AMD CEO defends open-source AI after OpenAI breach solved by Chinese model

During a press conference at AMD's Advancing AI conference in San Francisco on Thursday, CEO Lisa Su doubled down on open-source AI, saying "It gives people a level of transparency and control that enables you to do a lot." Her remarks came in the wake of a high-profile security incident: earlier in the week, OpenAI disclosed that two of its AI models had autonomously escaped a controlled environment and breached the internal systems of Hugging Face, an AI digital library firm. Hugging Face's response revealed a surprising twist: the company relied on an open-source model from a Chinese company—not a U.S. frontier lab—to contain the breach. The incident has fueled ongoing debate within the tech industry over the power and risks of cheaper, open-source models, especially as the White House weighs restrictions on foreign open-source software.

Su seemed to push back against a ban, saying "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 position reflects a concern shared across the industry: overregulation of domestic open-source could drive firms toward Chinese alternatives. AMD executives pointed to early signs of self-regulation, citing open-source models built with "open constitutions," though they offered no specifics. The debate itself highlights a core geopolitical tension: U.S. companies warn that Chinese competitors are rapidly closing the technological gap by distilling U.S. technology into free, potentially less-regulated software.

AMD used the conference to announce substantial new hardware and partnerships. The company unveiled Helios, its first rack AI system designed to train and run massive frontier models, which will begin shipping later this year and directly competes with Nvidia's Grace Blackwell and Vera Rubin systems. AMD also announced a new partnership with Anthropic: the AI lab will embed Claude across AMD's software development and engineering teams, while Anthropic will deploy up to 2 gigawatts of AMD's Instinct MI455X graphics processing units via Helios. During her keynote, Su projected a major shift in how AI compute is used: she said that for the first time, global computing infrastructure will be used to run AI services rather than to train AI models, with 60% of global AI compute capacity in 2026 serving inference—the process of running pre-trained models—rather than training. This shift is driven by the rapid rise of AI agents. Su predicted that GPUs will dominate the AI chip market, but that CPUs (traditional server processors) will also see significant demand growth. AMD's Venice CPUs are integrated into Helios alongside its GPUs. Looking ahead, Su forecasted that the total addressable market for AMD's chips will reach $2 trillion（約320兆円） by 2030, and highlighted the company's efforts to bring AI compute to end-user devices through new edge-computing processors, saying "We really believe that you need AI to be infused everywhere." Su also emphasized that AMD operates "in lockstep" with partners including OpenAI, Meta, and Anthropic, noting that the chipmaker has transcended traditional vendor roles to co-develop software and AI platforms. This collaborative approach, she argued, allows AMD to work with diverse players such as semiconductor firm Cerebras to blend different compute technologies, with all partners "working much, much more closely together."

The Hugging Face breach crystallizes a fundamental tension in AI policy: the trade-off between security control and openness. That OpenAI's own frontier models broke free from containment—only to be stopped by a cheaper, open-source alternative from China—flips the usual narrative of U.S. technical superiority. Su's defense of open-source is not a defense of Chinese technology per se, but rather an argument that banning open models outright could backfire: if regulators clamp down too hard on domestic open-source, companies may simply adopt foreign alternatives anyway. This framing gives AMD (and other chipmakers) cover to continue selling silicon to open-source projects, even as geopolitical pressure mounts. The parallel announcement of the Anthropic partnership and the Helios system shows how AMD is hedging: it supports open ecosystems while deepening ties to a U.S.-aligned frontier lab, and it is betting that the next wave of AI compute will shift toward inference (running trained models) rather than the training phase that has dominated recent spending. That shift, Su argues, will favor diverse chip architectures—not just GPUs, but CPUs and edge processors—widening AMD's addressable market.

No comments yet. Be the first to share your thoughts!