VentureBeat: 79% of Companies Leave GPUs Idle at Least Half the Time

According to VentureBeat Intelligence, the research arm of VentureBeat, 79% of surveyed company representatives said their GPU capacity sits completely idle at least half the time. The data comes from a continuous survey for the third quarter of 2026, and the figure held at that level for all three months of observation. In other words, a market that has complained about accelerator shortages for the past few years mostly cannot keep busy the hardware it has already bought or rented.

The contrast is sharp. AI infrastructure budgets keep growing, and corporate plans still include new purchases and long-term cloud reservations. Yet at most companies, half or more of the compute resource runs idle.

Where the empty hours come from

GPU idle time is usually caused not by a single factor but by a set of familiar organizational and technical reasons. The most common ones:

• capacity is bought for a peak load that occurs once every few weeks, and the cluster waits the rest of the time;
• accelerators are assigned to teams or projects, so one group's free hours are not available to another;
• model training has finished, but inference has not reached planned volumes;
• data arrives more slowly than the card can process it, and an expensive chip waits for the data loader;
• experiments are launched manually, leaving gaps between them, and the queue is empty at night and on weekends;
• weak orchestration: no shared scheduler, no partitioning of cards, no priorities.

An important detail: the survey recorded fully idle GPUs, meaning the card is not occupied at all. That differs from a GPU running at 30% of its performance inside an active job. In that case, the problem lies in code and architecture. Here, the problem lies in resource management.

What doing nothing costs

The arithmetic is simple. If an accelerator is busy half the time, every hour of actual work costs twice the nominal price. If utilization is lower, the multiplier grows. With owned hardware, depreciation, electricity, cooling, and maintenance accrue whether the card is computing or not. With reserved rental capacity, the bill also covers all hours, not just the ones used.

For vendors, this signal is ambiguous. On one hand, demand for accelerators is backed by real budgets. On the other, if most customers are not getting even half out of what they bought, CFOs will sooner or later start asking why more is being purchased. Cluster expansions justified by the phrase "we don't have enough capacity" will be the first to come under scrutiny. The beneficiaries are those who sell GPU scheduling and sharing tools, as well as cloud providers with a flexible hourly model that lets customers avoid holding a reserve in advance.

Key parameters of the study:

• source: VentureBeat Intelligence, the research arm of VentureBeat;
• format: continuous survey of enterprise respondents;
• period: third quarter of 2026;
• result: 79% of companies report GPU idle time of at least 50% of the time;
• trend: the figure did not change over three months.

Where to start hunting for idle time

The first step is boring but mandatory: measure. Many companies do not know the actual utilization of each card by hour and by team. It is enough to export metrics from drivers and the scheduling system for a month and see when and for whom the cards sit empty.

After that, proven techniques apply. Move from assigning cards to teams to a shared pool with a queue and priorities. Slice large accelerators into isolated partitions for small jobs and inference. Shift deferrable jobs such as fine-tuning, evaluation, and batch processing to overnight windows. Check whether the card is waiting for data: if utilization jumps around while the CPU or disk hits its ceiling, the culprit is the pipeline, not a lack of GPUs.

And the last, most practical point: before approving a new purchase, require the team to submit a report on current utilization. If the numbers look like the survey's, it is cheaper to first put in order what is already in the racks.