🎯🚀 Why OpenAI just torched their own API business

In partnership with

Preface

My article on Pricepoint last week was, BY FAR, the most popular thing I’ve ever written. It required building up sources at Pricepoint’s clients over months but it’s clear you want to hear more inside scoops, on Pricepoint and potentially others.

I’ll be working more and more to cover how they work.

Main Article

Last week, OpenAI's CFO mentioned that the company thinks about becoming a mini-hyperscaler like AWS "as a business down the line." A few days later, they released GPT-5 at $10 per million tokens, roughly 10x cheaper than Anthropic's flagship model while matching its performance benchmarks.

Both moves make more sense when you understand how differently these companies make money, and what OpenAI can afford to do that Anthropic cannot.

How Anthropic Built a Profitable API Business

For about 18-24 months, Anthropic consistently had the best coding model available. This matters because AI coding became one of the most successful applications of AI, with fast revenue growth and real adoption among software engineers. The result was that coding-focused applications like Cursor, Windsurf, Bolt, Lovable, and Replit all built their products around Claude, becoming the clear market leader.

Unlike most AI companies burning money on infrastructure, Anthropic found actual profitability. They kept pricing steady at $75 per million tokens for their premium model and $15 per million for their standard offering. As Anthropic's CEO told Alex Kantrowitz, "we make improvements all the time that make the models, like, 50% more efficient than they are before... for every dollar the model makes, it costs a certain amount. that is actually already fairly profitable."

But about 85-90% of Anthropic's revenue comes from their API business. When your entire business depends on one market, you can't afford to sacrifice margins in that market, even temporarily. OpenAI generates roughly $13 billion in annual recurring revenue, with about 75% coming from consumer subscriptions ($9.75 billion), 20% from API services ($2.6 billion), and 5% from other sources. This revenue mix creates entirely different strategic options. OpenAI can lose money on API pricing indefinitely because most of their revenue comes from elsewhere. Anthropic cannot.

The Advertising Revenue Engine

OpenAI's recent hiring decisions reveal where they're headed next. In December, they brought on Kate Rouch as their first CMO - she ran ads at Meta for 11 years and built Instagram's advertising system.

The advertising market generates $334 billion annually, with Google's ads business doing $207 billion and Meta another $127 billion. These companies have learned something important about competition: anything that reduces user engagement needs to be eliminated or made cheaper.

Google built Chrome and gave it away for free because every Chrome user funnels into their $207 billion advertising machine. Facebook paid Samsung and other manufacturers to preinstall Facebook apps, then subsidized phone distribution across Africa with cheap devices and wifi because cheaper phones and connectivity meant more Facebook users, which meant more advertising revenue. Microsoft didn't build PCs - they made PC hardware so cheap that Compaq and Dell competed on razor-thin margins while Microsoft collected Windows licenses from every sale.

The pattern is always the same: pick one profit center with substantial margins, then burn everything adjacent to the ground to protect and expand that profit center. If OpenAI launches advertising - and the Kate Rouch hire strongly suggests they will - they join Google and Meta in having powerful economic incentives to make AI inference as cheap as possible. More AI usage equals more opportunities to serve ads.

Commoditizing the Competition

OpenAI's $10 per million token pricing for GPT-5 flipped the entire coding ecosystem. Cursor, Bolt, Lovable, Devin, and Windsurf all became OpenAI launch partners overnight. When equivalent capabilities suddenly cost 90% less, switching becomes necessary for business model reasons, not just cost savings.

But there's a deeper technical dynamic that makes Anthropic's position even more precarious. Model distillation, where you train smaller, cheaper models to mimic the outputs of larger, expensive ones, keeps improving rapidly. This means that even if Anthropic launches a breakthrough coding model, competitors can feasibly create "good enough" versions within 6-9 months at much lower cost. As AI models make more incremental rather than breakthrough improvements, maintaining a significant quality advantage becomes harder to sustain.

Anthropic's coding edge was genuine and sustained, but OpenAI doesn't need to match their quality perfectly, they just need to achieve similar performance at dramatically lower cost. When the economic advantage becomes this large, minor quality differences might become irrelevant. This also explains OpenAI's apparent interest in optimizing for inference rather than training infrastructure. If you plan to be a major inference provider by volume, you want the entire hardware ecosystem optimizing for cost-effective inference chips rather than premium training infrastructure.

The Infrastructure Endgame Funded by Advertising

Once OpenAI uses advertising revenue to subsidize their way back to market leadership in AI inference, they can start building the operational complexity services that command higher margins: fine-tuning, model versioning, enterprise compliance, monitoring, analytics, and all the "AWS layer" services that enterprises pay substantial markups to avoid thinking about. The basic inference becomes commodity-priced, but the operational complexity around it generates sustainable margins.

This is where their plans to become a "mini-hyperscaler like AWS" make sense. AWS doesn't make most of its money on raw compute - that's commodity-priced. They make money on load balancing, auto-scaling, security compliance, and hundreds of other operational services that enterprises pay premium margins for. OpenAI can apply the same playbook to AI infrastructure, but only after they've used advertising revenue to eliminate the pricing competition that would make this impossible.

The sequencing matters tremendously. You can't build a hyperscaler business while competitors are charging premium margins that make customers shop around. First you use cross-subsidization to recapture market share and eliminate pricing competition, then you rebuild margins through operational complexity services while competitors can no longer afford to match your infrastructure pricing.

Why Revenue Diversification Wins

We've seen this competitive dynamic before across the tech industry. Companies with advertising revenue streams can subsidize adjacent markets indefinitely because cheaper access to those markets drives more usage, which creates more advertising opportunities. Pure-play companies in those adjacent markets need margins to survive, but they're competing against businesses that can afford to lose money forever.

The advertising giants understand that their $334 billion profit pool can fund the commoditization of almost any adjacent market. Google subsidizes browsers, email, cloud services, and mobile operating systems. Facebook subsidized mobile infrastructure across developing markets. Microsoft used software revenue to commoditize PC hardware. OpenAI appears to be using current consumer revenue and future advertising revenue to commoditize AI inference, making it economically impossible for companies like Anthropic to maintain their business models.

What This Means for Everyone Else

For application companies like Cursor, Bolt, and Lovable, this is actually excellent news in the short term. Cheaper inference means they can build much more AI-intensive products, experiment with multi-step agent workflows, and potentially achieve gross margin positive unit economics for the first time. While some observers worry these companies lose differentiation when their core AI capabilities become commodity inputs, there's substantial complexity around evals, fine-tuning, prompt engineering, and multi-step agents that aren’t easily replicated. Brand power and execution advantages also matter more when the underlying technology becomes commoditized.

The hardware ecosystem will likely split into two markets: training infrastructure remains premium and specialized (still dominated by NVIDIA), while inference infrastructure becomes a cost optimization race where every chip company not named NVIDIA probably benefits from increased competition and volume.

For Anthropic, the path forward requires moving up the stack quickly into productivity applications, systems of record, or specialized enterprise workflows where their technical capabilities can translate into sustainable competitive advantages that aren't directly exposed to inference pricing competition.

Enterprise software companies should pay attention to the emerging opportunities in building operational complexity layers for AI in specific verticals. As basic AI capabilities become commoditized, value will accrue to companies that can handle the integration, compliance, monitoring, and workflow complexity that enterprises will pay premium margins to avoid managing themselves.

The Broader Competitive Dynamic

What's happening with OpenAI and Anthropic resembles how web search and social media became dominated by advertising-funded platforms. The companies with the largest advertising businesses have both the economic incentive and financial resources to commoditize adjacent markets that might reduce user engagement or increase user acquisition costs.

OpenAI's strategy positions them to be both the low-cost infrastructure provider and the premium operational service layer, using advertising revenue to fund the transition between these roles while competitors with less diversified revenue streams cannot afford to match their pricing indefinitely.

Whether Anthropic can successfully move up-stack, whether other AI companies can find sustainable business models outside of advertising-funded commoditization, and whether this market concentration benefits innovation and competition remains to be seen. But the economic logic is straightforward: in technology markets, companies with the most diversified revenue streams can afford to lose money in markets where their competitors need to make money, and that asymmetry usually determines long-term market structure.

Big investors are buying this “unlisted” stock

When the founder who sold his last company to Zillow for $120M starts a new venture, people notice. That’s why the same VCs who backed Uber, Venmo, and eBay also invested in Pacaso.

Disrupting the real estate industry once again, Pacaso’s streamlined platform offers co-ownership of premier properties, revamping the $1.3T vacation home market.

And it works. By handing keys to 2,000+ happy homeowners, Pacaso has already made $110M+ in gross profits in their operating history.

Now, after 41% YoY gross profit growth last year alone, they recently reserved the Nasdaq ticker PCSO.

Invest for $2.90/Share

_{Paid advertisement for Pacaso’s Regulation A offering. Read the offering circular at invest.pacaso.com. Reserving a ticker symbol is not a guarantee that the company will go public. Listing on the NASDAQ is subject to approvals.}