One AI Lab Turned a Profit. The Other Lost $12 Billion.

Two AI labs closed their books this month, and the numbers read like two different industries. Anthropic posted its first operating profit: $559 million on $11.6 billion in second-quarter revenue. OpenAI took in $6.7 billion and lost $12.3 billion running the business ^[1]. Same technology, same year, opposite directions. The split isn't about whose models are better. It's about which business each lab chose to run. OpenAI has spent the year stripping itself down for a public offering. It shut down Sora, the video generator that was losing $15 million a day, and cancelled a $1 billion deal with Disney ^[2]. Fidji Simo, who runs its applications, put the reasoning plainly.

We cannot miss this moment because we are distracted by side quests. — Fidji Simo

Ads are now running in 38 countries, described by the company as revenue diversification ahead of a planned 2027 IPO ^[3]. It cut its Luna model's fees by 80% ^[4]. Every move points the same direction: get the revenue lines looking broad and steady before the roadshow. Anthropic took the other road. It sold to enterprises, kept Amazon and Google each under a 33% ownership cap to preserve its independence ^[5][6], and now claims a $30 trillion addressable market to justify what is reported to be a $2 trillion offering ^[7]. The discipline shows in the margin. The consumer experiments OpenAI is now killing were never Anthropic's to begin with. Both are sprinting toward the same window, and the window has a problem. MIT's NANDA initiative found that 95% of enterprise generative AI pilots show no measurable profit-and-loss impact on $30-40 billion in spending ^[8]. Prices are collapsing across the board: Anthropic halved a high-performance model's price, xAI and Meta followed, and Chinese rivals undercut them all at roughly a tenth of the cost ^[9][4]. The customers who are supposed to justify the valuations keep failing to show a return. The supply side is still expanding, which is exactly why the demand-side gap is dangerous. Nvidia projects hyperscaler capital spending will hit $1.3 trillion in 2027 ^[10], and Morgan Stanley still models 20-60% returns on the infrastructure ^[11]. The money keeps flowing in while the buyers keep failing to show a profit — a gap that can hold only as long as nobody looks at it too closely. Anthropic's own analysts put the timing in terms that amount to an admission: the offering must catch the market's current enthusiasm before doubts about cost and scalability set in ^[7]. One lab is running toward that window in the black. The other is bleeding $12 billion a year while it sprints.