79% of companies admit their GPUs sit idle at least half the time, yet most plan new AI platforms

VentureBeat Intelligence has published data showing that most surveyed companies running AI in production have GPUs that sit idle at least half the time. Even so, most of them plan to add or switch AI compute platforms in the next 12 months. The article, written by Matt Marshall and dated October 23, 2026, pools three monthly VB Pulse surveys from July, August and September 2026.

The headline figures: 79% of the 218 respondents who reported their GPU utilization (172 of 218) said their GPUs are busy half the time or less. The share held between 77% and 81% in each of the three months. Among the companies with idle GPUs, 70% (120 of 172) plan to adopt a new, additional or replacement AI compute platform in the next 12 months. Among companies whose GPUs are busy more than half the time, the figure is 74% (34 of 46). The outlet describes the difference as not statistically significant.

The backdrop is the debate over whether AI is a bubble. The piece notes that on October 13 the International Monetary Fund warned that a sharp drop in AI stocks could threaten global financial stability. According to the author, the discussion has focused on compute sellers (Nvidia, Microsoft, Google, AWS, CoreWeave), while this survey looks at what their customers are doing. His thesis is that "the bubble is inside the enterprise." Demand for AI is real and still growing, especially for frontier models from Anthropic and OpenAI. But companies are paying for two kinds of AI that cannot easily be swapped for each other.

The first is capacity. The author believes many companies were spooked by reports of GPU shortages in 2024 and 2025 and secured capacity so they would not come up short. He notes that the surveys did not ask how companies obtained that capacity or whether they controlled it directly. The second is frontier models from Anthropic, OpenAI and Google, paid for per token via API, by subscription or through the major clouds. GPT and Claude models are closed: the provider or the model maker runs them on its own infrastructure, so a company cannot load them onto GPUs it already controls. Putting those GPUs to use would mean switching to an open-weight model and running it in-house, which takes real engineering work. In VentureBeat's own tests in October, DeepSeek was the cheapest of five open models, but many companies are reluctant to rely on Chinese models as their workhorse.

As an example of cost pressure, the article cites Uber CTO Praveen Neppalli Naga. According to Fortune, he said in August that "we are reaching the end of the so-called tokenmaxxing era" after the company used up its 2026 AI budget within a few months. An AI lead at a Fortune 500 company, speaking on condition of anonymity, said costs keep rising and the company hopes to offset them with productivity gains, though it does not yet know whether it will.

The data fit that reading on several points. Among companies with idle GPUs that plan a new platform, 90% (108 of 120) listed a cloud platform or a model provider among their options. Only 8% (9 of 120) mentioned an in-house or colocated GPU cluster or a self-managed open-source stack. Companies running their own stack also report better utilization: 56% (18 of 32) have GPUs idle at least half the time, compared with 83% (154 of 186) of the rest. The survey did not ask why. The author's guess is that teams managing their own stack see every idle hour and have engineers to fill it. Finally, in the September survey, 91% (58 of 64) of companies planning a new platform had no return-on-investment figure reviewed by finance. The author acknowledges that some idle capacity may be deliberate, held in reserve for demand spikes, and that a new platform may replace an old one.

The article argues that the next savings will come from tokens, context and agents. It expects a growing share of spending to go to model usage, because agents plan and carry out multi-step tasks. It cites a Gartner prediction that the inference cost per agentic workflow will rise more than fivefold through 2028. Price per token is falling, but agents consume far more tokens. Most of a coding agent's tokens are context that gets re-read at every step, which is why caching matters. A technical guide published in March by two Nvidia engineers says that after the first call in a Claude Code session, later calls reuse between 85% and 97% of the context from cache. Anthropic charges between one-fortieth and one-tenth of the normal input price for cached tokens, depending on the model. On agent teams, Vals AI reports that a lead agent with up to five subagents costs between 1.8 and 5.1 times as much as a single agent. Only one of four configurations improved scores by a statistically significant margin.

The article makes six recommendations. Measure cost per completed task, covering both GPU capacity and API spend. Test which model and effort level to set as the default, judging by cost per finished task rather than price per token. Send each agent only the context it needs and cache whatever is repeated. Add a second agent only when the task splits into parallel parts. Move routine, high-volume work such as classification, extraction and summarization from frontier APIs to open models running on idle capacity. And fill the remaining GPU hours with deferrable work such as batch inference, evaluations and fine-tuning.

The piece also promotes VentureBeat's Agentic Infrastructure Stack Summit, held October 28 in San Francisco under the tagline "Smarter agents need more memory. The best teams will use less." Speakers include representatives of T-Mobile, WEKA, PayPal, Box, Adobe, Cisco (Outshift), Artificial Analysis, Uber, Upwork and Microsoft, among others. Sponsors are WEKA, Outshift by Cisco and OneTrust.

Methodological caveats: the analysis pools 385 qualified respondents from organizations with 100 or more employees. Each month a different, self-selected group of VentureBeat readers and panelists took part, so the results describe those respondents and not all companies. The utilization data exclude respondents who do not measure GPU usage or do not run GPUs. Some differences between groups, such as 56% versus 83%, rest on small samples. Statistical significance was set at p < 0.05. The conclusions about causes are the author's interpretations, not results the survey measured directly.