OpenAI's $10B Broadcom Chip Deal: The Power Game of AI Computing Infrastructure

Late one night in September 2025, Silicon Valley investment circles exploded with news: OpenAI had signed a custom chip deal worth over $10 billion with Broadcom, with the first batch scheduled for mass production in 2026. Someone immediately joked: "Now Nvidia will lose sleep." But those truly in the know understood this wasn't just a contract—it was a "coup" regarding the future of AI infrastructure.

Why Call It a Coup?#

Because for the past decade, artificial intelligence computing power has been almost monopolized by Nvidia. Training GPT-4 and GPT-5 relies on computing matrices built from thousands of A100 and H100 cards. Every AI company must bow to Jensen Huang—queuing, paying premiums, even bundled purchasing. Computing power is no longer market-priced but a power relationship under scarce resources. OpenAI's move to build chips is essentially about taking back control of its own destiny.

Broadcom's Rise to Power#

The story traces back to 2013, when Avago acquired LSI Logic. This name might be unfamiliar, but it was one of the global pioneers in custom ASIC design. In 2016, Avago swallowed the veteran Broadcom and went global under the Broadcom name. In subsequent years, its appetite grew larger:

2018: Acquired CA Technologies
2019: Absorbed Symantec's enterprise security division
2023: Acquired VMware for a staggering $69 billion

Chips and software became bound under one empire.

Ruthless Business Logic#

Broadcom's business logic is extremely ruthless: it doesn't pursue all market share, only serving giants who can afford to pay. Google's TPU chips are one of its creations. Behind every generation of TPU, you can see Broadcom engineers' fingerprints—from logic design to 3.5D packaging to HBM stacking. Meta and ByteDance have also approached it for inference accelerators.

It's not a "card-selling" company, but an "arms dealer crafting weapons for the few."

Why OpenAI?#

Because GPT series inference costs have approached an unsustainable critical point. External estimates suggest GPT-4's single conversation inference cost ranges from a few cents to a dime. With billions of daily calls, costs surge like a flood.

The Critical Point of Cost Pressure#

GPU versatility is good, but the price is soaring power consumption and costs. Broadcom's ASIC is different: it sacrifices versatility for ultimate energy efficiency. Once model architecture becomes relatively stable, ASIC's per-token cost can be 30% to 50% lower than GPUs. For OpenAI, this means survival.

System-Level Integration Capability#

More critically, Broadcom can build not just chips but "racks." Its Tomahawk series switching chips have become the implicit standard for AI data centers:

Single chip provides 51.2Tbps switching capacity
Supports 400G to 800G cluster interconnection
Combined with 3.5D XDSiP packaging technology
Stacks logic chips, IO chips, and HBM on one substrate
Creates "chiplet puzzles" exceeding 6000 square millimeters

In other words, Broadcom can deliver integrated chips, memory, networking, packaging, and complete systems. Nvidia builds a "card empire," while Broadcom creates a "system arsenal."

The Software Ecosystem Shortcoming#

Of course, its shortcoming is obvious: software ecosystem.

Nvidia's CUDA is the de facto standard for the entire deep learning industry, with almost all frameworks, operators, and compilers prioritizing CUDA support. Broadcom can only rely on open-source compilation stacks like XLA and PyTorch-XLA, patching kernels and adapting operators themselves. For developers, this is an incompletely cultivated wilderness.

Whether this shortcoming can be addressed determines Broadcom's future.

Market Feedback#

The market has already responded. After news of the OpenAI partnership broke, Broadcom's stock price surged, with analysts stating: "It's replaying Nvidia's growth trajectory."

Financial Data Speaks#

Latest earnings show:

Broadcom's AI chip revenue surged 63% year-over-year
Single quarter: $5.2 billion
Next quarter projected to grow to $6.2 billion

Behind these numbers aren't just OpenAI, but heavyweight customers like Google and Meta.

Don't Rush to Write Nvidia's Obituary#

GPU versatility remains irreplaceable. Large model training iterates extremely fast, and ASICs might not keep up. The real landscape is hybrid:

Training: Still completed on GPUs
Inference: Gradually shifting to ASICs to reduce TCO (Total Cost of Ownership)

This is like WWII air force equipment: strategic bombers still need versatile platforms, but battlefield fighters can be highly customized.

The Control Game#

This game has another layer of metaphor: control.

If OpenAI completely depends on Nvidia, its business model is constrained by external hardware vendors' pricing and capacity. Broadcom's model is "symbiotic" with customers: customers place orders, Broadcom helps build chips, binding life cycles together.

For Sam Altman, this is strategically inevitable. After all, he's still planning the "Stargate" super data center, aiming to build the world's largest computing facility over the next decade. Without hardware-level independence, his computing empire is just a rented palace.

Risks Ahead#

Timeline Challenges#

The 2026 mass production target isn't easy:

Advanced packaging capacity is tight
TSMC's CoWoS production lines are already at full capacity
Yield control is also a hurdle

Model Evolution Uncertainty#

If future mainstream forms rapidly shift toward:

Mixture of Experts (MoE) models
Retrieval-Augmented Generation (RAG)
Video generation

Current ASIC architectures might be "born at the wrong time."

Regulatory Risks#

Finally, regulation—OpenAI's previous investment in Rain AI faced CFIUS review. Cross-border AI chip cooperation could become a geopolitical bargaining chip at any time.

Ultimate Significance: Multipolar Computing Landscape#

So what's the ultimate significance of this deal?

It's not about killing Nvidia, but telling the world: GPUs are no longer the only path.

AI's computing landscape will shift from single monopoly to multipolarity. In coming years, we'll see:

Structural division of "GPU training + ASIC inference"
Compilers and distributed frameworks becoming new battlegrounds
Computing sovereignty becoming a strategic resource contested by both companies and nations

Historical Echoes#

In the 1980s, Japan's NEC once dominated globally with custom chips but ultimately lost to the PC universal architecture wave. Today, can Broadcom avoid repeating history? The answer isn't necessarily optimistic.

But one thing is certain: When OpenAI and Broadcom shook hands, the old order of AI infrastructure began to crack. What emerges from the cracks isn't sunlight, but another, more brutal computing war.

Conclusion#

To conclude, I'd like to quote Adorno: "Technology is not neutral; it carries history." Broadcom's custom chips represent giants' competition for future computing power—a triple game of cost, power, and survival.

We may be standing at the threshold of a new era—the GPU empire still stands tall, but ASIC guerrilla warfare has quietly begun.

This article is a deep reflection on AI infrastructure transformation. If you're interested in AI industry technical trends and business landscapes, follow my AI100 Challenge for more insights.