Two numbers define the AI industry in 2026.
$690 billion. That’s what Microsoft, Alphabet, Amazon, Meta, and Oracle will spend on capital expenditure this year. Roughly $450 billion targets AI infrastructure—datacenters, GPUs, networking, power systems.
$0.10. That’s what Salesforce charges per agent action. Not per user. Not per seat. Per autonomous step an AI takes on your behalf.
These numbers should not coexist. The infrastructure layer is making the largest capital bet in technology history while the application layer is racing toward marginal cost pricing. Someone is wrong about the math.
The Infrastructure Sprint
The 2026 hyperscaler capex numbers are staggering. The Big Five will spend more on infrastructure this year than the GDP of Thailand. Capital intensity—the percentage of revenue devoted to capex—has reached 45-57% for some players. For context, traditional software companies spent 5-10%.
This is not equity-funded expansion. The industry raised $108 billion in debt during 2025 alone, with projections suggesting $1.5 trillion in debt issuance over the coming years. When Microsoft issues bonds to buy NVIDIA chips, you know the financing model has fundamentally shifted.
What’s driving this? Two factors. First, training compute demand continues to double every six months for frontier models. Second—and more importantly—inference is becoming the dominant workload. Every chat, every agent action, every search query requires GPU cycles. The $0.10 per action model presumes near-infinite inference capacity at near-zero marginal cost.
Here’s the problem: inference at scale is not near-zero cost. It’s the opposite.
The Energy Constraint
For two years, GPU availability was the bottleneck. You couldn’t get H100s. Lead times stretched to months. That constraint has largely resolved—NVIDIA shipped over 3 million datacenter GPUs in 2025.
The new bottleneck is energy.
A modern AI datacenter consumes 50-100 megawatts. The next generation—designed for reasoning models and multi-agent orchestration—will require 200-500MW. That’s a small city. The electrical grid cannot scale at the pace of GPU deployment.
Consider this: the total power consumption of all US datacenters was about 17 gigawatts in 2022. Projections for 2027 exceed 35 gigawatts. The industry is attempting to double its energy footprint in five years while the grid was built for decades of incremental growth.
This matters for agent economics. Every $0.10 action consumes energy. If energy becomes the binding constraint—and all evidence suggests it will—then inference costs cannot fall to zero. They may rise.
The Enterprise Agent War
While hyperscalers build infrastructure, enterprise vendors are deploying agents at an unprecedented pace.
Salesforce Agentforce has 8,000+ customers and generated $900 million in AI revenue within six months of launch. ServiceNow ranks #1 in Gartner’s AI Agent evaluation. Microsoft Copilot now orchestrates agents across the entire 365 suite.
The pricing model is the real story. Salesforce’s Flex Credits at $0.10 per action represents a fundamental break from per-seat SaaS pricing. Why charge $50/month/user when an agent might take 100 actions worth $10? Why charge $50 when the same agent might take 1,000 actions worth $100?
This is consumption-based pricing applied to autonomous systems. It’s AWS for cognitive work.
But here’s the uncomfortable question: what happens when the cost of providing those actions exceeds the price?
The Margin Squeeze
Current economics work because:
- Hyperscalers have subsidized inference to capture market share
- Model efficiency improvements have outpaced demand growth
- Energy costs have been manageable relative to compute costs
None of these conditions are guaranteed to persist.
If energy becomes scarce, inference prices rise. If model efficiency plateaus, infrastructure utilization degrades. If hyperscalers need to service $1.5 trillion in debt, subsidies end.
The enterprise agent market is projected to reach $52.6 billion by 2030. The infrastructure being deployed on its behalf will cost over $1 trillion. That’s a 20:1 infrastructure-to-revenue ratio. No industry has sustained that math long-term.
Three Scenarios
Scenario 1: The Efficiency Miracle
Model architecture improvements reduce inference costs by 90%+ while maintaining capability. Energy per query drops dramatically. The $0.10 action becomes $0.01, then $0.001. Infrastructure is overbuilt but utilization catches up as agent adoption accelerates. Hyperscalers earn returns on their capex.
Scenario 2: The Margin Collapse
Inference costs remain stubborn. Energy constraints limit deployment. Hyperscalers cannot raise prices because enterprise contracts lock in consumption rates. The $1.5 trillion debt load becomes a balance sheet crisis. Infrastructure buildout slows dramatically by 2027.
Scenario 3: The Consolidation
Only the largest hyperscalers can sustain the infrastructure burden. Smaller players exit or are acquired. AI becomes a utility—centralized, regulated, with pricing power. The $0.10 action becomes $1.00 as providers consolidate market power. Enterprise customers have no alternatives.
What to Watch
Energy capacity announcements: If hyperscalers start buying power plants, the constraint is real.
Inference pricing trends: If AWS/ Azure/ GCP start raising per-token prices, the subsidy era is ending.
Enterprise agent ROI metrics: If $0.10 actions don’t deliver productivity gains, adoption stalls and the revenue side of the equation weakens.
Debt market signals: If AI infrastructure bonds trade at widening spreads, the financing model is under stress.
The Bottom Line
The AI industry is engaged in the largest capital deployment in technology history. The numbers are remarkable. The assumptions are aggressive. The risks are asymmetric.
If infrastructure providers are right, AI becomes the platform for all economic activity—justifying every dollar spent. If they’re wrong, the $1.5 trillion debt load becomes the defining crisis of the next decade.
The $0.10 action is not a pricing strategy. It’s a bet on the efficiency miracle. The infrastructure layer is betting that inference costs collapse while the application layer is betting that pricing power enables margins.
Both cannot be right. By 2027, we’ll know which side miscalculated.
The gap between infrastructure investment and application revenue will define AI economics for the next decade. The $690 billion question is whether efficiency gains can close it.