AI chip rivals Nvidia and Broadcom diverge as data-center architectures overlap after latest earnings

Nvidia's data-center segment stretches beyond GPUs

Nvidia's fiscal 2027 second-quarter data-center revenue reached $89 billion, up 117% from a year earlier, anchored by GPU sales but increasingly accompanied by adjacent silicon and software. Grace CPUs, launched in 2021, have generated more than $5 billion in sales over the last year, and management expects the new Vera Rubin CPUs to add roughly $20 billion in revenue by next quarter, with fiscal 2028 revenue forecast to more than double as the platform ramps. The company's software stack has expanded through CUDA, Nemotron, Cosmos and Isaac, and Nvidia's $12.9 billion acquisition of Hugging Face was framed as the clearest signal that the company is positioning itself as a destination where generative models are shared, not only where they run.

Broadcom's custom silicon and networking pull ahead

Broadcom's fiscal third-quarter AI semiconductor revenue hit $16.7 billion, up 221% year over year and accounting for 56% of total revenue, with custom accelerators, or XPUs, making up 73% of that AI chip total. Management guided AI semiconductor revenue to $115 billion in fiscal 2027 and $230 billion in fiscal 2028, while CEO Hock Tan has said customer demand already exceeds those targets, with supply, not demand, identified as the binding constraint. Networking solutions are expected to grow at a comparable pace in the upcoming quarter, and infrastructure software generated $8.8 billion in revenue, up 29% year over year, reinforcing the VMware acquisition as a complement to the semiconductor business. Custom chips for Alphabet's Google Cloud, OpenAI, Meta Platforms and Anthropic sit atop Broadcom's Ethernet switches, optics and software.

Two architectures competing for the same AI capex

The framing inside Nvidia and Broadcom's earnings reports has shifted from GPUs versus custom XPUs to competing visions for the entire data-center stack. Nvidia is positioning itself as a reusable AI factory that any cloud provider or enterprise can plug into, while Broadcom is building purpose-built engines for a handful of giant customers ordering capacity by the gigawatt. Broadcom's pitch from Tan is that chips designed around one customer's model requirements can be cheaper and more power-efficient than a one-size-fits-all GPU, a calculus now driving hyperscaler procurement. Nvidia's revenue mix is also broadening: AI cloud, industrial and enterprise customer revenue rose 138% year over year in the second quarter, outpacing the 102% growth in its hyperscale business.

Inference workloads emerge as the contested ground

OpenAI's Jalapeño chip, co-developed with Broadcom, was built to reduce reliance on Nvidia GPUs and lower operating costs, and OpenAI has said the chip delivers 1.5 to 1.9 times more AI work per watt than Nvidia's GB200 and GB300 chips, with lower latency across several benchmarks. That positions Broadcom's custom silicon in the inference stage, the operational deployment of trained AI models, while Nvidia's hardware remains the standard for training the largest models. OpenAI has deployed roughly 100,000 Grace Blackwell NVLink72 systems to build Astra, with CEO Jensen Huang indicating an additional 400,000 chips are being brought online, reinforcing Nvidia's role in training even as custom chips target inference economics.

Valuation gap keeps Nvidia in the larger AI trade

Nvidia traded at roughly 14.4 times forward one-year earnings as of September 10, compared with about 18.8 times for Broadcom, according to analyst commentary accompanying the earnings comparisons. Combined with Nvidia's larger revenue base, total revenue of $96.2 billion, up 106% year over year in fiscal 2027's second quarter, the valuation gap has framed Nvidia as offering greater AI scale, broader customer exposure and a more reasonable entry point relative to Broadcom's faster-growing but more richly priced custom-silicon franchise.