$725 Billion AI Spending Breakdown:
Where the Money Actually Goes

The Big Picture: Who's Spending What

The spending is concentrated in five companies.

Amazon leads the pack with a projected $200 billion in 2026 capex. CEO Andy Jassy defended the plan by noting that AI capacity is being monetized as fast as it is installed. AWS reached a $142 billion annualized revenue run rate.

Alphabet is spending $175 to $185 billion. It has already revised its capex guidance three times upward from an initial $71 to $73 billion range for 2025.

Microsoft set its calendar-2026 capex at $190 billion. That's well above the $152 billion average analyst estimate. Two-thirds of Microsoft's capex last quarter went toward short-lived assets such as GPUs and data center equipment.

Meta plans $115 to $135 billion.

Oracle is targeting $50 billion.

Collectively, these five companies are spending roughly 75% of their aggregate capex on AI infrastructure. GPUs, high-bandwidth servers, networking equipment, and purpose-built data centers.

CreditSights now projects total spending for the top five hyperscalers at around $750 billion. That's a 67% year-over-year increase. Bank of America forecasts global hyperscale cloud provider capex will exceed $800 billion in 2026.

Where the Money Goes: Training vs. Inference

The most significant shift in AI spending in 2026 is the transition from training to inference.

Training — building the model — is expensive but bounded. It happens once, or periodically when a model is updated. Inference — running the model in production — is a recurring, usage-based cost that never stops.

Inference workloads are set to overtake training revenue by 2026. As one analysis put it, "many companies now spend more on running AI than on building it." Industry analyses put inference at roughly 80 to 90% of total compute dollars over a model's lifecycle. Training accounts for 10 to 20%.

Deloitte estimated that inference workloads accounted for half of all AI compute in 2025. That jumped to two-thirds in 2026. By 2030, inference is projected to account for 80% of AI critical IT load. Infrastructure for live services will be four times larger than infrastructure for training.

This matters because inference is a recurring cost. One analysis estimated the inference market to be 10 to 50 times larger than the one-time training market. The rise of agentic AI accelerates this. Every user query, every API call, every agent loop burns tokens. Each token has a cost.

The Infrastructure Breakdown: What $725 Billion Buys

Breaking down the $725 billion reveals a pyramid of costs.

At the base is the network. Hundreds of billions go to building the pipeline that moves data into and out of data centers. Pipeline costs include undersea cables, fiber backbone capacity, last-mile connectivity, and cross-connect fees.

Above that sits the data center structure itself. Land acquisition, construction, electrical systems, cooling infrastructure, and security. Power and cooling alone can account for 30 to 50% of total data center operating costs.

At the top of the pyramid is compute hardware. GPUs, TPUs, and custom ASICs — the chips that actually run the AI workloads. These represent the largest single category of spending, but they cannot function without everything below them.

Each hyperscaler's capex also includes layer-specific allocations. AWS builds its own networking chips (Nitrov3, 800 Gbps per device). Meta has custom-designed GPU clusters optimized for recommendation models. Microsoft has invested in modular data center designs that reduce construction timelines.

NVIDIA captured the majority of GPU revenue in 2025. Its data center revenue was $130.5 billion for fiscal 2026 — a testament to how much of this spending flows through a single supplier. But AMD is making inroads with the MI400 series, and the hyperscalers' custom silicon (Trainium, TPU, Axion) is reducing NVIDIA's share of incremental spend.

The Hidden Costs Nobody Accounts For

Beyond the headline numbers, several categories of spending are systematically undercounted.

Software licensing for AI platforms is one. Enterprise AI platforms from Databricks, Snowflake, and Palantir cost $500,000 to $5 million annually per deployment. Gartner found that organizations underestimated AI software costs by 30 to 55% in their initial budgets.

Data preparation and labeling account for an estimated 30 to 50% of total AI project costs. Data scientists spend up to 80% of their time on data preparation, not model building. This labor cost rarely appears in the "AI budget."

Talent costs are escalating. The median salary for an AI engineer at a hyperscaler exceeds $400,000. At startups and mid-market firms, AI specialists command 30 to 50% premiums over traditional software engineers. The total cost of an AI engineering team — including recruiting, retention, and churn — frequently doubles the direct salary line item.

Compliance and governance add another layer. 77% of organizations believe AI adoption has outpaced their governance capabilities. Building internal guardrails, audit systems, and compliance frameworks costs $2 to $5 million per enterprise deployment. New regulations in the EU (AI Act) and potential US federal frameworks add recurring compliance costs that few budgets anticipate.

The hidden costs — data, talent, software, governance, and energy — may add 40 to 80% on top of the reported capex figures. That would push the true cost of AI infrastructure well above $1 trillion in 2026.

The Enterprise Reality Check

While hyperscalers spend billions, enterprise AI adoption tells a different story.

98% of FinOps practitioners now manage AI spend. That's up from just 31% two years ago. The era of unchecked AI spending is ending. The era of FinOps for AI has begun.

CloudZero's survey found that 71% of organizations have little control over AI costs. Engineering and product teams are often paying for the same AI capabilities through separate budgets. A consolidated AI budget can save 30 to 40% without cutting a single feature.

Enterprises are also finding that AI cost optimization requires a fundamentally different approach from traditional cloud cost management. AI costs are driven by model complexity and token consumption — not just compute instances. Optimizing prompt structure, caching inference results, and batching requests can cut inference costs by 50 to 90%.

Shadow AI spending is another blind spot. Employees using unsanctioned AI tools cost enterprises an estimated 15 to 30% of their official AI budget. Not because the tools are expensive individually — but because the costs are invisible and uncapped.

The Bottom Line

The $725 billion question is not just "where is the money going?" It's "is the money being spent wisely?"

Hyperscalers are betting that every dollar of AI capex generates a multiplier in future cloud revenue. Amazon's cloud backlog is $464 billion. Microsoft's is $627 billion. Google's is $460 billion. These are signed contracts with enterprise customers who have already committed to paying for the cloud capacity these companies are building.

But the enterprise side tells a different story. 98% of FinOps practitioners now manage AI spend. That's up from just 31% two years ago. The era of unchecked AI spending is ending. The era of FinOps for AI has begun.