The build-out of artificial intelligence compute and data centre infrastructure is the single largest physical and capital-intensive technology programme of this decade. ABI Analytics sizes the in-scope global industry at $431 billion in 2024, rising to $608 billion in 2025, $1.25 trillion in 2030 and $1.84 trillion in 2035 under our Base scenario – a compound annual growth rate of 11.7% over the eleven-year forecast window.
Six segments make up the in-scope industry: AI compute hardware (GPUs, custom silicon, HBM); hyperscale and colocation real estate; power and grid infrastructure; cooling and mechanical; neocloud / GPU-as-a-Service; sovereign AI programmes. The defining constraint is no longer demand – that question is settled – but physical capacity to deliver electrical power. ERCOT, PJM, CAISO and Dominion Virginia interconnect queues have grown to multi-year backlogs.
Five regional deep-dives anchor this report: the United States (46% of global capex; saturating Northern Virginia overflows into Texas, Arizona, Ohio); China (parallel stack under United States export controls, anchored on Huawei Ascend and SMIC); India (fastest pipeline expansion globally, IndiaAI Mission anchor); the Middle East (G42 UAE, HUMAIN Saudi, cheap gas-fired electricity, sovereign capital depth); and the European Union plus United Kingdom (constrained on power, burdened by AI Act compliance, AI Factories programme).
Cumulative industry capex 2025-2030 totals $4.6 trillion across all regions under our Base scenario. The top-15 listed competitors captured $273 billion of in-scope revenue in 2024 – 63% of TAM – leaving a $158 billion long-tail dominated by Chinese cloud providers, sovereign programmes, hundreds of regional colocation operators, and the still-mostly-private major neoclouds.
Top-of-class equity exposure: Nvidia (NASDAQ: NVDA) as the dominant AI silicon platform; the integrated United States hyperscalers (MSFT, GOOGL, AMZN, META, ORCL); Equinix (NASDAQ: EQIX) and Digital Realty (NYSE: DLR) for colocation REIT; TSMC (NYSE: TSM) and SK Hynix (KRX: 000660) for upstream foundry and memory; Vertiv (NYSE: VRT) for cooling and power-distribution; CoreWeave (NASDAQ: CRWV) as public-market neocloud beta. Private-market opportunities concentrate in Indian operators (Yotta, CtrlS, Sify) and Middle East sovereign-linked vehicles.
| Step | Variable | 2025 | 2030 | 2035 | Source |
|---|---|---|---|---|---|
| 1 | Global DC power capacity (GW) | 295 | 585 | 820 | IEA Electricity 2024 + grid interconnect queue data |
| 2 | AI-purpose capacity share (%) | 28% | 45% | 55% | Hyperscaler capex guidance + GPU shipment tracking |
| 3 | AI capacity (GW) | 82 | 263 | 451 | Computed; reconciled to hyperscaler disclosures |
| 4 | Avg capex per GW ($B/GW) | $5.0 | $5.5 | $5.7 | Hyperscaler dollars deployed / MW commissioned |
| 5 | Hyperscale + colocation revenue ($B) | $225 | $395 | $565 | Hyperscaler infra + Equinix/DLR/GDS/VNET |
| 6 | AI compute hardware revenue ($B) | $185 | $350 | $475 | Nvidia DC + AMD DC + Broadcom AI + TSMC + HBM |
| 7 | Power & grid equipment ($B) | $80 | $182 | $282 | Vertiv + Schneider + Eaton + utility-attributable capex |
| 8 | Cooling & mechanical ($B) | $48 | $104 | $165 | Vertiv + Schneider + Asetek + CoolIT + Submer + GRC |
| 9 | Neocloud / GPU-aaS ($B) | $42 | $135 | $222 | CoreWeave + Lambda + Crusoe + Nebius + sovereign neoclouds |
| 10 | Sovereign AI programmes ($B) | $28 | $82 | $128 | US Stargate + UAE G42 + Saudi HUMAIN + India + EU |
| 11 | TOTAL Industry TAM ($B, Base) | $608 | $1,248 | $1,837 | = sum of 5,6,7,8,9,10 |
| Rank | Company | Country | In-scope ($B) | % of TAM |
|---|---|---|---|---|
| 1 | Nvidia Corp. | USA | $56.0B | 13.0% |
| 2 | AMD | USA | $12.9B | 3.0% |
| 3 | Broadcom (compute & networking) | USA | $21.7B | 5.0% |
| 4 | SK Hynix (HBM) | Korea | $17.0B | 3.9% |
| 5 | Samsung (memory/DC) | Korea | $16.8B | 3.9% |
| 6 | Microsoft (DC value-add) | USA | $29.4B | 6.8% |
| 7 | Alphabet (DC value-add) | USA | $31.5B | 7.3% |
| 8 | Amazon (DC value-add) | USA | $31.9B | 7.4% |
| 9 | Meta Platforms (DC value-add) | USA | $16.5B | 3.8% |
| 10 | Oracle (DC value-add) | USA | $9.5B | 2.2% |
| 11 | Equinix Inc. | USA | $8.7B | 2.0% |
| 12 | Digital Realty Trust | USA | $5.5B | 1.3% |
| 13 | GDS Holdings | China | $1.4B | 0.3% |
| 14 | Vertiv Holdings | USA | $6.8B | 1.6% |
| 15 | Schneider Electric (DC) | France | $7.2B | 1.7% |
| – | Sum top-15 in-scope | $272.8B | 63.3% | |
| – | ABI 2024 Industry TAM | $431B | 100% | |
| – | Long-tail / private gap | Chinese cloud providers + sovereign programmes + private neoclouds + regional colocation operators | $158.2B (36.7%) | |
| Rank | Company | Country | Ticker | In-scope ($B) | % of TAM |
|---|---|---|---|---|---|
| 1 | Nvidia Corp. | USA | NASDAQ: NVDA | $56.0B | 13.0% |
| 2 | AMD | USA | NASDAQ: AMD | $12.9B | 3.0% |
| 3 | Broadcom (compute & networking) | USA | NASDAQ: AVGO | $21.7B | 5.0% |
| 4 | SK Hynix (HBM) | Korea | KRX: 000660 | $17.0B | 3.9% |
| 5 | Samsung (memory/DC) | Korea | KRX: 005930 | $16.8B | 3.9% |
| 6 | Microsoft (DC value-add) | USA | NASDAQ: MSFT | $29.4B | 6.8% |
| 7 | Alphabet (DC value-add) | USA | NASDAQ: GOOGL | $31.5B | 7.3% |
| 8 | Amazon (DC value-add) | USA | NASDAQ: AMZN | $31.9B | 7.4% |
| 9 | Meta Platforms (DC value-add) | USA | NASDAQ: META | $16.5B | 3.8% |
| 10 | Oracle (DC value-add) | USA | NYSE: ORCL | $9.5B | 2.2% |
| 11 | Equinix Inc. | USA | NASDAQ: EQIX | $8.7B | 2.0% |
| 12 | Digital Realty Trust | USA | NYSE: DLR | $5.5B | 1.3% |
| 13 | GDS Holdings | China | NASDAQ: GDS | $1.4B | 0.3% |
| 14 | Vertiv Holdings | USA | NYSE: VRT | $6.8B | 1.6% |
| 15 | Schneider Electric (DC) | France | EPA: SU | $7.2B | 1.7% |
Dominant AI accelerator share globally (~88%). CUDA + NVLink + InfiniBand ecosystem moat. H100/H200/B200/GB200; Rubin 2026. TSMC CoWoS dependency. CORE LONG.
Widest AI moat. OpenAI $14B partnership + Stargate JV $100B initial + G42 UAE framework. Capex $32B (FY23) → $80B (FY25). CORE LONG.
Sole leading-edge supplier. CoWoS monopoly. Every major AI chip fabbed here. HPC+AI share rising 30% → 50%+. Taiwan concentration = dominant risk.
TPU v6 (Trillium) most mature non-Nvidia accelerator. Capex $52B (FY24) → $75-85B guide (FY25). Gemini competitive with GPT-4. CORE LONG.
~270 colocation DCs in 33 countries. Interconnection (Fabric, Network Edge) = 90%+ gross-margin recurring revenue. CORE LONG (REIT).
Leading neocloud. ~250k Nvidia GPUs. IPO March 2024 ($1.5B / $19B valuation). Microsoft anchor (~60% revenue). SECONDARY LONG (volatile).
| Layer | Standards / methods | Top vendors | Maturity |
|---|---|---|---|
| L7 I Application & model layer LLM training, inference, agentic AI | PyTorch, JAX, CUDA, Triton | OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral, xAI | Emerging |
| L6 I Orchestration & ops Cluster mgmt, scheduler, observability | Kubernetes, Slurm, Ray, MLflow, W&B | Run:ai (Nvidia), CoreWeave, NVIDIA Base Command | Pilot |
| L5 I AI compute GPUs, custom silicon, accelerators | CUDA, HBM3/HBM4, NVLink, PCIe 5/6 | Nvidia (H100/B200/Rubin), AMD (MI300), Google TPU | Commercial |
| L4 I Networking & interconnect Intra-DC + WAN | InfiniBand (NDR/XDR), Ethernet/Ultra Ethernet | Broadcom (Tomahawk), Nvidia (Quantum), Marvell, Cisco | Commercial |
| L3 I Power & cooling Generation, distribution, heat reject | Behind-the-meter gas, on-site solar, SMR PPAs | Vertiv, Schneider Electric, Eaton, Cummins, GE Vernova | Commercial |
| L2 I Physical infra Real estate, construction, civil | TIA-942, ASHRAE TC 9.9 | Skanska, AECOM, Turner, Jacobs, Holder, DPR | Mature |
| L1 I Grid interconnect Utility, transmission, substation | IEEE 519, NERC reliability, regional ISO rules | Local utilities + PJM, ERCOT, CAISO, Dominion VA | Mature |
| Technology | TRL 1 | TRL 2 | TRL 3 | TRL 4 | TRL 5 | TRL 6 | TRL 7 | TRL 8 | TRL 9 |
|---|---|---|---|---|---|---|---|---|---|
| GPU compute (Nvidia Blackwell/Rubin) | ● | ||||||||
| HBM3E memory packaging | ● | ||||||||
| CoWoS advanced packaging (TSMC) | ● | ||||||||
| InfiniBand / Ultra Ethernet fabric | ● | ||||||||
| Air cooling (CRAH) | ● | ||||||||
| Direct-to-chip liquid cooling | ● | ||||||||
| Single-phase immersion cooling | ● | ||||||||
| Two-phase immersion cooling | ● | ||||||||
| SMR for DCs (NuScale, X-energy, Oklo) | ● | ||||||||
| Behind-the-meter gas turbines | ● | ||||||||
| AI-optimized power chains (HVDC) | ● | ||||||||
| Photonic compute / silicon photonics | ● |
SK Hynix HBM4 commercial 2026; Samsung six months behind; Micron twelve months behind. Determines AI accelerator bandwidth + capacity per package.
TSMC CoWoS capacity tripled 2023-2025; doubling again to 2026. Customer demand continues to exceed supply. SoIC successor coming.
NuScale, X-energy, Oklo, Kairos – first commercial operation 2028-2030. Structural answer to grid-interconnect queue pressure for AI compute.
Power density per rack rising 15 kW (traditional) → 60-150 kW (AI). Vertiv, Schneider, Asetek, CoolIT, Submer, GRC.
Google TPU v6, AWS Trainium 2, Meta MTIA 2, Microsoft Maia 100/200. Internal threat to Nvidia 88% accelerator share over 2027-2030.
The era of demand-driven capacity is over for the US, EU and increasingly APAC. Behind-the-meter gas, nuclear restarts (Constellation TMI, NextEra Duane Arnold), SMR commitments (Oklo, X-energy, NuScale) and unconventional PPAs are the operator responses. Companies with secured power positions trade at structural premia.
Google TPU v6, AWS Trainium 2, Meta MTIA 2, Microsoft Maia 200 collectively threaten Nvidia's 88% accelerator share. Economics favour custom silicon (hyperscalers avoid Nvidia's 75%+ gross margin). Currently serves first-party workloads but rarely cost-competitive for third-party demand. Trajectory matters for 2027-2030.
UAE G42, Saudi HUMAIN, India IndiaAI Mission, EU AI Factories, US Stargate, METI Japan – most coordinated sovereign-capital deployment since post-war defence build-out. Cumulative commitments >$200B through 2030. Project-finance, PPP, equity stakes in named operators.
CoreWeave, Lambda, Crusoe, Nebius, Voltage Park – ~500k GPUs operated. Exposed to GPU resale-value collapse and hyperscaler in-sourcing. CoreWeave IPO March 2024 was proof-of-concept; next 18 months tests durability.
Air cooling handles racks to 15-25 kW. AI racks 30-70 kW today, 130-150 kW with Blackwell GB200. Direct-to-chip liquid now standard for new AI buildouts. Single-phase immersion TRL 7-8; two-phase TRL 5-6.
Northern Virginia, Loudoun County, Frankfurt, Amsterdam, Singapore, Dublin all show meaningful local pushback. Operators with water-recovery + renewable-PPA + community-benefit agreements obtain permitting faster.
NVLink, InfiniBand NDR/XDR, Ultra Ethernet determine training-cluster scale. Broadcom, Nvidia, Marvell, Cisco, Arista compete in-DC. Submarine cables (SubCom, NEC, ASN + hyperscaler private) determine inter-DC bandwidth.
EU AI Act phased 2025-2027; China data localisation; India DPDP Act; Saudi NDMO rules. Patchwork creates physically-in-country workload concentration and sovereign-cloud vehicle opportunities.
| Driver | Base 2030 assumption | +20% input | -20% input |
|---|---|---|---|
| AI workload growth (tokens/yr) | Doubling annually 2025-28 | +$250B (+20%) | -$225B (-18%) |
| GPU supply ceiling (CoWoS wafer-out) | TSMC CoWoS 2x by 2026 | +$150B (+12%) | -$185B (-15%) |
| Power availability (GW/yr added) | ~50 GW/yr 2025-30 | +$150B (+12%) | -$150B (-12%) |
| Compute efficiency (FLOPs/W) | ~2x per Nvidia generation | +$110B (+9%) | -$100B (-8%) |
| Hyperscaler capex/revenue ratio | ~28% 2025-26 | +$125B (+10%) | -$125B (-10%) |
| Sovereign AI disbursement | $82B/yr by 2030 | +$75B (+6%) | -$60B (-5%) |
| Region | Power & infra | Policy | Cost | Capital depth | ESG | AVG |
|---|---|---|---|---|---|---|
| United States | 3 | 4 | 3 | 5 | 4 | 3.8 |
| India | 5 | 4 | 4 | 3 | 3 | 3.8 |
| Middle East | 5 | 4 | 4 | 4 | 3 | 4.0 |
| Japan | 3 | 4 | 3 | 4 | 4 | 3.6 |
| Korea | 3 | 3 | 3 | 4 | 4 | 3.4 |
| Singapore | 2 | 4 | 3 | 4 | 5 | 3.6 |
| EU & UK | 2 | 3 | 3 | 4 | 5 | 3.4 |
| China | 3 | 3 | 5 | 4 | 2 | 3.4 |
| Brazil | 5 | 3 | 3 | 2 | 4 | 3.4 |
| Indonesia | 4 | 3 | 4 | 3 | 3 | 3.4 |
Best: NVDA, MSFT, GOOGL, AMZN, META, ORCL – integrated hyperscaler + chip-leader basket.
Worst: sub-scale pure-plays with weak balance sheets.
Best: TSMC, SK Hynix, Broadcom, AMD – highest gross margin tier.
Worst: commodity DRAM exposure outside HBM premium.
Best: EQIX, DLR, VRT, GEV, ETN – predictable cash flows + infrastructure beneficiaries.
Worst: over-leveraged secondary REITs.
Best: Indian operators (Yotta, CtrlS, Sify); MENA sovereign vehicles (G42, HUMAIN co-invest); project finance with secured PPAs.
Worst: speculative greenfield without power secured.
| # | Risk | Likelihood | Evidence | Mitigation |
|---|---|---|---|---|
| 1 | Power availability binds harder than IEA reference | HIGH | ERCOT/PJM/Dominion queues 4-7 yr; transformer lead times 60-90 wk | SMR commercial deployment 2028-30; behind-the-meter gas; nuclear restarts |
| 2 | TSMC geopolitical shock or capacity failure | MED-HIGH | Taiwan concentration; CoWoS bottleneck | TSMC Arizona ramp; Samsung Foundry expansion; Intel Foundry |
| 3 | AI workload demand resets – compression event | MED | Enterprise AI ROI questions; agentic AI productisation | Hyperscaler capex discipline; multi-year customer contracts |
| 4 | Hyperscaler custom silicon displaces Nvidia faster | MED | Google TPU v6 maturity; AWS Trainium 2 scaling | Nvidia software ecosystem moat; customer switching cost |
| 5 | Neocloud business-model fragility | MED-HIGH | CoreWeave Microsoft concentration; GPU resale-value risk | Diversified customer base; tighter underwriting; longer-tenor contracts |
| 6 | Sovereign AI programme execution misfires | MED | Procurement delays; political volatility | Multi-jurisdiction diversification; partnership-led delivery |
| 7 | Regulatory tightening (AI Act, antitrust) | MED | EU AI Act enforcement; antitrust on hyperscaler concentration | Compliance investments; multi-region cloud strategies |
