Why Only AI Training Can Save the Economy

What It Covers

AI infrastructure spending now drives 39% of marginal US GDP growth, but enterprise budget caps and token scarcity are threatening lab revenue growth. The argument: mass-scale AI training is the only mechanism that can simultaneously satisfy lab token consumption needs and deliver enterprise ROI justifying increased spend.

Key Questions Answered

• •AI Infrastructure as GDP Engine: AI data centers, hardware, and networking reached 1.4% of US GDP in Q1 2026, doubling from 0.7% the prior year. Excluding AI investment entirely, US economic growth in the first half of 2025 would have been 0.1% annualized. Big Tech AI capex alone is projected to exceed $800B in 2026.
• •Seat-to-Agent Economics Shift: Per-user AI economics have moved from $20–$200 per month seat pricing to potentially thousands of dollars monthly under agentic, usage-based consumption models. Anthropic's revenue run rate jumped from $30B to $47B annually in weeks, driven almost entirely by Claude Code's agentic token consumption rather than new subscriber growth.
• •Token Scarcity Reality Check: Enterprises built 2025 AI budgets around assisted-AI assumptions, then collided with agentic-AI costs. Uber exhausted its entire annual AI budget in four months and imposed a $1,500 monthly per-employee cap. Companies like Ramp are routing to DeepSeek, while Cursor's Composer 2.5 delivers comparable performance to top models at one-tenth the cost.
• •Known ROI Bias Risk: Budget caps and CFO scrutiny push employees toward incremental productivity use cases—doing existing work slightly faster—rather than exploratory agent experiments that generate new economic value. Organizations must deliberately create structured sandboxes and explicit permission frameworks to encourage high-uncertainty agentic experimentation, or they will systematically underutilize AI's transformative potential.
• •Agent Management as New Work Primitive: Managing agents is a fundamentally different knowledge work skill than prompting assisted AI—closer to management training than software training. Only 28% of organizations have enabled employees to use AI to change actual business processes. Video courses produce awareness without confidence. Labs launching forward-deployed engineering teams address only centralized use cases, missing the bottoms-up experimentation required for full value capture.