Compute

So you come up with a killer idea for a company – think any industry – healthcare, finance, etc. You build the killer prototype, launch, and because of great planning and testing – customers love it. The business then proceeds to grow like wildfire. A dream come true! Now you have the problem every founder hopes to have: keeping up with demand.

But wait. In this new world of ours you have a new problem. You need compute to keep the product running, and more of it as customers continue to sign up. For a business built around AI, that computing power starts to look a lot like air you breathe to grow and stay alive. Whether you rent the hardware or pay an AI provider each time your product does something, you need access to stay alive and to grow.

Now the key questions regardless of the business: Is the capacity there? Can you get it? And even if someone will sell it to you, can you raise the money to secure it?

I have come to start to worry that we could have companies doing well by all the measures we usually look at, with paying customers and a product that works, that still cannot stay in business. The thing they need to deliver the product is controlled by someone else. So you built a unicorn – but can you keep it alive?

Consider what happened to Windsurf. In June 2025, the AI coding company said Anthropic had given it less than a week’s notice before cutting off nearly all its direct capacity for Claude 3.x models. Windsurf had access through other providers, but they could not immediately cover demand. It was totally willing to pay for what they needed. By July, paid users had direct access to Claude Sonnet 4. The company found a way through – this time. [1]

Now imagine being the founder trying to explain that disruption to your customers. They bought your product and love it. They fully expect it to work and for you to keep it up and running. The relationship between you and the model provider is your problem, until it becomes theirs.

This is what makes the access question more complicated than just counting chips or adding up new data centers. A company needs the right capacity at the time it has promised to deliver it. Having compute somewhere in the world does not settle whether that company can use it.

CoreWeave, which rents out computing capacity for AI, said on its August 2026 earnings call that its near-term capacity was effectively sold out. Its CFO described deciding who gets capacity based on strategic importance, customer relationships, and financial returns. He also reported July price increases of approximately 25% across its product configurations. [2]

If you are coming back for more capacity, your growth forecast now depends partly on how your supplier ranks you against its other customers. Being willing to pay may get you into the conversation. It does not necessarily mean you will get what you need.

And securing capacity can mean paying well before the revenue arrives.

CoreWeave’s 2025 annual filing reported weighted-average prepayments of 15% to 25% of total contract value across its active contracts. Its committed contracts averaged about five years. It also sells pay-as-you-go access, but reserving capacity ahead of time can involve a substantial commitment. [3]

Take a hypothetical $10 million contract. At those percentages, you need $1.5 million to $2.5 million upfront. You could make money on every customer you serve and still be unable to write that check. The next round of financing may have to pay for the capacity behind sales you already know how to win.

Even the product getting better can complicate the bill. When Cursor explained its pricing changes in July 2025, its CEO said the hardest AI requests cost roughly ten times as much as simple ones. Most users’ costs had stayed fairly constant. But longer tasks consumed more, which made a request count a poor measure of expense. [4] If customers ask the product to do more work, the company needs to understand what that does to its margin.

There is a reasonable answer to all of this: prices can fall, and more capacity can become available. AWS announced a 44% cut in on-demand prices for its H100-based P5 instances in June 2025. [5] And large reservations can help smaller customers get access. In May 2026, Anthropic said new compute agreements had allowed it to raise usage limits. [6]

Those improvements matter. So does the time it takes to benefit from them. A cheaper model has to do the job your customers are paying for. Moving to another provider may require rebuilding and testing parts of the product. New capacity arriving next year does little for a company that needs it this month.

Windsurf got through its disruption. Cursor changed how it charged. My concern is what happens to the company that cannot make the adjustment before the cash runs out.

If I were looking at a growing AI business, I would want to see how its promise to customers lines up with what its compute suppliers have actually agreed to provide. On what terms, for how long, and with how much cash required before the next customer pays? A sales forecast is incomplete without that understanding.

The larger concern is which companies get the chance to grow. If reliable access requires commitments a young company cannot finance, customers can love its product and still lose it. We could end up rewarding the ability to secure computing capacity as much as the ability to build something people want.

That is a very different test for a founder who thought winning the customer was the hard part.

References

1. Windsurf. “Statement on Anthropic Model Availability.” June 3, 2025; updated July 16, 2025. Now hosted on Devin’s website.

2. CoreWeave. Q2 2026 earnings call, corrected transcript. August 11, 2026, pp. 3–4 and 7.

3. CoreWeave. 2025 Form 10-K. Business and customer-payment risk disclosures; fiscal year ended December 31, 2025.

4. Michael Truell, Cursor. “Clarifying our pricing.” July 4, 2025.

5. AWS. “Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances.” June 5, 2025.

6. Anthropic. “Higher usage limits for Claude and a compute deal with SpaceX.” May 6, 2026.