The AI race is driving up costs for GPUs, memory, and software

OpenAI, Anthropic, and the largest cloud providers are expanding computing capacity, while companies embedding AI in their products and processes are also spending more. Memory, reserved GPUs, and software are getting more expensive: memory prices have more than doubled over the past 12 months, AWS has raised prices for reserved GPUs four times, and Gartner expects inference costs per AI-agent workflow to increase more than fivefold by 2028.

Rising computing costs are making AI more expensive for companies

Computing commitments are growing

In its IPO prospectus, updated on November 2, Anthropic disclosed approximately $895.4 billion in purchase and lease commitments. Almost all of that is tied to computing infrastructure. About $214 billion of the commitments can be terminated or depend on certain conditions; the agreements run through 2045. The company reported revenue of $33.2 billion for the first nine months of 2026.

Some agreements require Anthropic to pay for resources whether or not it uses the reserved capacity. Anthropic said it needs to rapidly expand its customer base and revenue to generate the cash flow required to meet these commitments.

On October 28, Alphabet reported negative free cash flow for the second consecutive quarter. The company’s capital expenditures for the third quarter totaled $51.4 billion. In August, as it expanded its AI infrastructure, Alphabet also sold $25 billion in bonds, Bloomberg reported.

The spending is already showing up in buyers’ budgets. In VentureBeat Intelligence’s October survey on AI infrastructure and computing, 56% of respondents said they had faced at least one software vendor change in the past 12 months: an additional charge for AI, a price increase at renewal, or a feature moved into a more expensive tier. Another 38% said budget overruns had forced them to delay, scale back, or cancel an AI project during the same period. Only 12% cited a lack of computing capacity.

Software prices are rising

Among respondents who had encountered at least one of these changes, 40% said the extra costs amounted to at least 10% of their organization’s annual software spending.

On July 1, Microsoft raised list prices for some commercial Microsoft 365 subscriptions. Prices for enterprise plans with Teams went up by 5% to 13%, while Microsoft 365 F1 without Teams rose by up to 43%. Microsoft linked the changes to new AI, security, and management features. For existing commercial customers on affected plans, the new prices take effect at the first subscription renewal after July 1, regardless of how the subscription was purchased.

Starting August 1, 2025, Salesforce raised list prices by an average of 6% for the Enterprise and Unlimited editions of Sales Cloud, Service Cloud, Field Service, and some industry cloud products. The company attributed the increase to ongoing product development and customer value, while also changing its Agentforce offerings.

Cloud providers have built up $2.48 trillion in orders

The four providers’ combined backlog in their latest reports totaled about $2.48 trillion:

Companies calculate this figure differently, and the amounts include contracts unrelated to AI. In January, Microsoft CFO Amy Hood told analysts that OpenAI accounted for about 45% of the company’s commercial orders at the time. According to The Wall Street Journal, OpenAI signed a five-year, $300 billion computing capacity contract with Oracle. Amazon’s latest filing with the SEC also lists $138 billion in AWS commitments to OpenAI.

Alphabet forecasts capital expenditures of $195–205 billion in 2026. The company’s CFO, Anat Ashkenazi, expects them to rise significantly in 2027.

Memory prices have more than doubled

On October 21, Gartner said demand from data center systems had more than doubled memory prices over the past 12 months. The cost increases have spread across supply chains for storage systems, servers, networking equipment, computers, and devices. Gartner analyst John-David Lovelock expects “memflation” to continue through 2027, after which hardware prices will begin to stabilize.

On April 29, Hood told investors that Microsoft expected about $190 billion in capital expenditures for calendar year 2026, with roughly $25 billion of that amount attributable to more expensive components. In July, the company lowered its forecast to about $175 billion after reclassifying data center leases. In October, it lowered the forecast again, to about $170 billion, citing delays in getting capacity from third-party providers.

AWS has raised prices for reserved GPUs four times

AWS changed prices for EC2 Capacity Blocks for ML, a product for reserving GPUs in advance, several times in 2026. Prices rose about 15% on January 4, about 10% for some configurations in April, and about 20% on July 1. On October 7, AWS raised prices again, by about 15% for most GPU configurations in the US. AWS’s pricing page says reservation prices are updated regularly to reflect supply and demand.

Of the 128 survey participants whose organizations use GPUs or accelerators, 73 estimated average utilization in production at 50% or less, 13 reported utilization above 50%, and 42 do not measure it. Among the 114 organizations that reserve capacity in advance, 31% spend more than a quarter of their AI computing budget on unused reservations. Another 22% do not know what share of their spending goes toward unused capacity.

Workflow costs could rise

Gartner expects token prices to fall by 95% by 2030. But on August 17, the company also forecast that inference costs for a single AI-agent workflow would increase more than fivefold by 2028.

Will Sommer, a senior director analyst at Gartner, warned that cheaper tokens alone would not justify AI spending. Each new generation of AI capabilities, he said, will require more tokens, which may also cost more.

Infrastructure requires engineering resources

Working with AI infrastructure requires specialized engineering expertise. At VentureBeat’s Agentic Infrastructure Stack Summit, held on October 28 in San Francisco, Upwork Vice President of Platform Infrastructure Engineering Sudhanshu Bist called GPU capacity planning a difficult task. The latest chips are hard to obtain on demand, so capacity has to be reserved in advance. Ordering equipment to meet peak demand means it sits idle during slower periods.

Discussing data on idle GPUs, WEKA Chief AI Officer Val Bercovici said operating this kind of infrastructure is a significant engineering challenge for companies whose core business is unrelated to it. Such organizations may not have the necessary specialists, he said, and working with AI and GPU infrastructure is outside their area of expertise.

Changing models can also require additional engineering work. In VentureBeat Intelligence’s September report on AI-agent orchestration, 59% of respondents whose organizations had changed the model in a live application reported a drop in performance. They had to rework the application or return to the previous model.

Most expect AI budgets to grow

Of the 156 survey participants, 124 expect their organizations’ AI budgets to increase over the next 12 months; 29 of them forecast growth of more than 50%. Among those 124 respondents, 27% assess whether spending is justified based on measurable financial returns they can substantiate to the finance department.

Across the full sample, 40 people use financial returns, 69 use measurable operational outcomes such as time saved, and 40 rely on management’s assessment without formal measurement.

Respondents were also asked how they assess returns on AI infrastructure. Only 21% track them rigorously. Of the 91 participants who measure returns rigorously or partially, only 34% have a metric that has been reviewed and approved by the finance department.

What companies should check

The VentureBeat Intelligence survey was conducted in October 2026. Participants included employees of organizations with at least 100 staff and consultants working with such organizations. It received 172 responses, 156 of which met the study criteria. Among eligible participants, 87% came through a research panel, and 10% were consultants answering about their primary client. Participation was voluntary, so the results describe only this sample. The July and August editions of this series have also been published on VentureBeat.

Dataist’s take: Rising costs stem from several factors: expensive components, payments for unused reservations, and the additional expense of supporting AI applications. The survey shows that companies more often cite budget overruns than a lack of computing capacity as the reason for delaying or canceling projects. At the same time, the findings apply only to survey participants, and many organizations still cannot substantiate returns on infrastructure with figures approved by the finance department.