Broadcom: Broadcom Designs Chips Challenging NVIDIA for Hyperscalers

Broadcom controls most of the custom ASIC market for AI, locking in multi-year contracts with Google, Meta, ByteDance, and now OpenAI. Custom chips deliver 2-10x better efficiency than general-purpose GPUs for inference workloads, the part of AI that processes user queries after models are trained. As inference scales to 70% of all AI compute by 2027, hyperscalers need chips built for cost per query, not raw speed. This positions Broadcom as the infrastructure winner in a $60-$90 billion addressable market by fiscal 2027.

Google's token processing surged from 480 trillion in April 2025 to 1,300 trillion by October, a 171% jump driven by Gemini adoption. Every token runs on Broadcom-designed TPUs, now entering their seventh generation. As Google expands multimodal AI capabilities, chip volumes are expected to become "substantially more significant" through 2027. This single customer relationship validates the custom silicon moat and shows how fast inference demand can scale when models go mainstream.

Broadcom is on track to end fiscal 2025 with nearly $20 billion in AI revenue, growing 63% year-over-year, with projections to double annually and exceed $60 billion by 2030. The $110 billion backlog, half tied to semiconductors, provides multi-year revenue visibility as hyperscalers commit billions to custom silicon. With gross margins at 67% and operating cash flow at 45%, Broadcom generates more cash per dollar of revenue than NVIDIA while capturing spending that previously went to off-the-shelf GPUs.

OpenAI Custom Chip Design Sign-Off: OpenAI's custom AI accelerator design, co-developed with Broadcom, enters final validation ahead of mass production starting H2 2026. Any announcement confirming chip specifications, volume commitments, or early production milestones would validate the $10 billion contract and signal Broadcom's ability to scale custom silicon for a fourth hyperscale customer, reducing concentration risk.

AI Inference Market Doubles to $255 Billion by 2030: The global AI inference market is projected to grow from $106 billion in 2025 to $255 billion by 2030, driven by generative AI adoption and real-time decision-making infrastructure. As inference becomes 70% of AI compute workloads, hyperscalers will prioritize custom ASICs optimized for cost per query over training GPUs. Broadcom's 70% market share positions it to capture $35-50 billion in annual revenue by 2028 if penetration holds.

AI Infrastructure Power Constraints Drive ASIC Adoption: Data centers face energy bottlenecks as AI workloads consume 10-20 megawatts per facility by 2028. Custom ASICs deliver 5-10x better performance per watt than GPUs, making them the only viable path to scale inference without building new power plants. Regulatory pressure or utility rate hikes could force rapid adoption, turning Broadcom's efficiency advantage into the critical moat that determines which hyperscalers can afford to grow AI services profitably.

Broadcom generates most of its $20 billion AI revenue from just three hyperscalers, Google, Meta, and ByteDance. If even one customer delays chip orders, pivots to in-house design, or slows AI infrastructure spending, revenue could drop significantly within a single quarter. This concentration risk means Broadcom's growth is tightly linked to hyperscaler capital expenditure cycles, which can shift quickly based on demand, energy costs, or strategic priorities. Unlike NVIDIA, which sells to thousands of customers, Broadcom's fate depends on a handful of decisions made by three CFOs.

NVIDIA launched a custom chip design unit in early 2024 to compete directly with Broadcom's $10 billion ASIC business, while Intel won contracts with NVIDIA for custom Xeon CPUs in September 2025. Both moves threaten Broadcom's monopoly on hyperscaler chip design. If NVIDIA successfully bundles custom silicon with its dominant GPU software ecosystem, or if Intel undercuts Broadcom on price using its own fabs, customers could shift wallet share away. Marvell is also teaming with NVIDIA to offer NVLink-enabled custom chips, adding another competitor to a market Broadcom once owned outright.

Broadcom hiked VMware pricing 500-600% after its 2024 acquisition, forcing customers into bundled subscriptions and eliminating free products. The EU flagged the licensing changes as anticompetitive in June 2025, while customers are actively migrating to Nutanix and open-source alternatives. VMware generated $3.8 billion in Q3 2025, but if enterprises accelerate cloud repatriation or switch to competitors, software revenue, which funds semiconductor R&D, could shrink. This would pressure Broadcom's ability to invest in next-gen ASIC designs while competitors close the technology gap.