Ep. 020 - Anthropic vs OpenAI Usage, Margins, Meta Compute, Future of MSL (Tokenomics) | Crystual Huang, Max Kan, Joey Brookhart, Jordan Nanos

Summary

The discussion opens with enterprise token budgeting and quickly lands on the idea that most employees are nowhere near their caps; the real pressure is on a small cohort of power users, especially in coding-heavy teams. The guests argue that coding/software engineering dominates token consumption, while many budgeting policies overfocus on low-cost tasks like email. They compare subscription plans versus API billing, framing consumer plans as heavily subsidized and enterprise plans as deliberately pushed toward per-token pricing because that is where the margin sits. The episode then contrasts Anthropic’s API-heavy business mix and margin profile with OpenAI’s much larger consumer base and low paid conversion, while suggesting recent model gains are shifting OpenAI toward a healthier enterprise/API mix. The back half turns to Meta compute and the “clawback” logic of renting capacity at high rates, then to RL environments as a new data market where frontier labs will pay extremely high prices for tasks that improve models.

Notes

• Enterprises are tightening token budgets, but the speakers think the pain is concentrated in a small number of power users rather than most employees.
• Some companies route work through cheaper models first, then compress prompts before sending them to expensive frontier models.
• Budget enforcement is inconsistent: some companies count internal models, others exclude them.
• Coding and software engineering are described as the most token-hungry workload.
• Email writing is framed as basically negligible in token terms, making some budgeting rules look misguided.
• The market is portrayed as coding-first, with most API revenue driven by software engineering workloads.
• The 99th percentile of companies may spend roughly $100k per employee per year on AI, implying very high ROI at the margin.
• Consumer subscriptions are described as heavily subsidized relative to API usage.
• Enterprise plans intentionally force token-based billing because that is where the margin is.
• Some startups sidestep enterprise plans by buying many consumer subscriptions and charging them to the company card.
• A $200 monthly Codex plan is said to map to roughly $12k of API credits.
• A $200 monthly Anthropic plan is said to map to roughly $8k of API credits.
• Anthropic enterprise subscription plans reportedly include no usage and bill entirely on API pricing.
• Anthropic is portrayed as mostly API revenue and already profitable on an operating basis in Q2.
• The panel says Anthropic could be profitable by more than $1 billion in Q3.
• OpenAI is described as having had a much larger consumer mix, with only a small share of weekly active users paying for a plan in the cited period.
• The cited OpenAI user base is 950 million weekly active users, with 6% paying for a plan.
• Recent OpenAI model releases are described as improving its business mix toward enterprise and API usage.
• OpenAI’s 5.5 and 5.6 releases are framed as inflection points, with net-new ARR now looking closer to Anthropic’s.
• Codex workflow preferences are constrained by practical UX issues like tabs and remote SSH.
• Meta, XAI, and SpaceX-style compute deals are discussed as high-rent rentals with clawback clauses.
• Google is criticized for signing long-term TPU commitments that cannot be clawed back.
• Selling frontier tokens is described as the best business model, though compute rental at above-market rates can also work.
• Meta Compute is framed as a backstop that can justify much higher CapEx if MSL does not work out.
• Revenue per megawatt or gigawatt is rising quickly as models and pricing improve.
• Anthropic revenue per megawatt is cited at about 16 million last year, with more than a 2x increase in three quarters.
• Token-as-a-service is presented as a major channel for enterprise consumption through AWS and Azure.
• The market for token-as-a-service is expected to keep expanding quickly.
• Token spend is now concentrated enough that hyperscaler distribution and existing enterprise relationships matter more than pure model quality.
• The episode pivots from budget controls to a broader tokenomics view centered on revenue per unit of compute and distribution.
• Token-as-a-service providers can help chip startups by supplying engineering expertise for Trainium and TPU serving.
• If a new accelerator reaches scale, frontier labs will likely optimize for it directly rather than rely on an intermediary.
• RL is framed as the most important current scaling law for model capability improvements.
• RL environments are described as a way for models to repeatedly practice white-collar tasks until they learn them.
• Frontier-lab data budgets are said to exceed $10B this year and could grow 10x again in 2027.
• Demand from frontier labs is described as effectively unlimited; the bottleneck is scaling high-quality data creation.
• Anthropic was said to be the most aggressive buyer of coding data from RL-environment startups before 56 SOL.
• Good RL data requires a real engineer, a verifier, and a realistic prompt.
• The highest-quality coding tasks can reportedly cost well over five figures apiece.
• Strong RL-task creators can reportedly make seven figures annually.
• Creating RL data has become harder in the last eight months and now often means targeting a specific model failure mode.
• A practical heuristic is to let the model try a task 10 times; if it is your first attempt, make the task about 10x harder.