Big Tech Is Spending $725 Billion on AI in 2026 — A 77% Jump That Dwarfs Any Prior Infrastructure Boom

The AI infrastructure arms race has reached a number that defies easy comparison to any prior technology build-out in history. Google, Microsoft, Meta, and Amazon are collectively on track to spend approximately $725 billion on capital expenditures in 2026 — up 77% from the already-record $410 billion spent in 2025. To put that in perspective: the entire US interstate highway system, built over four decades, cost roughly $530 billion in today’s dollars. Big Tech is planning to spend more on AI infrastructure in a single year.

Who Is Spending What

Microsoft leads the pack with $190 billion in projected capex for calendar year 2026 — significantly above the $152 billion analyst consensus heading into the year. The company’s CFO disclosed that $25 billion of that figure reflects higher costs from rising memory chip and component prices rather than additional compute acquisition, an inflationary pressure affecting the entire sector. Despite this historic spend, Microsoft has indicated it expects to remain capacity-constrained throughout 2026.

Amazon is approaching $200 billion in annual capex, cementing AWS’s position as the largest cloud infrastructure buildout in the world. The company has been investing heavily in custom silicon — its Trainium 3 chip for training and Inferentia series for inference — to reduce its dependence on Nvidia and manage unit economics at scale.

Alphabet has guided to full-year capex in the range of $180 billion to $190 billion, reflecting Google Cloud’s explosive 63% year-over-year revenue growth and a backlog that has nearly doubled quarter-over-quarter to $460 billion. That backlog figure is arguably the most significant validation that enterprise AI demand is real and durable.

Meta raised its capex forecast to between $125 billion and $145 billion, hiking both ends of guidance by $10 billion, driven by the same memory and GPU cost inflation affecting its peers. Meta’s investment is more concentrated in training infrastructure for its open-weight Llama model family and its proprietary AI recommendation and ad systems than in external cloud services.

Why Costs Are Exploding

Several structural factors are pushing AI infrastructure costs beyond what even optimistic scenarios projected a year ago.

Memory is the new GPU. High-bandwidth memory (HBM) has emerged as the binding constraint for AI workloads, and demand has so outstripped supply that prices are at sustained premium levels. Microsoft’s disclosure that $25 billion of its budget reflects component cost inflation — not additional capacity — suggests the memory crunch is forcing companies to pay significantly more for the same compute footprint.

Cooling infrastructure. The shift to liquid cooling for dense GPU clusters, required for the thermal management of H100 and GB200 racks, adds substantial cost and construction complexity compared to earlier air-cooled data center designs. New purpose-built AI data centers are substantially more expensive per square foot than general-purpose facilities.

Power and land. In markets with constrained grid capacity — Northern Virginia, Ireland, Singapore — the competition for power access has pushed energy procurement costs sharply higher. Several major projects have faced multi-year delays due to interconnection queue backlogs, creating pressure to pay significant premiums for secured capacity.

Custom silicon offset. All four companies are investing heavily in proprietary AI chips (Google’s TPU v5, Amazon Trainium 3, Microsoft Azure Maia 2, Meta’s MTIA v2) specifically to reduce their per-compute-unit costs and dependency on Nvidia. But these programs require their own substantial R&D and manufacturing investment before they can generate meaningful capex savings.

Revenue Is Growing — Just Not as Fast as the Spend

The spending surge is not happening in a vacuum. Enterprise AI demand is translating into meaningful revenue at the hyperscalers, and the growth rates are impressive by any standard.

Microsoft’s AI business is on an annual revenue run rate of $37 billion, representing 123% year-over-year growth. Google Cloud saw revenue surge more than 60% year-over-year to $20 billion, with the cloud backlog at $460 billion. Meta’s ad business continues to generate the cash flows that fund its AI buildout, with AI-optimized ad targeting driving per-user revenue improvement.

But Jefferies analyst Brent Thill — who coined the phrase “the AI economy is healthy” in a widely circulated note — is essentially making a faith-based argument: that the infrastructure being built today will be the foundation for AI revenue streams in 2027, 2028, and beyond that don’t yet exist in their full form. His dismissal of spending skepticism as “garbage” reflects the conviction of the bull case, but the revenue-to-capex gap is real and widening.

The core tension is timing. Infrastructure spending must be committed one to three years before the compute capacity is available, and before it’s clear which specific AI workloads will consume it. If AI adoption accelerates as the bulls project, companies that build now will have structural capacity advantages. If adoption is slower or more concentrated among fewer use cases, the write-downs and overcapacity problems could be severe.

The Semiconductor Supply Chain Is Straining

The $725 billion number is not just a statement about AI demand — it’s a test of whether the global semiconductor supply chain can actually deliver at this scale. TSMC, Samsung, and SK Hynix are collectively running at or near full capacity on the processes that matter most for AI chips (3nm and 5nm logic, HBM3E memory). New capacity, even with aggressive investment, takes 18 to 36 months to come online.

Nvidia remains the dominant GPU provider, with its GB200 NVL72 rack systems the most sought-after infrastructure in the world. The company’s ability to maintain quality and yield at the volumes implied by this spending level is a genuine operational risk — and explains why AMD, Intel, Amazon, Google, and Microsoft are all running parallel programs to develop alternatives.

Analyst Splits: Validation vs. Warning

Wall Street is not of one mind on whether the $725 billion is justified. The bull case, articulated by analysts at Jefferies, Deutsche Bank, and Goldman Sachs, points to the backlog growth at hyperscalers as evidence that supply is being absorbed as fast as it can be built. If Google Cloud has $460 billion in contracted backlog, the infrastructure spending to serve that backlog is arguably mandatory, not speculative.

The bear case, articulated with more frequency in private than in print, centers on concentration risk: a disproportionate share of AI cloud revenue is coming from a small number of very large enterprise customers who are themselves in the early phases of AI deployment. If those customers consolidate spending, hit budget cycles, or decide to run models on their own infrastructure, the cloud absorption rates could shift significantly.

The $725 billion number will be one of the defining economic stories of 2026 — either as the largest successful infrastructure bet in corporate history, or as the opening chapter of a cautionary tale about technology investment cycles.

Either way, the AI economy has already passed the point of no return. The question now is whether the applications catch up to the infrastructure.