epyc is the only CPU portfolio that leads across all use cases,” was clearly aimed at both Intel’s latest Xeons and the growing crop of Arm server CPUs. She claimed that Venice delivers more than twice the agents per watt for agent sandboxes versus x86 and up to 3.3 times more performance per watt at the rack than unnamed Arm competitors in a 100-kilowatt rack. She also pointed out that x86 software compatibility still matters when you’re “adding thousands of employees to your enterprise” in the form of agents. The critical angle: AMD is right that agentic AI gives CPUs a second life, but this space is now intensely contested. Intel is not standing still on core count, memory bandwidth, or AI offload, and Arm vendors are pushing hard on efficiency and custom silicon for cloud providers. Venice’s generational gains are impressive on paper, but AMD still has to win OEM designs, cloud footprints and independent software vendor certifications at scale, all while customers are also considering Arm and custom accelerators tailored to their agent workloads. ### 3. Jeetu Patel and Santosh Janardhan: CPUs and GPUs become ‘conjoined things’ If Su provided the product narrative, Jeetu Patel from [Cisco Systems Inc.](https://www.cisco.com/) and Santosh Janardhan from [Meta Platforms Inc.](https://www.meta.com/) provided the architectural reality check, and they largely backed AMD’s thesis that CPUs are back in the spotlight. Patel argued that “it’s not just a GPU game anymore. It’s CPUs and GPUs. If anything, I think CPUs are becoming at least as important, if not more.” His view is grounded in enterprise knowledge, which Cisco has more of than all the other companies on stage combined. Long-running agent workflows, tools, databases and networks all must be orchestrated around frontier models. He framed compute as a heterogeneous fabric where “you must think about CPUs and GPUs as conjoined things. You hand off workloads depending on other workloads. You employ the right hardware.” That is exactly the kind of messaging AMD needs enterprise CIOs to hear to become a more strategic vendor. Meta’s Janardhan pushed the same idea at the data center scale. For him, the AI problem is no longer about squeezing every percentage point out of a single chip; it’s about “the whole data center as one integrated system — servers, hardware, networking, cooling, power.” He noted that data centers and silicon “take years to build,” and argued that the industry needs to be “sitting down in a room, co-designing today for what we need to deploy in 2027 and 2028.” In other words, CPUs, GPUs, memory, networking and power are now co-equal design levers. It was great to see these companies on stage with Su, because they aren’t second-tier logos; they’re two of the largest AI and networking players on the planet, validating AMD as a co-design partner, not a backup supplier. For the established players, that’s a warning that their traditional lock-in at the CPU and GPU levels may not survive an era in which hyperscalers want multi-vendor, co-designed systems to manage risk, cost and power. ### 4. ROCm.AI and Hyperloom: AMD tries to leapfrog on AI-native tooling On software, AMD went straight at its perceived weakness, the software gap versus Nvidia’s stack, with a different approach: Let AI write and optimize more of the GPU code. Senior Vice President Vamsi Boppana introduced ROCm.AI as “an agentic AI platform that brings the capabilities of AI-assisted GPU programming to developers.” Built on AMD’s open software stack, ROCm.AI adds an AI optimization layer called Hyperloom that can “analyze the workload, tune configurations, select and tune kernels, adjust parallelism strategies and iterate towards performance goals.” Internally, AMD has already pushed “a suite of 14,000 models through Hyperloom,” generating optimizations that would have been “impossible even with a large team of engineers before.” In a live example, an AI agent targeting MiniMax M3 on MI355s with VLLM identified an opportunity to write a more optimized Mixture-of-Experts GEMM kernel, delivering a 38 percent improvement in tokens per second. Engineers onstage admitted that some AI-generated kernels are “shockingly good, sometimes better than the most manually tuned versions.” AMD’s competitive bet is that an open stack plus AI-assisted optimization can close the software-ecosystem gap with Nvidia faster than traditional hand-tuning ever could. The risk is that this is still early-stage technology: automatic code generation must be safe, reproducible and debuggable at scale, and customers will want to see real-world workloads, not staged demos. Nvidia will not sit still here either; it has every incentive to build its own AI-native tooling on top of CUDA and its closed ecosystem. ### 5. Co-designed, open systems: AMD plays the long game against lock-in The final through line in Su’s keynote was an embrace of co-design and openness as AMD’s strategic wedge against incumbents. OpenAI’s infrastructure chief, Sachin Katti, described a future where AI becomes “a problem at a data center scale,” not just a “rack-level problem,” and stressed the need to co-design “from CPUs to GPUs to memory, networking, storage, power distribution and the cooling systems that go with it.” Cerebras CEO Andrew Feldman, announcing a joint solution that marries AMD’s Helios racks with Cerebras’ wafer-scale engine, argued that customers who used to choose between “high throughput” and “extraordinary speed” can now get both — “five times the throughput while continuing to deliver this extraordinary speed” for ultra-low-latency inference. Su tied these threads together by leaning hard into open software. Because AMD’s compiler stack and drivers are largely open, partners and even AI agents can see all the way down to the ISA. OpenAI’s Philippe Tillet credited that openness with enabling “very, very significant performance gains” and faster portability of GPT-class models to AMD hardware. In a world where “recursion,” that is AI helping design the next generation of AI systems, is becoming a reality, AMD is betting that openness will let it harness that flywheel more effectively than a closed stack. The competitive implication is that while Nvidia still has the deepest, most entrenched software ecosystem, AMD aims to be more open to attract a broader base of partners and developers. Intel has the x86 incumbency but has struggled to convert it into AI mindshare. AMD is positioning itself as the third pole: not just cheaper GPUs, but an open, co-designed CPU-plus-GPU platform tuned for the agentic, data-center-as-a-system future. Whether that bet pays off will depend on execution in silicon delivery, software quality, and ecosystem traction, not just on keynotes. But if Advancing AI 2026 is any indication, AMD has stopped talking like a fast follower and started acting like a company that expects to set the terms of the AI infrastructure debate. ### Final thoughts AMD used Advancing AI 2026 to make a statement: It is no longer content to be “the alternative” to Nvidia, and it now has credible hardware at the rack level, a resurgent CPU portfolio tuned for agentic AI, and a software story that leans into AI-assisted optimization and open ecosystems. Helios plus MI450 gives hyperscalers and frontier labs a rack-scale option that can be argued on performance per watt and tokens per dollar, while the Venice Epyc family targets the emerging three-tier AI compute stack, comprised of GPU hosts, dense agent sandboxes, and general-purpose enterprise, at a moment when CPUs are quietly becoming the control plane and workhorse for agentic workflows. On software, AMD has stopped pretending it can out-CUDA Nvidia head-on; instead, it is betting that an open stack plus AI-native tools like ROCm.AI and Hyperloom can compress the time it takes to get real workloads performant on its silicon. The uncomfortable reality for AMD is that none of this makes Nvidia “trail” in the broader AI race. Nvidia still owns the dominant software ecosystem, the bulk of the deployed AI accelerator footprint, and a deeply integrated toolchain. In most large AI shops, AMD remains the second platform to be qualified rather than the assumed default. Openness is AMD’s best strategic tool against that incumbency, lowering switching costs and inviting partners and, eventually, AI agents to co-design and tune all the way down to the instruction set, which differs from Nvidia’s model. The question is whether that openness, combined with competitive performance and TCO, is enough to move AMD from “necessary diversification” to “first choice” for a meaningful share of new deployments. Right now, openness looks necessary but not sufficient, and AMD will need multiple cycles of flawless execution on silicon, software and the ecosystem to turn this impressive keynote positioning into durable market power. *Zeus Kerravala is a principal analyst at ZK Research, a division of Kerravala Consulting. He wrote this article for SiliconANGLE.* ##### Photo: Zeus Kerravala # A message from John Furrier, co-founder of SiliconANGLE: Support our mission to keep content open and free by engaging with theCUBE community. **Join theCUBE’s Alumni Trust Network**, where technology leaders connect, share intelligence and create opportunities. **15M+ viewers of theCUBE videos**, powering conversations across AI, cloud, cybersecurity and more** 11.4k+ theCUBE alumni**— Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network. # Are you AWS customer? Support SiliconANGLE Financially by buying your AWS services from our Marketplace portal page and links. **About SiliconANGLE Media** [SiliconANGLE](https://cts.businesswire.com/ct/CT?id=smartlink&url=https%3A%2F%2Fsiliconangle.com%2F&esheet=54119777&newsitemid=20240910506833&lan=en-US&anchor=SiliconANGLE&index=9&md5=646b1b564e2259100a2b8638aab0a552), [theCUBE Network](https://cts.businesswire.com/ct/CT?id=smartlink&url=https%3A%2F%2Fwww.thecube.net%2F&esheet=54119777&newsitemid=20240910506833&lan=en-US&anchor=theCUBE+Network&index=10&md5=7de2a85f95ab4a4a495cede20b8cb1da), [theCUBE Research](https://cts.businesswire.com/ct/CT?id=smartlink&url=https%3A%2F%2Fthecuberesearch.com%2F&esheet=54119777&newsitemid=20240910506833&lan=en-US&anchor=theCUBE+Research&index=11&md5=7bb33676722925eb57d588ec343e4f6f), [CUBE365](https://cts.businesswire.com/ct/CT?id=smartlink&url=https%3A%2F%2Fwww.cube365.net%2F&esheet=54119777&newsitemid=20240910506833&lan=en-US&anchor=CUBE365&index=12&md5=d310fb35919714e66ad8d42c9c0c1bc6), [theCUBE AI](https://cts.businesswire.com/ct/CT?id=smartlink&url=https%3A%2F%2Fwww.thecubeai.com%2F&esheet=54119777&newsitemid=20240910506833&lan=en-US&anchor=theCUBE+AI&index=13&md5=b8b98472f8071b23ebb10ab9a8dd0683)and theCUBE SuperStudios — with flagship locations in Silicon Valley and the New York Stock Exchange — SiliconANGLE Media operates at the intersection of media, technology and AI. Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.