The New Chip War is Fought in the Rack

The AI race is silently moving away from simply procuring chips to what NVIDIA calls ‘AI factories’. Broadcom recently presented its Q1 FY26 results, and they show how quickly that transition is happening.

The semiconductor company’s revenue rose 29% year-on-year to $19.3 billion, driven largely by demand for AI hardware. AI semiconductor revenue alone reached $8.4 billion—a 106% increase from a year ago.

Behind that surge lies a structural change in how AI infrastructure is being built. Instead of simply buying GPUs or accelerators, hyperscalers now want fully integrated systems that combine compute, networking, and software. Broadcom has moved quickly to supply exactly that.

The result is a subtle shift in power in the AI hardware market. While NVIDIA still dominates GPUs, companies such as Google, Meta, and OpenAI are increasingly designing their own accelerators while relying on Broadcom to turn those designs into large-scale systems.

Custom Silicon

The biggest reason for Broadcom’s growth is the rise of custom accelerators.

Hyperscalers that once relied heavily on NVIDIA GPUs for training models now favour specialised chips tailored to specific workloads. For example, Google continues to deploy successive generations of its Tensor Processing Units (TPUs) via Broadcom, while other developers are planning similar chips.

During the earnings call, Broadcom CEO Hock Tan said, “In 2027 and beyond, we expect to see even stronger demand from next generations of TPU. For Google, we continue our trajectory of growth in 2026 with strong demand for the seventh-generation Ironwood TPU.” 

Announced in April last year, Google’s TPU v7 Ironwood is designed for large-scale AI training and inference, with a 9,216-chip footprint per liquid-cooled pod.

Additionally, demand for Anthropic TPU clusters could grow from 1 GW of compute in 2026 to over 3 GW by 2027, he noted. Anthropic has signed orders worth $21 billion with Broadcom for custom chips.

Broadcom designs and manufactures these chips using its XPU architecture and advanced packaging technologies. It expects OpenAI to deploy its first-generation XPU in 2027 at over 1 GW of compute capacity.

Broadcom predicts that the custom silicon market will expand rapidly. It expects visibility into more than $100 billion in AI chip revenue by 2027 and confirmed that Google TPU demand will accelerate next year despite some insourcing. 

Tan added that the company’s XPU franchise extends beyond TPUs. “Contrary to recent analyst reports, Meta’s custom accelerator MTIA (Meta Training and Inference Accelerator) roadmap is alive and well. We’re shipping now. In fact, for the next generation XPUs, we will scale to multiple gigawatts in 2027 and beyond.” 

That projection highlights a growing shift away from general-purpose GPUs toward application-specific accelerators. It also explains why Broadcom’s role in the AI ecosystem is expanding even as NVIDIA remains dominant in compute hardware.

The Rise of the Rack

For years, semiconductor vendors sold components. GPUs came from one supplier, networking from another, and system integrators assembled everything inside data centres.

That model is now changing.

Today’s AI clusters contain tens of thousands of accelerators and must behave as a single machine. Networking speed, memory bandwidth, and power management have become as important as the chips themselves.

For that, Broadcom focuses on rack-scale infrastructure. Instead of supplying individual chips, the company helps hyperscalers build entire clusters around custom accelerators and networking fabrics.

During its earnings call, Broadcom said it works closely with a small group of AI companies—six customers—designing their own chips, as well as building large language models and agent-based AI platforms.

The company expects demand to continue growing as its partners expand training and inference capacity. “What is very interesting and surprising to us is that inference is driving a substantial amount of compute capacity,” Tan observed. Google Gemini models, too, now process more than 10 billion tokens per minute.

He also confirmed that the company signed a sixth custom ASIC customer and expects chip shipments to double in 2027. 

Networking Could be a New Ground

If custom chips are the engines of AI clusters, networking is the wiring that keeps them running. Training a large language model requires thousands of processors constantly exchanging data. Slow networks quickly become a bottleneck.

Broadcom has built a major business around solving that problem.

“AI networking revenue grew 60% year-on-year and represented one-third of total AI revenue,” Tan said. The company projects AI networking to grow to 40% of total AI revenue.

Its Tomahawk family of Ethernet switches, designed specifically for large AI clusters, connect accelerators at a massive scale while maintaining low latency. One recent product, the Tomahawk Ultra, competes directly with NVIDIA’s NVLink switching technology. Unlike NVIDIA’s proprietary interconnect, Broadcom relies on high-speed Ethernet standards to link processors. 

In Q4 FY26, networking generated $11 billion in revenue for NVIDIA, up more than 3.5x year over year. That strategy appeals to hyperscalers that prefer open systems over tightly integrated platforms. 

Broadcom said it has “fully secured capacity of leading-edge wafers, HBM (high-bandwidth memory), and substrates for 2026 through 2028.”

Demand for AI networking is already growing faster than the accelerator business itself. In other words, the pipes that move data between chips may soon be as valuable as the chips themselves.

NVIDIA Leads, but Landscape is Shifting

Despite Broadcom’s rapid rise, NVIDIA remains the dominant force in AI hardware. 

Analyst expectations for NVIDIA’s 2026 capital expenditure across the top five cloud providers and hyperscalers have increased by nearly $120 billion since the start of the year, bringing projected spending close to $700 billion, CFO Colette Kress had said during the company’s Q4 earnings call. 

Yet the market is becoming more complex. Kress said, “Every data center is power-constrained, and customers make architectural decisions based on performance per watt to maximise AI factory revenue.” 

NVIDIA offers a vertically integrated stack that includes GPUs, networking hardware and software tools. Broadcom, by contrast, provides the infrastructure that allows hyperscalers to build their own systems around custom silicon.

These two approaches are not mutually exclusive.

Many AI clusters still combine NVIDIA GPUs with Broadcom networking. At the same time, several large technology companies are pursuing alternatives based on in-house accelerators.

That hybrid model could define the next phase of the AI race.