Nvidia Is Quietly Buying CoreWeave's Unsold GPUs. What If Tokens Don't Get Cheaper for Your YC F26 AI Startup?

Every AI startup applying to YC F26 has built its financial model on one assumption: tokens get cheaper. Not "might." Will. It is the closest thing the 2026 application pool has to a religion, and it is doing a lot of quiet work in a lot of spreadsheets.

Sitting near the top of Hacker News today is an I/O Fund breakdown of the circular financing behind the GPU boom, published June 12 and resurfaced this week with roughly 290 points. The short version: Nvidia invested $2 billion each into CoreWeave and Nebius, those companies spend that money buying Nvidia GPUs, and Nvidia separately agreed to purchase CoreWeave's unsold capacity under a backstop with an initial value of $6.3 billion, running through April 2032.

If you are writing your F26 application right now, the deadline is July 27 at 8pm PT. This piece is about the one question that story should make you ask, and what a strong answer to it looks like.

What the story actually says

Hyperscalers do not want AI capex on their balance sheets. Microsoft has committed roughly $60 billion to neoclouds. Meta has committed up to $62.2 billion across CoreWeave and Nebius. Those commitments are recognized as operating expense over the life of the contracts rather than capex today.

The neoclouds absorb the capex instead. CoreWeave's Q1 2026: revenue of $2.08 billion, up 112% year over year, against $7.7 billion of capex, free cash flow of negative $4.71 billion, a cash balance of $2.27 billion, and total debt of $24.86 billion. Interest expense alone was $536 million, or 25.8% of revenue. Full-year capex guidance is $31 to $35 billion against roughly $8.68 billion of expected operating cash flow.

That 112% is the number bulls will point at, and they are not wrong to. This is a company growing extraordinarily fast into real demand. But it is funding that growth with debt, and the rate environment is moving the wrong way: the 3-year Treasury has gone from under 3.6% to roughly 4.16% this year.

Does this actually touch your token bill?

Be honest about the chain, because a YC partner will be.

You do not rent from CoreWeave. You buy tokens from OpenAI, Anthropic, or Google. Nothing in a neocloud's balance sheet appears on your invoice.

But OpenAI and Anthropic are themselves CoreWeave customers. The labs' compute costs run partly through this system, and the price you pay is set by labs competing for your usage while their own capacity is procured from companies burning cash at scale. Cheap inference in 2026 is not purely a physics story about better chips. It is partly a capital-markets story, and capital markets change their minds faster than transistors do.

The link is real but it is indirect and slow. Anyone who tells you the GPU financing structure will hit your API bill next quarter is selling something. What it does mean is that "tokens get cheaper" is an assumption, not a law, and F26 is the first application cycle where a partner might actually poke at it.

The question to be ready for

Not "what do you think about the AI bubble." Nobody wants your macro thesis, and every applicant has one.

The question you should be able to answer cold is narrower: if your inference costs tripled, what would you change, and how fast?

The bad answer is "costs are falling exponentially, so this only gets better for us." That is the assumption restated as a defense. The other bad answer is "we would switch to a cheaper open model." Fine, but you have clearly never done it, or you would know it breaks your evals, your prompts, and your latency budget.

A good answer sounds like this: "Inference is 12% of our COGS. At 3x token prices we are still at 62% gross margin. At 10x, we move classification to a fine-tuned 8B model we already benchmarked at 94% of current quality, which takes about two weeks."

That is a strong answer because it proves you ran the experiment instead of trusting the market.

The part nobody is writing about: your contracts

Here is where this gets practical, and where most F26 applicants are quietly exposed.

If you believe costs only fall, you make a series of pricing and contract decisions that are all fine in that world and all terrible in the other one. Look at yours before you submit.

Are you selling unlimited-usage seats? Flat "$50 per seat, use it as much as you want" pricing is a short position on token prices. It is a fantastic deal for you if inference gets cheaper and a margin-eating trap if it does not. Your power users are the ones who will destroy you, and they are the ones you are currently celebrating.

Did you sign an annual contract with no cost pass-through? A twelve-month enterprise deal at a fixed price, signed in July 2026, means you have agreed to absorb any input cost movement until July 2027. That may be the right call to win a logo. Know that you made the bet.

Does your pricing move when your costs move? Usage-based and outcome-based pricing pass cost changes through naturally. Seat-based pricing does not. Neither is wrong, but only one of them survives an input-price shock without a renegotiation.

None of this requires you to predict interest rates. It requires you to know which side of the bet you are on. Knowing what bets you have made is the job. Making them without noticing is not.

The three numbers to have before July 27

You have two weeks.

• Gross margin with COGS itemized. Inference, vector DB, orchestration, and the humans. Most founders claiming 80% margins have never counted the humans.
• Cost per unit of value, monthly, for the last three months. Per resolved ticket, per generated test, per agent run. Falling because of something you did (caching, model routing, cutting a retrieval hop), not because a lab cut prices for you.
• Your fallback, benchmarked. Name the cheaper model you actually tested, the quality delta you measured, and the margin it preserves.

The honest counterpoint

The compute glut may simply continue. Nvidia generated $119 billion of free cash flow over the last twelve months, which buys a lot of patience, though the CoreWeave backstop is a specific instrument with an initial value of $6.3 billion rather than an open-ended promise. Reasonable people read this both ways.

But you do not get funded for being right about macro. You get funded for being unusually clear about your own business. A founder who can describe their cost structure under stress reads as serious. A founder who cannot reads as a tourist, and this news cycle is a very easy way for anyone reading your application to tell the difference.

If you want to know which one you sound like before you hit submit, the cheapest fix is to have someone who has sat on the other side of the table read your answers. That is what YC Roaster does: YC alumni tear into your F26 application and tell you exactly where you are hand-waving, while there is still time to fix it.

Two weeks. Go find your gross margin.