The only figure with a denominator: 5.70 dollars a month for one Cloud Run instance

5.70 dollars a month

Cloud Run Instances gives an agent a dedicated, singleton runtime that does not shut down when it goes idle. The price in Google Cloud's August round-up is explicit: running one with 1 vCPU and 1 GiB of memory continuously for 30 days costs 5.70 dollars. Among the figures in the post this is the only quantity that carries its own denominator; the machine size, the period and the amount sit in one sentence. With those three pieces, the monthly cost of an always-on agent can be worked out.^[1]

The second change touches the shape of the bill. Filestore gained a new backend storage layer built directly on Colossus, Google's foundational distributed storage system, and that layer lets input-output operations per second be provisioned independently from storage capacity; it is deeply integrated with GKE. A quantity you can dial separately becomes a measurable cost only when a rate figure or a price per rate is published. The round-up supplies neither, so the storage layer does not sit on the same denominator as the Cloud Run line.^[1]

The same bottleneck from two sides

The picture April Li of TSMC gave at SEMICON Taiwan 2026 showed where the loss accumulates: data movement takes up to 60 per cent of typical system activity and AI accelerators run at utilisation below 40 per cent. Two changes in the August round-up target the software-side counterpart of that same loss: separating operation rates from storage capacity works on the width of the data path, and an instance that does not shut down when idle removes the warm-up. It is equally possible that the same changes come from the ordinary road map of storage and serverless products, with no relation to accelerator utilisation.^{[2], [1]}

The third item is isolation. An experimental library developed with Anyscale brings gVisor, Google's open-source application kernel, into distributed Ray clusters on GKE. The description given for these environments is stronger isolation, fast startup times and low memory overhead; none of the three comes with measurement conditions. Until the workload, the container being compared against and the memory baseline are written down, the three phrases stay the vendor's own description and do not become a performance result.^[1]

What would become measurable?

On August 31 this column recorded Li's figures and noted that TSMC's answer, 3DFabric and COUPE, is still in development, and that the 1 trillion transistor mark stands as a projection for 2030. What was missing on the manufacturing side that day is missing on the cloud side a week later: most of the machinery said to improve utilisation or widen the data path is announced without a measured number.^{[3], [2]}

The Colossus-backed Filestore tier moves from a described capability to a priced one the moment a price per unit of operation rate is published. Where a reader should look is clear: whether a figure per operations-per-second appears in Filestore pricing. When that figure arrives, the storage layer can be compared in the same sentence as the Cloud Run line; until it does, 5.70 dollars a month stands on its own.^[1]