NVIDIA's reign as the undisputed leader of AI computing is entering its final chapter, with custom chips from competitors like Broadcom eroding its dominance within the next 18 to 24 months. Both companies just posted blockbuster earnings, but the divergence in their growth trajectories and customer strategies reveals a fundamental shift in how the world's biggest tech companies are building AI infrastructure.

Why Are Hyperscalers Suddenly Betting Against GPUs?

NVIDIA delivered a stunning quarter with data center revenue of $89.02 billion, up 117 percent year-over-year, powered largely by its Blackwell Ultra GPU architecture. The company's total revenue reached $96.22 billion, up 106 percent year-over-year. CEO Jensen Huang told investors that "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable," and guided fiscal 2028 revenue growth at approximately 70 percent.

Jensen Huang

But Broadcom's story is even more dramatic. The company's AI semiconductor revenue hit $16.70 billion, up 221 percent year-over-year and 54 percent sequentially. More importantly, Broadcom is shipping custom accelerators, called XPUs (specialized processing units), designed specifically for individual hyperscalers. Google's Ironwood TPU v7 shipped in volume, OpenAI's Jalapeno accelerator started shipping, and Anthropic is expected to deploy five gigawatts of TPU v8i in 2027.

The reason hyperscalers are moving away from general-purpose GPUs is economics. Broadcom's Hock Tan claimed that Jalapeno runs OpenAI workloads at "half the cost of a GPU," a massive advantage when you're operating at the scale of a trillion-dollar company. Tan also stated that TPU v8i is "comparable if not surpasses the Vera Rubin GPU," NVIDIA's latest flagship processor.

What Does This Mean for NVIDIA's Moat?

NVIDIA isn't defenseless. The company has built what Huang calls a "full-stack AI factory platform," not just a chip company. This includes CUDA, a software ecosystem that locks developers into NVIDIA's hardware, plus integrated networking solutions and financing partnerships with major investment firms like Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR.

However, these advantages are eroding faster than most investors realize. Broadcom has secured supply commitments for roughly $115 billion of AI revenue in fiscal 2027 and $230 billion in fiscal 2028, according to company guidance. That's a clear signal that hyperscalers have already made their bets and are locking in custom silicon production.

How to Track the Shift From GPUs to Custom Chips

• Monitor Quarterly Revenue Growth Rates: NVIDIA's data center growth is slowing from triple-digit percentages, while Broadcom's AI semiconductor revenue is accelerating. When Broadcom's growth rate exceeds NVIDIA's for two consecutive quarters, the inflection point has arrived.
• Watch Supply Chain Announcements: Broadcom is opening a Singapore substrate factory in fiscal 2027 and tripling laser capacity. Track whether these manufacturing expansions meet demand or signal supply constraints that could slow custom chip deployments.
• Follow Hyperscaler Capital Expenditure Guidance: When Google, OpenAI, Meta, and Anthropic disclose that custom chips represent more than 50 percent of their AI infrastructure spending, NVIDIA's dominance has officially shifted.
• Track CUDA Adoption Among Startups: If emerging AI companies begin building on custom accelerators instead of NVIDIA GPUs, the ecosystem lock-in that protects NVIDIA will weaken significantly.

Both companies face supply constraints. NVIDIA warned of "extreme pricing conditions in memory," while Broadcom is racing to expand manufacturing capacity. The real test will be whether Vera Rubin, NVIDIA's next-generation GPU, maintains its "fastest product ramp in NVIDIA's history" promise, and whether Broadcom can convert its 10-gigawatt 2028 Anthropic pipeline into actual shipped silicon.

Vera Rubin, NVIDIA's next-generation GPU

Valuation multiples tell part of the story. Broadcom trades at a forward price-to-earnings ratio of 18, compared to NVIDIA's 23, despite having a clearer path from $58 billion in fiscal 2026 AI revenue toward $230 billion in fiscal 2028. For investors seeking defensive exposure to AI infrastructure, NVIDIA remains the safer bet. But for those betting on the inference era, where custom chips dominate, Broadcom's XPU strategy offers the more compelling opportunity.

The timeline is tighter than most realize. Within 18 to 24 months, custom accelerators will likely capture the majority of hyperscaler AI spending, fundamentally reshaping the semiconductor industry's power structure. NVIDIA's dominance isn't ending overnight, but the era of GPU-centric AI infrastructure is entering its final chapter.