Farewell to Single-Chip Dependence! OpenAI Plans to Release Cross-Platform AI Optimization Tools, Targeting NVIDIA CUDA

OpenAI is considering making its internally developed cross-chip software optimization tools publicly available. If implemented, this move would directly challenge the moat NVIDIA has long built through its CUDA software ecosystem.

On June 1, tech media outlet The Information reported that Sachin Katti, OpenAI's Senior Vice President of Compute and Infrastructure, stated in a public discussion that the company is developing a software abstraction layer, enabling researchers and product teams to run AI workloads without concern for which vendor supplies the underlying hardware.

When asked whether this capability would be made available externally, Katti explicitly stated that "it is under consideration," describing it as an "agentic optimization capability." He added, "We want to provide this capability to the world."

Analysts suggest this statement is significant. NVIDIA's market dominance has long relied on CUDA—a proprietary suite of compilers, libraries, and optimization tools that serves as the core dependency for mainstream AI developers running software on NVIDIA chips. Once OpenAI's cross-platform tools are publicly released, it will further erode CUDA's differentiated advantages and accelerate diversification in the AI computing power market.

Multi-Chip Strategy Accelerates as OpenAI Rushes to Reduce Dependence on NVIDIA

According to reports, Katti stated bluntly during the discussion that the AI industry will move toward "high heterogeneity," with companies simultaneously using AI chips from multiple vendors. This judgment reflects a profound strategic shift within OpenAI itself.

OpenAI previously relied almost entirely on NVIDIA chips but has recently signed agreements with Amazon, Cerebras, and AMD to introduce their AI chip resources, while also developing custom AI chips in-house.

Katti did not disclose during the discussion whether OpenAI would adopt Google's custom chips, as Anthropic and Meta have done.

This trend is not unique to OpenAI. Anthropic and Meta are equally unwilling to depend on a single supplier for such a core business function, and no single vendor can individually meet their massive computing power demands.

Software Abstraction Layer: The AI Version of Google's Borg Model

The report states that Katti compared the software system OpenAI is building to Google's famous Borg compute management system—the key infrastructure that allowed Google to scale its products massively across heterogeneous hardware. "This is the path we are taking in the AI field," he said.

Even more disruptive is Katti's implication that AI itself will become the tool to break the CUDA monopoly. "We expect to leverage AI to generate optimized kernels, thereby truly supporting all these different chip options," he stated.

Anjney Midha, founder of Amp, pointed out in the same discussion that if developers like OpenAI publicly release such internal tools, enabling AI to run efficiently on chips from NVIDIA, Google, AMD, and others, it would constitute a substantial impact on NVIDIA.

In fact, CUDA's moat is quietly narrowing. Meta's PyTorch framework has long allowed developers to write AI code for multiple chips more conveniently, and some startups are selling AI tools that translate PyTorch code into low-level code that can run directly on chips.

Vera Rubin Chip Deployment Imminent; Bottlenecks Shift to Power and Engineering

In addition to software strategy, Katti disclosed OpenAI's progress in deploying NVIDIA's next-generation Vera Rubin chip system. He stated that OpenAI has received early samples of the chip and expects to put it into use for AI training by the end of this year.

Katti gave positive remarks regarding the issues NVIDIA exposed during the launch of the Blackwell system, believing that NVIDIA has learned from them. The first-generation Blackwell system caused headaches for many cloud service providers due to network, firmware, and cabling complexities during scaled deployment, but the new version has improved significantly. "NVIDIA has indeed learned from many growing pains," he said.

Katti did not reveal which cloud service provider would first host OpenAI's Vera Rubin cluster, stating only that there is "healthy competition" among the parties. OpenAI's current major cloud service providers include Microsoft, Oracle, and Amazon.

Notably, Katti identified power supply and engineering capabilities, rather than the chips themselves, as the biggest bottleneck in current computing power expansion.

"What is constraining us more at present is power and engineering capabilities, rather than other factors," he said. This judgment has direct reference value for AI infrastructure investors regarding resource allocation directions.