Accel is reportedly in talks to lead a USD 1 billion round in Thinking Machines at a USD 40 billion valuation, according to TechCrunch. The same report says the company has an annual run rate of over USD 100 million. That's every hard number available: no margins, no inference cost, no published retention. The rest is interpretation, mine included.
Why it matters
Do the division and you land on a multiple in the range of several hundred times annualized revenue. That isn't the price of a software company: it's the price of a call option on a future. What's interesting isn't the figure itself, but what it gets measured against this very week.
Crusoe reportedly raised USD 3 billion at a USD 30 billion valuation, and the piece notes the round came together after the data center developer secured a reported USD 13 billion contract with Jane Street. Wonderful more than doubled its valuation to USD 5 billion in under six months, with a USD 550 million Series C aimed at accelerating product, expanding its FDE teams and covering demand.
Three prices, three different logics: a signed contract at Crusoe, traction speed at Wonderful, a promise at Thinking Machines. The signal for anyone writing code isn't "there's a bubble" or "this is going to blow up." It's more concrete: a good chunk of what you consume today as a cheap API is funded by venture capital, not by its own unit economics. That makes your present cheaper and leaves you exposed in the future.
And in the same news cycle, Vercel made exactly the opposite move: it enabled Basic build machines on Pro and Enterprise, 2 vCPUs and 8 GB, as a cheaper option for small apps and agents that compile with few resources. Up top, billions are raised on expectations; down below, the platform you use every day hands you a knob to pay less per build minute. That contrast is the post.
What changes in practice
The first thing is architectural, and it isn't new but now it has fresh evidence behind it: if you're going to build on a model provider valued at promise-level multiples, treat its API as an infrastructure detail, not as the center of your domain. A port, an adapter, your own input and output contract, and tests against your contract rather than the provider's format. The day the pricing gets reshuffled or the model changes its name, the change has to be cheap for you.
The second thing is about money, and it's measurable today. Basic machines cost USD 0.0035 per vCPU per minute, that is USD 0.007 per build minute, according to the changelog. Hobby accounts were already running on those same 2 vCPU machines, which are now being renamed Basic. For new Pro and Enterprise projects the default is still Elastic, which scales automatically and is the recommended option for most cases.
| Criterion | Basic | Elastic |
|---|---|---|
| Resources | 2 vCPUs, 8 GB | Scale automatically |
| Default on new projects | No | Yes, on Pro and Enterprise |
| Use case it covers | Small apps and agents that compile with few resources | Recommended for most |
| Published price | USD 0.007 per build minute | Not stated in the changelog |
To turn them on you have to update the project configuration from team settings, project settings or with Vercel CLI 59.6.0 or later. If you manage several small projects —landing pages, internal dashboards, content sites—, that's where the change pays for itself.
When NOT to use it
Don't drop down to Basic out of a savings reflex. The price is per vCPU minute: if your build goes from three minutes to nine because it lost parallelism, you saved nothing and made your feedback cycle worse on top of it. And here's the honest part: the changelog doesn't publish the price of Elastic machines, so I can't give you the break-even point. Measure it on your own project before flipping the switch.
Don't touch it either if your pipeline does a full typecheck, tests and page generation in the same step. Eight gigs of memory is a real ceiling: a monorepo with heavy TypeScript or a site with many static routes can run out of memory during the build, and that failure shows up intermittently and at night, not when you're testing it. If the recommended default is Elastic, stepping outside the default is a decision you have to be able to justify with a number of your own.
And about the rounds: don't rearrange your stack over a valuation. All three pieces describe reported deals, not confirmed by the companies. An annualized run rate is not annual revenue and says nothing about margins. Picking a model provider because it raised capital is exactly the same mistake as picking it because it came out on top of a benchmark.
What I'd do today
Two small, boring things. One: review how many of my Pro projects have lightweight builds and try Basic on a single one, with the previous week's build times written down for comparison. Two: look at where a model provider's SDK is imported directly into the domain layer and give it the adapter it should have had from the start. Neither of those shows up in a USD 40 billion headline, and both will matter more to me than the round a year from now.