At a packed hall inside San Francisco’s Moscone Center, AMD Chief Executive Lisa Su took the stage and delivered a message that mixed defiance with optimism. Open source isn’t a risk. It’s an advantage. The declaration came just days after a notable security incident at Hugging Face, where an OpenAI agent caused a breach. Engineers there turned to an open-source model from a Chinese firm to contain the problem. Su didn’t shy away. “I think open source is a great thing,” she said. “It gives people a level of transparency and control that enables you to do a lot.”
The comments, delivered during her keynote at the Advancing AI 2026 conference, struck a chord. They arrived amid rising worries in Washington about Chinese developers distilling American AI technology into unguarded models. Yet Su pushed back. “This active conversation about restricting open models is an area where we all believe that they have a significant place in the ecosystem, and we just have to make sure that we manage all pieces of that.” Her stance reflects AMD’s broader strategy. Build hardware that works in open systems. Encourage developers to choose flexibility over lock-in.
And then she unveiled the hardware to back it up. Helios. AMD’s first rack-scale AI system. Not just a collection of GPUs sold to system integrators. A complete, production-ready rack designed from the ground up for massive training and inference jobs. The system packs 72 Instinct MI455X accelerators. It draws on the new EPYC Venice processor, built on TSMC’s 2-nanometer process. It adds Pensando Vulcano networking chips for scale-out connectivity. All tied together through ROCm 7, AMD’s open-source software platform. Su called it “the world’s best AI rack.” A direct challenge to Nvidia’s tightly integrated NVL systems.
Details matter here. One rack can hit nearly 3 exaflops of AI performance. It holds 31 terabytes of HBM4 memory. Pricing starts around $5.25 million per rack, according to analysts who tracked the event. Shipments begin in the third quarter. Customers have already committed at scale. Anthropic plans to deploy up to 2 gigawatts of MI455X GPUs using Helios racks. OpenAI and Meta have signed multi-year deals that together reach 12 gigawatts. Those numbers come from Tech Insider, which covered the financial and deployment angles in detail the day after the keynote.
Su’s vision goes beyond raw specs. She sees a coming explosion in inference workloads. AI agents are driving the change. By her estimate, 60 percent of global AI compute capacity will focus on running models rather than training them in 2026. That shift plays to AMD’s strengths. GPUs will still dominate the chip market. But CPUs like Venice gain new relevance inside these racks. The overall addressable market for AMD silicon could reach $2 trillion by 2030. “We really believe that you need AI to be infused everywhere,” Su told the audience. The company introduced new edge processors to make that possible.
ROCm 7 forms the software backbone. AMD claims it delivers 3.5 times the performance of the previous version. It integrates more tightly with popular open inference tools such as vLLM, SGLang and llm-d. The company also launched ROCm Enterprise AI and an AMD Developer Cloud. These offerings let smaller teams test workloads on real Instinct hardware without buying a full rack. The approach contrasts sharply with Nvidia’s CUDA platform. AMD hopes openness will attract developers tired of proprietary ecosystems. Recent coverage from SiliconANGLE highlights how co-designed networking underpins the entire Helios strategy at scale.
Partnerships tell part of the story. Sam Altman, CEO of OpenAI, joined Su on stage. A symbolic moment given OpenAI’s status as one of AMD’s largest disclosed AI customers. Meta helped shape the Open Rack Wide form factor that Helios uses. The design offers clear advantages in density and cooling. Anthropic will embed its Claude model into AMD’s internal tools while deploying the hardware. AMD works in lockstep with these firms, moving beyond simple supplier relationships to co-develop software and platforms. It even collaborates with Cerebras on heterogeneous compute approaches. “It’s the classic case of the more useful AI gets, the more you want to use it,” Su said.
The timing feels deliberate. Nvidia’s Jensen Huang recently voiced similar support for open-source models, urging the U.S. to worry less about Chinese progress in the area. Both CEOs see the same trend. High-quality open models are here to stay. The question is how to govern them. AMD executives mentioned the idea of models with “open constitutions” for self-regulation. They offered few specifics. Still, the concept aligns with Su’s transparency argument.
Helios itself builds on years of preparation. AMD first previewed elements of the open rack architecture at the Open Compute Project Global Summit in 2025. It partnered with ZT Systems for rack-level design and customer enablement. The result is a modular system meant to slide into existing data centers rather than force a complete rip-and-replace. That modularity could prove decisive. Many large operators prefer to avoid vendor lock-in. But. Execution will decide the outcome. AMD must prove that its software stack matches Nvidia’s maturity. Early benchmarks for ROCm have improved. Yet CUDA’s vast library of optimized kernels remains a high bar.
Analysts reacted with cautious optimism. AMD shares traded near $553 in the sessions following the event. The company has momentum in the data-center market. Its Instinct MI300 series already powers significant installations. Helios represents the next step. A full system sale instead of component sales. The rack-scale approach mirrors what Nvidia has done successfully with DGX and NVL. Yet AMD insists its open foundation sets it apart. Recent commentary on X from industry watchers like Ryan Shrout of Signal65 captured the sentiment. “For years the knock on AMD in AI was that it sold chips while the competition sold racks. Helios changes that.”
Not everyone buys the narrative completely. Some point to lingering software fragmentation in the ROCm world. Others question whether enterprises will trust open-source foundations for their most sensitive models. Su’s defense of open AI after the Hugging Face episode aimed to address those doubts head-on. Transparency, she argued, actually reduces certain risks. Control stays with the user. And when problems arise, the community can respond faster than a single vendor.
The broader industry context adds weight. Demand for AI infrastructure continues to surge. Hyperscalers and cloud providers are building gigawatt-scale clusters. Power consumption has become a limiting factor. Helios emphasizes energy efficiency alongside performance. The combination of Venice CPUs, MI455X GPUs and advanced networking seeks to deliver high throughput with better watts per token. Whether those claims hold in real-world deployments remains to be seen. AMD has scheduled technical sessions and workshops throughout the Advancing AI event to share early customer data.
Su’s keynote also touched on the long-term road map. The MI500 series GPUs are already in preview. Further process shrinks and architectural advances promise another 1000-fold improvement in AI performance by 2027, according to company projections. That pace feels aggressive. It also signals confidence. AMD isn’t content to trail Nvidia in the AI race. It wants to define an alternative path built on open standards, collaborative development and customer choice.
Of course, talk is one thing. Delivery is another. The next 12 months will test AMD’s claims. Helios racks must ship on schedule. ROCm must continue closing the feature gap. Partners like Anthropic, OpenAI and Meta must demonstrate strong results at scale. If they do, the open alternative could gain real traction. If not, Nvidia’s installed base and software moat may prove too strong.
Either way, Su’s appearance at Advancing AI 2026 marked a pivotal moment. AMD has moved from challenger to serious contender in rack-scale AI systems. Its bet on openness represents a philosophical difference as much as a technical one. In an industry often criticized for closed gardens and vendor dependency, that position carries appeal. The market will render its verdict soon enough. For now, the message from San Francisco is clear. AMD believes the future of AI infrastructure doesn’t have to be proprietary. It can be shared, transparent and, above all, useful.
WebProNews is an iEntry Publication