AI Industry Revenue and Usage Surge in Mid-2026, but Four Structural Paradoxes Emerge

By mid-2026, the AI industry's primary operating metrics—revenue and usage (token calls)—both sustained rapid growth. However, when breaking down individual companies' financial and business structures, the same set of growth figures masks different business models. Currently, the AI business model faces four structural paradoxes: cost, hierarchy, responsibility, and open-vs-closed source, pointing to the rapid commoditization of intelligence, with profits only staying in a few segments.

When GPT-4 was released in March 2023, the price per million tokens was $30 for input and $60 for output. Three years later, the price for equivalent intelligence has fallen over 95%. In March 2026, global weekly token usage reached 20.4 trillion; China's daily usage exceeded 140 trillion, more than a thousand-fold increase from early 2024. The decline in token unit price did not lead to lower total bills, because the demand structure changed: at sufficiently low prices, many tasks that were previously uneconomical entered the scope of AI processing, and enterprise monthly spending rose from $50,000 to $500,000. Facing bill pressure, the industry shifted from saving unit price to saving tasks: simple tasks are routed to small models, and task budgets are set for agents. Claude Code exited low-priced subscriptions, and Copilot switched to usage-based billing.

Databricks co-founder Ghodsi once described a case: a data connector historically took nine months to develop. The first attempt with a large language model saved only one and a half months, until someone decomposed and re-engineered the process, achieving delivery in one quarter.

Jensen Huang once publicly stated that economic value would eventually be realized at the application layer. But in the summer of 2026, the mainstream view in the primary market was that applications were dead, with capital flowing first to models and infrastructure. Breaking down the industry stack, when the generative AI ecosystem reached an annualized revenue of about $400 billion, the chip layer captured roughly 70% of revenue and nearly 80% of gross profit; the application layer generated about $60 billion in revenue, with gross margins mostly between 0% and 30%, while chip companies like Nvidia achieved gross margins around 70%. The model layer itself is also diverging: companies that focus on enterprise APIs and rely on expanding usage from existing customers can improve inference gross margins from deep losses to around 70%; while companies burdened with massive free users have operating performance far below their revenue headlines. AI products no longer compete on user retention time, but on whether users are willing to entrust an important task to them. General-purpose chatbots and general-purpose agents are the main battlefield for leading model companies, while the direction where startups can still be viable is embedding intelligence into specific scenarios.

Harvey has deep roots in the legal field, accumulating review records and case data that short-term investment cannot replicate.

Anthropic's annualized revenue grew from about $1 billion in December 2024 to over $47 billion by May 2026. OpenAI's official figures: ARR of $2 billion in 2023, $6 billion in 2024, over $20 billion in 2025, and over $25 billion by end of February 2026. Anthropic's inference infrastructure gross margin improved from 38% a year ago to over 70%; an internal document updated by OpenAI in November 2025 pushed its free cash flow breakeven point to around 2030. The two years of fastest cost decline were precisely the two years of most pronounced profit divergence. Intercom launched Fin, charging per resolved issue at $0.99 per transaction, with no charge if unresolved. The heavier the responsibility a vendor is willing to bear for results, the higher the gross margin and the larger the budget. Selling by token competes for IT budgets; selling by results cuts into labor budgets.

Segments with high volume but low responsibility—such as customer service, product descriptions, and voice-to-ticket—are most easily covered by foundational models. Segments that are low-frequency, highly regulated, and have high error costs—such as legal and healthcare—involve heavy delivery and long sales cycles, but what is sold is process, data, and responsibility.

Looking at usage alone, open source has taken the lead. On the OpenRouter platform, the top five models by traffic are all open-weight models; in terms of developer adoption, open source stands at 79% versus 71% for closed source. Chinese models' share on the platform rose from less than 2% a year ago to over 45% in April 2026, with the top ten models accounting for 61% of token usage. Looking only at revenue, closed source has a clear advantage. A survey of Global 2000 enterprise CIOs by a16z shows that enterprise spending on open-source models fell from 19% of total model budget last year to 11% this year, with closed source accounting for 89%, and the average enterprise annual large-model budget is about $7 million. Within the OpenRouter platform, Anthropic captured 46% of dollar share with only 12% of token share. Enterprises choose closed source for reasons including reliability, technical support, compliance assurance, and a clear responsible party when problems arise. Total cost of ownership for open source and closed source is converging, and hybrid deployment is becoming the norm: open source for experimentation and edge cases, closed source for critical paths.

The open-source camp itself does not rely on models for profit—some aim for platform ecosystems, others treat open-source models as customer acquisition channels for cloud businesses. The model layer is rapidly commoditizing, with profit pools migrating to the upstream (computing power) and downstream (orchestration, data, and services). The premium of the closed-source camp is strongly correlated with the capability gap; as open source approaches, the pricing power of per-token billing weakens.