Anthropic and OpenAI take 89% of AI startup revenue

Anthropic and OpenAI captured 89 per cent of revenue generated by 34 leading AI startups, according to The Information’s latest revenue snapshot. The 34 companies tracked are now producing nearly $US80 billion in annualised revenue. Scale is coming quickly but not evenly. For all the noise around hundreds of model labs, copilots and agent startups, most of the money keeps flowing to two companies that already dominate attention, capital raising and enterprise mindshare. Gossip column aside, that matters. Revenue is concentrating this early — and if the pattern holds, the fight for the next layer of value in AI may be less about who trains the best general model and more about who can build a durable business without being trapped underneath the market’s two largest platforms.

Look past the hype and the concentration figure tells a sharper story. Generative AI still looks crowded from the outside, but the inside resembles a classic infrastructure market: extreme upfront costs, heavy dependence on cloud partners and a strong tendency for customers to gravitate toward a small set of perceived safe bets. The Information reported that the 34-company cohort lifted combined annualised revenue by 112 per cent in six months. Rapid growth usually creates room for challengers. Instead it is widening the gap between the leaders and everybody else. That is a different signal from the old SaaS playbook, where adjacent startups carved out healthy niches before a handful of firms consolidated the category.

Part of the dynamic is mechanical. Training frontier models costs billions. Inference remains capital-intensive. Enterprise buyers prefer signing with vendors they think will still be standing in three years. Reuters reported in March that OpenAI had crossed $US25 billion in annualised revenue — a scale few software companies reach this early in a platform shift. A later Reuters story said OpenAI was courting private equity groups to support an enterprise AI venture, a sign it wants to reach beyond model access and deeper into deployment. Anthropic has followed a similar enterprise track, leaning on large model contracts and strategic backing. Once the cycle starts turning, scale attracts more scale: more revenue buys more compute, better compute delivers better service levels, better service levels help close larger contracts.

The rest of the field is not irrelevant. But many smaller AI startups are working in a narrower lane than fundraising decks imply. They build useful products while buying core capability from the same firms they are meant to disrupt. Brookings warned that AI platform companies can end up competing with the customers building on top of them, especially when the platform owner can see which use cases gain traction and then fold similar features into its own stack. In ordinary cloud software that risk exists. In generative AI, where the model provider may control both the core technology and the price of inference, the squeeze is tighter.

For smaller builders, that is the real warning.

Strip away the model-race framing and the concentration story is about something simpler: which part of the stack is actually investable for everyone else. Application startups can still matter. Regulated sectors, vertical workflows and distribution-heavy niches where local trust counts for more than raw model performance — those are real moats. But the business model has to survive what comes next. The underlying model gets cheaper. A flagship feature gets copied. The platform owner bundles a similar tool into a broader enterprise contract. Reuters Breakingviews noted in March that the leading groups still face heavy cost and financing pressures of their own. Concentration at the revenue line does not guarantee durable profits. What it does show is where customer spend is landing right now.

For Australian founders and investors, the framing flips. The obvious temptation in any AI boom is to back the company with the strongest benchmark scores or the slickest assistant. The harder question — the one that matters over a five-year horizon — is whether a local startup can own enough workflow, data, distribution or compliance expertise to stay valuable if Anthropic or OpenAI cuts prices, launches a competing feature or signs directly with the enterprise customer. Australian software buyers are running the same calculation in reverse. If the market is consolidating around a couple of model vendors, procurement teams need to think seriously about portability, data governance and how much of their future IT architecture gets locked into one provider’s commercial roadmap.

The Information’s sample covers leading private AI startups, not the whole software economy. Revenue share is not the same as long-term market power. The application layer remains busy, and some of the best businesses in AI may sit above the foundation model rather than beside it. Still, an 89 per cent share for two companies is a striking number this early in the cycle. It suggests the generative AI market is not maturing like an open bazaar where dozens of players steadily find their lane. It looks more like a platform market where the top vendors are pulling capital, customers and developer dependence toward themselves at the same time.

The revenue snapshot does not settle whether Google or another deep-pocketed rival can still redraw the leaderboard, and it says little about which applications will capture the most end-user loyalty. What it does show is that the first big wave of generative AI spending is already clustering around a tiny number of firms. For startups building on top of that layer, and for Australian enterprises deciding where to place long-term bets, the practical question has shifted. It is no longer whether AI will create a large software market. It is whether enough of that market will be left for anyone outside the dominant model platforms.