Nvidia's Revenue-Per-Gigawatt Model Redefines AI Infrastructure

Nvidia's Revenue-Per-Gigawatt Model Redefines AI Infrastructure

Nvidia is no longer selling chips—it's selling the entire physical foundation of AI, and pricing it by the gigawatt. The company disclosed in its August earnings call that its revenue opportunity has climbed from roughly $18 billion per gigawatt of AI infrastructure with its Hopper generation to $25 billion with Blackwell, and now $40 billion per gigawatt with the upcoming Vera Rubin platform, according to details first reported by 24/7 Wall St.

The shift reflects a fundamental change in what Nvidia actually delivers. Management described the offering as "a full-stack AI factory platform" that bundles the Vera CPU, Rubin GPU, NVLink and InfiniBand interconnects, Ethernet networking, physical rack systems, algorithms, and the CUDA software layer that developers code against. In practical terms, hyperscalers now buy the compute, the wiring that connects it, the enclosures it sits in, and the software environment—all from one vendor.

Nvidia reported second-quarter fiscal 2027 revenue of $96.2 billion, up 106% year over year, with Data Center revenue of $89.0 billion. The company guided third-quarter revenue to $108 billion. CEO Jensen Huang told investors the fiscal 2028 outlook of approximately 70% growth is "supply-constrained," adding that "our entire supply chain is challenged. And everybody is really running flat out."

The company also announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion in third-party capital for AI infrastructure. A SoftBank Energy campus in Ohio will host Nvidia compute for OpenAI under 20-year leases. Management noted that "Nvidia Compute is fully utilized across every cloud we serve," and that hyperscaler cloud backlogs exceed $2 trillion.

Why it matters

Nvidia's expanding revenue per gigawatt demonstrates how platform control—not just silicon performance—drives margin expansion in AI infrastructure. The CUDA software moat keeps developers locked in, while the full-stack model lets Nvidia capture value from networking, systems integration, and software that competitors historically left on the table. For enterprises evaluating AI infrastructure investments, this pricing structure signals that total cost of ownership now extends far beyond chip acquisition to encompass the entire deployment footprint.

Rivals Target Custom Silicon and Interconnect Gaps

Broadcom posted third-quarter fiscal 2026 revenue of $29.6 billion, up 86% year over year, with AI semiconductor revenue of $16.7 billion. The company co-designs custom AI accelerators—called XPUs—with hyperscalers who want silicon tuned to specific workloads rather than general-purpose GPUs. CEO Hock Tan said custom accelerators can run frontier models at "half the cost of a GPU." Named deployments include Google's TPU v8i, one gigawatt of capacity for Anthropic in 2026, and OpenAI's first-generation custom accelerator. Broadcom also supplies the Ethernet switch chips and optical components that move data between accelerators.

AMD reported second-quarter 2026 revenue of $11.5 billion, up 50% year over year, with Data Center revenue of $6.7 billion. The company's Helios rack-scale platform bundles EPYC Venice CPUs, MI450 series GPUs, Pensando networking, and software into pre-integrated racks. Anthropic committed to deploy up to two gigawatts of MI450 GPUs in Helios racks, while Meta plans up to six gigawatts of Instinct capacity. CEO Lisa Su said "customer pull for Helios is very strong and tracking ahead of our initial forecasts."

Marvell Technology delivered second-quarter fiscal 2027 revenue of $2.7 billion, with Data Center revenue of $2.2 billion representing 79% of total revenue. The company designs custom chips and interconnect silicon for hyperscalers, including high-speed optical components and Ethernet switch silicon that connect AI clusters at 800G and 1.6T speeds. Management said the custom business will "more than double year over year in fiscal 2028." Marvell also disclosed an expanded partnership with Google that includes a warrant allowing Google to acquire up to 7% of Marvell shares tied to revenue milestones.

Intel reported second-quarter 2026 revenue of $16.1 billion, up 25% year over year, with Data Center and AI revenue of $6.3 billion. Xeon 6 was selected as the host CPU for Nvidia's DGX Rubin NVL8 system. CEO Lip-Bu Tan said server CPU demand "continues to far outpace available supply." Intel's purpose-built silicon revenue nearly tripled year over year, approaching a $2 billion run rate with a target of $4 billion. Intel Foundry posted a $2.1 billion quarterly operating loss but has Intel 18A in volume production.

All five companies reported supply constraints heading into 2027, with bottlenecks in wafers, high-bandwidth memory, substrates, and power delivery limiting fulfillment of existing orders.

This analysis draws on earnings details and management commentary first reported by 24/7 Wall St.