The largest coordinated capital deployment in the history of technology is underway. The winners will not be whoever builds the most compute. They will be whoever solves power, cooling, and the silicon disruption that is quietly eroding the market's most dominant moat.
The numbers have moved beyond superlatives. Understanding what is being built, at what cost, and who is funding it is the essential starting point for anyone making decisions in or around this sector.
In the first quarter of 2026, the four largest technology companies reported quarterly capital expenditures that, taken together, exceeded $130 billion in a single three-month period. Amazon alone guided to $200 billion in annual capex for 2026, more than doubling its 2025 spend. Google committed $175–185 billion. Meta raised its full-year guidance to $125–145 billion, citing higher component pricing and additional data centre costs. Microsoft set its calendar-year 2026 capex at $190 billion, well above the $152 billion consensus analyst estimate. The combined 2026 commitment from the four largest hyperscalers approaches $700 billion, nearly double the record $410 billion spent in 2025, and roughly 77% higher year-on-year.
These are not projections. They are disclosed commitments, backed by earnings calls, CFO guidance, and in several cases pre-signed construction and equipment contracts that cannot be unwound without significant financial penalty. The infrastructure supercycle that analysts have discussed since 2023 has arrived, and it is arriving at a velocity that is straining every physical constraint in the system simultaneously: GPU supply, construction timelines, skilled labour, and most consequentially, electrical power.
"The sector is experiencing an infrastructure investment supercycle requiring up to $3 trillion by 2030. Roughly 100 GW of new capacity is anticipated to come online between 2026 and 2030, equating to $1.2 trillion in real estate asset value creation alone."
JLL Research, 2026 Data Centre Market OutlookThe global data centre market was valued at $383.82 billion in 2025 and is projected to reach $902 billion by 2033, growing at an 11.3% CAGR (Grand View Research, March 2026). The AI-specific data centre segment, which carries materially higher capital intensity per megawatt, is growing faster: valued at $147.28 billion in 2025 and projected to reach $810.61 billion by 2033 at a 23.9% CAGR. These two figures tell the structural story clearly. General data centre infrastructure is compounding at a solid rate. AI data centre infrastructure is compounding at more than double that rate, driven by workloads that require fundamentally different hardware density, cooling architecture, and power delivery than anything the industry has previously built at scale.
The four primary layers of this buildout have distinct economics and distinct competitive dynamics:
Shell, core, and mechanical build. AI tech fit-out adds up to $25M per MW on top. Average global construction cost grew 6% in 2026. Speed to power now the primary site selection criterion above location or cost.
NVIDIA holds approximately 80% of the AI accelerator market by revenue with $194B in FY2026 data centre revenue. Custom ASIC shipments growing at 44.6% vs 16.1% for merchant GPUs in 2026, the first year custom silicon has outpaced GPUs in growth rate.
Electricity demand from data centres soared 17% in 2025. AI-focused consumption poised to triple by 2030. Small modular reactor pipeline grew from 25 GW to 45 GW in the 12 months to April 2026 as operators pursue behind-the-meter generation.
NVIDIA's Vera Rubin is its first 100% liquid-cooled system. Air cooling led with 38% market share in 2025, but liquid cooling is the fastest-growing segment as rack densities driven by AI accelerators make traditional air management physically inadequate.
For the first time in the history of the technology industry, the growth of a software-driven sector is being limited not by capital, not by talent, and not by demand, but by access to electrical power. This is the most consequential structural fact in AI infrastructure today.
The technology industry built its business model on the assumption that physical infrastructure was largely irrelevant. Software scaled at near-zero marginal cost, data centres sat in industrial parks drawing power from a grid that someone else built and maintained, and capital was the binding constraint on growth. That model is being unmade in real time.
Data centre electricity consumption soared 17% in 2025, far outpacing global electricity demand growth of 3% (IEA, April 2026). By one estimate, global data centre energy consumption could approach 1,050 TWh by 2026, which would make the sector the fifth-largest electricity consumer in the world if measured as a country, between Japan and Russia. By 2030, the IEA projects data centre electricity consumption will double from its current level, with AI-focused consumption poised to triple. This is not a projection that assumes continued exponential growth in model complexity. It assumes the current trajectory of inference workload expansion and agent-based AI deployment, both of which are accelerating.
The consequences are already materializing in ways that are directly reshaping where and how data centres are built. PJM capacity market clearing prices for the 2026–2027 delivery year increased to $329.17 per MW, more than ten times the $28.92 per MW price in 2024–2025, with rapid data centre growth identified as a major contributing factor. Grid connection timelines in the traditional hyperscale clusters of Virginia, Dublin, Frankfurt, Amsterdam, and London have extended to multi-year waits in some cases. Texas passed Senate Bill 6 in 2025, redefining the interconnection process for large electrical loads exceeding 75 MW, a direct legislative response to data centre demand overwhelming the ERCOT grid.
The industry's response has been to stop waiting for the grid and start building its own generation. The tech sector accounted for approximately 40% of all corporate power purchase agreements for renewables signed in 2025. More significantly, the pipeline of conditional offtake agreements between data centre operators and small modular reactor projects grew from 25 GW to 45 GW in the twelve months to April 2026 alone. Microsoft has signed long-term power purchase agreements with nuclear operators including Three Mile Island specifically for AI data centre load. AWS invested $650 million in a data centre co-located with a nuclear facility. The "Bring Your Own Generation" model, once a niche workaround, is becoming standard infrastructure strategy for any facility above 100 MW.
For founders and operators, the power crisis is simultaneously a constraint and a market signal. It is a constraint because any business dependent on available cloud compute capacity faces costs and latency that are rising as demand outstrips supply. It is a market signal because every element of the power delivery chain, from transmission equipment to grid software to cooling systems to SMR development, is undersupplied relative to the demand being committed to today.
NVIDIA is the most consequential semiconductor company in a generation. It is also facing a structural challenge from inside its own customer base that no previous dominant chipmaker has encountered at this scale.
NVIDIA's position in AI infrastructure is, by any objective measure, extraordinary. The company holds approximately 80% of the AI accelerator market by revenue. Its data centre segment generated $194 billion in FY2026, with data centre now representing over 80% of total company revenue. CUDA, its proprietary programming framework, has over 5 million developers building on it, creating a software ecosystem lock-in that is structurally independent of any single hardware generation. Jensen Huang cited $1 trillion in committed orders through 2027 at GTC 2026 and stated simply that supply would remain constrained. NVIDIA's gross margins of 85–88% dwarf competitors and fund an R&D pipeline and TSMC capacity reservation that makes the competitive response from AMD and Intel structurally difficult.
Vera Rubin, NVIDIA's next architecture after Blackwell, is expected to ship in the second half of 2026. Each Rubin R100 GPU draws approximately 2,300 watts of thermal design power, nearly double its predecessor, forcing a total redesign of data centre electrical systems toward 800-volt power delivery and mandatory liquid cooling. The rack price is estimated at $3.5–4 million, a 25% increase over Grace Blackwell. This is not a product that is losing the performance race. It is a product that is raising the barrier to entry for every subsequent generation of infrastructure deployment.
The disruption, however, is structural rather than competitive in the traditional sense. It is not AMD or Intel that threatens NVIDIA's dominance. It is NVIDIA's own customers. TrendForce projects custom ASIC shipments from cloud providers will grow 44.6% in 2026, against just 16.1% growth for merchant GPU shipments. That gap, nearly three to one in growth rate, marks the first year in the AI era that custom silicon has meaningfully outpaced general-purpose GPUs, and it signals a deliberate strategic decision by the world's largest AI spenders.
| Player | Position | Market Standing | Strategic Edge | Key Risk |
|---|---|---|---|---|
| NVIDIA | Dominant | ~80% AI accelerator market share; $194B FY2026 data centre revenue; Vera Rubin shipping H2 2026 | CUDA software ecosystem; 5M+ developers; TSMC capacity secured; annual architecture cadence | Custom silicon growth at hyperscalers; export restrictions limiting China revenue; valuation concentration |
| Broadcom | Challenger | Leading ASIC designer for Google TPUs, Meta MTIA, and ByteDance; custom AI chip revenue growing rapidly | Deep hyperscaler relationships; networking silicon (Tomahawk); custom silicon compounds with inference growth | Concentrated in a small number of large hyperscaler relationships |
| AMD | Challenger | ~5–7% AI GPU share; MI350X matching B200 on FP8 TFLOPS; OpenAI 6GW and Meta multi-year contracts | First-to-2nm advantage with MI400; ROCm improving; pricing flexibility vs NVIDIA | Software ecosystem gap vs CUDA remains material; utilisation requires significant kernel engineering |
| Google (TPU) | Custom Silicon | Tensor Processing Units deployed at scale internally; Cloud TPU available externally | Purpose-built for Google's workloads; cost efficiency at scale; full vertical integration | Limited third-party ecosystem; not a merchant silicon business |
| Amazon (Trainium) | Custom Silicon | Trainium 2 deployed in AWS data centres; Trainium chips powering OpenAI 6GW agreement | Cost reduction vs NVIDIA for training workloads; integrated AWS ecosystem | Software tooling immature; customer adoption requires migration from CUDA workflows |
| Microsoft (Maia) | Custom Silicon | Maia AI accelerator chips deployed in Azure; Cobalt CPU for general cloud workloads | Deeply integrated with Azure and OpenAI workloads; cost optimisation for inference at scale | Early-stage deployment; customer-facing availability limited |
The strategic logic driving custom silicon adoption is straightforward: at the scale hyperscalers are now operating, even a 10–15% improvement in performance per watt or cost per inference translates to billions of dollars in annual savings. Google's TPUs, Amazon's Trainium, and Microsoft's Maia are not built to sell to third parties. They are built to reduce dependency on a single supplier and capture the margin that NVIDIA currently earns on every chip they purchase. NVIDIA's moat is deep and durable at the training layer. The inference layer, where workloads grow as AI adoption matures, is where the competitive erosion will be most visible over the next three years.
The AI infrastructure buildout has been underway since 2022. What makes 2026 the inflection point is the simultaneous arrival of six forces that compress timelines, raise stakes, and create asymmetric opportunity for those positioned ahead of the wave.
AI represented approximately a quarter of all data centre workloads in 2025, with training driving most of the demand. JLL projects a significant shift in 2027, when inference workloads are expected to overtake training as the dominant AI requirement. By 2030, AI could represent half of all data centre workloads. Inference is a fundamentally different business from training: it requires lower-cost, higher-efficiency compute deployed at the edge and at massive scale. The firms building inference-optimised infrastructure, from chip design to facility architecture, are positioning for the larger and more durable market.
Hyperscaler capex has reached 45–57% of revenue, historically unthinkable levels for businesses of this scale. The five largest hyperscalers have committed to adding approximately $2 trillion of AI-related assets to their balance sheets by 2030. Given AI hardware depreciates at approximately 20% annually, this implies $400 billion in annual depreciation, exceeding their combined 2025 profits. The market is now bifurcating between operators with the balance sheet to sustain this cycle and those that face stranded asset risk if AI adoption timelines slip. Identifying which side of that bifurcation each player sits on is the core analytical question for infrastructure investors.
The traditional data centre clusters, Northern Virginia, Dublin, Amsterdam, Frankfurt, and London, are capacity-constrained by grid limitations that cannot be resolved in the near term. New capacity is shifting to regions with available grid headroom: Texas, the Pacific Northwest, the Nordic countries, the Middle East, and Southeast Asia. This geographic repricing is creating new real estate markets, new utility relationships, and new regulatory frameworks simultaneously. The operators who secured grid capacity or behind-the-meter generation in secondary markets two years ago are now holding the most valuable positions in global infrastructure.
The 45 GW pipeline of conditional offtake agreements between data centre operators and SMR projects is not simply speculative enthusiasm. It represents the financial anchor that small modular reactor developers have lacked for a decade: a creditworthy, long-duration off-taker with contractual commitment to purchase power at a price that makes the project economics work. Several SMR developers are now within 18–24 months of breaking ground on first commercial facilities, funded in part by hyperscaler agreements. The firms that positioned in the SMR supply chain, from nuclear engineering to component manufacturing, will see demand that was theoretical become contractual.
The IEA notes that while power consumption per AI task is declining rapidly due to efficiency improvements, more people are using AI and energy-intensive uses such as AI agents are rising. Jensen Huang described the current moment as "the agentic AI inflection point" in his Q4 2025 earnings commentary. Agentic workflows, where AI systems take sequences of actions autonomously rather than responding to single queries, are dramatically more compute-intensive than single-turn inference. As agent deployment scales, the energy demand curve will accelerate rather than moderate, extending the capital deployment supercycle further than current consensus projections suggest.
Saudi Arabia's HUMAIN initiative, awarded a $2.7 billion contract for a 480 MW data centre in Riyadh in January 2026, is one of dozens of sovereign AI programmes globally that are building national compute infrastructure independent of the hyperscaler ecosystem. France, the UAE, India, and Japan have all committed sovereign AI infrastructure budgets in the hundreds of millions to low billions. These sovereign buyers are not part of the hyperscaler capex figures. They represent an incremental demand layer for the same constrained supply of AI-ready power, compute, and construction capacity, further tightening every physical bottleneck in the system.
The AI infrastructure buildout is large enough that broad exposure captures the theme but diffuse enough that position selection determines returns. These three themes represent the most structurally defensible entry points available in 2026.
The risk of undifferentiated enthusiasm about AI infrastructure is that it leads to expensive positions in already-crowded trades. NVIDIA at 30x forward earnings and the major hyperscalers trading at historically elevated capital intensity are not obvious value opportunities. The more interesting question is where the market is mispricing a structural position within the buildout, and what specific characteristic makes that position difficult to displace once established. Each of the three themes below rests on a durable structural advantage rather than simply a growth rate.
Every other bottleneck in AI infrastructure has a capital solution: you can buy more GPUs, hire more engineers, expand construction timelines. Power is different. Grid connections cannot be expedited by spending more money. Transmission infrastructure takes years to permit and build. The operators, equipment providers, and energy developers who hold contracted power capacity in high-demand markets today hold a position that is structurally irreplaceable in any relevant planning horizon. PJM capacity prices rising 10x in a single year is not noise. It is the market repricing the scarcity of a critical input that cannot be manufactured on demand.
Training a frontier AI model requires extraordinary compute concentrated over a short period in a small number of facilities. Inference, running that model for millions or billions of users continuously, requires compute deployed at massive scale, with lower latency requirements, higher efficiency demands, and far broader geographic distribution. JLL projects inference will overtake training as the dominant AI workload by 2027. The infrastructure optimised for inference, lower-cost custom silicon, edge deployment, liquid-cooled high-density racks, and the networking that connects them, is structurally different from what the current buildout primarily produces. The firms positioning for inference infrastructure today are building for the next cycle before the current one peaks.
The AI infrastructure buildout requires an enormous volume of physical work that software cannot replace: site acquisition, facility construction, mechanical and electrical systems integration, liquid cooling installation, and network fibre deployment. Average global construction cost per MW reached $11.3 million in 2026, with the AI tech fit-out adding up to $25 million per MW on top. The 100 GW of new capacity planned for 2026–2030 equates to $1.2 trillion in real estate asset value creation and hundreds of billions in construction and integration services. The firms providing these services are benefiting from demand that is contracted years in advance, not exposed to model performance uncertainty, and not subject to the silicon obsolescence cycle that makes chip investments binary.
The scale of capital commitment in AI infrastructure creates a correspondingly large set of risks. Three are structural and worth building into any investment or strategic framework. Three are cyclical and require monitoring rather than immediate action.
The consensus view of AI infrastructure risk centres on two questions: whether AI revenues will justify the infrastructure investment, and whether NVIDIA's dominance is sustainable. Both are legitimate. Neither is the most important risk for operators and investors who are building positions in the physical infrastructure layer rather than the chip or application layer. The most material risks for the physical infrastructure buildout are different in character and less frequently discussed.
AI hardware depreciates at approximately 20% annually, and each new GPU generation requires meaningfully different infrastructure: higher power density, different cooling architecture, new electrical systems. A facility optimised for Blackwell GPUs at 120 kW per rack faces a redesign challenge when Vera Rubin arrives at higher rack densities, and another when the generation after that arrives. Data centre real estate built for today's workloads may be partially stranded by workloads two generations ahead. The firms building with architectural flexibility and upgrade pathways are more defensible than those optimising for current hardware specifications.
The five largest hyperscalers have committed to adding approximately $2 trillion in AI assets by 2030, at a combined depreciation run rate that will exceed their collective 2025 profits by the end of the decade. Amazon is projected to turn free cash flow negative in 2026 as capex commitments bite. The infrastructure is being built for an AI revenue curve that has not yet arrived at the scale being assumed. If enterprise AI adoption is slower or more concentrated than current projections, the free cash flow impact on the hyperscalers is significant, and the downstream impact on their supplier base is direct.
US export restrictions on advanced AI chips to China have already cost NVIDIA an estimated $15–20 billion in potential revenue and have accelerated Chinese domestic chip development at Huawei and emerging domestic foundries. As the geopolitical competition around AI infrastructure intensifies, further restrictions on chip exports, equipment sales, and data centre construction in strategic markets could create a bifurcated global infrastructure market where supply chains, software ecosystems, and regulatory frameworks are incompatible. Any infrastructure strategy with cross-border components needs to account for this fragmentation risk explicitly.
Approximately 40% of announced AI data centre projects face construction delays due to power infrastructure bottlenecks, not chip supply. Adding to this, specialised electrical and mechanical contractors capable of building gigawatt-scale AI facilities are in structurally short supply. Data centre construction timelines have extended from 18–24 months to 36–48 months in some markets due to labour and materials constraints. This creates a timing risk for operators whose AI revenue commitments depend on facility completion schedules that the construction industry cannot currently meet reliably.
Research shows 70% of Americans now oppose data centres near their homes, making them less popular than nuclear power plants as a neighbourhood amenity. Local government opposition to large-scale data centre development has delayed or blocked projects in Ireland, the Netherlands, and several US states. As data centres grow to gigawatt-scale campuses with significant visual, noise, and grid impact, the permitting and community relations dimension of infrastructure development is becoming a material constraint that capital alone cannot resolve.
The IEA notes that power consumption per AI task is declining rapidly, with efficiency improving at a rate unprecedented in energy history. If algorithmic improvements, model compression, and inference optimisation compound faster than anticipated, the energy demand curve could moderate meaningfully from current projections. DeepSeek's January 2025 demonstration that competitive model performance was achievable at a fraction of assumed compute cost was a genuine signal that efficiency gains are non-linear and unpredictable. This risk is asymmetric: it would reduce energy demand but simultaneously expand AI adoption by lowering the cost of deployment, potentially offsetting the efficiency gains through volume.
Each position below reflects a different structural characteristic of the AI infrastructure buildout. The analytical question in each case is what the market is mispricing, and what specific development closes that gap.
The most straightforward AI infrastructure trade, long NVIDIA and long hyperscalers, is also the most crowded and the most exposed to the specific risks described in this report. The more interesting analytical positions sit in the physical and energy infrastructure layer, where the market has been slower to price structural scarcity, and in the inference transition, where the shift from training to inference workloads will reshape demand in ways that are not yet reflected in consensus infrastructure assumptions.
The AI infrastructure supercycle is real, it is funded, and it is accelerating. The $700 billion in hyperscaler capex committed for 2026 is not speculative enthusiasm; it is disclosed guidance backed by signed contracts and construction timelines that cannot be unwound. The 100 GW of new data centre capacity planned for 2026–2030 will happen because the underlying demand for compute is not a projection. It is a present reality reflected in every quarterly earnings call from every major technology company simultaneously.
What is less certain is which layer of the infrastructure stack captures the most durable value. The chip layer is where the current consensus is concentrated, and it is also the layer most exposed to the custom silicon disruption, the architecture obsolescence cycle, and the export control fragmentation risk. The physical and energy infrastructure layer, power, cooling, construction, and interconnection, is where scarcity is most structural and where the market has been slowest to price the constraint correctly.
For founders building in or adjacent to AI infrastructure: the opportunity is not in competing with hyperscalers on compute. It is in solving the physical constraints they cannot solve internally. Energy procurement, facility operations, cooling engineering, and grid interconnection are not problems that can be solved by writing better software. They require specialised expertise, physical assets, and regulatory relationships that are being repriced in real time.
For investors: the inference transition, the custom silicon shift, and the nuclear power renaissance are the three developments that are most mispriced relative to their structural significance in the current market. Positions built around these themes before the consensus fully forms will reflect the asymmetric returns that the infrastructure buildout ultimately offers to those who understand it with precision rather than enthusiasm.
Virtus Orbis advises founders, growth-stage companies, and investors on capital strategy, market entry, and cross-border expansion. We bring institutional-grade analysis to the decisions made at the intersection of technology, infrastructure, and capital markets.
Disclaimer: This research note is produced by Virtus Orbis LLC for informational purposes only. It does not constitute financial advice, an investment recommendation, or an offer to buy or sell any security or financial instrument. References to specific companies, securities, or investment themes are for illustrative and analytical purposes only. All market data and projections are sourced from publicly available third-party research (JLL Research, Grand View Research, Precedence Research, MarketsandMarkets, IEA, BloombergNEF, Financial Times, Futurum Group, TrendForce, Silicon Analysts, Brookings Institution, Morgan Stanley, CNBC, IEEE ComSoc). Forward-looking projections involve inherent uncertainty and actual outcomes may differ materially. Virtus Orbis does not hold investment positions in any securities mentioned herein. Readers should conduct independent analysis and seek professional advice before making any investment or strategic decision. © 2026 Virtus Orbis LLC. All rights reserved.