AI Inference Costs Could Rise Fivefold by 2028 Despite Cheaper Tokens

AI is supposed to make businesses more productive and less expensive to operate. Yet a growing number of companies are discovering an uncomfortable reality: adopting AI can increase technology costs before it delivers measurable savings.

The pressure is coming from several directions at once. Memory prices have more than doubled in 12 months. AWS has raised reserved GPU prices four times in 2026. Microsoft 365 subscriptions have increased by as much as 43% on selected plans. Meanwhile, cloud providers are committing enormous sums to infrastructure, and AI developers are locking themselves into long-term computing contracts.

The result is a cost problem that extends far beyond the companies building AI models. Businesses that simply want to use AI are increasingly paying for the infrastructure race through higher software subscriptions, cloud capacity and hardware expenses.

The AI Infrastructure Race Is Becoming a Corporate Budget Problem

The scale of investment is extraordinary.

In an amended IPO prospectus filed November 2, Anthropic disclosed approximately $895.4 billion in purchase and lease commitments, substantially all related to computing infrastructure. Around $214 billion is subject to termination rights or conditions, and the commitments extend through 2045.

Anthropic reported $33.2 billion in revenue during the first nine months of 2026. However, some agreements require payment regardless of actual resource utilization. The company acknowledged that it must expand its customer base and revenue rapidly to generate enough cash flow to meet those obligations.

Alphabet is also experiencing the financial pressure. After spending $51.4 billion on capital expenditure in the third quarter, it reported negative free cash flow for a second consecutive quarter. It also sold $25 billion in bonds in August amid its infrastructure expansion.

These figures illustrate the economics behind the AI race. Building advanced models requires enormous upfront investment, and the infrastructure must be financed before all of its capacity can generate revenue.

Cloud Providers Are Sitting on Nearly $2.5 Trillion in Backlog

Microsoft, Oracle, Amazon and Google Cloud reported approximately $2.48 trillion in combined backlog in their latest results.

These figures use different backlog definitions and include contracts unrelated to AI. Nevertheless, they show how much future business major cloud providers have already committed to deliver.

The relationships between AI developers and infrastructure providers are becoming particularly important. Microsoft CFO Amy Hood said OpenAI represented about 45% of Microsoft's commercial backlog at the time of her January comments. The Wall Street Journal reported a five-year, $300 billion computing contract between OpenAI and Oracle. Amazon's latest SEC filing also identifies $138 billion in AWS commitments from OpenAI.

This creates a powerful commercial cycle. AI developers need more computing capacity, cloud providers invest to supply it, and long-term contracts help justify the investment.

But those commitments also create pressure to recover costs.

Microsoft 365 and Salesforce Are Passing More Costs to Customers

The price increases are no longer limited to cloud infrastructure.

Microsoft raised list prices on selected commercial Microsoft 365 subscriptions starting July 1, 2026. Enterprise suites that include Teams increased between 5% and 13%, while Microsoft 365 F1 without Teams rose by as much as 43%.

Microsoft justified the changes through new AI, security and management features. Existing commercial customers on affected plans encounter the new prices at their next renewal after July 1, regardless of purchasing channel.

Salesforce had already increased list prices for Enterprise and Unlimited editions across several cloud products by an average of 6% in August 2025, citing ongoing innovation and customer value as it introduced Agentforce packaging.

VentureBeat Intelligence's October survey shows how widespread this trend has become. Fifty-six percent of respondents said a software vendor had added an AI surcharge, increased renewal prices citing AI or moved previously included features into a more expensive tier during the past year.

Among respondents reporting at least one such change, 40% said the additional cost represented at least 10% of their annual software spending.

For businesses, AI is increasingly becoming part of the price of software they already use.

Memory and Reserved GPUs Are Becoming More Expensive

Hardware costs add another layer of pressure.

Gartner reported that memory prices had more than doubled over the previous 12 months, driven by demand from data-center systems. The increases are spreading through storage, servers, networking equipment, PCs and other devices. Gartner expects this pressure, described as “memflation,” to continue through 2027 before hardware prices begin to stabilize.

GPU access is also becoming more expensive.

AWS increased prices for its EC2 Capacity Blocks for Machine Learning reserved GPU product approximately four times in 2026. The increases included around 15% in January, roughly 10% in April for some configurations, about 20% in July and approximately 15% in October for most U.S. GPU configurations.

This matters because companies often reserve capacity in advance to ensure that computing resources are available when needed.

However, reservation does not guarantee efficient utilization.

Among 114 survey respondents whose organizations reserved computing capacity in advance, 31% said more than a quarter of their AI compute spending went toward reserved capacity they did not use. Another 22% did not know how much capacity was unused.

The problem is not simply expensive GPUs. It is paying for expensive GPUs that remain idle.

Cheaper Tokens Will Not Necessarily Make AI Agents Cheaper

One of the most important cost assumptions in AI is that inference will become cheaper as models and hardware improve.

Token prices are expected to fall substantially. Gartner predicts token costs could decline 95% by 2030.

Yet Gartner also forecasts that inference costs per agentic workflow will increase more than fivefold by 2028.

The explanation is straightforward. More capable AI agents may consume more tokens, perform more reasoning steps and use more expensive models to complete increasingly complex tasks.

A lower price per token does not automatically mean a lower cost per successful business outcome.

Companies must account for token consumption, workflow complexity and task success rates. Measuring only the headline price of an AI model can create a misleading picture of operating expenses.

The Hidden Engineering Costs Are Just as Important

AI infrastructure requires specialized expertise.

At VentureBeat's Agentic Infrastructure Stack Summit, Upwork's Sudhanshu Bist described GPU planning as a complex problem. Companies often need to commit to capacity before they know exactly how much they will use. Reserving for peak demand can leave expensive resources idle during quieter periods.

WEKA's Val Bercovici highlighted another challenge: businesses whose core operations are not infrastructure management may lack the talent needed to operate GPU environments efficiently.

Model changes create additional costs. In VentureBeat Intelligence's September Agentic Orchestration report, 59% of respondents whose organizations switched models in live applications said performance declined, forcing them to rework the application or switch back.

The cost of AI therefore includes more than subscriptions, hardware and tokens. It also includes engineering time, integration work, monitoring and operational recovery.

Growing AI Budgets Do Not Guarantee Financial Returns

Despite rising expenses, AI investment continues to expand.

Of 156 survey respondents, 124 expected their AI budgets to increase over the following 12 months. Twenty-nine expected growth of more than 50%.

But financial accountability remains weak.

Only 27% of those expecting higher AI budgets judged the value of spending through a measured financial return they could defend to their finance team. Across the full sample, 69 respondents used operational outcomes such as time saved, while 40 relied on leadership judgment without a formal measurement.

Infrastructure spending was even harder to justify. Only 21% of respondents tracked infrastructure returns rigorously. Among the 91 who tracked returns rigorously or partially, just 34% had a figure reviewed and approved by their finance organization.

That is a serious management gap. Companies are increasing investment faster than they are improving their ability to prove what that investment produces.

Businesses Need to Manage AI Like an Operating Cost

For large enterprises, the immediate priority is cost visibility. For small and midsize businesses, it is avoiding unnecessary complexity and commitments.

Several actions can make a practical difference:

• Review Microsoft 365 renewals. Identify affected subscriptions and calculate the impact before the next renewal.

• Track every vendor increase. Record AI surcharges, renewal changes and features moved into higher-priced tiers.

• Measure GPU utilization. Do not reserve additional capacity without understanding how much existing capacity is actually used.

• Check AWS reservation prices. Capacity Block rates have changed repeatedly in 2026.

• Calculate cost per successful task. Include token consumption, workflow execution and failure-related rework.

• Require finance-reviewed returns. Connect AI spending to measurable savings, revenue or operational improvements.

Smaller companies should be particularly careful about buying infrastructure they cannot efficiently operate. Using managed AI services or simpler workflows may be more economical than building and maintaining a dedicated GPU environment.

The objective should not be to use the largest model or deploy the most agents. It should be to solve a business problem at a cost that makes commercial sense.

The Next AI Advantage May Belong to Companies That Control Costs

The AI industry is entering a phase in which capability is only part of the competitive equation. Infrastructure commitments, memory shortages, cloud pricing and software renewals are reshaping the economics of adoption.

For technology providers, higher prices may help finance the enormous investment required to build AI systems. For customers, however, those increases create pressure to demand clearer value.

The central question for businesses is changing. It is no longer simply whether AI can improve productivity. It is whether the productivity gains exceed the full cost of deploying, maintaining and scaling the technology.

Companies that measure utilization, control vendor commitments and connect AI spending to financial outcomes will be better positioned to benefit from the next wave of automation.

In an increasingly expensive AI market, disciplined execution may become a more durable advantage than access to the most powerful model.

AI Inference Costs Could Rise Fivefold by 2028 Despite Cheaper Tokens

AI infrastructure costs, enterprise AI spending, GPU pricing, memory chip prices, Microsoft 365 pricing, AWS reserved GPUs, AI inference costs, agentic AI workflows, cloud infrastructure, AI software surcharges, Anthropic infrastructure, AI budget management, GPU utilization, AI return on investment