Video summary
AI Is More Expensive Than Humans
Main summary
Key takeaways
Summary of the Video’s Main Points (AI Costs + “Token Maxing” Incentives)
Uber’s AI spend burned out faster than planned
Uber’s CTO said the company had exhausted its full 12-month AI budget by April 2026. The main driver was aggressive internal adoption of Anthropic’s Claude Code:
- Adoption rose from 32% to 84% within months (across ~5,000 engineers)
- 95% of engineers used AI tools monthly
- 70% of committed code was allegedly AI-generated
The video argues this wasn’t purely organic adoption. Uber allegedly introduced internal leaderboards ranking teams by AI tool usage, so higher consumption brought higher status. Individual engineers were reportedly incurring $500–$2,000/month in AI usage costs.
The presenter’s framing: the company effectively rewarded spending, then treated the oversized bill as a surprise.
The economics: cheaper tokens can still produce higher total costs (Jevons paradox)
The presenter explains that even if per-token costs fall, total costs can rise because agentic AI consumes more tokens per task:
- The per-token cost drops (citing Gartner: inference on frontier models could be 90%+ cheaper by 2030)
- But agentic workloads may use 5–30x more tokens per task
Lower unit prices can therefore cause far more total token usage, increasing aggregate spend. This is presented as the Jevons paradox applied to compute: efficiency leads to more consumption, not less.
Meta’s “Cloudonomics” turns token use into a competitive game
The video claims Meta employees used a leaderboard system called Clodonomics to track token consumption across 85,000+ employees, with gamified titles for top users.
Allegations cited include:
- In a 30-day period, employees consumed 60 trillion tokens
- The top user consumed 281 billion tokens in a month
The presenter argues this points to idle or engineered token inflation—for example, employees allegedly left agents running idle to improve rankings. Even after changes, Meta’s CTO was quoted as endorsing the idea that spending more tokens was “easy money,” implying no perceived limit.
Amazon and Microsoft reinforce the same incentive trap
The video portrays both companies as sustaining similar incentives:
- Amazon: employees allegedly “token max” by using internal AI tools for trivial tasks to climb rankings.
- Microsoft: the presenter says Microsoft initially offered Claude Code alongside Copilot CLI. Engineers allegedly preferred Claude Code, but Microsoft later cancelled Claude Code licenses for some divisions (framed as “tool chain unification,” with timing suggesting cost-cutting).
The presenter highlights the irony: Microsoft removed the competitor its engineers liked, and the shift aligns with Copilot moving toward usage/token-based billing—restructuring costs rather than solving the underlying problem.
Nvidia’s “contradiction” is framed as incentives, not errors
The presenter claims:
- Nvidia CEO Jensen Huang wanted engineers to consume extremely large token amounts (e.g., $250k tokens per $500k engineer)
- Nvidia VP Bryan Catanzaro noted compute costs are already beyond employee costs
The presenter argues both statements can be “true” under different incentives:
- Nvidia benefits from GPU demand (more tokens → more GPU revenue)
- Engineering leadership manages within a cost envelope
Broader claim: supply-chain incentives reward consumption, which then propagates downstream to customers and employees.
Is it a bubble? The presenter rejects certainty, but flags structural risk
The presenter refuses to label the situation a bubble “in real time,” arguing that certainty arrives only after outcomes. However, they cite instability signals:
- 2026 capex across major firms reportedly reaching ~$740B (Amazon, Microsoft, Alphabet, Meta)
- Tech layoffs in 2026 already exceeding last year’s pace
- MIT study: AI automation is economically viable in only 23% of roles with primary visual tasks (implying much work still requires humans)
Framing: the technology works, but the question is whether spending rates are sustainable relative to returns.
Core critique: measuring “inputs” (tokens/usage) instead of “outputs” (value)
The presenter argues token/leaderboard systems treat consumption as a proxy for value. But there’s no guaranteed relationship between “more tokens” and “better business outcomes.”
They use an analogy: rewarding salespeople for fuel burned, not for houses sold. This creates competitive AI waste, where teams optimize easily measured metrics instead of real results.
Proposed alternative: redirect compute toward problems that generate durable demand
The video argues the industry’s demand-generation strategy is wrong. Instead of pushing existing AI users to consume more, the presenter suggests funding major compute access for universities and cross-disciplinary researchers (e.g., chemistry, physics, medicine, economics, climate science, etc.).
The goal is to use AI for breakthroughs that naturally create sustained demand, such as:
- Drug discovery
- Climate modeling
- Protein folding
The presenter emphasizes this isn’t meant to replace researchers, but to empower expert-led work, tying AI usage to meaningful outputs rather than gamified consumption.
Closing implication
The presenter claims Uber is adjusting course (a form of course correction), but argues the bigger issue is widespread:
- Without a coherent theory of what AI is supposed to accomplish,
- organizations default to consumption metrics,
- leading to burned budgets and potentially fragile deployment strategies.
Presenters / Contributors Mentioned
- Primary presenter (narrator): “I’m ML” (described as having a PhD in computer science)
- Uber CTO: (name not given in subtitles)
- Meta CTO: Andrew Bosworth
- Gartner analyst: Will Summer
- Firms referenced: The Information, Forbes, The Verge (Tom Warren), Financial Times, Axios
- Nvidia CEO: Jensen Huang
- Nvidia vice president (applied deep learning): Bryan Catanzaro