Ep. 020 - Anthropic vs OpenAI Usage, Margins, Meta Compute, Future of MSL (Tokenomics)

Summary

This episode is a detailed discussion of AI token budgeting, enterprise access policies, and the economics of subscription versus API pricing. The hosts argue that most companies are overreacting to token consumption, because only a small set of power users—especially engineers doing coding and software work—drive meaningful spend, while email and other lightweight tasks are essentially negligible in token terms. They also describe how firms route around budget caps by using cheaper models for preprocessing, or by exploiting inconsistencies in whether internal models are counted. The second half shifts to company strategy and margins: Anthropic is presented as API-heavy and highly profitable, OpenAI as still burdened by a large free consumer base, and Meta as potentially able to monetize spare compute through neocloud-style or clawback-based deals. The overarching takeaway is that the winners will be the companies that control frontier inference, enterprise distribution, and compute flexibility—not those that merely maximize consumer usage.

Notes

• The episode is anchored in a point-in-time discussion of token budgeting, Meta compute strategy, MSL futures, product releases, and Anthropic profit margins.
• Enterprises are cracking down on AI budgets because they think employees are spending too many tokens.
• Most employees are not near their token limits; the restriction mainly hits a few power users.
• Per-company monthly budgets are favored over per-person caps because usage varies widely across employees.
• Smaller firms conserve expensive tokens by running cheap models first, then sending condensed prompts to frontier models.
• Budget accounting is inconsistent across firms, including whether internal models count toward the cap.
• Uncapped or uncounted systems tend to get heavily overused.
• Coding and software engineering are described as the most token-hungry workloads.
• The hosts think many token-budget rules are poorly targeted because they ignore where the spend really comes from.
• Email generation is treated as token-cheap and a bad target for strict model restrictions.
• Some firms restrict advanced coding tools to engineering teams only, which the hosts call shortsighted.
• The hosts argue employees should broadly get access to AI tools to learn where they are most productive.
• A cited comparison says a $200/month plan can map to roughly $12K of API credits for Codex and about $8K for Anthropic.
• The subscription-versus-API question is framed as a subsidy problem: subscriptions are cheap relative to usage, while API billing captures margin.
• Startups sometimes buy multiple individual subscriptions on a company card instead of using enterprise API billing.
• The hosts estimate Claude Max 20x break-even utilization at about 10%.
• They estimate break-even utilization at 20% for Pro/5x, 11.4% for OpenAI Plus/Pro 5x, and 5.7% for OpenAI Pro 20x.
• Average utilization is thought to exceed break-even, especially on 20x plans, implying negative-margin subscriptions.
• The hosts say subscription businesses want to push as much volume as possible toward API billing.
• Anthropic enterprise subscriptions are described as having no usage allowance; usage is metered at API pricing.
• Anthropic is described as having more than 80% of ARR on the API side.
• The hosts say Anthropic is already profitable on an operating basis in Q2 and could exceed $1B of operating profit in Q3.
• Anthropic’s gross-profit advantage is being reinvested into training to widen the model-capability lead.
• The hosts explicitly argue against maximizing EBITDA too quickly and in favor of reinvesting in training.
• A large API customer can dominate the revenue mix even if many users are on subscriptions.
• Anthropic is framed as historically enterprise-heavy, with more than 90% enterprise mix.
• OpenAI is described as starting consumer-heavy, with about 60% of ARR in consumer in Q1.
• Only 6% of OpenAI’s 950 million weekly active users are said to pay for a plan.
• OpenAI’s free-user burden is said to lower blended gross margins by about 20 points.
• OpenAI’s mix is said to have flipped in Q2 toward more enterprise, helped by API growth since late March.
• The hosts say OpenAI’s net-new ARR is starting to look more comparable to Anthropic’s.
• There is still a very large free-user base at OpenAI, giving room for monetization expansion.
• The hosts suspect most paying OpenAI users are using Codex rather than the core paid product.
• Anthropic consumer conversion is described as around 9% of free users paying.
• Consumer AI use is likened to Google search replacement behavior.
• Ads are treated as a hard-to-scale monetization model because retention curves are poor.
• The hosts say OpenAI may have ambitious 2030 advertising revenue targets, but ads are still a tough business model.
• Enterprise API monetization is presented as the better-margin business.
• Anthropic is said to have an early lead because it was heavily leveraged to decoding in the first half of the year.
• The market is characterized as moving toward a two-horse race in mid-July.
• Consumer adoption is spilling into enterprise usage, with prior Claude Code power users now testing Codex.
• OpenAI looked weak earlier in the year, with ARR growth likely flat in March and April.
• GPT-5.5 is described as a major inflection point for OpenAI.
• GPT-5.6 is described as roughly as good as Fable while costing about half as much.
• OpenAI’s net-new ARR is now said to be comparable to Anthropic’s.
• The speaker says they are basically defaulting to OpenAI now for the tasks they care about.
• The Codex app is criticized for awkward multi-tab workflow and poor remote SSH support.
• The speaker says the target market for Codex is vibe coders and B2B SaaS builders.
• The conversation shifts to who will come in third, including xAI and Meta compute.
• Frontier-model quality is treated as the main condition for API relevance.
• If a model is frontier, the hosts expect its token-as-a-service margins to be very strong.
• Grok 4.5 and Spark 1.1 are treated as scaling-curve proof points rather than frontier systems.
• The speaker says they will move their token usage to Grok 4.5 or 1.1.
• Gemini is described as clearly in fifth place.
• Model quality is tied directly to compute availability in the hosts’ framing.
• Elon’s strategy is framed as renting out excess compute at 3x-4x market rates with 90-day clawbacks.
• The speakers contrast inference monetization with short-term compute deals.
• Meta compute is presented as a way to emulate SpaceX-style economics with clawback optionality.
• The speakers argue Meta can buy more compute now if it retains the ability to claw it back later.
• Google is criticized for signing long-term TPU deals that cannot be clawed back.
• Token-as-a-service is described as a very popular enterprise consumption model.
• AWS and Azure benefit from existing enterprise distribution and procurement relationships.
• The speaker estimates Anthropic’s indirect token-as-a-service share rose from about 5% six months ago to about 20% now.
• The hosts think financial services still has a lot of unrealized AI adoption potential.
• Together, Fireworks, and Base10 are expected to grow, but frontier token volumes are expected to grow faster.
• The startup thesis is that capturing even a tiny slice of a massive inference market can still be a strong business.
• Anthropic revenue per megawatt is described as more than doubling in three quarters.
• The speaker expects revenue per megawatt to double again within two to three quarters.
• OpenAI API success is expected to lift Microsoft Foundry estimates.