Broadcom Won't Build OpenAI Chip Without Microsoft

Broadcom has reportedly refused to move forward with manufacturing OpenAI's custom AI chip unless Microsoft commits to purchasing at least 40 percent of the output — a demand that has thrown one of the most ambitious chip projects in the AI industry into serious doubt. The standoff highlights the enormous financial risks involved in building custom silicon and the complex power dynamics between three of tech's most influential players.

OpenAI's internal leadership has already signaled frustration with the arrangement. Sachin Katti, a manager overseeing the chip initiative at OpenAI, described the Microsoft dependency as 'financially unattractive' in an internal message, according to reporting from The Decoder.

Key Facts at a Glance

• Broadcom will not finance production of OpenAI's custom AI chip unless Microsoft agrees to buy 40% of the chips produced
• Microsoft has not yet agreed to the purchase commitment
• The first phase of the chip project alone costs approximately $18 billion
• OpenAI manager Sachin Katti called the arrangement 'financially unattractive' internally
• The impasse threatens to delay OpenAI's ambitions to reduce its dependence on Nvidia
• The situation exposes the financial fragility behind custom chip development in AI

An $18 Billion Gamble With No Buyer Locked In

The scale of investment required for custom chip development is staggering. At roughly $18 billion for just the first phase, OpenAI's chip project rivals the cost of building an entirely new semiconductor fabrication facility — a figure that would make even the largest tech companies pause.

For context, Nvidia spent approximately $3.8 billion on research and development in its most recent fiscal year. Intel's planned Ohio fabrication facility carries an estimated price tag of $20 billion. OpenAI's custom chip ambition sits squarely in that rarefied territory of mega-scale infrastructure investment.

Broadcom's insistence on a guaranteed buyer is not unusual in the semiconductor industry, where the economics of chip fabrication demand massive upfront capital with long lead times before any return on investment. What makes this situation remarkable is the identity of the parties involved and the strategic implications of each potential outcome.

Why Broadcom Is Drawing a Hard Line

Broadcom's position is rooted in basic risk management. As a fabless semiconductor company that designs chips and outsources manufacturing, Broadcom would need to commit enormous resources — engineering talent, design time, and manufacturing contracts with foundries like TSMC — to bring OpenAI's custom chip to life.

Without a guaranteed buyer for a significant portion of the output, Broadcom faces the possibility of producing billions of dollars' worth of specialized silicon with no market. Custom AI chips, unlike general-purpose processors, have extremely limited resale value outside their intended use case.

• Design costs for advanced chips at 3nm or 5nm nodes can exceed $500 million per design
• Manufacturing commitments with foundries typically require multi-year, multi-billion-dollar contracts
• Custom silicon cannot easily be repurposed if the primary customer walks away
• Yield risks during early production runs can drive costs even higher
• Time-to-market pressures mean delays could render the chip obsolete before it ships

Broadcom has experience building custom chips for hyperscalers — it already produces custom TPU (Tensor Processing Unit) chips for Google and has worked with Meta on similar projects. In those cases, however, the customer and the buyer are the same entity. The OpenAI situation introduces a middleman problem that complicates the financial calculus considerably.

Microsoft's Complicated Position in the Deal

Microsoft's hesitation to commit to the 40 percent purchase threshold reveals the increasingly complex nature of its relationship with OpenAI. While Microsoft has invested more than $13 billion in OpenAI and remains its most important commercial partner, the two companies' strategic interests are not perfectly aligned — particularly when it comes to infrastructure.

Microsoft already operates one of the world's largest AI infrastructure deployments, built primarily on Nvidia's H100 and upcoming B200 GPUs. Committing to purchase 40 percent of a new, unproven custom chip would represent a significant strategic pivot and a massive financial commitment on top of its existing Nvidia investments.

There are several reasons Microsoft might be reluctant:

• Nvidia's ecosystem is deeply embedded in Microsoft's Azure cloud platform and switching costs are enormous
• Unproven performance — OpenAI's custom chip has no track record compared to Nvidia's battle-tested hardware
• Financial exposure — 40 percent of an $18 billion first phase means roughly $7.2 billion in chip purchases
• Strategic leverage — committing early could weaken Microsoft's negotiating position with both OpenAI and Nvidia
• Market uncertainty — the AI chip landscape is evolving rapidly, and locking into custom silicon carries long-term risk

Microsoft is also reportedly developing its own custom AI chip, Maia 100, which launched in late 2023. Investing heavily in OpenAI's separate custom chip effort could create internal conflicts and redundancies within Microsoft's own hardware strategy.

OpenAI's Chip Independence Dream Faces Reality

OpenAI's push for custom silicon reflects a broader industry trend among major AI companies seeking to reduce their dependence on Nvidia, which currently controls an estimated 80-90 percent of the AI training chip market. The logic is straightforward: controlling your own chip supply reduces costs, improves supply chain security, and allows for hardware optimized specifically for your models.

Google pioneered this approach with its TPU chips, now in their 6th generation. Amazon has its Trainium and Inferentia chips for AWS. Meta has invested heavily in custom silicon for its AI workloads. Even Tesla has designed its own Dojo chip for training autonomous driving models.

But OpenAI's situation is fundamentally different from these companies. Unlike Google, Amazon, or Meta, OpenAI does not operate its own cloud infrastructure at scale. It relies on Microsoft's Azure for the vast majority of its compute needs. This means OpenAI cannot simply build chips and deploy them in its own data centers — it needs a partner willing to buy, deploy, and maintain the hardware.

This structural dependency is precisely what Katti described as 'financially unattractive.' OpenAI finds itself in the awkward position of designing a chip it cannot deploy independently, manufactured by a company that will not build it without a buyer, intended for a partner that has not agreed to purchase it.

The Broader AI Chip Landscape Adds Pressure

The timing of this standoff is particularly significant given the explosive growth in AI chip demand. Nvidia reported record revenue of $26 billion in its most recent quarter, driven almost entirely by AI chip sales. The company's Blackwell architecture is already generating massive pre-orders from hyperscalers.

Meanwhile, the custom chip market is heating up. Broadcom itself projected that its custom AI chip revenue could reach $60-90 billion by 2027, reflecting the growing appetite among tech giants for alternatives to Nvidia's off-the-shelf solutions.

Other players are also making moves:

• AMD continues to gain ground with its MI300X accelerator, offering a more cost-effective alternative to Nvidia
• Intel is pushing its Gaudi 3 accelerator for AI training workloads
• Startups like Cerebras, Groq, and SambaNova are carving out niches in specialized AI hardware
• TSMC is expanding capacity to meet the surging demand for AI chip fabrication

In this environment, OpenAI's delay in securing custom silicon could prove costly. Every quarter without its own chip means continued dependence on Nvidia's pricing and supply allocation — both of which are heavily constrained.

What This Means for the AI Industry

The Broadcom-OpenAI-Microsoft triangle illustrates a fundamental tension in the AI industry: the companies building the most advanced AI models are not necessarily the ones with the financial infrastructure to support the hardware investments those models require.

For developers and businesses building on OpenAI's APIs, this standoff has indirect but meaningful implications. Custom chips optimized for OpenAI's architecture could dramatically reduce inference costs, potentially leading to lower API pricing. Delays in the chip program mean those cost reductions may take longer to materialize.

For the broader market, this situation underscores Nvidia's continued dominance. Despite widespread efforts to develop alternatives, no company has yet produced a chip that matches Nvidia's combination of performance, software ecosystem (CUDA), and developer adoption. Every failed or delayed custom chip project reinforces Nvidia's moat.

The financial dynamics also reveal the limits of OpenAI's current business model. Despite generating an estimated $3.4 billion in annualized revenue, OpenAI remains deeply unprofitable and dependent on external capital for major infrastructure investments. An $18 billion chip project is simply beyond its independent financial capacity.

Looking Ahead: Possible Outcomes and Timeline

Several scenarios could resolve the current impasse. Microsoft could eventually agree to the 40 percent commitment, perhaps with modified terms or a phased approach. Alternatively, OpenAI could seek additional buyers — potentially other cloud providers or even sovereign AI initiatives — to distribute the purchase commitment.

There is also the possibility that OpenAI abandons or significantly scales back its custom chip ambitions, choosing instead to rely on a combination of Nvidia hardware and Microsoft's Maia chips. This would be a strategic retreat, but it might be the most pragmatic path given the financial constraints.

The most likely timeline suggests that even if all parties reach agreement in the coming months, a production-ready custom chip is still 2-3 years away from deployment at scale. Chip design, validation, manufacturing ramp-up, and software optimization all require significant time — time during which Nvidia will continue to advance its own roadmap.

For now, the standoff serves as a powerful reminder that in the AI era, the most critical bottleneck is not algorithms or data — it is silicon. And the companies that control chip manufacturing and purchasing power hold the real leverage in shaping the industry's future.