Google's Compute Engine gets Custom Machine Types to let you tailor VM RAM and virtual CPU allotments
Frederic Lardinois / TechCrunch :
Context & Ripple Effects
This lands at the end of an aggressive 2015 for Compute Engine: in May Google cut cloud prices by up to 30 percent and launched preemptible instances, by July Windows Server reached general availability on the platform, and September brought general availability of the Autoscaler alongside 32-core VMs.
Custom Machine Types completes that arc on the sizing axis: where those earlier moves attacked price, automation, and OS coverage, this removes the fixed small/medium/large menu entirely, letting customers dial in exact vCPU and RAM combinations. It is Google's answer to competing on fit rather than just discounts.
First-order effects
- Compute Engine customers no longer have to over-provision to reach a needed memory-to-CPU ratio — they pay only for the allotment they specify instead of rounding up to a preset tier.
- The change sharpens Google's differentiation against rivals whose instance catalogs remain fixed-size, right after the May price cuts already reset cost expectations.
Second-order effects
- AWS and Azure face pressure to respond with equivalent custom-sizing options or justify why their preset instance families still make sense for mismatched workloads.
- Finer-grained VM definitions push cloud billing further toward pure resource metering, weakening the bundling power of standardized instance tiers across the market.
Third-order effects
- If tailoring the virtual machine shape proves out, the next step visible in Google's own trajectory is tailoring the underlying hardware itself — the path that later produced the Tensor Processing Unit for TensorFlow workloads and ARM-based Tau T2A instances with Ampere Altra.
- Cloud competition shifts from who has the cheapest standard box to who can match silicon, shape, and price to each workload — a structure that favors providers controlling their own chip designs.
The trend: Cloud compute is evolving from fixed instance menus toward fully customizable, workload-matched resources — first in software-defined sizes like these, then in purpose-built silicon.