AI Infrastructure Capex Hits $941.5 Billion — The Largest Capital Buildout in Tech History

The four largest US technology companies will spend more on AI infrastructure this year than Switzerland's entire GDP. That is not hyperbole. It is a budget line item.

This is the largest coordinated infrastructure buildout in the history of the global technology industry. It surpasses the combined inflation-adjusted spending of the Interstate Highway System and the Apollo program. And it raises a question that no earnings call has fully answered: what happens when the spending stops accelerating?

The Big Four: Who Is Spending What

Amazon leads in absolute terms with $200 billion in planned 2026 capex, up roughly 50% year-over-year. The bulk flows into AWS data centers, custom Trainium silicon, and logistics infrastructure. CEO Andy Jassy told investors the company is monetizing capacity as fast as it can install it.

Alphabet is close behind at $180–190 billion, nearly double its 2025 spend. About 60% goes to servers, 40% to data centers and networking. The company raised $80 billion through a combination of bond offerings and a $10 billion Berkshire Hathaway private placement, signaling that even Alphabet's operating cash flow cannot cover the buildout alone.

Microsoft is running at a $145–150 billion annualized rate, with $37.5 billion spent in a single quarter. The company disclosed an $80 billion backlog of Azure orders that cannot be fulfilled due to power constraints, a rare admission that demand is outpacing even the most aggressive buildout pace.

Meta, despite having no public cloud to monetize, is spending $115–135 billion, its highest capex year by a significant margin. The company is building gigawatt-scale data centers in Louisiana and Ohio and has become one of the largest corporate purchasers of nuclear energy.

Where the Money Actually Goes

The common assumption is that hyperscaler capex equals GPU purchases. That was true in 2023 and 2024. In 2026, the composition has shifted fundamentally.

More than 60% of AI infrastructure spending now goes to power, cooling, and data center construction, not compute hardware. NextWave Insights estimates that approximately 40% of announced AI data center projects face construction delays due to power infrastructure bottlenecks, not chip supply. The binding constraint has moved from silicon to electrons.

Microsoft has signed long-term power purchase agreements with nuclear operators including Three Mile Island. Meta partnered with Vistra, Oklo, and TerraPower. Amazon is building behind-the-meter gas plants. The hyperscalers are effectively becoming energy companies with a software side business.

The Inference Threshold

The shift from training to inference as the dominant compute load is the most important structural change in AI infrastructure economics. In 2023, roughly 80% of AI compute went to model training. By mid-2026, that ratio has inverted. Inference now consumes an estimated 70% of AI compute cycles across the hyperscalers.

This matters because inference economics are fundamentally different from training economics. Training is a fixed cost: you spend once and get a model. Inference is a variable cost: every user query, every API call, every agent loop consumes compute. The hyperscalers are effectively placing a multitrillion-dollar bet that inference demand will grow faster than the efficiency improvements that reduce per-token cost.

The bet has historical precedent. Cloud computing followed the same arc: early adopters paid a premium for raw capacity, then unit costs collapsed and volume exploded. Token costs have already dropped 280-fold in two years, according to Deloitte's Tech Trends 2026 analysis. Yet enterprise AI spending is still accelerating because usage is growing faster than prices are falling.

This dynamic explains why the hyperscalers are building custom silicon despite NVIDIA's dominant position. Trainium, TPU, Maia, and MTIA are not designed to outperform NVIDIA on raw training benchmarks. They are designed specifically to optimize the inference cost curve: to make each query cheaper than the hyperscaler can buy from NVIDIA. The winner in this race is not the company with the fastest chip. It is the company with the lowest cost per token at scale.

The Financing Shift

As we wrote in July, the question of whether AI capex generates proportional returns has no settled answer. What is clear is how the buildout is being funded, and the answer marks a structural shift in how the largest technology companies manage their balance sheets.

In 2021, the Big Four spent $149 billion buying back their own stock. In 2026, they are spending $700 billion building infrastructure. The swing from financial engineering to physical asset construction is the largest reallocation of corporate capital in modern history.

This transition has created an entirely new asset class. AI data center debt (bonds issued by hyperscalers, neocloud operators, and their financing vehicles) surpassed $200 billion in 2025 alone, according to CreditSights. Meta issued $30 billion in a single bond deal, among the largest corporate bond offerings on record. Oracle launched a $20 billion at-the-market share offering. Alphabet raised $85 billion across six currencies, including a rare 100-year GBP bond.

Goldman Sachs projects total hyperscaler capex from 2025 through 2027 will reach $1.15 trillion, more than double the $477 billion spent in the 2022-2024 period. The firm's latest forecast pegs 2026 alone at $755 billion, 83% above 2025.

The Downstream Effects

The scale of the buildout is reshaping multiple industries simultaneously. NVIDIA's data center revenue reached $81.6 billion in a single quarter, up 85% year-over-year. Custom silicon programs at all four hyperscalers (Trainium at Amazon, TPU at Google, Maia at Microsoft, MTIA at Meta) represent a direct challenge to NVIDIA's GPU dominance, though none have yet displaced it at scale.

Data center vacancy rates in North America have hit 1.4%, an historic low. Power developers are seeing demand that resembles the early days of the shale gas boom. Equinix and Digital Realty are trading at multiples that reflect data center scarcity rather than colocation margins.

The energy implications are staggering. AI data center power demand is projected to reach 1,000 TWh globally by 2026, equivalent to the entire electricity consumption of Germany. McKinsey estimates $1.3 trillion, or 25% of total AI investment through 2030, will flow to power and cooling infrastructure alone.

Five Bets, One Buildout

The AI infrastructure buildout is not a single investment theme. It is at least five, each with different risk profiles and return horizons: the chip suppliers (NVIDIA, AMD, Broadcom), the infrastructure operators (Equinix, Digital Realty, Vertiv), the power developers (Constellation Energy, Vistra), the hyperscalers themselves (Amazon, Microsoft, Alphabet, Meta), and the debt instruments that finance the buildout.

The risk is not that the buildout stops. It is that the buildout continues but the return on invested capital disappoints, and the market reprices the entire stack from growth to value multiples. The 2026 capex guidance already priced in 50% above analyst expectations. The bar for future upside is high.

The opportunity is that the buildout is supply-constrained, not demand-constrained. Every company we spoke to reports the same dynamic: they could spend more if they could get power permits faster, or secure more GPUs, or find enough skilled construction labor. That scarcity is a pricing signal, and it suggests the buildout has further to run than the current guidance implies. The question is not whether the spending continues, but which parts of the value chain ultimately capture the returns.

Sources