Say goodbye to single-chip dependence! OpenAI plans to launch a cross-platform AI optimization tool, aiming at Nvidia CUDA.

OpenAI is considering making its internally developed cross-chip software optimization tools publicly available; if this happens, it will directly impact the moat Nvidia has long built around its CUDA software ecosystem. On June 1, according to technology media outlet The Information, Sachin Katti, OpenAI’s Senior Vice President of Computation and Infrastructure, said in a public discussion that the company is developing a software abstraction layer that allows researchers and product teams to run AI workloads without worrying about which supplier's hardware is underneath. When asked whether this capability would be made available to others, Katti stated clearly, "It's within the scope of consideration," and described it as "agentic optimization capability," saying, "We want to provide this capability to the world." Analysts say this statement is highly significant. Nvidia’s market dominance has long relied on CUDA—this proprietary system of compilers, libraries, and optimization tools is the core dependency for mainstream AI developers to run software on Nvidia chips. If OpenAI’s cross-platform tool is publicly released, it will further weaken CUDA’s differentiated advantage and accelerate diversification in the AI computing power market. Multi-chip strategy speeds up: OpenAI accelerates move away from Nvidia dependence According to reports, Katti stated in the discussion that the AI industry will become "highly heterogeneous," with companies using AI chips from multiple suppliers simultaneously. Behind this view is a profound shift in OpenAI’s own strategy. OpenAI previously almost entirely relied on Nvidia chips, but has recently signed agreements with Amazon, Cerebras, and AMD, introducing their AI chip resources, and is also developing custom AI chips in-house. Katti did not disclose whether OpenAI would adopt Google’s custom chips like Anthropic and Meta. This trend is not unique to OpenAI. Anthropic and Meta are also unwilling to rely on a single supplier for such a core aspect of their business, and no single supplier can meet their enormous computing power needs alone. Software abstraction layer: AI version of Google’s Borg model The report says Katti compared OpenAI’s developing software system to Google’s famous Borg computation management system—the key infrastructure enabling Google to scale products across heterogeneous hardware. "This is exactly the path we are walking in the AI field," he said. Even more disruptive, Katti hinted that AI itself will become the tool to break CUDA’s monopoly. "We expect to use AI to generate optimized kernels to truly support all these different chip options," he said. Amp founder Anjney Midha pointed out in the same discussion that if developers like OpenAI make such internal tools public, enabling AI to run efficiently across Nvidia, Google, AMD, and other chips, it will pose a substantial challenge to Nvidia. In fact, CUDA’s moat has already begun to narrow. Meta’s PyTorch framework has long allowed developers to conveniently write AI code for multiple chip types, and some startups also sell AI tools that translate PyTorch code into low-level code that runs directly on chips. Vera Rubin chip deployment imminent; computing bottlenecks shift to power and engineering In addition to its software strategy, Katti also revealed OpenAI’s progress deploying Nvidia’s next-generation Vera Rubin chip system. He stated that OpenAI has obtained early samples of the chip and is expected to use it for AI training by the end of this year. Katti gave positive feedback on issues Nvidia exposed during the launch of the Blackwell system, saying Nvidia has learned from the experience. The first-generation Blackwell system caused headaches for many cloud service providers due to the complexity of networking, firmware, and wiring during large-scale deployment, but the new system has greatly improved. "Nvidia really learned from many growing pains," he said. Katti did not disclose which cloud provider will first host OpenAI’s Vera Rubin cluster, only stating there is "healthy competition" among the parties. OpenAI’s main cloud service providers currently include Microsoft, Oracle, and Amazon. Notably, Katti pointed to power supply and engineering capacity, rather than chips themselves, as the biggest bottleneck to computing power expansion. "Right now what limits us more is power and engineering capacity, not anything else," he said. This assessment offers direct reference value for resource allocation by AI infrastructure investors. Risk warning and disclaimer The market is risky; investment needs to be cautious. This article does not constitute personal investment advice, nor does it take into account individual users’ specific investment objectives, financial situations, or needs. Users should consider whether any opinions, views, or conclusions herein fit their particular circumstances. Investing accordingly is at your own risk.