根据VentureBeat Intelligence的最新数据，大多数在生产中运行人工智能的受访公司表示，他们使用的GPU至少有一半时间处于闲置状态，其中大多数公司计划添加或切换人工智能计算平台。
在报告GPU利用率的218名受访者中，高达79%的受访者表示，他们的GPU每月有三次占用一半或更少的时间
。在这些公司中， 70%的公司表示他们计划在未来12个月内采用新的、额外的或替代的人工智能计算平台。在三个月的每个月中，空闲GPU的份额在77%到81%之间。我们
；三个月的数据显示模式保持稳定。
这些数字很重要，因为华尔街正在争论人工智能是否是一个泡沫。10月13日，国际货币基金组织警告说，
人工智能股价大幅下跌可能威胁全球金融稳定
。大多数论点都集中在销售人工智能计算的公司，如英伟达、云巨头微软、谷歌和亚马逊网络服务，以及专业的GPU云，如CoreWeave ，以及他们是否在构建比客户支付更多的容量。我们研究了这些客户实际上在做什么。
乍一看，这些数字看起来像是一个矛盾。为什么有这么多未使用GPU容量的公司计划添加或切换AI计算平台？正如技术领导者告诉我的那样，他们在今年秋季的预算季节面临着削减人工智能账单的巨大压力，他们为什么要让GPU闲置？
对人工智能的需求是真实的，但它落在了与闲置的GPU不同的地方。专业人士和消费者继续使用更多的人工智能，其中大部分工作都通过Anthropic和OpenAI的前沿模型进行。这种需求将在一段时间内继续支持这些公司的计算支出。不匹配位于一个层次下方，在为两种不容易相互替代的人工智能付费的企业内部。
第一种是容量。我相信许多公司对2024年和2025年GPU短缺的报道感到害怕，并排列了GPU容量，这样他们就不会被短缺。我们的调查显示，公司使用的大部分产能都处于闲置状态，尽管调查没有询问公司是如何获得这种产能的，也没有询问他们是否直接控制了这种产能。
第二种是前沿模型。工程师和其他员工更喜欢Anthropic、OpenAI和Google最有能力的模型，通过这些公司的API、月度订阅或大型云提供商通过代币支付。当时的想法是，公司需要最好的智能来领先，他们会在以后解决经济问题。优步首席技术官Praveen Neppalli Naga在8月份表示，
“我们即将结束所谓的tokenmaxxing时代。”
据《财富》杂志报道，优步在几个月内就花掉了2026年的人工智能预算。一家财富500强公司的人工智能领导者在10月份告诉我，人工智能成本不断上升，虽然该公司预计生产力的提高最终会抵消这些成本，但它还不知道这一点。
亚马逊网络服务、微软和谷歌现在通过他们的云销售Anthropic或OpenAI的模型， OpenAI周三表示，
in the coming weeks. OpenAI’s frontier GPT models and Anthropic’s Claude models are closed, so the cloud provider or the model maker runs them on its own infrastructure and bills by the token or for capacity reserved for that one model. A company cannot load them onto GPUs it already controls. To shift work onto those GPUs, a company has to switch to an open-weight model and run it itself. That is real engineering work, and the cheapest options r…
Companies with idle GPUs are still planning new platforms. Within the next year, 70% of companies with idle GPUs plan to adopt a new, additional or replacement AI compute platform. That’s about the same as the 74% of companies whose GPUs are busy, a statistically insignificant difference.
Companies with idle GPUs are considering cloud platforms or model providers. Among those planning a new platform, 90% named a cloud platform or a model provider among their options, and only 8% named an on-premises or colocated GPU cluster or a self-managed open-source stack.
Companies that run their own stack report higher GPU utilization. Among companies running AI on an on-premises or colocated GPU cluster or a self-managed open-source stack, 56% said their GPUs sit idle at least half the time, against 83% of all other companies running AI in production. We did not ask why. My guess is that a team running its own stack sees every idle hour directly and has the engineers to fill it.
Few companies planning a new AI compute platform have a finance-reviewed ROI figure.
, 91% of all companies planning a new AI compute platform had no return-on-investment figure their finance team had reviewed. To be sure, some idle capacity may be deliberate, since a company may reserve GPUs for spikes in demand, and a planned platform may replace an older one.
I expect a growing share of companies’ AI spending to go to model usage. By this, I mean the per-token charges that model makers and cloud providers collect on every request. Those charges are likely to grow because companies are shifting to AI agents, which plan and carry out multi-step tasks. Inference costs per agentic workflow
, Gartner has predicted. Inference is the work a model does to answer each request. The price of each token is falling, but agents use so many more tokens, and more expensive ones, that the falling prices do not keep up.
Most of the tokens a coding agent processes are context, meaning the instructions, documents and history it rereads at every step. That is why caching matters so much. In a technical guide first published in March, two
that after the first call in a Claude Code session, later calls reuse 85% to 97% of their context from a cache. Reusing that context is far cheaper than processing it fresh:
between one-fortieth and one-tenth of its normal input price for cached tokens, depending on the model.
Adding agents multiplies the bill, too. Coding-agent teams, with a lead agent running up to five subagents at once, cost 1.8 to 5.1 times as much as a single agent, and only one of four team setups improved benchmark scores by a statistically significant margin,
Companies should get more finished work out of the compute and tokens they already pay for before they sign the next contract. In practice, that means six things:
measuring the cost of each completed task, across both GPU capacity and API spending;
testing which model and effort level to set as the company default, judged by the cost of each finished task rather than the price per token, since a cheaper model can use so many more tokens that it costs more per task;
sending each agent only the context it needs, and caching repeated context, whether through model-provider APIs or on companies’ own infrastructure;
adding a second agent only when the task splits into parallel pieces;
moving routine, high-volume work, such as classification, extraction and summarization, from frontier model APIs to open models running on idle capacity;
filling the remaining idle GPU hours with work that can wait, such as batch inference, model evaluations and fine-tuning.
VentureBeat’s October 28 summit takes on these choices
on October 28 in San Francisco, under the theme “Smarter Agents Need More Memory. The Best Teams Will Use Less.” The day runs in four parts:
Jeff Simon, chief information officer of T-Mobile; Val Bercovici, chief AI officer of WEKA; Deepika Khera, senior director of AI at PayPal; and Ben Kus, chief technology officer of Box, on coordinating agents while controlling cost and risk.
Anjul Bhambhri, senior vice president of engineering at Adobe, and Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, on memory, context and coordination.
Micah Hill-Smith, co-founder and CEO of Artificial Analysis, and Uday Kiran Medisetty, distinguished engineer at Uber, on inference economics, model routing and the cost of completed tasks. The track also includes a WEKA-sponsored panel with Sudhanshu Bist of Upwork and Val Bercovici of WEKA.
Reynold D’Sa, corporate vice president at Microsoft, on infrastructure for agentic workloads, alongside a session featuring Omni and Acxiom.
2026 VB Pulse surveys on AI infrastructure and compute. The surveys drew 385 qualified respondents at organizations with 100 or more employees, including consultants answering for their main client. Each month’s survey drew a different self-selected group of VentureBeat readers and panel respondents, so the results describe those respondents rather than every company.
Companies running AI in production whose GPUs are busy half the time or less
Companies with GPUs busy half the time or less that plan a new, additional or replacement AI compute platform within 12 months
Companies with GPUs busy more than half the time that plan a new, additional or replacement AI compute platform within 12 months
Companies with idle GPUs planning a new platform that named a cloud platform or a model provider among their options
Companies with idle GPUs planning a new platform that named an on-premises or colocated GPU cluster or a self-managed open-source stack
Companies running AI on an on-premises or colocated GPU cluster or a self-managed open-source stack whose GPUs are busy half the time or less
All other companies running AI in production whose GPUs are busy half the time or less
September only: companies planning a new AI compute platform with no return figure reviewed by their finance team
The utilization figures cover companies running AI in production that gave a utilization figure. They leave out companies that said they do not measure GPU use or do not operate GPUs, and July respondents who chose more than one utilization answer. Respondents could name several platforms they use and several they are considering. Cloud platforms are AWS, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure; model providers are OpenAI, A…
Applications to attend the Agentic Infrastructure Stack Summit close Monday, October 26.
Sponsors of the summit include WEKA, Outshift by Cisco and OneTrust.