This is a sample of CapRelay's research. CapRelay provides similar research and more for every US-listed company.
Start free trialNVIDIA's growth is still accelerating even at a scale most investors thought was already close to a ceiling. Revenue grew 106% year-over-year in Q2 FY27 to $96.2B, marking the fourth consecutive quarter of year-over-year growth acceleration. Management's own framing is that this is a supply-constrained business: guided FY28 revenue growth of approximately 70% sits well below the roughly 100% unconstrained demand growth customers are forecasting. This isn't a story of a company defending a peak; it's a company rationing allocation among willing buyers.
Key points supporting the case:
The clearest evidence that this isn't a maturing business is the guidance gap itself. Management has guided to approximately 70% revenue growth for FY28, explicitly describing this as a supply-constrained number, while simultaneously stating that customer demand forecasts point to unconstrained growth closer to 100%. Cloud industry backlog now exceeds $2T, and top-5 hyperscaler CapEx is expected to reach nearly $800B in 2026 and $1.3T in 2027.
Vera Rubin, which began production shipments in August 2026, already has purchase orders from every major hyperscaler, AI cloud, and system OEM ahead of what management expects to be the fastest product ramp in company history. Vera Rubin is expected to represent about 20% of data center revenue in Q3 FY27, just its first full quarter of availability — evidence that customers are not waiting to adopt each new generation, they are racing to it.
NVIDIA's dollar opportunity per gigawatt of AI data center capacity has grown from roughly $18B in the Hopper generation to $25B with Blackwell to $40B with Vera Rubin. This isn't simply a chip story — it reflects NVIDIA layering in NVLink scale-up networking, InfiniBand and Spectrum-X scale-out networking, standalone Vera CPUs, and now the Groq LPU for low-latency inference, alongside the GPU itself. The Q2 FY27 Groq 3 LPX launch — a rack-scale system built on NVIDIA's architecture that management says achieves nearly 4x the tokens-per-second of the next-best alternative — illustrates how NVIDIA is extending its addressable spend into premium, high-interactivity inference niches rather than ceding them to specialized competitors.
Standalone CPU (Vera) demand is a good proof point of this dollar-capture expansion: NVIDIA sees roughly $20B in total server CPU demand and expects CPU revenue to more than double in FY28, positioning NVIDIA as a leading server CPU supplier in a market it barely participated in three years ago.
The ACIE segment (AI clouds, industrial, and enterprise customers) grew 138% year-over-year and 25% sequentially in Q2 FY27, reaching $40.3B versus $48.7B from hyperscalers. Management now frames non-hyperscaler demand as durably capable of growing faster than hyperscale over time, expecting ACIE to eventually represent roughly half of the data center business. This category includes sovereign AI (which more than tripled year-over-year in Q2 FY27) and NeoClouds, which are expected to exit 2026 with 8 gigawatts of installed capacity, up from roughly 3 gigawatts at the end of 2025.
This diversification matters because it broadens the customer base beyond the five or six hyperscalers whose capital spending decisions have historically driven the bulk of the debate around NVIDIA's demand durability. It also reduces reliance on any single customer's balance sheet — countries, regional clouds, and enterprises across a much larger set of counterparties are now buying full AI factories rather than negotiating custom chip programs, a segment where NVIDIA's integrated, "buy it and operate it" platform has few credible competitors.
Even as hyperscalers and some frontier labs pursue custom silicon, NVIDIA's core differentiation — a single architecture that runs training, post-training, and inference across every major model architecture (dense, mixture-of-experts, diffusion, autoregressive) — continues to hold. Management's argument is that ASICs are workload- and cloud-specific, while NVIDIA's platform is rentable, financeable, and durable across the entire AI lifecycle. That fungibility is precisely what large infrastructure investors are underwriting: NVIDIA has now partnered with major capital providers (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR) to mobilize over $500B in third-party financing for AI infrastructure buildouts, arguing that NVIDIA's compute — unlike a single-purpose ASIC — can be redeployed if any one customer's demand falls short.
This same durability argument underpins NVIDIA's growing exposure to frontier AI labs. NVIDIA has invested nearly $50B in these labs and expects demand from labs where it extends balance-sheet support to represent roughly a quarter of its business in FY28. While this concentration and the associated credit exposure (including up to $105B in cumulative payment obligations tied to the SB Energy/OpenAI Portsmouth campus) is a genuine risk that has not been tested through a demand downturn, it is also evidence of how central NVIDIA's compute has become to the infrastructure plans of the companies most likely to define the AI era.
Memory pricing has increased more than NVIDIA anticipated, pushing gross margin guidance down to 74% for Q3 FY27, with a bottom of 71-72% expected in Q4 FY27 before settling at 72-73% in FY28 as executed price increases take effect. This is a real, near-term earnings headwind, and NVIDIA's move to an annual product cadence for data center systems raises the likelihood of recurring transition-related cost noise. But management has been direct that this pressure stems from the same AI-driven demand increase that is fueling NVIDIA's own growth, rather than from a loss of competitive positioning, and the reaffirmed 72-73% FY28 target suggests the pressure is viewed as manageable through pricing and mix rather than a permanent step-down in profitability.