[Adobe Stock]
Anthropic has been “very open from day one” about how capital intensive AI development is, said Daniela Amodei, the company’s president and co-founder in a recent interview with Bloomberg. “It’s a really big upfront cost to train the models and to serve inference on them,” she said. In fact, Goldman Sachs estimates that cumulative AI capex from 2026 to 2031 will reach around $7.6 trillion across compute, data centers and power. “As a scale comparison, that is equivalent to almost one-quarter of annual U.S. GDP and roughly 1.4 times Germany’s annual GDP.”
“My guess is that over time, the core set of companies that are working to advance the frontier are just going to need access to capital,” Amodei said. “The public market is very well suited to that.”
Floating an AI pause while preparing for quarterly revenue disclosure
Public markets also want to see quarterly revenue conversion and pressure to turn costly compute into revenue. To that end, Anthropic and its peers have been tightening usage limits and pushing more agentic work toward metered economics. Token-based billing is meant to provide a more predictable revenue stream. The methodology is facing something of a revolt. Uber apparently used up an entire year of its AI budget in four months, and is now capping AI use at $1,500 per employee per month. A senior Uber executive questioned whether rising token usage was producing features customers actually valued. Walmart has also instituted similar caps on AI spending. Meanwhile, Nvidia’s Vice President of Applied Deep Learning Bryan Catanzaro declared: “For my team, the cost of compute is far beyond the costs of the employees.”
Microsoft AI chief Mustafa Suleyman called Anthropic “extremely expensive”, said many people were urgently looking for alternatives. He added that Microsoft’s goal was to reduce and ultimately eliminate what it pays Anthropic. A few weeks ago, the Microsoft executive Michael Wetter shared that the company had invested more than $100 billion in its OpenAI alliance when factoring in operational costs.
Agent competition heats up
Meanwhile, the competitive landscape is heading up. Meta would start charging for Meta Business Agent across WhatsApp, Messenger and Instagram, with larger businesses paying based on token consumption, Bloomberg reported on June 3.
When GitHub Copilot moved from flat-rate subscriptions to token-based billing, TechCrunch reported that some users claimed their projected monthly costs jumped from $29 to nearly $750, or from about $50 to roughly $3,000.
Gartner projects that inference on a 1-trillion-parameter LLM will cost GenAI providers more than 90% less by 2030 than it did in 2025. Providers are expected to retain much of that saving, and agentic models require 5 to 30 times more tokens per task than standard chatbots.
While measuring volume of AI spending, which was later dubbed ‘tokenmaxxing,’ was trendy just a couple of months ago, critics say the practice is essentially done for. The aggressive spending appears to have played a role in pumping up Anthropic’s revenue. Anthropic has told investors it expects to report its first profitable quarter in June 2026, as revenue growth starts to outpace its steep compute costs, WSJ explains in an article teasing “mind-blowing growth” at the company. Its projected operating profit is around $559 million on expected Q2 revenue of $10.9 billion, more than doubling the $4.8 billion generated in Q1. On the top line, run-rate revenue reached roughly $47 billion as of late May 2026, up from $9 billion at the end of 2025, driven primarily by enterprise adoption and Claude Code.
In Anthropic’s post floating the idea of a pause on AI development, and the prospect of recursively self-improving models, the company also notes that “Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025.” It also said more than 80% of the code merged into Anthropic’s codebase as of May 2026 was authored by Claude.
Anthropic’s Mythos will likely come at a premium
The metric Anthropic is celebrating resemble the coding mode some of its customers have stopped trusting, i.e. measuring volume over quality. The post mentioning the metric acknowledges that it is “almost certainly an overstatement of the true productivity gain.” While Uber struggled to find the connection between more AI-generated code and more value for users, Anthropic’s long-term profitability will rely on finding a firm connection between tokens sold and value obtained.
Meanwhile, Claude Mythos, one of the most hyped models in recent memory, which is only available now as a gated release to members of the consortium Project Glasswing, remains subsidized. “Anthropic is committing up to $100M in usage credits and $4M in donations to open-source security organizations to support this work,” the company noted when announcing the gated launch of the model. Incidentally, the model is available to consortium members at five times as much as Opus, at $25/$125 per million input/output tokens.
