Strategic Shift in AI Hardware

The rapid advancement of generative AI models such as OpenAI’s ChatGPT and DALL-E has resulted in monumental increases in computational requirements. Traditionally, OpenAI has relied heavily on Nvidia’s GPUs—seen as the industry standard for AI workloads. However, as global demand for sophisticated AI models continues to surge, concerns over cost, supply chain bottlenecks, and over-reliance on a single vendor have accelerated OpenAI’s exploration of alternative chip solutions.

The Financial Times reported on Thursday that OpenAI will collaborate with Broadcom to design and fabricate its own AI chips, a project internally code-named “Project Tigris.” According to insiders, the chips will be tailored exclusively for OpenAI’s data centers and internal needs, rather than for sale to outside customers—a strategy also adopted by fellow tech giants Google, Amazon, and Meta in their custom silicon efforts.

Reducing Reliance on Nvidia & Industry-Wide Chip Diversification

OpenAI’s move signals an industry-wide trend: Big Tech’s determination to diversify supply chains and gain more control over the hardware driving the next generation of AI services. While Nvidia maintains its dominance—holding over 80% share of the AI chip market as of 2024—companies are eager for alternatives amidst persistent chip shortages and escalating pricing. According to global market intelligence firm IDC, enterprise spending on AI silicon soared past $50 billion in 2024, with custom chips making up an increasing portion of that investment.

OpenAI’s continued hardware collaboration with both Broadcom and Taiwan Semiconductor Manufacturing Co. (TSMC) demonstrates its commitment to a multi-pronged strategy. Sources indicate that OpenAI will continue to deploy a mix of Nvidia, AMD, and now in-house chips, optimizing across cost, performance, and availability. In February 2025, Reuters reported that OpenAI’s first generation silicon was nearing design completion, with plans to tap TSMC’s advanced fabrication plants—the same facilities used by global chip leaders—from 2025 onward.

Broadcom’s AI Ambitions and the Competitive Landscape

Broadcom is rapidly emerging as a critical force in the custom AI silicon market. The company’s CEO, Hock Tan, noted in a recent earnings call that Broadcom has secured over $10 billion in AI infrastructure orders for fiscal 2026. While he did not publicly identify its new customers, industry analysts widely believe that OpenAI is among the most high-profile entrants, joining Amazon, Google, and Meta, all of whom develop custom chips to better tailor AI workloads to their services.

Tan recently disclosed that Broadcom is collaborating with at least four new deeply engaged customers on next-generation custom processors, reflecting a strategic shift across Silicon Valley as hyperscalers and emerging AI leaders insource mission-critical intellectual property previously left to off-the-shelf providers.

What Will OpenAI’s Custom Chip Enable?

Designing proprietary silicon is a complex and capital-intensive proposition, especially as the demands of AI training infrastructure skyrocket. A single large language model (LLM) training run can require thousands of top-tier GPUs and consume megawatts of electricity.

By developing a custom AI chip optimized for its models—including inference and training workloads—OpenAI aims to bring greater efficiency, scalability, and cost control to its infrastructure. The first chips are expected to leverage bleeding-edge manufacturing nodes at TSMC, possibly 3-nanometer or 2-nanometer technologies, rivaling those of Nvidia’s H100 and Blackwell series. While technical specifications remain under wraps, industry rumors suggest that ‘Project Tigris’ will focus on maximizing memory bandwidth, low latency, and energy efficiency—three performance metrics critical for next-generation AI deployments.

For OpenAI, the ability to deploy tens of thousands of custom chips across its data centers will alleviate chronic GPU shortages and lower per-training costs. According to data from Mizuho Securities, training the latest GPT-5 model is projected to exceed $200 million in cloud compute spending if solely using Nvidia hardware. In-house silicon could significantly reduce this operational overhead, freeing up capital for other R&D and product investments.

Industry Implications and the Race for AI Infrastructure

OpenAI’s entry into the custom chip space not only reshapes its internal architecture but may also set a precedent for other AI-focused startups and enterprises. As tech leaders like Google (with its Tensor Processing Unit, TPU), Amazon (via AWS Inferentia and Trainium), and Meta (with MTIA) invest billions to optimize their AI hardware stacks, the sector is poised for an arms race where compute access becomes the new competitive moat.

The global AI chip market is projected to grow at a 30%+ CAGR through 2030, expected to surpass $250 billion in annual revenue according to Goldman Sachs. The push for vertical integration—controlling everything from software frameworks to custom hardware—positions companies like OpenAI to more rapidly iterate and protect their intellectual property in an AI-centric future.

In the near term, enterprises and cloud providers can expect short-term market disruptions as AI chip supply chains rebalance, but analysts forecast that custom silicon will eventually democratize high-performance AI and lower costs for end users.

Looking Ahead

While OpenAI and Broadcom have yet to publicly confirm technical specs or production targets for the new chip, the partnership underscores the strategic necessity of hardware innovation in the global AI race. In-house silicon remains a tough undertaking, fraught with both technical and financial risk, but the potential payoff—unmatched performance and strategic independence—could transform how AI models are built and delivered at scale.

With OpenAI’s debut chip scheduled to roll out in 2026, industry watchers are closely monitoring whether this move will redefine compute paradigms in generative AI and spur a new era of hardware-software co-innovation.

Reporting by Disha Mishra and industry analysis by Reuters, FT, and additional sources.