Fazlur Rahman
Quick Answer
Capital expenditures by the top five hyperscalers are expected to reach nearly $800 billion in 2026 and $1.3 trillion in 2027 per NVIDIA CFO Colette Kress. Cloud industry backlog exceeds $2 trillion. NVIDIA's hyperscale run-rate of about $195 billion implies the company captures roughly 24% of hyperscaler CapEx, equating to about $315 billion in hyperscaler revenue in calendar 2027.
Last verified: Sep 16, 2026.
At a glance
- Hyperscaler CapEx 2026: ~$800B
- Hyperscaler CapEx 2027: ~$1.3T
- Cloud industry backlog: >$2T
- NVIDIA hyperscaler run-rate: ~$195B/year
- NVIDIA share of hyperscaler spend: ~24%
- Vera Rubin revenue per gigawatt: $40B (vs $25B Blackwell, $18B Hopper)
- AI server price increases: >15% in some cases
What the $1.3 trillion actually means
The $1.3 trillion 2027 hyperscaler CapEx figure is a step-function increase from 2025 levels. Top-five hyperscaler CapEx was running at $400-$500 billion in 2025. The 2026 guidance of $800 billion is roughly 60-100% growth year-over-year. The 2027 guidance of $1.3 trillion implies another 60% growth on top of 2026. The trajectory is not linear — it reflects the compounding effect of AI training and inference demand, plus the buildout of sovereign AI infrastructure outside the traditional hyperscalers (NVIDIA Q2 FY27 earnings call, August 26, 2026).
For comparison: total US private-sector nonresidential fixed investment was roughly $3.7 trillion in 2025 (BEA data). Hyperscaler CapEx of $1.3 trillion in 2027 would be roughly 35% of that total — an unprecedented concentration of capital in a single industry segment.
Where the money goes
Hyperscaler CapEx flows to GPUs, custom silicon, networking, data-center construction, and power infrastructure. NVIDIA captures roughly 24% of the total — primarily the GPU and networking portion. Custom silicon (AWS Trainium/Inferentia, Google TPU, Microsoft Maia) captures another growing share, currently estimated at 8-12% of the total. Networking (InfiniBand, Ethernet switches, optical transceivers) is another 8-10%. Data-center construction and power infrastructure (cooling, transformers, generators) capture the remaining 50-55% (Motley Fool, September 3, 2026).
The 24% NVIDIA share is the headline number, but the construction and power infrastructure portion is where data-center labor demand comes from — see the August 2026 construction unemployment story (BLS data: construction at 3.1%, lowest in 26-year history). The full-stack CapEx figure is what drives both the GPU pricing opportunity and the construction labor squeeze.
The Vera Rubin cycle
Vera Rubin is the most important strategic update from Q2 FY2027. The platform is already running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius. Vera CPUs, Vera BlueField-4, and Spectrum-6 switch systems expand the platform beyond accelerator demand alone, which makes NVIDIA harder to evaluate as only a chip supplier (NVIDIA Q2 FY27 press release, August 26, 2026).
The Vera Rubin stack includes the Groq LPU announced earlier in the quarter. NVIDIA's "revenue per gigawatt" metric — the company's preferred measure of compute productivity — has grown from $18 billion (Hopper) to $25 billion (Blackwell) to $40 billion (Vera Rubin, including CPU, GPU, NVLink, networking, and LPU). CEO Jensen Huang's framing on the call: customers "want to race to the next generation as fast as they can" because NVIDIA's compute is so productive that the tokens the GPUs generate are "insanely profitable" for them (Carson Group earnings call transcript, August 26, 2026).
The supply ceiling and pricing power
NVIDIA's preliminary expectation is for FY2028 revenue to grow approximately 70% year-over-year, and management says supply is what limits that figure. CEO Jensen Huang put it more directly: "even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%" (Futurum Group, August 2026).
Memory cost increases have driven AI server price increases above 15% in some cases, but NVIDIA's pricing power has protected profitability while still meeting demand from customers with urgent capacity needs. NVIDIA notified customers of the AI server price increases specifically tied to higher memory costs — HBM (high-bandwidth memory) and high-capacity DRAM demand has outpaced supply through 2026 (Futurum Group, August 2026).
The AWS 2-million GPU deployment
The AWS 2-million GPU commitment is the largest disclosed single hyperscaler buildout. AWS is deploying an additional 2 million NVIDIA GPUs starting in the current quarter through Q2 FY2029, along with Vera CPUs. The expansion builds on AWS's already vast installed base of NVIDIA compute (Carson Group earnings call transcript, August 26, 2026).
AWS's NVIDIA spend is concentrated in Trainium and Inferentia silicon as well, but the 2-million-GPU commitment is the largest disclosed GPU buildout from a single hyperscaler. For neoclouds like CoreWeave and Nebius, the AWS scale-up increases the supply of alternative cloud capacity, which has historically pressured per-hour GPU rental rates on the secondary market.
What enterprise buyers should watch next
Three near-term datapoints.
- NVIDIA Q3 FY2027 earnings call (November 17, 2026): Will confirm whether Vera Rubin production is scaling to meet the implied 70% FY28 growth, and whether the AWS 2-million GPU deployment is on track.
- Goldman Sachs Communacopia + Technology Conference (September 10, 2026): Jensen Huang fireside keynote may surface next-generation product roadmap details.
- GTC Berlin (October 21, 2026): Jensen Huang keynote will likely surface European hyperscaler and sovereign AI infrastructure announcements.
What this means for the broader AI infrastructure market
The $1.3 trillion 2027 figure has implications well beyond NVIDIA. For GPU cloud buyers, the Vera Rubin rollout across 5 production partners means Blackwell Ultra pricing should ease over Q4 2026 as Vera Rubin absorbs the highest-priority workloads. For power and cooling infrastructure providers, the CapEx scale confirms a multi-year demand pipeline. For construction labor markets, the data-center buildout pipeline is the dominant driver of trade wage pressure (see the August 2026 construction unemployment story at 3.1%). For sovereign AI infrastructure builders, the scale validates the thesis that hyperscale AI infrastructure is the dominant capital deployment of the late 2020s (Motley Fool analysis, September 3, 2026).
For NVIDIA specifically, the supply ceiling is the constraint. The 70% FY2028 growth rate is what the supply chain can deliver, not what demand would absorb. Memory pricing pressure is the margin constraint. The Vera Rubin cycle is the productivity narrative. Each of these will be tested in the Q3 FY2027 print on November 17, 2026.
