Startups (still) aren’t businesses (yet?)

Why make a business that makes a little bit of money every day when you can make an asset that makes a ton of money all at once?

There are, I believe, exactly two laws that govern Silicon Valley. The first law is the bitter lesson, which we talked about last week. It comes from Richard Sutton, a Stanford-educated computer scientist, a Turing Award winner, and one of the founding fathers of a subfield of artificial intelligence. It says that in a competition between overwhelming compute and clever logic, overwhelming compute wins every time. Work smarter, not harder—unless the thing working harder is a exponentially accelerating computer.

The second law comes from a sitcom:

Richard (a naive startup founder): I just thought that, mainly the goal of companies is to make money.

Russ (Dan Bilzerian as a VC, a Zyn can in distressed Armani jeans): Yeah, no no no no, that’s not how it works. I don’t want to make a little bit of money every day. I want to make a fuckton of money all at once.

If you ask a venture capitalist or the Silicon Valley zeitgeist why it’s good for a startup to be profitable, they will give you a lot of answers. Being profitable shows that the company makes a good product, or at least one that can be effectively sold. It shows that the company is reasonably disciplined and well-managed, and doesn't waste lots of money on useless and unproductive things. It means the company is “default alive” and in control of its own destiny—it can stand on its own; it doesn’t need to sell board seats for operating cash; it can exist at the pleasure of its current owners and not outside venture capitalists. Being profitable likely means that the company’s financial metrics are healthy, and that investors and public markets—if and when the company chooses to take their money—will pay a premium to buy shares of the startup.¹

But one answer that they’re rarely give is that making money is good because you can just keep it. Despite businesses being “profit seeking enterprises or concerns”—that’s literally the definition²—profit in Silicon Valley is never the point. I’ve never heard of a venture capitalist walking into a board meeting, seeing a suitcase full of cash on the table and saying, “this is great, let’s divvy it up, and then just keep doing this, every quarter, forever.” No, to the venture capitalist, and often to the startup itself, profit is a means to another end: An blockbuster IPO or a big acquisition. Though the point of that is still to make money, it’s to make a bunch of money all at once, instead of a little bit every quarter.

In other words, despite the lofty language that people use to talk about Silicon Valley and its ambitions to build enduring companies, very few startups are, or even aspire to be, businesses that make money. Instead, they are assets that are meant to be sold.

We’ve talked about this before too:

It's not that startups are Ponzi schemes, but they are pretty Ponzish. …

Nobody—neither employees nor the early-stage VCs who invest in startups—value a startup’s equity by estimating the future dividends that the startup will pay its shareholders and plugging a bunch of numbers into a Black-Scholes model; they value the equity by guessing how much someone else will pay for it in the future. Their return isn’t funded by the operations of the business; it’s funded by a future investor. …

All of Silicon Valley is built around this scheme. Venture capitalists exist solely to fund the scheme: They invest in a company “until it reaches a sufficient size and credibility so that it can be sold to a corporation or so that the institutional public-equity markets can step in and provide liquidity.” Most founders start companies to get rich from the scheme; companies like 37signals are notable precisely because they don’t do the scheme. People say things like “from idea to IPO” because an IPO is when the scheme cashes out; they say “a startup is at least a seven-year commitment” because the scheme takes about seven years to run.

But, that post is from the ancient days of 2023. Then, the last line was an important detail: The scheme was slow and expensive to execute. Good software was hard to build. It took millions of lines of code to create something like Snapchat or Slack, and it took a long time and a lot of money to write all of it. That meant that most companies had to raise venture capital and operate a loss, often for years, before they could make even a small profit. No investor gets excited about potentially getting paid a little bit of money every quarter seven years after they wrote a check for tens of millions of dollars, so a ton of money, all at once, was the only palatable option.

Now, things are different? The fastest-growing AI startups are tiny teams making nauseating amounts of money. According to a tweet³ from two months ago, five AI companies—Cursor, Loveable, Bolt, Mercor, and Eleven Labs—employed a total of under 130 people and were making $280 million in revenue. Arcade AI makes $5 million and has five employees. Genspark Agent earned their first $10 million in nine days. Codeium (now Windsurf) hit $40 million with about 150 employees.

Admittedly, those figures are revenue, not profit. And a lot of AI products are wrappers⁴ around model providers like OpenAI and Anthropic, and a lot of their revenue⁵ gets passed through to those vendors. Still, the physics are different now. Companies that used to take a decade to build appear seemingly overnight, going from zero to millions in revenue in less time than it takes to mail something to France. It’s the era of the billion-dollar solopreneur; of the gritty startup. And so the mechanics of Silicon Valley are changing:

The old Silicon Valley model dictated that start-ups should raise a huge sum of money from venture capital investors and spend it hiring an army of employees to scale up fast. Profits would come much later. Until then, head count and fund-raising were badges of honor among founders, who philosophized that bigger was better.

But Gamma [ a profitable AI company that lets people create presentations and websites ] is among a growing cohort of start-ups, most of them working on A.I. products, that are also using A.I. to maximize efficiency. They make money and are growing fast without the funding or employees they would have needed before. The biggest bragging rights for these start-ups are for making the most revenue with the fewest workers.

But is the scheme changing? Are these new companies businesses, trying to make money, or are they still assets meant to be sold?

Still assets, it seems:

OpenAI is in talks to acquire Windsurf, an artificial intelligence-assisted coding tool formerly known as Codeium, for about $3 billion, according to a person familiar with the matter.

On one hand, of course Windsurf should sell itself for $3 billion! It’s a VS Code extension! It’s not even the biggest one!⁶ It has lots of competitors, including VS Code itself! Most of its revenue probably isn’t recurring, but monthly credit card swipes that are one competitive product release away from walking out the door! Sell, when you have people's attention! Top-tick that bubble!

On the other hand, if there were ever a startup that might be tempted to become a business—a company that, you know, makes money so that it can keep it—it seems like it’d be something like Windsurf. Its team is relatively small; its growth is vertical. It’s a profile of the modern AI startup: Young and efficient, with lots of revenue and few workers. If it wasn’t worried about its revenue base evaporating from underneath it—which, very reasonably, it might be—then why ever sell?

Eh. Because, I suspect, selling was always the point. Windsurf was never meant to be a business; it was meant to be an asset. The people who started it, joined it, and invested in it cared about how much Windsurf’s shares were worth, not about how much profit Windsurf made. Efficiently making money wasn’t inherently good, because it was never the final goal. The goal was to sell shares, and efficiency and profitability are good today because they’re the trendy scale on which shares are weighed.⁷

Which makes it seem unlikely that Silicon Valley’s “year of efficiency” and “era of grit” lasts for very long. So long as startups are built to be assets, there is no fundamental gravity around efficiency. It’s marketing, more or less. Businesses care about profit; assets care about perception.

When everyone is calming down from a historic fit of profligacy, efficiency is exciting. But now that everyone is trying to be disciplined, the preferred aesthetic could change—be the big spender; blitzscale everything; bully everyone else out. Because eventually venture capitalists will probably get sick of economical startups with doors that open like this, and will want the spectacle of startups with doors that open like that.

---

The industrialization of IT

The bitter lesson comes at you fast:

We’re excited to roll out an early version of Gemini 2.5 Flash today in preview in the Gemini API via Google AI Studio and Vertex AI. Building upon the popular foundation of 2.0 Flash, this new version delivers a major upgrade in reasoning capabilities, while still prioritizing speed and cost.

Our new 2.5 Flash model has an amazing performance to cost ratio, putting it on the pareto frontier.

According to their internal tests, Google’s new model, which was released yesterday, performs nearly as well as Gemini 2.5 Pro at code generation, but costs a fraction as much. Input tokens are 90 percent cheaper; output tokens are 60 percent cheaper.

Just last week, the choice every engineering team had to make was between hiring a junior software engineer, which cost about $250,000 a year, or sending 2 million annual requests—about one every 15 seconds—to Gemini, which cost $240,000 a year. Today, that choice is now between the same software engineer, or 9.5 million annual requests, which is one every 3.3 seconds.

How will we manage that volume of production? Well. Less than four hours before Google launched Gemini 2.5 Flash, Linear, a popular issue tracking application, pivoted into industrial production:

With Linear for Agents, we’re introducing a platform for a new model of collaboration. One where human and artificial intelligence work side by side. Agents become teammates: Assign them to issues, mention them in comments, and collaborate on projects together.

When asked if this meant that Linear expected AI agents to create product specs, write code, and do design work, Linear CEO Karri Saarinen said yes: “Linear can become a home for agents, tackling tasks across the entire product development workflow.”

Compete with a factory of overwhelming compute at your own peril, y’all.

---

For reasons, the final installment of the The White Lotus Power Rankings have been delayed until next week. The staff at benn.substack.com apologizes for the inconvenience, and welcome your feedback at benn.gripe.

1. It probably means other things too, but I have no idea what they are, because in 16 years of working, I’ve worked for a profitable company for a total of eleven months.

2. I was very tempted to start this post with a whole “Webster defines…” bit, but then I’d probably get fired for setting off some AI detector.

3. Sure?

4. Not derogatory.

5. Maybe all of it? Maybe more than all of it?

6. Before chasing Windsurf, OpenAI reportedly tried to get a date with Cursor. “See, this is interesting, cuz I don’t want to buy Windsurf, so can I take their yes and bank it, and use it on Cursor?”

7. The obvious counterpoint to this is Cursor, which was apparently offered a similar chance to make a ton of money all at once, and declined. (I’ve heard rumors of an $8 billion offer? Though my source is, like, loose gossip.) Still, I doubt that decision centered around models of future cashflows and potential shareholder dividends, and was probably instead something like, “eh, no, I bet we someone else will eventually offer us more. And if they don’t, we’ll do a banger of an IPO.”