A cheque from an investor is usually a simple trade: cash for equity. In May 2026, Sam Altman did something different: he offered every startup in Y Combinator's current cohort $2 million in OpenAI API tokens, the units that pay for using their GPT models, in exchange for equity. The offer reached roughly 169 startups across YC's Spring and Summer 2026 batches. No cash changed hands.

Each startup that accepts signs an uncapped SAFE, a contract that converts into equity when the startup raises its next round. Uncapped means the better the company performs, the smaller the slice OpenAI ends up with. The SAFE also carries no Most Favored Nation provision, the clause that lets an early investor pick up any better terms offered to investors who come later.

Set it next to YC's own SAFE, and the gap tells you everything. YC's own standard deal puts $375,000 into an uncapped SAFE too, but that one carries an MFN, and in an instrument with no cap, the MFN is the only thing standing between an early investor and everyone who follows. OpenAI's version has neither, offering a weaker instrument than the accelerator gives itself, across an entire batch.

Which means the equity is the smallest thing on the table. So, what is OpenAI trying to do here?

What OpenAI is really buying

Three things, and the token offering achieve all of them at a discount. That’s where the story starts. These tokens are priced at the retail rate developers pay, which covers research, training runs, and margin on top of the raw compute. OpenAI carries none of that markup when it spends its own supply. Its internal compute margin reportedly reached roughly 70% by October 2025, up from 52% a year earlier. The company has never confirmed the figure, but the direction is largely agreed by the industry: a token costs OpenAI far less to serve than it costs a startup to buy.  

Now watch what the discount buys, in ascending order of value.

Cheapest of the three is defaults: these are the youngest companies in the market, still choosing the foundations they will build everything else on. A large balance of pre-paid credits doesn't lock anyone in, but it quietly tilts every build-versus-switch decision toward the provider that is footing the bill. Among enterprises that have attempted a migration between AI providers, 58% say it failed outright or took far more effort than expected. Two million dollars buys a default at the moment it is cheapest to win.

More valuable is the closed loop. Circular financing already defines this cycle: a supplier invests, and the company spends the investment buying the supplier's product. Microsoft's $13 billion of funding commitments to OpenAI were reportedly delivered in substantial part as Azure compute rather than cash, but OpenAI has pulled the loop tighter still, since there is no cash to circulate at all. The investment can only be spent on the investor, on one product, and it generates returns later, when the credits run dry and the batch starts paying retail.

But the most valuable of all is a sneak peek into the next-generation of companies that no competitor can buy: for as long as the credits last, OpenAI can keep close tabs on an entire YC batch as they stress-test its models across every problem those founders are chasing, before anyone else sees what works.

OpenAI collects all three whether a single company in the batch succeeds. The equity is the only part that depends on the startups winning, and it is the only part OpenAI left unprotected. Read the MFN again in that light: waiving it costs OpenAI almost nothing, because the equity was never where the deal was designed to pay off.

The investor who is also your vendor

What looks like generosity from one side of the table looks like something else from the other.

A traditional investor plays one role: they own a piece of your company and want it to grow. OpenAI holds two at once, as investor and as the vendor supplying the infrastructure the startup runs on.

The upside is immediate: API spend is a direct, recurring cost, and removing $2 million of it for a small, deferred stake is close to free capital. At seed, when the alternative is three months spent raising the same amount in cash, most founders will take it.

The cost is a conflict no ordinary cap table contains. A normal investor profits only when you profit, but OpenAI would hold equity while also being the vendor the startup runs on, two roles that do not always pull the same way, and the terms don't say whether OpenAI might one day build in a startup's own market. None of this is unique to startups. Any enterprise whose AI supplier also holds a stake in its business would face the same divided interests. The YC batch just shows the arrangement in its purest form.

No bill, no brakes

That same night, Altman put a name to the behaviour he's hoping the credits produce: “i am excited to see what will happen with tokenmaxxing startups”. Tokenmaxxing means leaning into AI usage as hard as possible, treating heavy consumption as the way a lean team moves fast. Sound advice for such a team, and also the behaviour that turns a startup's credits into usage on OpenAI's platform. Every token burned pulls the default a little further toward OpenAI, which is the return the deal was built to produce.

Two companies have already run the experiment on what happens when that spending goes unwatched. Meta employees burned 73.7 trillion tokens in a month chasing an internal leaderboard, until CTO Andrew Bosworth had to tell staff that token usage “is not a measure of impact of any kind”. Uber spent its entire 2026 AI coding budget in four months, and despite near-total adoption its COO said the link between the spending and the product “is not there yet.”

Both companies were spending real money and watching it leave, and the invoice itself was the discipline. A startup that takes $2 million in prepaid credits has no invoice to watch, and the party that issued them gains from every one it burns. Paid usage carries its own brake: the bill. Free credits from a backer that benefits when they are burned carry none.

What this means the next time a vendor offers you a discount  

Zoom out, and the lesson travels well past Silicon Valley. The specifics are a startup story, but the structure is not quite as fragile. Committed-spend discounts, migration credits, GPU capacity all offered below cost: enterprise AI is already sold this way, and every one of these offers arrive looking like an attractive discount. OpenAI just showed everyone their version of how to use these discounts to gain a competitive edge. A supplier that funds your usage is buying its way into your startup, and the ones who do it well structure the deal, so they win whether or not you do.

So, the question the “tokenmaxxing” YC batch is answering first is the one every AI buyer inherits next. When your vendor offers you something that looks free, evaluate the price of what you are buying. The credits are often the most obvious part. Look beyond the tokens and you will see what you are giving up.