Vercel changed the billing model for Fluid compute on Thursday, moving from a 50 millisecond minimum billing increment to true per-millisecond measurement with no floor. For an application whose typical response takes 8 milliseconds of compute, which describes a lot of API routes that mostly read from a cache, the old model billed six times more than the work consumed. The change is retroactive to July 1 and Vercel is crediting the difference. Chief executive Guillermo Rauch framed it as removing an artifact of how serverless billing evolved rather than a discount, which is technically accurate and also the kind of framing you use when a competitor forced your hand.
Where the minimum came from
AWS Lambda launched in 2014 billing in 100 millisecond increments, a granularity chosen because the underlying metering and the accounting overhead made anything finer uneconomical at the time. Lambda moved to 1 millisecond in 2020. Every serverless platform that came afterward inherited some version of the minimum, and Vercel's 50 millisecond floor was the residue of building on top of Lambda before moving to its own infrastructure.
Fluid compute, which Vercel introduced in 2025, already changed the model in an important way by allowing a single instance to handle concurrent requests rather than one at a time. That eliminated the pathology where a function waiting on a database call billed for wall clock time while doing nothing. Removing the floor completes that work: you pay for CPU time actually consumed, concurrently shared across requests, measured to the millisecond.
What it does to real bills
The teams who benefit most are the ones running many small requests. A documentation site with API routes that return cached content, a webhook receiver, an authentication callback: all of these do a few milliseconds of work and previously paid for fifty. Three companies shared before and after figures with us. A Berlin SaaS company's monthly Vercel compute line dropped from $2,840 to $1,190. A US media site went from $9,200 to $5,400. A smaller agency saw almost no change because its workload is dominated by server-rendered pages that take 120 milliseconds each.
That last case is the honest caveat. If your compute time already exceeded the floor, nothing changes. The pattern that benefits is high request volume with low per-request work, and applications built with heavy server components and database round trips are the opposite. Vercel's own estimate is that the median customer sees a 15 to 20 percent reduction and the top decile sees more than half.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The competitive pressure
Cloudflare Workers has billed CPU time rather than wall clock time since 2021 and has no minimum, which has been the single most effective point in its sales conversations against Vercel. Workers also charges nothing for time spent waiting on network calls, a model Vercel now approximates through concurrency but does not exactly match. Cloudflare's pricing has been the reference point every Vercel customer brings to a renewal negotiation.
Netlify moved to a similar model in the spring. AWS Lambda still bills wall clock with a 1 millisecond granularity, which means it charges for waiting, and this is the gap that keeps pushing web workloads toward the platforms built after Lambda. Amazon has been visibly aware of this and has not changed the model, presumably because doing so would reduce revenue on an enormous installed base without winning back customers who have already left.
The parts of the bill that did not change
Compute is one line among several. Vercel's bandwidth pricing, at $0.15 per gigabyte for fast data transfer beyond the included allowance, remains far above what a customer would pay at a raw CDN, and for content-heavy sites bandwidth dominates the invoice entirely. Image optimization is billed per source image and has caused several publicized bill shocks. Edge middleware invocations carry their own rate.
The pattern across platform-as-a-service pricing is that the headline metric gets competitive while adjacent metrics carry the margin. That is not unique to Vercel and it is worth understanding before celebrating a compute reduction. A customer whose bill is 70 percent bandwidth saved nothing this week. The practical advice, which several consultants repeated, is to pull a line-item breakdown before assuming a pricing announcement applies to you.
What this signals about the market
Serverless pricing has been converging toward measuring the thing that actually costs the provider money. That took a decade because the early models were built around technical constraints that no longer exist and because vendors had no incentive to move faster than competition required. The remaining distortions, mostly around bandwidth and around charging for waiting, will erode the same way.
For architects the practical consequence is that the cost argument for running your own containers keeps weakening for spiky workloads. A Kubernetes cluster sized for peak sits idle most of the time and you pay for the idle capacity. Per-millisecond serverless with concurrency comes close to paying only for work performed. The remaining reasons to run your own infrastructure are control, data residency, and workloads with sustained high utilization, which is a shorter list than it was two years ago.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- Edge Computing
