MAIN CONTENT
A revenue gulf wider than the technical divide
According to Reuters Breakingviews, Anthropic generated $16.3 billion in revenue during the six months ending June 2026, compared with about $1.2 billion a year earlier. That sum was 63 times Z.AI and MiniMax’s combined sales over the same period. The comparison underlines why model rankings alone do not measure the commercial strength of an AI developer.
Anthropic draws considerable support from enterprises, mainly in the United States, willing to pay for coding, analysis and workflow automation. Some customers spend as much as $10 million annually on its Claude models. Integration into recurring business processes can support durable demand and make revenues more predictable.
Compute is expensive; efficiency matters more
Its October 26 IPO filing showed that compute and infrastructure costs rose from roughly $2.3 billion to $8.9 billion year on year. Revenue, however, grew faster than that expense.
Anthropic recorded around $1.6 billion in adjusted operating income, while still reporting an operating loss of approximately $693 million under conventional accounting.
That distinction matters: adjusted income excludes specified costs and does not mean the entire business has become profitable. AP also reported a very large net loss, mainly shaped by noncash accounting adjustments linked to financing instruments. Investors need to separate the economics of serving AI customers from exceptional financial-accounting effects.
Chinese laboratories pursue a different route
Z.AI and MiniMax are building increasingly competitive models while emphasizing affordable access. Some models can be downloaded without charge, with income coming mainly from APIs, customization and enterprise services. Reuters Breakingviews estimates Z.AI’s annualized first-half revenue at about $284 million, close to the scale Anthropic reached in 2024.
But being at a similar point in a revenue timeline does not guarantee an identical future trajectory. US restrictions on advanced chip exports complicate model training in China, while fierce local competition and slower enterprise adoption constrain prices.
Z.AI’s GLM 5.2 is viewed as a frontier-class model, yet its average usage price is materially below Claude Opus 5.5, according to Artificial Analysis data cited by Reuters Breakingviews. Lower prices broaden access, but developers must show that usage volumes can cover research and delivery costs.
Open-model opportunities and the market test
Chinese companies could differentiate through open models, Chinese-language services, industrial applications and private customer deployments. Cost advantages may accelerate adoption among smaller businesses, although an extensive user base should not be confused with profit.
For investors, useful questions include the cost of delivering each request, customer retention and the proportion of repeat contractual revenue. The market value of an industry leader can inspire competitors, but also raises the standard for transparent reporting on cash flows, infrastructure and customer concentration.
Anthropic’s filing thus sends two messages at once: customers will pay for genuinely useful AI, and the path to that result is neither short nor cheap nor easily replicated across economies.
















