OpenAI-Broadcom Chip Deal Hits $180B Funding Wall

OpenAI's ambitious plan to develop its own custom AI chips with Broadcom has hit a major financing roadblock, with the full project estimated to cost a staggering $180 billion for chip production alone — before even accounting for data center construction and other infrastructure expenses. The partnership, once considered a done deal when announced last year, is now mired in complex negotiations over who foots the bill for one of the most expensive semiconductor ventures ever conceived.

According to an internal memo and 2 people familiar with the negotiations, the 2 companies are currently discussing terms for Broadcom to finance the first phase of production, which would require 1.3 gigawatts of data center compute capacity at an estimated cost of approximately $18 billion.

Key Facts at a Glance

• Project codename: 'Nexus' — a 10-gigawatt custom chip initiative
• Phase 1 cost: ~$18 billion for 1.3 GW of data center capacity
• Full project chip costs: ~$180 billion (chip production only)
• Key demand: Broadcom reportedly wants Microsoft to commit to purchasing 40% of production capacity
• Scope: Does not include data center construction or supporting infrastructure
• Status: Financing negotiations ongoing with no confirmed resolution

Broadcom Demands Microsoft Buy 40% of Output

One of the most contentious sticking points in the negotiations centers on Broadcom's reported demand that Microsoft — OpenAI's largest investor and cloud computing partner — commit to purchasing roughly 40% of the chip production capacity. This requirement reflects Broadcom's desire to de-risk what would be an enormous financial commitment by ensuring guaranteed demand from a deep-pocketed buyer.

Microsoft has already invested more than $13 billion in OpenAI and serves as the AI company's primary cloud infrastructure provider through Azure. However, asking Microsoft to absorb nearly half the output of a custom chip program adds another layer of financial entanglement to an already complex relationship.

The demand also raises strategic questions. Microsoft has been diversifying its own chip strategy, developing its Maia 100 AI accelerator internally while maintaining its massive purchasing relationship with Nvidia. Taking on 40% of a new custom chip's production capacity would represent a significant shift in Microsoft's silicon strategy.

The Staggering Economics of Project Nexus

The numbers behind Project Nexus are breathtaking even by Big Tech standards. At 10 gigawatts of total planned capacity, the project dwarfs most existing data center deployments worldwide. To put this in perspective, the entire U.S. data center industry consumed an estimated 17 gigawatts in 2023, meaning Nexus alone would represent a substantial fraction of current national capacity.

The $180 billion price tag for chip production alone would make it one of the most expensive technology infrastructure projects in history. For comparison:

• TSMC's Arizona fab complex: ~$65 billion across 3 facilities
• Intel's Ohio fab project: ~$28 billion for initial phase
• Saudi Arabia's NEOM: ~$500 billion (entire city project)
• The International Space Station: ~$150 billion over its lifetime
• Global semiconductor capital expenditure in 2024: ~$190 billion total

When data center construction, cooling systems, power infrastructure, networking equipment, and operational costs are added, the true total cost of Nexus could easily exceed $300 billion — a figure that would strain even the combined balance sheets of OpenAI, Broadcom, and Microsoft.

Why OpenAI Wants Its Own Chips

OpenAI's push into custom silicon reflects a broader industry trend of AI companies seeking to reduce their dependence on Nvidia, which currently dominates the AI accelerator market with an estimated 80-90% market share. Nvidia's H100 and newer B200 GPUs have become the gold standard for training and running large language models, but their scarcity and premium pricing have created bottlenecks across the industry.

By developing custom chips tailored specifically to its model architectures, OpenAI could theoretically achieve several advantages:

• Cost reduction: Custom ASICs can be 3-5x more power-efficient than general-purpose GPUs for specific workloads
• Supply security: Reduced vulnerability to Nvidia's allocation decisions and supply constraints
• Performance optimization: Chips designed specifically for transformer architectures could deliver superior inference performance
• Competitive moat: Proprietary hardware creates a structural advantage that competitors cannot easily replicate
• Vertical integration: Greater control over the full technology stack from silicon to software

This strategy mirrors moves by Google (with its TPU chips, now in their 6th generation), Amazon (with Trainium and Inferentia), and Meta (with its MTIA accelerator). However, none of these companies have attempted anything approaching the scale of Project Nexus.

The Broadcom Factor: Why This Partnership Matters

Broadcom is not a typical chip manufacturer — it is primarily a chip design company that outsources fabrication to foundries like TSMC and Samsung. The company has deep expertise in designing custom ASICs for hyperscale customers, having previously developed custom chips for Google's TPU program and other major cloud providers.

For Broadcom, the OpenAI partnership represents both an enormous opportunity and a significant risk. The company's semiconductor revenue reached $30.1 billion in fiscal 2024, with AI-related revenue growing rapidly. A successful Nexus project could cement Broadcom's position as the go-to partner for custom AI silicon.

However, the financing structure being discussed — where Broadcom would fund the first phase of production — would require the company to take on substantial balance sheet risk. Broadcom's total assets stood at approximately $165 billion as of late 2024, meaning the $18 billion first phase alone would represent a material commitment relative to the company's size.

Industry Context: The AI Chip Arms Race Intensifies

The OpenAI-Broadcom saga unfolds against a backdrop of unprecedented investment in AI infrastructure. Major technology companies have collectively committed hundreds of billions of dollars to AI data centers, chip development, and related infrastructure over the next several years.

President Trump's Stargate project, announced in January 2025, envisions up to $500 billion in AI infrastructure investment — with OpenAI, SoftBank, and Oracle as lead partners. The relationship between Stargate and Project Nexus remains unclear, but the overlapping ambitions suggest that OpenAI is pursuing multiple paths to securing the compute resources it needs.

Meanwhile, Nvidia continues to extend its lead in the GPU market. The company's upcoming Rubin architecture, expected in 2026, promises another generational leap in AI training and inference performance. This creates a moving target for custom chip projects like Nexus, which could take 2-3 years to reach production — by which time Nvidia's offerings may have advanced significantly.

The competitive landscape also includes several well-funded startups:

• Cerebras Systems: Recently filed for IPO with its wafer-scale AI chips
• Groq: Focused on ultra-fast inference with its LPU architecture
• d-Matrix: Developing in-memory computing chips for AI inference
• Tenstorrent: Led by legendary chip architect Jim Keller

What This Means for the AI Industry

The financing challenges facing Project Nexus highlight a fundamental tension in the AI industry: the gap between ambition and economic reality. While AI models continue to grow in capability and compute requirements, the infrastructure needed to support them is becoming prohibitively expensive — even for the world's wealthiest technology companies.

For developers and businesses building on OpenAI's platform, the outcome of these negotiations could have significant long-term implications. Custom chips optimized for OpenAI's models could dramatically reduce inference costs, potentially leading to lower API prices for end users. Conversely, if the project stalls or fails, OpenAI will remain dependent on Nvidia's pricing and supply decisions.

The Microsoft angle adds another dimension of complexity. If Microsoft does agree to purchase 40% of Nexus production capacity, it could signal a deeper integration between Microsoft and OpenAI's infrastructure — potentially making Azure an even more attractive platform for AI workloads while raising questions about competitive neutrality.

Looking Ahead: Uncertain Timeline, Massive Stakes

The path forward for Project Nexus remains deeply uncertain. Several key questions will determine the project's fate in the coming months.

First, will Broadcom agree to finance the $18 billion first phase, and under what terms? The company's willingness to take on this risk will depend heavily on the demand guarantees it can secure from Microsoft and potentially other customers.

Second, can OpenAI secure the additional capital needed for the full $180 billion-plus project? Even with its recent $6.6 billion funding round — which valued the company at $157 billion — OpenAI's resources fall far short of what Nexus requires. The company will likely need to raise significantly more capital, potentially through a combination of equity financing, debt, and strategic partnerships.

Third, how will Nvidia respond? The GPU giant has shown a willingness to compete aggressively on pricing and performance when threatened by custom chip initiatives. A more aggressive Nvidia pricing strategy could undermine the economic case for Project Nexus.

What is clear is that the AI infrastructure race is entering a new phase — one defined not just by technological ambition but by the hard realities of financing, manufacturing, and supply chain economics. The OpenAI-Broadcom negotiations may well become a defining moment in determining how the AI industry's compute foundation is built and who controls it.