Behind the video
Tokenmaxxing Madness Brings Anthropic To Profitability
Overall Thesis
The video argues Anthropic's claimed Q2 2026 adjusted profitability is likely a temporary artifact of reduced SpaceX compute fees and reallocated training compute, and will reverse once full SpaceX payments begin in July. It frames the broader 'token maxing' phenomenon across big tech as a wasteful, unsustainable corporate fad driven by fear rather than demonstrated productivity gains, predicting AI company revenues will eventually collapse when the fad ends.
Narratives
NVDANvidiaThe video quotes Nvidia CEO Jensen Huang's claim that engineers who don't spend heavily on AI tokens are a red flag, comparing AI skepticism to refusing CAD tools. The host critiques this as an unsupported rhetorical tactic rather than offering any view on Nvidia's stock or business outlook.
Key Arguments
- Jensen Huang says he would be 'deeply alarmed' if a $500k engineer didn't spend at least $250,000 on tokens
- Host argues Huang provides no evidence that token consumption correlates with productivity and uses a manipulative CAD-tools analogy
METAMeta PlatformsDescribes an internal Meta leaderboard called 'Claude-onomics' where employees compete to burn the most AI tokens, and notes the irony that Meta's own Llama models are reportedly unused internally despite massive investment. The host does not express a view on Meta's stock.
Key Arguments
- A Meta employee created an internal leaderboard tracking token consumption across staff
- The top-ranked employee consumed 328 billion tokens in one month, costing over $100,000 at minimum
- Despite billions spent on Llama models, Meta employees reportedly prefer using Claude rather than Meta's own AI
UBERUberUber is cited as having exhausted its entire 2026 AI budget within the first four months of the year, mostly on Claude Code spending, used as evidence of rapidly escalating and possibly unsustainable enterprise AI costs rather than a comment on Uber's business or stock.
Key Arguments
- Uber admitted burning through its full-year 2026 AI budget in just four months
- Most of this spending went toward Anthropic's Claude Code product
MSFTMicrosoftDiscusses Microsoft's historical OS/2 lines-of-code productivity metric failure and draws a parallel to Microsoft's current mandate that AI usage factor into performance reviews amid ongoing layoffs. No stock or financial outlook is expressed for Microsoft.
Key Arguments
- Microsoft now treats AI usage as non-optional and ties it to performance reviews during a period of layoffs
- An anonymous Microsoft employee admits to unproductive 'token maxing' just to inflate reported AI usage
- The video draws a direct parallel to IBM's flawed lines-of-code metric from the 1980s, which produced buggy, bloated software
CRMSalesforceSalesforce is cited as requiring engineers to spend a minimum amount on AI tokens weekly, leading some employees to build throwaway features purely to meet quotas. This is presented as an example of wasteful corporate AI spending rather than as a stock thesis.
Key Arguments
- Salesforce mandates at least $170 per week in Claude Code/Cursor token spend per engineer, roughly $9,000 per year minimum
- Many employees exceed the minimum to look better to managers, including building features they know will never ship
The host cites Coca-Cola CEO James Quincey's retirement, reportedly linked to insufficient understanding of AI, as a bizarre example of AI-driven corporate paranoia given AI's limited actual relevance to Coca-Cola's core business beyond generating poorly received ad commercials.
Key Arguments
- Coca-Cola's long-time CEO is retiring reportedly because he doesn't understand AI well enough
- The host notes Coca-Cola's only real AI usage is in generating commercials that are widely disliked by the public, calling the CEO change 'absolutely bizarre'
Hedges & Caveats
- We don't know how much Anthropic's reduced SpaceX fee for May and June actually is
- Anthropic probably pays more than $500 million in stock-based compensation, so it is probably still losing money on a GAAP basis despite the adjusted profit
- The hypothesis that AI dependency erodes unassisted developer skill is explicitly called 'just a hypothesis' by the narrator
- The second METR study's productivity improvements had confidence intervals that crossed zero, meaning they were not statistically significant