ex-Meta Principal Explains the AI Token Game
In a Nutshell
The AI industry has built an elaborate circular economy where hyperscalers, model companies, and investors all profit from increased token consumption, with enterprise customers ultimately funding the entire system through premium API pricing. Token maxing is heavily incentivized through subsidies, leaderboards, and FOMO marketing, but evidence shows it often produces twice the output at ten times the cost without correlating to actual business value. Enterprise leaders should replace token metrics with adoption dashboards and outcome tracking, while implementing cost guardrails and removing organizational bottlenecks rather than rewarding raw token volume.
These notes were generated by AI and may contain inaccuracies.
Token maxing is viewed by some as evidence of being an expert in agentic engineering who lives on the bleeding edge and achieves massively higher productivity than peers. Others see it as the most wasteful spending and the dumbest idea of the time. One perspective states: "If you can token max it's like I mean you could buy millions of years of consciousness." Engineers commonly run hundreds of agents, with some running a few thousand agents nightly for deeper work.
AI tokens are very costly. Companies are burning through budgets faster than ever before and receiving shocking bills. Industry leaders continue telling employees to use more tokens, and many companies incentivize AI usage by counting tokens. One executive noted they will probably use $300 million of Anthropic this year. Some companies effectively fund unlimited tokens for everybody.
The ecosystem begins with end users sending prompts to large language models, which generate tokens to complete work. Most people use frontier models and pay companies like OpenAI or Anthropic for tokens used. Large language models run in massive data centers operated by hyperscalers such as AWS, Google Cloud, and Microsoft Azure. Generating tokens requires significant computation, so OpenAI and Anthropic pay hyperscalers for the compute used. Data centers host large numbers of GPUs from chip makers including Nvidia and AMD. Google is the only company that controls the entire chain from chips through models.
Chip makers have the simplest and safest business model because they take purchase orders and sell chips with every transaction generating revenue. Hyperscalers face more complexity. When Microsoft spends money buying GPUs from Nvidia and placing them in Azure, they only recover costs if others pay to use those GPUs. Microsoft invested $13 billion in OpenAI, primarily in the form of Azure compute credits. Microsoft gives money to OpenAI, which returns it when running GPT models on Azure, recorded as revenue for Microsoft. This creates a circular game where Microsoft uses its own money to buy compute from itself through OpenAI's hands, increasing revenue numbers without net gain or loss.
OpenAI adopted the same approach by offering to invest $2 million in every YC company in the current batch, delivered as token credits. The same pattern occurs between AWS, Google, Microsoft, and Anthropic. All hyperscalers are investors in Anthropic. Microsoft, Nvidia, and Anthropic announced strategic partnerships with Anthropic committing to scale on Azure, adopt Nvidia architecture, and purchase $30 billion of Azure compute capacity. When users pay Anthropic for tokens, Anthropic's revenue increases, Microsoft Azure revenue increases as Anthropic pays for compute, and demand grows for Nvidia GPUs.
Sequoia Capital has been vocal about token maxing, with partner Sonia Huang stating everyone should be token maxing. Sequoia maintains its own leaderboard, as do many portfolio companies. Sequoia has invested in Anthropic, Nvidia, and OpenAI. The entire industry structure creates incentives for increased token usage across tech giants and capital providers.
AI companies lead by example by requiring heavy AI usage without limits from their own employees, then publicly discussing how far behind other companies are. One statement noted: "You should token max just giving engineers as many tokens as possible and that's a lot of fun. I hit 250 million one day." Another claimed: "You deploy a billion tokens of intelligence a day."
Small startups and individuals also engage in heavy token usage due to subsidization rather than direct business incentives. Monthly subscriptions from Anthropic and OpenAI at $200 per month are heavily subsidized. Heavy users can receive thousands of dollars worth of tokens for the fixed price. After diminishing returns begin, users continue burning tokens until hitting limits because there is no additional cost.
Subsidized subscriptions create the largest marketing campaign ever by generating FOMO. When AI companies advocate token maxing directly, they appear as salespeople. When fast-moving startups launch products while claiming massive productivity gains through AI, it creates different credibility. Enterprise leaders see big tech token maxing, observe small startups shipping faster through heavy AI use, and conclude their own companies must catch up.
AI companies lose money or break even when burning their own tokens and when subsidized users exceed plan values. Profit comes from enterprise companies charged through API pricing that is token-based and an order of magnitude more expensive than subsidized subscriptions. Enterprise customers effectively pay for everyone else in the chain. The profit of the entire food chain depends on enterprises adopting AI and token maxing.
The narrative that more token usage is required to survive has been carefully constructed, though AI does represent a genuine technological revolution. Leaders should avoid building token usage leaderboards because they reward gaming the metric and token usage does not measure productivity. A Jellyfish study of over 7,000 engineers found developers spending the most tokens generated roughly twice the output at ten times the cost.
Instead, track adoption through a dashboard showing aggregate percentages of people not using AI, using it occasionally, using it regularly, or using it as their primary work method. Focus on business outcomes including how many teams hit OKRs ahead of schedule, exceeded targets, or set more aggressive goals than previously imagined. Identify AI usage patterns correlated with team success.
Implement cost guardrails with generous usage quotas for experimentation rather than unlimited access. Quotas incentivize power users to find efficient approaches. Invest in existing employees who understand company systems rather than conducting AI layoffs. Provide training from internal AI-proficient colleagues to those who have not yet adopted AI regularly. Remove outdated processes, ownership boundaries, politics, quarterly planning requirements, cross-team dependencies, and approval bottlenecks that prevent AI-native engineers from operating at startup speed.
Those who have not yet adopted AI should learn and experiment despite initial productivity drops during the learning curve. Historical technological revolutions show that refusing adoption creates lasting disadvantage even if hype cycles burst. Those already using AI heavily should reject token counts as success metrics because this leads to producing low-quality work that harms reputation among peers. Companies ultimately value business success and cost efficiency over token volume.
Create documented success stories showing AI adoption producing real business results such as teams exceeding goals ahead of schedule, improved engineering efficiency for colleagues, or completion of previously avoided tech debt projects. Escalate organizational bottlenecks that limit velocity to leadership for resolution.
Keep Kun Chen in your library
Save the videos and channels worth coming back to, and find them again in one place.
