Global AI Investment: Hardware Surge And Application Shortfall

In 2026, global investment in artificial intelligence remains at elevated levels, with the four largest technology companies projected to spend more than $600 billion in annual capital expenditures. A pronounced structural imbalance has emerged: capital is overwhelmingly concentrated in hardware segments such as power, chips and data centers, while commercial returns at the application layer have yet to materialize. Concurrently, valuation leverage is rising, circular trading patterns are appearing, and early signs of labor substitution risk are accumulating. This analysis adopts the five‑layer framework popularized by Nvidia—energy, chips, infrastructure, models and applications—to examine investment flows, identify systemic risks and assess China’s differentiated advantages and policy options.

Electricity supply has become a binding constraint that is driving technology firms toward heavy‑asset strategies. Forecasts from Lawrence Berkeley National Laboratory indicate that U.S. data center electricity consumption could account for 6.7%–12% of national demand by 2028, yet grid aging, a shortfall in permitted capacity of roughly 50% relative to demand and community opposition have delayed or shelved projects worth hundreds of billions of dollars. In response, major cloud and AI operators have progressed from signing power purchase agreements to building their own generation assets and exploring nuclear options, signaling a structural shift from light‑asset cloud providers to energy operators. Nuclear power purchase agreements for AI data centers totaled about 7,400 MW by early 2026, typically with 20‑year terms, while policy initiatives such as emergency reliability auctions and industry commitments to fund generation and transmission upgrades have accelerated self‑provisioning. Longer‑term bets on small modular reactors have entered the investment phase, but commercialization faces approval, cost and concentrated demand risks and is not expected to scale before the 2030s. Some firms have also begun exploring space‑based data center concepts to exploit space solar power and vacuum cooling, though current demonstrations remain at single‑satellite, single‑GPU levels and face high launch, maintenance and latency challenges.

The demand profile for AI chips is shifting from training to inference. Training represents a largely one‑off capital expenditure, whereas inference is an ongoing operational cost that scales with usage; industry forecasts indicate inference’s share of AI compute demand rose from roughly one‑third in 2023 to two‑thirds in 2026 and could exceed 80% over time. This transition favors specialized inference accelerators (ASICs) that can deliver substantially lower per‑token costs than general‑purpose GPUs. Major cloud providers and hyperscalers have accelerated deployment of in‑house inference chips, and market projections anticipate the AI accelerator market reaching approximately $380 billion by 2028, with ASICs capturing a significant share. In China, IDC projects that training workloads in GenAI IaaS will decline from 76% in 2024 to 23% by 2029, while inference demand is expected to grow at a compound annual rate of 103%, far outpacing training. Domestic chip shipments exceeded 2.7 million units in 2024, with local brands growing rapidly and increasing their share of the domestic AI server market from 17% to about 42%.

Data center investment is expanding at an unprecedented pace, serving as the primary vehicle for hardware deployment. Gartner estimates data center systems spending reached about $489.5 billion in 2025, up 46.8% year‑on‑year, while Microsoft, Google, Amazon and Meta collectively invested $357.5 billion in 2025, a 65% increase. Wall Street forecasts project capital expenditures to rise to roughly $608.2 billion in 2026, and multiple projects with budgets in the tens to hundreds of billions are underway. Despite this acceleration, short‑term compute capacity remains tight: peak data center utilization is projected to reach 94% in 2026 and remain near 90% through 2028, with supply additions expected to ease utilization only gradually toward 2030. Power constraints are reshaping site selection logic, with high‑power training workloads migrating to power‑abundant remote locations while latency‑sensitive inference workloads must remain close to end users in core urban centers, producing a bifurcated “training in the hinterland, inference in the city” pattern.

Large model companies have raised capital at record valuations, yet their commercial trajectories reveal growing stress. Anthropic completed a $30 billion financing at a $380 billion valuation in February 2026, and OpenAI closed $122 billion at an $852 billion valuation in March 2026, with both firms planning public listings. Other model developers have also attracted substantial funding, including xAI’s $20 billion Series E and Mistral AI’s $2 billion Series C. Despite rapid revenue growth for some players, losses have expanded faster than revenues, driven by escalating training and inference costs. Estimates indicate that to justify current infrastructure spending, application‑side revenues must scale far beyond present levels; for example, Nvidia’s $193.7 billion data center revenue in 2025 implies an application revenue requirement on the order of $775 billion, while actual application revenues are estimated at roughly $150–200 billion, leaving a multi‑hundred‑billion‑dollar gap. Internal forecasts and reporting suggest that some model firms may not reach profitability until the late 2020s, contingent on substantial revenue expansion and cost control.

Financial risks are amplified by elevated equity valuations, rising leverage across the AI value chain and circular transactions that obscure genuine demand. Key valuation metrics for major U.S. indices have reached levels that exceed decade‑long thresholds, and risk premia have compressed to near or below zero, reducing safety margins. Circular capital flows among chipmakers, cloud providers and model firms have increased reported revenues but also concentrated systemic risk: investment and revenue linkages among platform companies, accelerator vendors and new cloud entrants create pathways for rapid contagion if one node underperforms. Debt issuance across hardware and cloud vendors surged in 2025 and continued into 2026, with quarterly issuance in early 2026 approaching prior annual totals, heightening the potential for leverage‑driven stress to propagate through the ecosystem.

Labor market impacts are beginning to surface as AI adoption accelerates. Since 2025, major technology employers have announced large workforce reductions, and AI‑related layoffs rose sharply in early 2026, with AI cited as the primary cause for a significant share of monthly job cuts. Capital markets have at times rewarded firms that reduce headcount, reinforcing incentives to pursue labor substitution as a near‑term efficiency lever. Analysts warn of a potential “intelligent substitution spiral” in which productivity gains from headcount reductions feed further capital‑driven layoffs, erode consumer purchasing power and produce a divergence between GDP growth and household income, with attendant risks for credit performance and systemic stability. While this dynamic remains nascent, policymakers face a window to design mitigating measures before localized disruptions evolve into broader structural imbalances.

China’s AI investment profile exhibits several differentiated advantages. Low electricity costs are a core competitive edge: power accounts for an estimated 60%–70% of data center operating costs, and abundant clean energy in western provinces has driven green power capacity to substantial levels. By the end of 2025, solar and wind installations in regions such as Inner Mongolia, Ningxia and Gansu reached significant gigawatt‑scale figures, enabling western green power prices to fall to levels that yield per‑million‑token compute costs materially below U.S. benchmarks. China’s 2026 government work report elevated compute‑power coordination to a national strategic priority, reinforcing institutional support for cross‑regional dispatch and green power direct‑connect models. Domestic chipmakers have accelerated commercialization and financing, with companies such as Moore Threads, Biren Technology, Cambricon and Suiyuan Technology advancing listings and capacity expansion while several firms reported substantial revenue growth and improved profitability in 2025. Structural advantages also include scale in mature process nodes, improving global share of mature process capacity, stable delivery timelines relative to overseas suppliers and deepening hardware‑software co‑optimization with domestic large models, which has enabled rapid “Day‑0” compatibility for new model releases.

China’s data center expansion is proceeding in parallel with global trends but along a distinct path. The domestic market reached an estimated RMB 318 billion in 2025 and is projected to grow to RMB 362.1 billion in 2026, supported by major public investments and substantial private capital. National initiatives that enable cross‑regional compute allocation, combined with liquid‑cooling adoption and advanced architectures, are improving energy efficiency and reinforcing a long‑term cost advantage. The success of low‑cost model strategies has been demonstrated by firms such as DeepSeek, whose V3 model reportedly achieved top‑tier performance at a fraction of the training cost of comparable Western models, prompting reassessments of high‑input, high‑performance assumptions among leading hardware vendors. China’s application ecosystem is also large and rapidly commercializing: domestic AI‑native applications reported monthly active users in the hundreds of millions, and several model providers have shown strong revenue acceleration following price adjustments and product maturation.

Policy recommendations to accelerate China’s AI development emphasize consolidating these comparative advantages while managing systemic risks. Strengthening coordination between compute and power supply through green‑power direct connections, long‑term power purchase mechanisms and market‑based joint ventures between power and compute firms can preserve low‑cost operation. Reinforcing the low‑cost model route requires targeted support for algorithmic efficiency, open‑source ecosystems and talent development to sustain competitive advantages in model efficiency. Promoting high‑quality application adoption calls for demand‑side pilots in priority sectors, expanded skills training to broaden the user base and financing innovations such as patient capital and compute‑as‑a‑service to lower adoption barriers. To prevent adverse labor‑market dynamics, policymakers should consider incentives for firms that raise productivity without reducing headcount, explore enhanced social protection pilots for affected workers and subsidize inclusive access to AI services to broaden benefits.

In sum, global AI investment is at a critical inflection point characterized by rapid hardware deployment and lagging application monetization. China’s structural advantages in power cost, model efficiency, domestic chip supply and application scale create a valuable window of opportunity, but that window is finite. Balancing the risks of global valuation excess and leverage accumulation with the opportunity to shape an intelligent economy will be the central policy challenge in the next phase of AI development.