Khan Capital | August 2025
Key Takeaways
- Combined AI capital expenditure from the top five hyperscalers is on track to exceed $400 billion in 2025 and approach $530 billion in 2026, the largest corporate infrastructure buildout since the 1990s telecommunications boom.
- The gap between AI capex and AI-specific revenue remains the central risk: cloud revenue growth is accelerating but not at the pace needed to justify the current rate of spending on a near-term basis.
- The DeepSeek shock of January 2025 demonstrated that efficiency improvements expand the addressable market rather than reducing spending (Jevons’ Paradox), and subsequent hyperscaler earnings confirmed accelerating demand.
- The beneficiary map extends well beyond Nvidia: data storage, power and utilities, networking infrastructure, and industrial companies are capturing downstream spending with less crowded positioning and more attractive valuations.
- The difference between the AI capex cycle and the 1990s telecom bubble is not the quality of the underlying technology (both were genuinely transformational) but the timing of revenue materialisation, the same risk that investors must monitor obsessively.
The numbers are staggering. Combined capital expenditure on AI infrastructure from the five largest hyperscalers (Alphabet, Amazon, Meta, Microsoft, and Oracle) is on track to exceed $400 billion in 2025 and approach $530 billion in 2026. Meta alone is spending more on AI infrastructure in a single year than the entire US airline industry spent on aircraft in the previous decade. Microsoft’s capital expenditure has more than doubled in two years. Nvidia’s data centre revenue has grown from $15 billion in fiscal year 2023 to a projected run rate exceeding $200 billion. The AI capital expenditure boom is the largest corporate infrastructure buildout since the telecommunications bubble of the late 1990s, and the critical question for investors is whether it will produce the same outcome.
The Scale: What $500 Billion Buys
Understanding the magnitude of the AI capex cycle requires placing it in context. Global data centre capital expenditure was approximately $150 billion in 2022. By 2025, just the top five hyperscalers are spending nearly three times that figure. The spend is concentrated in three categories: semiconductor procurement (dominated by Nvidia’s H100, H200, and Blackwell GPU families), data centre construction (including land, buildings, cooling systems, and electrical infrastructure), and power generation and procurement (since a single large AI training cluster can consume as much electricity as a small city).
The investment is not limited to the hyperscalers. The $500 billion Stargate joint venture announced in January 2025, involving OpenAI, SoftBank, Oracle, and others, committed to building a network of massive AI data centres across the United States. Sovereign wealth funds, private equity firms, and infrastructure investors are competing for data centre sites. Power utilities are signing long-term contracts with tech companies at premium rates. The entire energy, construction, and industrial supply chain is being reshaped by AI’s insatiable demand for compute and electricity.
The Revenue Question: Capex Is Running Ahead of Monetisation
The bull case for AI capex is that it represents investment in a transformational general-purpose technology whose productivity gains will eventually justify, and far exceed, the upfront cost. The bear case is simpler: the companies are spending $500 billion per year but generating a fraction of that in AI-specific revenue. The gap between investment and monetisation is the central tension in the AI trade.
Cloud revenue, which is the primary monetisation vehicle for AI infrastructure, is growing rapidly. Microsoft Azure’s AI services revenue growth has exceeded 60% year-over-year. Google Cloud’s AI revenue contribution is accelerating. Amazon Web Services’ AI-optimised instance revenue is the fastest-growing segment. But cloud revenue growth and AI capex growth are not the same metric, and the latter is dramatically outpacing the former.
The historical parallel is instructive. In the late 1990s, telecommunications companies invested approximately $1 trillion in fibre-optic cable, switches, and network infrastructure. The investment thesis was correct: internet traffic did grow exponentially, exactly as predicted. But the timing was wrong: the revenue to justify the investment arrived years later than the market had priced, and many of the companies that built the infrastructure went bankrupt before the demand materialised. Cisco, the “picks and shovels” play of the telecom boom (analogous to Nvidia’s role today), saw its stock rise 1,000% before losing 80% of its value when the bubble burst.
This is not a prediction that AI will follow the same path. The differences between 2025 and 1999 are significant: the hyperscalers funding today’s buildout have far stronger balance sheets, far more diversified revenue bases, and far greater visibility into customer demand than the telecoms of the late 1990s. But the structural pattern, massive capital investment in advance of proven revenue at scale, carries the same fundamental risk: the investment may prove correct in direction but wrong in timing.
The DeepSeek Disruption: Efficiency as a Threat and an Opportunity
The January 2025 DeepSeek shock added a new dimension to the capex debate. DeepSeek’s demonstration that competitive AI models could be trained at a fraction of the cost challenged the assumption that AI’s trajectory would inevitably require ever-increasing compute expenditure. If AI can be done cheaply, the moat created by massive capital expenditure narrows, and the returns on that investment compress.
In practice, the market’s response to DeepSeek revealed a more nuanced dynamic. Efficiency improvements do not reduce total AI spending; they expand the addressable market by making AI economically viable for a broader set of use cases and customers. This is a version of Jevons’ Paradox: as the cost of a resource (in this case, AI inference) falls, total consumption of that resource increases rather than decreases, driving net higher demand for infrastructure. The hyperscalers’ subsequent earnings reports confirmed that AI infrastructure demand had, if anything, accelerated after DeepSeek, as the prospect of cheaper inference encouraged more experimentation and adoption.
The Beneficiary Map: Beyond the Obvious Names
The AI capex cycle creates a layered set of investment opportunities that extends well beyond Nvidia and the hyperscalers.
Semiconductors: Nvidia remains the dominant beneficiary, but the supply chain extends to TSMC (fabrication), ASML (lithography equipment), SK Hynix and Samsung (high-bandwidth memory), and Broadcom (custom silicon and networking). The data storage companies (SanDisk, Western Digital) have emerged as surprise beneficiaries in 2025 as AI workloads generate unprecedented storage demand.
Power and utilities: AI data centres are energy-intensive, and the hyperscalers are securing power supply through long-term agreements with utilities, nuclear plant restarts, and investments in natural gas generation. Constellation Energy, Vistra, and other power generators with capacity near data centre clusters have seen significant stock price appreciation. The intersection of AI and energy is one of the most important emerging investment themes.
Industrial and construction: Data centre construction requires steel, concrete, cooling equipment, electrical components, and specialised engineering. Vertiv, Eaton, and Schneider Electric are among the industrial companies capturing the infrastructure spend downstream of the semiconductor and cloud layers.
Networking: As AI clusters scale, the networking infrastructure connecting GPUs, storage, and servers becomes a critical bottleneck. Arista Networks, Cisco, and specialist fibre-optic component makers are benefiting from the exponential growth in data centre networking bandwidth.
What the Market Is Misunderstanding
The durability of capex is more certain than the durability of margins. The hyperscalers have publicly committed to multi-year capex plans that are unlikely to be reversed: the competitive dynamics of AI make underinvestment more dangerous than overinvestment. But the returns on that investment depend on the pace of AI monetisation, the competitive intensity of the cloud market, and the risk that efficiency improvements compress the value of each dollar of capex faster than revenue can scale. Investors should be confident in the volume of spending but cautious about the returns on that spending.
The second derivative matters more than the first. Markets are priced for continued acceleration of AI capex. If the pace of spending increases from $400 billion to $530 billion (a 32% growth rate), that is bullish. But if spending grows from $530 billion to $600 billion (a 13% growth rate), the deceleration in the growth rate could trigger a multiple compression even as absolute spending continues to rise. The AI trade has been driven by positive earnings surprises; the moment surprises turn neutral, the stocks are vulnerable.
Not all capex is created equal. Training capex (building new models) is lumpy and potentially subject to efficiency gains that reduce future requirements. Inference capex (running models at scale for end-users) is more recurring and grows with adoption. As the AI industry matures from a training-dominated phase to an inference-dominated phase, the composition of capex will shift, with different beneficiaries in each category.
Implications for Investors
The AI infrastructure buildout is a multi-year cycle with significant remaining upside, but selectivity is paramount. Favour companies with demonstrated revenue acceleration from AI workloads, not just capex plans. Nvidia, TSMC, and Broadcom have the clearest revenue visibility. Data storage, power, and networking companies offer leveraged exposure with less crowded positioning.
Monitor the capex-to-revenue conversion rate obsessively. The moment hyperscaler earnings calls stop reporting accelerating AI revenue growth, the capex thesis will be challenged. Watch for any language suggesting capex moderation, efficiency-driven spending reductions, or slowing cloud demand growth.
The power and energy infrastructure trade is underfollowed and underowned. The intersection of AI and energy is creating a multi-decade investment theme that extends well beyond the semiconductor cycle. Utilities, nuclear operators, and gas-fired power generators with data centre proximity offer defensible, long-duration earnings streams.
Valuation discipline remains essential. The 1990s telecom analogy is not a prediction but a reminder: the best investment thesis in the world can produce catastrophic returns if the price paid is wrong. Investors should own the AI theme but size positions with awareness that a deceleration in revenue growth, an efficiency breakthrough that reduces capex requirements, or a broader market correction could produce significant drawdowns even if the long-term thesis remains intact.
Conclusion
The AI capital expenditure boom is real, structural, and accelerating. It is also, at $500 billion per year and growing, the largest corporate bet on a single technology in economic history. The investment thesis that AI will transform productivity, reshape industries, and create enormous value is almost certainly correct. The question that will determine whether AI capex creates or destroys shareholder value over the next three to five years is not whether the technology works but whether the revenue arrives at the pace that current valuations require. The difference between a revolution and a bubble is timing.
Related Reading
The AI capex boom persisted despite the disruption covered in DeepSeek Shock. For the stock that led the spending wave, see Nvidia’s AI Supercycle. For the valuation questions this spending raised, see The AI Bubble Debate.
Related Reading: see Tesla Q1 2026 and the Mag 7 capex reckoning. For continuing coverage on this theme, see our analysis of Cisco’s 25% Networking Surge: The Other AI Trade Comes Into Focus. For continuing coverage on this theme, see our analysis of The Great Tech Divergence: Software Breaks While Silicon Soars.
Leave a Reply