Background: how we arrived at the largest private investment cycle in history
To understand what $690 billion in annual capital expenditure actually means, you need a point of comparison. The entire US energy sector — oil, gas, utilities, renewables — invests approximately $170 billion in capital annually. The entire US interstate highway system, built over four decades, cost roughly $500 billion in today’s dollars. The Apollo programme that put humans on the moon cost approximately $280 billion in today’s money. AI infrastructure spending by five US technology companies in a single year is now larger than all of these.
The five largest US cloud and AI companies — Microsoft, Amazon, Alphabet, Meta, and Oracle — are guiding toward $635–690 billion in combined 2026 capital expenditure, a 67–74% increase over 2025. Approximately 75% of hyperscaler capex is now allocated to AI infrastructure — GPUs, high-bandwidth memory, networking, data centres, and power systems. Goldman Sachs projects total hyperscaler capex from 2025 to 2027 will reach $1.15 trillion — more than double the $477 billion spent from 2022 to 2024. Intellectia.AI
Hyperscalers are the companies that operate the world’s largest cloud computing infrastructure — primarily Amazon Web Services, Microsoft Azure, Google Cloud, and Meta. A GPU (Graphics Processing Unit) is the specific type of semiconductor chip that AI model training and inference workloads run on most efficiently; NVIDIA’s H100 and B200 series GPUs are the primary target of this spending wave.
The bull case: infrastructure supercycle, not speculation
The case for this level of spending being justified rests on historical precedent and forward demand projections. KKR’s infrastructure research team drew an explicit comparison to the electrification cycle of the early 20th century — a period of massive upfront capital deployment that appeared overbuilt relative to immediate demand but that underpinned decades of productivity growth as the economy adapted to a new energy infrastructure. McKinsey estimates companies will invest almost $7 trillion in global data centre infrastructure capital expenditures by 2030 — roughly the size of the combined GDP of Japan and Germany. In the first half of 2025, AI-related capex contributed materially to US GDP growth — more than consumer spending. Substack
Goldman Sachs Research notes that equity gains have been concentrated in AI infrastructure companies — semiconductors, hyperscalers, technology hardware providers, and power companies — with the average stock in Goldman Sachs’s basket of infrastructure companies returning 44% year-to-date, compared with a 9% increase in consensus two-year forward earnings-per-share estimates for the group. The inference from this gap is that markets are pricing in demand materialisation that has not yet appeared in earnings — which can be either justified optimism or overvaluation, depending on how that demand develops. Futurum Group
The bear case: free cash flow is already going negative, and $662 billion in commitments is sitting off balance sheet
Amazon’s capital expenditure has now reached $151 billion in the last twelve months, exceeding its operating cash flow entirely and pushing free cash flow into negative territory. Hyperscalers now spend 45–57% of revenue on capex — ratios previously unthinkable for technology companies and resembling industrial or utility companies more than traditional tech. Morgan Stanley and JP Morgan project the technology sector may need to issue $1.5 trillion in new debt over the next few years to finance AI infrastructure construction. MUFG Americas
The most striking bear-case data point is not in the income statement but off it. Moody’s reported in early 2026 that hyperscalers have approximately $662 billion in data centre lease commitments that have been signed but not yet commenced. These obligations sit off balance sheet under GAAP’s lease commencement standard, meaning they do not appear in the capital expenditure figures that analysts and investors typically scrutinise. This off-balance-sheet liability is larger than the combined on-balance-sheet debt of the same companies.
The revenue comparison is sobering. OpenAI’s $20 billion annual recurring revenue (ARR) — impressive for a company barely three years old as a consumer product — represents roughly 3% of the projected 2026 hyperscaler capex total. Anthropic’s $9 billion run rate occupies a similar position. The entire cohort of pure-play AI vendors — including Cohere, Mistral, Perplexity, and others — likely accounts for less than $35 billion in projected combined 2026 revenue against a $690 billion infrastructure investment. The revenue-to-capital-deployment ratio at this stage of the cycle is more extreme than at any comparable point in the dotcom era. Introl
The structural question: inference versus training
A specific technical shift is further complicating the bull case for infrastructure suppliers. The transition from AI training workloads to AI inference workloads could reshape the competitive landscape — and the two are not equivalent. AI training requires massive GPU clusters running for months to build models, which justifies enormous data centre buildout. AI inference — actually running the trained model to answer a question or generate content — is less compute-intensive per query and is increasingly being optimised to run on smaller, cheaper hardware. If the dominant workload shifts from training to inference faster than the market expects, the demand justification for $690 billion in annual infrastructure spending weakens materially. Goldman Sachs
NVIDIA’s own data, while projecting continued GPU demand, has shown revenue guidance misses in recent quarters as enterprise AI adoption proves slower than consumer AI adoption — a leading indicator that the B2B commercialisation of AI is on a longer timeline than the B2C usage curves suggest.
The power constraint nobody is adequately discussing
AI data centre power demand is projected to reach 156 GW by 2030, requiring roughly $5.2 trillion in cumulative data centre investment through the end of the decade. That power demand figure is the binding constraint that limits how quickly even well-funded hyperscalers can actually build: grid connection queues in the US, UK, and Europe now extend five to ten years in many markets, electricity infrastructure permitting is slow, and nuclear power — which several hyperscalers have targeted as a reliable, carbon-free energy source for AI data centres — carries decade-long construction timelines even in favourable regulatory environments. Intellectia.AI
This creates an ironic structural bottleneck: the hyperscalers have the capital, the chips are (just about) available from NVIDIA and its competitors, but the physical electricity to run these data centres at the projected scale is constrained by real-world infrastructure that cannot be accelerated by software or capital expenditure alone.
The honest verdict — not a bubble in the 2000 sense, but not without serious risk either
The comparison most frequently made is to the 1990s dotcom era, when vast fibre-optic infrastructure was built that exceeded immediate demand by an order of magnitude — causing a crash — but that ultimately underpinned the internet economy of the 2000s and 2010s. The structural parallel holds: genuine transformative technology, real long-run demand, but short-run supply-demand mismatch and speculative excess layered on top. Goldman Sachs noted directly: “The timing of an eventual slowdown in capex growth poses a risk to these companies’ valuations.” Futurum Group
The difference from 2000 is that the companies doing this spending — Microsoft, Amazon, Alphabet, Meta — are generating hundreds of billions in operating cash flow annually and have balance sheets that can absorb the risk if the AI revenue materialisation takes longer than expected. The risk is concentrated not in these giants’ solvency but in the valuations of the smaller ecosystem — semiconductor equipment companies, data centre construction firms, power infrastructure providers — whose stocks have priced in a decade of AI-driven demand with far less financial cushion if the cycle slows.