OpenAI Custom Chip Deal Stalls as Broadcom Demands Microsoft Buy 40%

OpenAI's ambitious plan to build its own custom AI chips has hit a major roadblock, with partner Broadcom reportedly demanding that Microsoft commit to purchasing 40% of the first batch of production capacity before it will invest in the project. The standoff threatens to delay a massive initiative codenamed 'Nexus' that could ultimately cost $180 billion in chip production alone — a project designed to free OpenAI from its expensive reliance on Nvidia hardware.

The revelation, based on an internal memo obtained by The Information and corroborated by 2 people involved in negotiations, exposes deep tensions in what was once heralded as a landmark partnership between the AI leader and the chip design giant.

Key Takeaways

• Broadcom is requiring Microsoft to purchase 40% of the first phase's chip production capacity as a precondition for funding
• The first production phase requires 1.3 gigawatts of data center capacity and costs approximately $18 billion
• The full 'Nexus' project targets 10 gigawatts of power consumption by 2030 — equivalent to 5 Hoover Dams
• Total chip production costs alone could reach $180 billion, excluding data center construction
• OpenAI projects it will burn through more than $200 billion in operating costs by 2029
• The project aims to reduce OpenAI's costly dependence on Nvidia's GPU ecosystem

A Partnership That Promised to Reshape AI Hardware

When OpenAI and Broadcom announced their joint AI chip venture last fall, the deal appeared to be a done deal. Both companies framed the collaboration as a strategic masterstroke — OpenAI would gain access to custom silicon tailored to its specific AI workloads, while Broadcom would cement its position as a critical player in the AI chip supply chain.

The logic was compelling. OpenAI currently spends billions annually on Nvidia's H100 and newer Blackwell GPUs, which dominate the AI training and inference market. Custom chips, designed specifically for OpenAI's transformer-based architectures, could deliver significant cost savings at scale while reducing the company's vulnerability to Nvidia's pricing power and supply constraints.

However, months of negotiations have revealed fundamental disagreements about who should bear the financial risk of such an enormous undertaking. The project's sheer scale — dwarfing most semiconductor initiatives in history — has made both sides cautious about committing capital without guarantees.

Broadcom's $18 Billion Ultimatum

At the heart of the impasse is Broadcom's insistence that Microsoft, OpenAI's largest investor and cloud computing partner, serve as a guaranteed buyer for a substantial portion of the initial chip output. Specifically, Broadcom wants Microsoft to commit to purchasing the first 40% of production capacity from the project's opening phase.

This first phase alone is staggering in scope. It requires building out 1.3 gigawatts of data center computing capacity at an estimated cost of $18 billion. To put that in perspective, a single gigawatt can power roughly 750,000 homes, and most large data centers today operate in the 50 to 200 megawatt range.

Broadcom's demand is not unreasonable from a business standpoint. The chip design company would be taking on significant financial exposure by funding the initial production ramp, and securing a blue-chip customer like Microsoft reduces that risk considerably. But the requirement adds another layer of complexity to an already intricate three-way relationship between OpenAI, Microsoft, and their hardware partners.

The Staggering Economics of Project Nexus

The full scope of Project Nexus is breathtaking even by Big Tech standards. When fully built out, the initiative aims to consume 10 gigawatts of electrical power — the equivalent of 5 Hoover Dams operating simultaneously. At the projected cost ratios from the first phase, chip production alone could run to $180 billion.

Critically, that figure does not include several major cost categories:

• Data center construction and land acquisition
• Power infrastructure including substations, transmission lines, and potentially dedicated power plants
• Cooling systems required for chips operating at this density
• Networking equipment to connect the massive compute clusters
• Ongoing operational costs including electricity, maintenance, and staffing

When these additional expenses are factored in, the true cost of Project Nexus could easily exceed $300 billion — placing it among the most expensive infrastructure projects in human history, rivaling national-scale energy and transportation initiatives.

The timeline is equally ambitious. OpenAI and Broadcom's original plan called for having enough custom chips online by 2030 to consume the full 10 gigawatts of power. That leaves roughly 5 years to design, fabricate, test, and deploy chips at a scale that would normally take a decade or more.

Why OpenAI Desperately Needs Custom Silicon

OpenAI's push for custom chips is driven by cold financial reality. The company projects that its operating costs will exceed $200 billion by 2029, with a massive portion of that spending flowing directly to Nvidia for GPU hardware and to cloud providers for compute time.

The economics of AI training and inference are brutal. Training a single frontier model like GPT-5 or its successors requires tens of thousands of high-end GPUs running for months, with each chip costing $25,000 to $40,000. Inference — actually running trained models for end users — is even more expensive at scale because it never stops.

Custom chips offer several potential advantages over general-purpose GPUs:

• Lower per-unit costs when produced at scale, since they eliminate features OpenAI doesn't need
• Higher efficiency for specific AI workloads, potentially delivering more compute per watt
• Supply chain independence from Nvidia, which currently controls roughly 80% of the AI chip market
• Architectural optimization tailored to OpenAI's specific model architectures and training techniques
• Long-term pricing stability rather than being subject to Nvidia's market-driven pricing

Companies like Google (with its TPU chips) and Amazon (with Trainium and Inferentia) have already demonstrated that custom silicon can dramatically reduce AI compute costs. Google's TPU program, now in its 6th generation, has been credited with giving the company a significant cost advantage in training and serving its Gemini models.

Microsoft's Delicate Balancing Act

The demand that Microsoft serve as a guaranteed buyer for 40% of initial capacity places the software giant in a complex position. Microsoft has invested more than $13 billion in OpenAI and relies heavily on OpenAI's models to power its Copilot AI products across Office, Windows, Azure, and other platforms.

At the same time, Microsoft has its own custom chip program. The company unveiled Maia 100, its first custom AI accelerator, in late 2023 and has been ramping up production. Committing billions to purchase OpenAI-Broadcom chips could potentially conflict with Microsoft's internal chip strategy.

Microsoft also maintains a deep and lucrative partnership with Nvidia. Azure is one of the largest customers for Nvidia's data center GPUs, and any move that appears to undermine that relationship requires careful navigation. The company must balance its investment in OpenAI's future against its existing hardware partnerships and its own silicon ambitions.

Industry Context: The Custom Chip Arms Race Intensifies

OpenAI's chip struggles play out against a backdrop of unprecedented investment in custom AI silicon across the tech industry. The race to reduce dependence on Nvidia has become one of the defining themes of 2025.

Google continues to advance its TPU lineup and recently announced plans to invest $75 billion in AI infrastructure this year. Amazon Web Services is aggressively pushing its Trainium2 chips as cost-effective alternatives to Nvidia GPUs. Meta has developed its own MTIA chips for inference workloads, and even Apple is reportedly exploring custom AI training hardware.

Meanwhile, Nvidia is not standing still. The company's Blackwell architecture has raised the performance bar significantly, and CEO Jensen Huang has outlined a roadmap that includes annual chip architecture refreshes — a pace designed to make it harder for custom chip projects to catch up.

The broader semiconductor industry is also grappling with capacity constraints. TSMC, which would likely manufacture any Broadcom-designed chips for OpenAI, is already struggling to meet demand from its existing customers. Securing 10 gigawatts worth of chip production capacity would require a massive expansion of fabrication facilities.

What This Means for the AI Industry

The stalled negotiations carry significant implications beyond the 3 companies directly involved. If Project Nexus fails or is substantially delayed, it would reinforce Nvidia's dominance over the AI hardware market for years to come. This could lead to higher costs across the entire AI ecosystem, from startups building applications to enterprises deploying AI solutions.

For developers and businesses building on OpenAI's platform, the outcome will eventually affect pricing. If OpenAI cannot reduce its hardware costs through custom chips, those expenses will inevitably flow through to API pricing and subscription fees. OpenAI's current pricing has already come under pressure from competitors like Anthropic, Google, and open-source alternatives.

The situation also raises questions about OpenAI's financial sustainability. With projected operating costs exceeding $200 billion by 2029 and the custom chip project requiring hundreds of billions more, the company's path to profitability depends heavily on either dramatically increasing revenue or finding ways to slash hardware expenses.

Looking Ahead: Critical Months for Project Nexus

The next few months will likely determine whether Project Nexus moves forward in its current form, gets restructured, or collapses entirely. Several key factors will shape the outcome.

First, Microsoft's willingness to commit to the 40% capacity purchase will be the most immediate variable. If Microsoft agrees, the first phase could begin moving toward production. If it balks, Broadcom may seek alternative anchor customers or walk away from the deal.

Second, OpenAI's ongoing fundraising efforts will play a role. The company has been exploring various financing structures, including debt instruments and strategic investments, to fund its massive infrastructure needs. Securing additional capital could give OpenAI more leverage in negotiations with Broadcom.

Finally, the competitive landscape will continue to evolve. If Nvidia's next-generation chips deliver the efficiency gains Jensen Huang has promised, the economic case for custom silicon could weaken. Conversely, if Nvidia's pricing continues to climb, the urgency of Project Nexus only grows.

What's clear is that the era of building AI at this scale has entered uncharted financial territory. The sums involved — hundreds of billions of dollars for a single chip program — would have seemed absurd just 3 years ago. Today, they reflect the staggering computational appetite of frontier AI systems and the high-stakes race to power them.