The Path to Profitability - Distribution, Productization, and Frontier Cost Deceleration
My response to Dwarkesh's open question on model profitability
Disclaimer: The views expressed in this piece are solely my own and do not represent those of my employer. I wrote this piece about technology and industry dynamics for educational purposes only, not to promote any specific products or services.
Context:
Back in April Dwarkesh announced a blog prize to inspire deeper thought on big outstanding questions in AI, and the following one caught my eye:
What’s the most plausible story where foundation model companies actually start making money? If you consider each individual model as a company, then its profits may be able to pay back the training cost. But of course, if you don’t train a bigger, more expensive model immediately, then you stop making money after 3 months. So when does the profit start? Maybe at some point scaling will plateau, but if progress at the frontier has slowed down, then the combination of distillation and low switching costs (cloud margins result from high switching costs) makes it really easy for open source to catch up to the labs, eating into their margins. So how do the labs actually start making money?
I spent ~2 weeks answering this question in 1000 words or less. In the month or so since I submitted durable profitability of AI companies has only become more important to consider with rising memory and other input costs.
I hope you can take as much away from reading this as I took away from writing it - enjoy!
My Answer:
The labs will become profitable when their products approach the limits of market demand for training scaling, not the technical limits of data centers. Once the labs reach the S-curve deceleration point, they can maintain user and revenue growth as training investment - the current incinerator of profits - decelerates.
We already see this bending the consumer chatbot market toward profitability. ChatGPT spent 15 months with the same default model (GPT-4o) and roughly tripled subscription ARR (estimated ~$3bn -> ~$10bn). The next default model, GPT-5, decelerated final training run costs versus historical trends. Per Epoch AI, GPT-5 used roughly 2.5x the training compute of GPT-4 - a sharp break from OpenAI’s historical pattern of scaling compute by around 100x with each integer GPT generation.
Why don’t most people expect this?
Narrative Violation 1 - Market Limits, Not Technical Limits:
Belief in scaling laws has blinded most to a basic fact - scaling investments requires proportional reward, and not every market endlessly rewards endless investments.
May 2024 to August 2025 - the aforementioned 15 months - proved this for consumer chatbots. OpenAI lost the top spot on Artificial Analysis to 3 other labs (Google, xAI, DeepSeek) and was widely seen as worse than Claude 3.5 Sonnet on most non-math tasks. Despite this, ChatGPT grew weekly active users from ~180m to ~700m. Competitors’ superior models did not reward their chatbot businesses with comparable growth.
OpenAI did release other models and features during this period. However, ChatGPT still grew users and revenue ~3.5x even though 95% of its users (the free tier) had limited access to better ChatGPT models, and competitors with existing distribution (X, Google) offered superior models.
Cheap open alternatives didn’t slow OpenAI either. DeepSeek R1 briefly beat GPT-4o on Artificial Analysis benchmarks while costing ~80% less at the API level. On January 27 2025, it overtook ChatGPT’s #1 spot on the US Apple App Store. Building a ChatGPT clone with free tools like Streamlit was trivial. Yet, users still didn’t find open or proprietary alternatives better enough to change their habits.
GPT-4.5 was the ultimate proof that the chatbot market had stopped rewarding massive incremental training steps. Despite reportedly using ~10x more training compute than GPT-4, it flopped and was pulled before being promoted to the default model. GPT-5 reportedly used less training compute than GPT-4.5 despite shipping later, and was structured as a prompt router (a decision widely believed to prioritize lowering inference costs over raw scaling performance).
OpenAI is still pursuing massive training investments, however those target the enterprise and coding markets where investment is still being rewarded with growth. Of the 14 model releases since GPT-5, 7 were Codex-only and the 3 ChatGPT-only releases were “-instant” variants (cheaper, faster responses), and only 4 were shared. Anthropic’s Q1 2026 ARR growth signaled that incremental investment in these markets still pays off - which is why both labs see it as rational to incinerate profits chasing them.
With chatbot training costs decelerating, inference costs collapsing, and revenue growing, ChatGPT is on track to be product-level profitable as soon as 2027.
Narrative Violation 2 - Revenue Models:
People also assume labs are pursuing revenue primarily through API consumption. In fact, they appear to be pivoting to use frontier API performance as marketing to pull users into subscription products.
Cost-plus pricing is not the path to riches for information goods - value-based pricing is. Windows does not sell virtualized RAM. It sells the latent potential of computers to people who couldn’t extract it themselves, priced at a fraction of the value delivered. An employee with a working computer is worth far more than a $22/month Microsoft license fee. Microsoft doesn’t price by marking up watts per clock cycle and doesn’t let itself be price-shopped that way. Model API businesses often fall into the cost-plus trap because they produce directly comparable outputs (tokens), and direct competition on tokens drives a race to the bottom.
Anthropic has taken this lesson to heart and is pivoting from selling tokens to selling the Claude Platform bundle. Claude Code started API-only, but Anthropic quickly folded it into the subscription platform (sharing a quota with Claude Chat, Cowork and Design). In April 2026, Anthropic pushed further toward subscription over API: it revamped the Claude Code Desktop app and ran an A/B test restricting CLI access to Max-tier subscribers. The subscription model weans users off thinking in tokens. Anthropic obscures token counts in quotas and changes limits frequently - which makes price-shopping harder.
The Windows analogy only holds if Anthropic can build equivalent IT-level lock in. However, its locked-in chat histories (which Claude can search more easily than users), personalized user profiles/memory and Anthropic-built connectors to enterprise data sources show a conscious push in that direction.
The enterprise and coding markets still reward incremental investment. Codex gained meaningful traction with GPT-5.5 (despite Opus 4.6 dominating the month prior) because users report Codex being better at solving the most complex problems. If Anthropic or OpenAI can retain users when they are no longer #1 on industry benchmarks or being perceived as ‘always having the #1 model’, we’ll know that market is ready to bend toward profitability too.
Conclusion:
As long as labs are investing in models that unlock new capabilities end users will pay for, they’ll remain unprofitable, and rationally so, in pursuit of growth.
We can’t know when the market will stop rewarding incremental training for coding and enterprise before we get there - OpenAI themselves could not predict it for the chatbot market until GPT-4.5 overshot demand. Nor can we predict if and when new markets will emerge that demand even more incremental training.
But the chatbot market shows profitability can come as soon as 3 years after that point - likely faster as margins improve. SemiAnalysis estimates Anthropic’s inference margins rose from 38% to 70% over the past 12 months, and there is no sign of inference costs improvements slowing down.