TheChek News Flow

The Takeaway

OpenAI, recently valued at $500 billion in an employee share sale, has projected it will generate $13 billion in revenue this year—nearly triple Anthropic’s most optimistic forecast of $4.7 billion. But OpenAI will burn more than triple as much cash as Anthropic this year due to its massive computing costs for developing new AI, or research and development compute.

In 2027, OpenAI’s cash burn will amount to about $35 billion versus free cash flow of $3 billion for Anthropic that year, according to forecasts from both companies.

In 2028, Anthropic projects it could generate as much as $17 billion in cash, compared to the nearly $47 billion of cash burn OpenAI has projected. (In a more achievable scenario, Anthropic projects about $3.6 billion in free cash flow in 2028 on $32.5 billion in revenue.)

Anthropic, started by ex-OpenAI senior executives, has focused on selling its Claude AI models through an API to business customers, such as Anysphere, maker of Cursor, and Harvey, a legal AI startup. It also sells subscriptions priced between $17 to $150 per month for its premium features, including its coding copilot Claude Code.

That’s a narrower scope than that of OpenAI, whose ambitions range from advertising to e-commerce to hardware AI devices.

While OpenAI has projected that revenue from ChatGPT subscriptions will power much of its growth in the coming years, Anthropic leaders believe the API will continue to be the company’s revenue workhorse. Its projected API and related revenue of around $3.8 billion this year compares to the $1.8 billion OpenAI earlier projected for API sales this year.

Microsoft also generates revenue from selling OpenAI models through an API to its Azure cloud customers, and it shares a small percentage of that revenue with OpenAI.

The company forecasts that sales to businesses through its API and related apps will continue to generate more than 80% of its revenue through 2028. Business customers, which Anthropic primarily focuses on, tend to be less likely to churn and more likely to renew or even expand their spending, investors in software startups say. But sales of its nine-month-old coding assistant, Claude Code, will propel sales to individual developers.

Claude Code is close to generating annualized revenue of $1 billion, up from roughly $400 million in July, according to one of the people. Anthropic’s annualized revenue, or the last month’s revenue multiplied by 12, was nearing $7 billion last month.

Anthropic raised its most optimistic revenue forecasts for the year by roughly 26% to $4.7 billion from a projection it gave investors ahead of its March round. It lifted its 2026 forecasts 28% to $15.2 billion and its 2027 forecasts 13% to $38.9 billion. Anthropic ended last year at roughly $381 million in revenue.

The updated revenue projections are great news for Google and Amazon, which rent out specialized servers to Anthropic.

Anthropic expects its gross profit margin, which measures how much revenue it makes compared to the cost of producing that revenue—largely from running servers—to swing from negative 94% last year to as much as 50% this year and 77% in 2028.

However, Anthropic’s calculation only accounts for the costs of running its AI models for its paying users. If it were to include the cost of running its models for nonpaying users, as OpenAI does, Anthropic’s gross margins would be lower: negative 109% last year, 47% this year and 75% in 2028 in its most optimistic projections. The most efficient enterprise software startups typically have gross margins of 70% or higher.

Solo VC investor extraordinaire Elad Gil said onstage at TechCrunch Disrupt 2025 that AI has been one of the least predictable tech booms he’s ever seen.

Gil is on the cap table of virtually every hit company of the past decade, including many of today’s leading AI companies.

Still, he thinks that over the last year, certain AI markets appear to be nearly sewn up by market leaders. Beyond these areas, a vast swath of AI remains anyone’s game.

“I started investing in generative AI in 2021 … [A]t the time, not very many people were paying that much attention to it,” Gil said. But he had seen the massive leap in capability between GPT-2, launched in 2019, and GPT-3, launched in 2021. “The step between 2 and 3 was so large that if you just extrapolated out the scaling laws, or the curve, then you could really assume that this was going to be incredibly important,” he said.

That convinced him to start backing early-stage startups building products powered by large language models. His bets included both foundational model makers like OpenAI and Mistral, as well as application companies like Perplexity, Harvey, Character.ai, Decagon, and Abridge. Yet throughout 2024 and much of 2025, the capabilities of foundational models leaped with every release, upending AI every few months.

“I used to say at the time that AI was the one market where the more I learn, the less I know. Usually, the more you learn about something, the better you know it, the easier you can predict the future, etc. But AI was just hazy. There’s just too much uncertainty. And I think there’s still markets like that in AI,” he said.

However, he’s also now seeing markets with clear winners. The most obvious example is with foundational models themselves. Even though hundreds of models exist, and some countries like South Korea are still working to develop sovereign models by local companies, leaders have emerged. “Google, Anthropic, OpenAI, maybe xAI, maybe Meta, maybe Mistral — it’s like a handful,” he predicts of the winners.

After models, he thinks AI-assisted coding has runaway winners that will make it hard for new entrants to catch up. Not only have the foundational model makers moved in (Anthropic with Claude Code, OpenAI with Codex) but also startup leaders like Anysphere’s Cursor and Cognition’s Devin (which acquired Windsurf) will be hard to beat. And there are well-funded startups like Magic (which Gil called a possible “outlier”) and Poolside on their tails.

He sees medical transcription as being cornered, with Abridge a front-runner and a handful of others like Ambience being “important.”

He names customer support — which was an early target of both traditional AI and the new crop of AI agent startups — as having hard-to-catch market leaders, such as his portfolio company Decagon. (It raised $131 million at a $1.5 billion valuation in June.) OpenAI chairman Bret Taylor’s startup, Sierra, competes in this space. This is also an area where the incumbents — Salesforce, HubSpot, and many others — are adding AI offerings.

So which markets seem wide open? Gil says financial tooling (fintech), accounting, AI security, and “other markets that we know are by default very interesting. We just don’t know who’s going to do it.”

Ironically, fast growth isn’t the signal it once was that a company is going to be a breakout hit. “The CEOs of every big company are basically telling their teams, hey, we have an edict. We need to figure out our AI strategy,” Gil said. “These giant enterprises are willing to try things that two years ago they never would have tried, and it’s only because of AI.”

So new AI markets can land a lot of revenue from big-name, enterprise customers quickly, “but that doesn’t mean they’re going to stick,” Gil points out.

It is only after a market goes through its trial-phase boom cycle that a startup and investors can see if this revenue will stay and grow. “There’s false signal, and then there’s stuff that is just working,” Gil said. He calls out legal AI startup Harvey as one of the market leaders that’s “just working.” It raised three massive rounds in 2025, leaping from a $3 billion valuation to $5 billion to $8 billion, in just a few months.

Gapping up
In reaction to earnings/guidance:

Other news:

Gapping down
In reaction to earnings/guidance:

Other news: