OpenAI Eyes Custom AI Chips with Broadcom, Diversifying Beyond Nvidia and AMD

Broadcom and OpenAI have formalized their burgeoning partnership, signaling a significant escalation in the race to build out the infrastructure necessary for advanced artificial intelligence.

The companies announced on Monday a collaborative effort to jointly develop and deploy 10 gigawatts of custom AI accelerators. This ambitious initiative reflects a broader industry trend toward bespoke silicon solutions tailored to the unique demands of AI workloads. While financial details of the agreement remain undisclosed, the sheer scale of the project underscores the strategic importance attributed to this collaboration.

News of the partnership triggered a positive market reaction, with Broadcom shares experiencing a notable surge of 9.88%.

While Broadcom and OpenAI have been quietly working together for the past 18 months, the public announcement signals a new phase, characterized by plans to develop and deploy racks of OpenAI-designed chips, with the deployment slated to begin late next year. This announcement comes on the heels of OpenAI’s recent high-profile partnerships with Nvidia, Oracle, and Advanced Micro Devices, all designed to bolster its access to the capital and compute resources critical for realizing its ambitious AI development roadmap. These strategic alliances point to OpenAI’s strategy of diversifying its supply chain and mitigating the risks associated with relying on a single vendor for its crucial computing needs.

“These things have gotten so complex you need the whole thing,” OpenAI CEO Sam Altman remarked in a podcast featuring executives from both OpenAI and Broadcom, coinciding with the news release. This statement highlights the complexity of modern AI infrastructure, necessitating a holistic approach that encompasses not just processing power, but also networking, memory, and cooling solutions.

The envisioned systems encompass a comprehensive solution, integrating networking, memory, and compute components, all meticulously customized for OpenAI’s specific workloads and built upon Broadcom’s robust Ethernet stack. By venturing into custom chip design, OpenAI aims to significantly reduce its compute costs and optimize the utilization of its infrastructure investments. Industry experts estimate the capital expenditure for a 1-gigawatt data center to be approximately $50 billion with the chips potentially accounting for approximately $35 billion, based on current pricing dynamics largely influenced by Nvidia.

The collaboration with Broadcom promises “a gigantic amount of computing infrastructure to serve the needs of the world to use advanced intelligence,” Altman stated, emphasizing its potential to unlock significant efficiency gains. “We can get huge efficiency gains, and that will lead to much better performance, faster models, cheaper models — all of that.” Such efficiency gains can be crucial for unlocking wider accessibility and adoption of advanced AI models.

Broadcom has emerged as a primary beneficiary of the generative AI boom, experiencing surging demand for its custom AI chips, internally designated as XPUs. Although Broadcom maintains confidentiality regarding its large-scale web clients, industry analysts have speculated that Google, Meta, and ByteDance are among its key customers. The OpenAI deal underscores the growing importance of custom silicon in the AI space, where companies with unique workloads are prioritizing performance and efficiency above off-the-shelf solutions.

Broadcom’s stock valuation has reflected this burgeoning demand, with shares up over 50% year-to-date, building on a more than doubling in value in 2024, pushing the company’s market capitalization beyond $1.5 trillion. Last month the chipmaker’s stock experienced a 9% rise following its earnings call, during which they revealed securing a new $10 billion customer, a deal widely speculated by analysts as pertaining to OpenAI. This consistent growth underscores Broadcom’s strategic positioning within the rapidly evolving AI infrastructure landscape.

However, Broadcom’s Semiconductor Solutions Group President, Charlie Kawwas, clarified that OpenAI is not the mystery $10 billion customer disclosed in the report, leaving the specific identity of that customer still shrouded in speculation. This subtle distinction highlights the intricate dynamics of customer relationships in the highly competitive semiconductor industry.

According to OpenAI President Greg Brockman, the company has employed its own models to accelerate chip design and improve efficiency. This further exemplifies OpenAI’s commitment to innovation across multiple levels of AI development.

“We’ve been able to get massive area reductions,” he stated in the podcast. “You take components that humans have already optimized and just pour compute into it, and the model comes out with its own optimizations.” This capability has implications far beyond chip design, potentially influencing materials science and manufacturing processes.

Broadcom CEO Hock Tan asserted that OpenAI is currently developing “the most-advanced” frontier models. This resonates with industry analysts that believe some AI companies will focus on building proprietary silicon, creating a divide from AI companies leveraging off-the shelf hardware.

“You continue to need compute capacity — the best, latest compute capacity — as you progress in a road map towards a better and better frontier model and towards superintelligence,” he noted. “If you do your own chips, you control your destiny.” This sentiment aligns with the fundamental principle of vertical integration, allowing companies to exert greater control over critical components of their value chain.

Altman implied that 10 gigawatts is just the beginning of a much larger infrastructure expansion.

“Even though it’s vastly more than the world has today, we expect that very high-quality intelligence delivered very fast and at a very low price — the world will absorb it super fast and just find incredible new things to use it for,” he predicted. This reflects the long-term vision that underpins OpenAI’s ambitious strategy.

OpenAI currently operates on slightly more than 2 gigawatts of compute power.

According to Altman, this level of infrastructure has been sufficient to scale ChatGPT to its current capabilities, develop and launch the Sora video creation service, and conduct extensive AI research. However, rapidly escalating demand is driving the need for substantially greater resources. OpenAI has announced that it has committed to roughly 33 gigawatts of compute with Nvidia, Oracle, AMD and Broadcom.

“If we had 30 gigawatts today with today’s quality of models,” he added, “I think you would still saturate that relatively quickly in terms of what people would do.” This statement reveals the insatiable appetite for computing power that is characteristic of cutting-edge AI research and development.