Summary

  • Colab’s pricing page gives compute units a 90-day life; the access obtained from a unit depends on the virtual machine and GPU or TPU configuration.
  • Eligible Google AI and Colab subscriptions can deposit units into the same balance on different monthly schedules. Turning that balance into useful work still requires a ready experiment and available hardware.

The purchasing department is ready to run the experiment. Unfortunately, it does not run experiments. The researchers are still cleaning the data, waiting for a colleague and changing the test. Meanwhile, the compute budget they bought has started ageing.

That is the small but consequential clock inside Colab. Google’s pricing page lists 100 monthly compute units for Pro and 600 in total for Pro+, with units expiring after 90 days. Its explanation of a compute unit adds another moving part: the amount of access depends on the virtual machine configuration and use of dedicated accelerators such as GPUs or TPUs.

A balance therefore cannot be copied into a project plan as a fixed number of GPU-hours. It is a time-limited allowance for using a service. A finished experiment requires someone to convert that allowance into the right machine, a completed run and a result worth keeping. Buying the allowance is the easy part.

A dated ticket to an undated seat

Google is quite explicit about keeping room to manoeuvre. The Colab FAQ says low-cost access to powerful GPUs depends on flexibility to adjust usage limits and hardware availability. Paid plans share the dynamics of changing resource availability. Premium GPU access remains subject to availability.

There is a sensible bargain here. Google retains flexibility across its machines; users avoid much of the work of setting up and maintaining their own. Short experiments, teaching notebooks and prototypes that tolerate different configurations can make good use of that convenience.

A deadline changes the bargain. Consider a test that needs substantial GPU memory, with its data arriving just before a scheduled presentation. Adding units expands the spending allowance. It does not book a particular GPU for the evening before the presentation. Adapting the test to another configuration may mean splitting work or rearranging runs, adding staff time and scheduling effort to the cost. That is a planning risk implied by the published rules, rather than a measured customer loss.

The expiry and the availability condition matter together. A team that can move its experiments has something valuable to trade for convenience. A team with one narrow execution window has already spent much of that flexibility before it buys anything.

Two subscriptions can feed one hungry balance

Google’s 22 September announcement brought premium Colab benefits into eligible Google AI plans. Existing Colab subscriptions continue, and the benefits can be stacked.

The FAQ supplies the practical detail. Google AI and a native Colab Pro or Pro+ subscription can coexist on the same account. Their compute-unit grants enter the same balance, potentially on different monthly dates. Each subscription is still managed where it was purchased.

For someone already running plenty of experiments, the extra units may find immediate employment. Someone who bought Google AI for other products may acquire a compute allowance before acquiring a reason to use it. The balance grows; the supply of clean data and researcher attention stays put.

The origin of each grant therefore matters. The 90-day rule is explicit on Colab’s pricing page; the conditions of a particular Google AI benefit need checking against that offer. A combined total is useful for seeing what remains, but weak as a calendar of what arrived, under which terms and for which work.

The tempting response is to run something, anything, before units disappear. An unnecessary parameter sweep can produce an impressive consumption rate and a pile of files somebody must interpret. The credit gets fully used while the researcher gets less useful work done. An expiry notification has no qualifications to run a laboratory.

The run needs a funded ending

Colab Pro+ supports continuous code execution for up to 24 hours with sufficient units. Google also says the backend terminates when units are exhausted on Pro, Pro+ or Pay As You Go. A long-session feature still needs a balance capable of carrying the session through.

The useful measure is a completed experiment with a saved, usable result. A configuration burning more units per hour may finish important work sooner. A lower hourly burn can become expensive if it leaves people waiting or repeating preparation. Consumption and productivity are related, but the direction has to be established for the actual job.

Account-level records connecting grants, expiry, available hardware and useful completed experiments are missing; those records would quantify unused credit and the effective cost of results.

Buyers can nevertheless improve the order of operations. Identify executable work first, prepare its inputs and estimate consumption on the intended configuration. Then size the allowance, including room to save outputs and repeat worthwhile tests. Buying first and finding a use later makes preparation compete with a clock that is already running.

Keep the notebook, take responsibility for the machine

Google’s local-runtime guide allows the Colab interface to connect to a user’s own hardware. Sharing the notebook does not share that local runtime. Familiar working habits can remain while responsibility for compute moves to the user.

That responsibility includes capacity, maintenance and rebuilding the environment. It puts a price on a more predictable execution window. Flexible experiments may belong in a changing shared pool; immovable commitments may justify a different resource arrangement.

The comparison that matters is the cost of finishing the work on time against the cost of rearranging it. Units are an input to that calculation. Spending every last one is a very tidy achievement for an account, and an entirely optional achievement for a scientist.