The strangeness of token revenue

The first half of 2026 belonged to Anthropic. Their revenue went vertical ($9B at the end of 2025 to $47B in May). The company raised back-to-back financing rounds, each larger than basically every IPO ever, at unimaginable valuations. And they just filed confidentially to go public. For late-stage investors, interest in Anthropic has drowned out interest in pretty much any other private offering out there. With growth like that, why put a dollar in anything other than Anthropic? What could possibly get a better return?

As an inception stage investor, I’m not concerned with whether or not Anthropic is actually worth $1T. That’s not my market. But I'm very interested in that astounding revenue curve, because token revenue is stranger than it looks, and that strangeness might just be where the opportunities for startups live.

Token revenue borrows from SaaS, because access is often packaged as a subscription. And it borrows from cloud, because usage is metered programmatically and priced per unit. But it is not quite either one. And I think that difference comes down to a single fact: not all tokens ought to be valued the same.

The token revenue equation

One way to think about the revenue of any token-seller is a simple model: 

revenue = tasks × tokens per task × price per token

The problem is that tokens are not homogeneous. Some tokens are for exploration. Some are for execution. Some are just waste. And each should be valued differentially.

Here’s what I mean: 

Exploration tokens are for research and development. You're using the latest frontier model to explore whether or not a task is doable at all. This is premium-priced token revenue, but dependent on the lab constantly opening up new domains. 

Execution tokens are the opposite. They are what you spend once you know the task works and you want to run it in production. This is durable, usage-based revenue, but it’s also highly competitive.

Waste tokens are the flip side of the exploration coin. When intelligence gets cheap enough to waste, people waste it. Agents read emails and draft replies that no one will send. They write code no one will ever deploy. Some of this is absurd, and will be seen as such. But a lot of this is just the tax you pay when you don’t yet know what is possible.

The tension here is obvious. Exploration token revenue is what pays the labs’ bills, but exploration tokens are always decaying into cheap, highly contested execution tokens. That’s a scary proposition for the labs.

What somebody at one of the labs might argue, however, is that nobody actually buys tokens. They buy completed work. And that’s why the frontier premium exists. 

A task that burns fifty thousand tokens and succeeds is cheaper than one that burns ten thousand tokens and fails four times, once you count the pain of retries and human review. For genuinely hard work, the expensive model may be the cheap one.

But that premium only applies to what is hard now. Not to what was hard last year and has since become easy. And the moment a task becomes legible, every term of the revenue equation comes under pressure. Open models attack price per token by getting good enough at work the frontier already figured out. Small, specialized models attack tokens per task by doing one job for a fraction of the general model's token cost.

And there’s a deeper reason to think that task commoditization is inevitable.The labs need distribution, because that’s how they can turn their novel capabilities into revenue. But the act of serving the model and solving problems is also the act of producing a record of how that work was done. Competitors don't need to match the frontier from scratch. They can learn from its output. The more a frontier capability gets used, the more raw material exists for a cheaper model to reproduce it. This is why the labs hate distillation, and why they can't do much about it.

I don't want to overstate this. I’m not arguing that the labs will lose. But it does mean the ground behind the frontier is always eroding, and the labs will have to keep sprinting to outrun it.

This is also where the startup opportunities come into focus. 

The opportunity for startups is to industrialize what the frontier has proven works.

I'll resist the temptation to list every possible layer, because they are not equally good businesses. One example from the portfolio is a company called Specific that makes small language models, distilled and trained for specialized use cases where speed, cost or security matter. To use the language of this essay, Specific helps companies turn expensive exploration tokens into cheap execution tokens. And, as frontier capabilities continue to advance and frontier cost continues to rise, Specific’s market is starting to appear right before our eyes.

Two slopes

I think it’s useful to imagine the fate of the labs as a race between two slopes. The rate at which the labs open new tasks only the frontier can do, against the rate at which cheaper models get good enough at tasks the frontier already proved out.

The bull case is that the frontier marches on. Every time yesterday's work commoditizes, the frontier uncovers something larger.

The skeptic's case is simply that the labs can't stop selling the thing that erodes their edge.

The first phase of AI was about capability. You asked the smartest model in the world because you didn't yet know what was possible. The next phase will be about speed and cost, and my bet is that most of that market will not belong to the labs, but startups.

Demand for intelligence can be infinite without the pricing power for frontier tokens being infinite too.

David Peterson

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

ENTERPRISE/TECH NEWS

March of the Trillion Dollar Titans. You couldn’t miss it. Anthropic is reportedly valued near $1 trillion following a massive $65 billion funding round. And as of last night, the company appears to have confidentially filed for IPO. Meanwhile, SpaceX has filed for an IPO targeting a staggering $1.75 trillion valuation, concurrently securing significant defense contracts. The company recently won a $4 billion contract from the U.S. Space Force for its Golden Dome satellites and additional billions for a satellite network. This public offering would be one of the largest in history, reflecting its dominance in space launch, internet, and defense. Not all analysts, however, are excited. Some are downright skeptical: “Morningstar analysts issued a more cautious endorsement on Monday. They value SpaceX at $780 billion, about half its private valuation, and they outlined several reasons investors might be better off sitting out the IPO. A relatively small share of the company’s stock will become available to trade when it debuts, and Morningstar expects “buoyant investor appetite for AI infrastructure bids” and the stock’s expedited entry to major market indexes will create exceptionally strong demand around the IPO.

“We expect SpaceX’s share price will likely survive separation and may even ascend, at least for a time,” the analysts wrote. But shares are likely to experience “max Q, the moment of greatest atmospheric pressure on a launch vehicle,” in the succeeding months as private investors and employees are given the opportunity to sell their stock in public markets. For that reason, they argue long-term investors will have opportunities to invest in SpaceX “with more margin of safety” down the line.”

Corporate AI ROI Disappoints as Venture Shifts to Hardware. A Bain survey reveals that corporate AI investments are delivering less cost reduction than anticipated, with many predicted returns yet to materialize. “In an April survey of 951 respondents from companies with more than $100 million in revenue, Bain found that 37% said they experienced cost reductions of between 10% and 20%, but a larger 40% saw improvements of 10% or less. Only 4% of global respondents achieved AI-related savings of more than 30%, the survey, shared exclusively with Bloomberg, found. Here’s the part that Bain found the most troubling: 44% of large companies that are funding their next wave of AI spending are basing those investments on the last round of savings — savings that haven’t yet materialized for some of them.”

HOW TO STARTUP

A wave of layoffs hits Israeli tech. The wave of tech layoffs that has rippled through the US has arrived in Tel Aviv. “Meta announced last week that it was laying off about 8,000 employees worldwide, including dozens in Israel. Microsoft announced voluntary departures and layoffs affecting about 7% of its U.S. workforce. Snap laid off about 16% of its employees worldwide, roughly 1,000 people. American companies ZoomInfo and Shutterfly closed their Israeli development centers and dismissed hundreds of workers. U.S. fintech companies PayPal and Intuit laid off dozens of employees in Israel. Israeli company Wix announced this week that it was cutting 20% of its workforce, about 1,000 employees, including 900 in Israel. Israeli AI video company Lightricks laid off dozens. Israeli sports platform Minute Media cut 12% of its workforce worldwide, including 60 workers in Israel. Israeli online advertising company Taboola laid off about 100 employees. The layoffs have also reached areas that once seemed especially resilient. Israeli fintech star Rapyd surprised the market by reporting hundreds of layoffs in Israel, though the final number has not been published. Israeli AI company AI21 shut down its AI model development operation and laid off 60% of its staff. Amdocs, one of the pillars of Israeli high-tech, reported layoffs affecting 10% of its workforce, including hundreds in Israel. Cybersecurity giant SentinelOne carried out a broad 10% workforce cut, affecting about 70 employees in Israel. Firebolt, cybersecurity company Axonius and fintech giant Nayax have also laid off dozens of employees each.”

Grindmaxxing. A tweet by Harry Stebbings went viral for highlighting the 24/7 work culture at insurance startup Corgi. Andreas Helbig at Atomico posted a strong counter-argument. “The issue isn’t the hard work - intensity is a necessary condition to success, and that means putting in at least a sufficient amount of hours. The issue isn’t that Corgi “does all of this, only for insurance.” Some people are passionate about rockets, others about more obscure things like insurance. The first group is bigger than the second. Doesn’t make the second group invalid. The issue I take is how patronizing the statements are. This is not the only path to success. Intensity is a necessary condition, but it comes in many shapes and forms….Find your own way, find a mission that’s motivating for you (even if some might be condescending to it), make it healthy for you (it’s a marathon, not a sprint), and don’t preach your own way of working as “the only way or you’ll lose””

HOW TO VENTURE

Bill Gurley on building God. The legendary Bill Gurley was recently a guest on The All-In Podcast. In a widely shared clip, he offered a new theory to explain what Anthropic thinks its doing. “Anthropic is a mystery to me. I've never, ever seen a company that is both leading their field and the most negatively outspoken commenter on what they do. And my initial theory was the regulatory capture theory. Quite frankly, I think they're very close to achieving that. But then they just got so loud that I've literally, in the past 30 days, read everything I can about Anthropic, and I've come up with a new theory. I call it the Dr. Frankenstein theory. The more I dig, I've met people who, I dare say, think it's their responsibility, and they're excited about, building a species that's superior to humans.”

The OG on 10x. Oren Zeev, one of the OGs of Israeli (and global) venture capital gave an interview to Michael Eisenberg of Aleph. The entire conversation is worth a listen, but especially this clip on the irrational obsession with 10x growth.

DualBird
Senior Software Engineer (Tel Aviv, Haifa, Hybrid)

Groundcover
Account Executive - East Coast (Boston)

Reco
AI Software Engineer (Tel Aviv)