Where $600B is going. Eight consecutive quarters of upward revisions later, the Big Five hyperscalers have committed $660â720B in 2026 AI infrastructure spend â and Wall Street consensus has been too low every single time.
That number didn't materialize overnight. It's the result of eight consecutive quarters of upward revisions, with Wall Street consensus running too low every single time. Goldman Sachs projected $500B. The actual figure blew past it. For two straight years, estimates implied ~20% annual growth. Actual spending exceeded 50% both times.
Here's how each company's guidance evolved â from initial FY2025 print to today:
| Hyperscaler | Initial FY2025 Guidance | FY2025 Actual | FY2026 Guidance |
|---|---|---|---|
| Microsoft (MSFT) | ~$80B | ~$88â90B | ~$120B+ |
| Alphabet (GOOGL) | ~$75B | ~$91â93B | $175â185B |
| Meta (META) | $60â65B | ~$70â72B | $115â135B |
| Amazon (AMZN) | ~$100B+ | ~$125B | ~$200B |
| Oracle (ORCL) | ~$16B | ~$35B | ~$50B |
| Combined | ~$430B | ~$400â420B | $660â720B |
Sources: Company earnings calls Q1âQ4 2025; CreditSights 2026 estimates; The Register Feb 2026
The single most important number: 75%. That's the share of 2026 hyperscaler capex â roughly $450â500 billion â that is AI-specific infrastructure. Not general cloud. Not office real estate. GPU clusters, AI servers, liquid cooling, and the power grid to run them.
Verbatim from the most recent earnings calls. Every major hyperscaler described the same condition: more demand than supply.
"We will continue to increase our infrastructure investments to support the growing AI demand we're seeing." â Satya Nadella, CEO
CFO Amy Hood acknowledged "capacity constraints" â Microsoft has more AI orders than it can fulfill. ~$80B in unfulfilled Azure orders sitting in backlog because the company lacks sufficient grid power to energize GPUs already purchased.
"We're seeing our AI investments and infrastructure drive revenue and growth across the board." â Sundar Pichai, CEO
Google Cloud revenue hit $17.7B in Q4 2025, up 48% YoY. FY2026 guidance is more than double 2025 actual, exceeding all peer hyperscalers. When asked what keeps him up at night, Pichai answered: compute capacity constraints and supply chain bottlenecks.
"Demands for compute resources across the company have increased even faster than our supply." â Meta CFO
Meta has signed ~$20B in external Oracle cloud capacity on top of its own buildout â hedging against schedule risk and peak training demand. Prometheus project: a 1 GW supercluster targeting 500,000 GPUs in 2026.
"As fast as we install this AI capacity, it's getting monetized." â Andy Jassy, CEO
Amazon is the single largest AI infrastructure spender. AWS now runs at a $142B annualized revenue rate. Free cash flow turned negative in Q3 2025; Amazon raised $12B in new debt and cut 30,000 corporate roles generating ~$6B annual savings to offset capital strain.
"We are ramping capital spending to $35 billion in 2026 to build data centers and secure GPU capacity." â Safra Catz, CFO
Customers include OpenAI, xAI, Meta. Oracle bumped FY2026 capex to $50B in December â a massive pivot for a historically asset-light company. GPU rental gross margins run at mid-teens %, against claims of 30â40% on an "adjusted" basis. This ROIC gap is the central bear case for the stock.
$600B+ is not a monolithic spend. Here's how it breaks down across the AI infrastructure stack:
| Category | Share | 2026 Figure | Key Context |
|---|---|---|---|
| AI Silicon (GPUs/accelerators) | 40â50% | $240â300B | H200 â Blackwell B200 transition; Blackwell Ultra (B300) ramping Q1 2026 |
| Networking | 10â15% | $60â90B | InfiniBand, NVLink, Spectrum-X, 800G optical transceivers |
| Datacenter shells & construction | 20â25% | $120â150B | Land acquisition, building construction, real estate |
| Power infrastructure | 10â15% | $60â90B | Grid connections, transformers (128-week lead times), nuclear deals |
| Cooling systems | 5â10% | $30â60B | Liquid cooling transition; Blackwell racks at 120kW vs. H100 at 40kW |
Sources: CreditSights, Cerno Capital, analyst estimates
The Silicon Layer: The transition to Blackwell (B200/GB200) is underway â each Blackwell rack (NVL72) draws 120kW vs. H100's ~40kW and requires liquid cooling, making every Blackwell deployment a facility renovation. Blackwell Ultra (B300) began shipping in volume Q1 2026: 288GB HBM3e, 7.5x dense throughput vs. H100. Next on deck: Vera Rubin (H2 2026) â 50 PFLOPS NVFP4 inference, 10x lower cost per token vs. Blackwell.
The GPU trade was 2023â2024. The infrastructure trade is 2025â2026. Here's where the money is actually flowing:
Every megawatt of AI capacity needs electrons, and hyperscalers are signing 20-year nuclear PPAs at a premium. Meta signed a massive PPA with Vistra for Comanche Peak. Microsoft reactivated Three Mile Island Unit 1 for 20 years (Constellation). Google acquired Intersect Power for $4.75B. The IEA expects U.S. electricity demand growth to be driven overwhelmingly by data centers through 2026. U.S. utility capex now exceeds $211B annually â more than every other U.S. industry.
Every custom silicon chip being built by hyperscalers runs on ARM architecture. Amazon's Graviton and Trainium, Google's TPU, Apple's chips, NVIDIA's Grace CPU â all ARM. The sovereign AI buildout (30+ nations) is also predominantly ARM-based. ARM's business model: license the architecture, collect royalties on every chip shipped. As custom silicon grows, ARM royalties compound. Consensus: $1.47B rev, $0.58 EPS.
AI networking at 800G+ is impossible without advanced optical transceivers, and Coherent is one of two dominant suppliers. Transceiver shipments for 800G-and-above are projected to grow 163% in 2026 (24M â 63M units). Coherent is also the contract manufacturer for Lumentum â it captures the photonic supply chain at multiple layers. The stock has lagged NVDA and VRT, making the valuation more interesting relative to the exposure.
The most direct play on the Blackwell transition. When GPU rack density jumped from ~40kW (H100) to 120kW (Blackwell NVL72), air cooling became physically impossible â every new deployment requires liquid cooling. Vertiv is one of three dominant providers. Q4 2025 backlog: $15B. Organic revenue growth: 28% YoY. Evercore ISI price target: $210. YTD 2026 return: +86%. The company stopped disclosing backlog after Q4 2025 â either because the number got too large or management didn't want to anchor expectations.
The "grid-to-chip" play. Eaton makes the switchgear, transformers, and power distribution equipment that connects utility grid to data center rack. Data center sales grew 64% in 2025. Deutsche Bank named it top pick for 2026 with a $400 price target. ETN underperformed in 2025 (consensus ran too high), which means the setup for 2026 is more favorable. Key metric: order backlog and lead times â Eaton disclosed record backlogs at the UBS industrials conference, December 2025.
This is not a balanced "on the other hand" section. It's the actual risk inventory that matters.
// Signals to watch for 2026 deceleration
Three structural dynamics underrepresented in mainstream AI capex coverage.
Everyone knows Google has TPUs and Amazon has Trainium. But the consensus treats custom silicon as a minor footnote. The actual number: custom ASICs now handle 15â25% of internal hyperscaler workloads (TrendForce, early 2026). That's approaching a quarter of all compute. Google's TPU v5 matches Blackwell on specific inference benchmarks. Amazon's Trainium 2 crossed $10B ARR.
The bear implication: NVIDIA's TAM growth from hyperscaler share gains is already decelerating, even as absolute dollar amounts grow. The bull case for NVIDIA increasingly depends on (a) sovereign AI, (b) enterprise deployments outside the Big Five, and (c) inference stack advantage from CUDA/NIM lock-in. NVIDIA's response: Groq acquisition ($20B), Vera Rubin architecture optimized for inference economics, and $2B equity investment in CoreWeave.
30+ nations are now building domestic AI infrastructure â Saudi Arabia, UAE, India, Japan, France, Canada. The key word is sovereign: these projects aren't price-sensitive the way enterprise contracts are. They're geopolitically motivated. A government building a national AI compute layer doesn't negotiate the price of Blackwell GPUs the way AWS does.
NVIDIA has visibility into $500B of cumulative Blackwell/Rubin revenue through end of 2026 â a significant portion sovereign. Telcos in 5 continents are partnering with NVIDIA to build sovereign AI infrastructure. Vertiv called out sovereign AI as a "significant and unexpected" growth driver in its EMEA and APAC order books in late 2025. The market models NVIDIA against hyperscaler capex. The sovereign channel is an entirely separate demand source that appears nowhere in most sell-side models.
2023â2024 was the training era. 2025â2026 is the inference era. OpenAI processes billions of requests daily. Meta serves AI recommendations to 3B+ users. That traffic is inference, not training.
Why this matters for capital allocation: Training is bursty, GPU-hungry, and benefits from NVIDIA's Blackwell. Inference is continuous, latency-sensitive, and benefits from networking (Arista, Coherent), memory bandwidth, and cost efficiency per token. The market still assigns most of the AI capex narrative to GPU suppliers. But if inference is now 75% of total AI compute (NVIDIA's own projection for 2030), the compounders are Arista (networking), Coherent (optical), Vertiv (thermal), and NVIDIA Dynamo. This transition is why NVIDIA launched Vera Rubin â its first architecture explicitly optimized for inference economics over training throughput.
Five tickers most directly exposed to the AI infrastructure capex cycle â across silicon, power, and thermal.
| Ticker | Our Take | Consensus | Next Print |
|---|---|---|---|
| NVDA | The toll booth on AI inference. Rubin launch H2 2026 is the next catalyst. Custom silicon risk is real but embedded; sovereign AI is the surprise upside. Watch: inference token economics vs. Groq/Rubin thesis. | ~$180B FY2026 rev · 68â73% gross margin | Late May 2026 |
| VRT | Most direct Blackwell-transition play. Every rack going from 40kW to 120kW passes through Vertiv's order book. $15B backlog (Q4 2025). Risk: stopped disclosing backlog; concern is whether growth is being pulled forward. | ~$9.5â10B FY2026 rev · ~25% EBITDA margin | Late Apr/May 2026 |
| AMD â | MI400 series gaining ground in sovereign AI and inference. AMD's "fastest-ramping product in company history." CUDA moat is real but inference workloads are more ASIC-agnostic than training. Next print is the setup for H2 custom silicon narrative. | $7.25B Q1 rev · $1.09 EPS · Data Center ~$3.3B | May 5, AMC |
| ETN | Grid-to-chip. Underperformed in 2025, setting up for 2026 with more realistic expectations. Deutsche Bank top pick. Watch Q1 order growth commentary vs. backlog expansion. Tariff drag easing. | ~$27â28B FY2026 rev · ~21% operating margin | Late April 2026 |
| ORCL | $455B AI contract backlog, cloud growing 28%. Bear case: margin compression on GPU rentals (mid-teens vs. claimed 30â40% adjusted). $50B capex in 2026 is a big bet for a historically asset-light company. Bull case: OCI becomes preferred sovereign cloud. | ~$75B FY2026 rev · $7.70 EPS | June 2026 (FY Q4) |
We update this tracker each quarter as guidance revisions come in. Enter your email and we'll send you the Q3 2025 update the moment it's live.
Companies, markets, and trends. Drop in any name and get a structured intelligence brief in 40 seconds â earnings previews, sector trackers, competitive deep dives.
No credit card. Cancel anytime.