Will $5 trillion in AI infrastructure investment be enough? Cloud providers facing that question must also yield a return, prompting questions about pricing hikes for users.
Cloud infrastructure providers are making massive investments in AI to meet seemingly insatiable demand, and some have recently raised prices quietly for some services -- will the continued spending spree mean more of the same?
That was the question some industry watchers were asking this week as Oracle unveiled a plan to finance its $300 billion deal with OpenAI through a $50 billion combination of stock sales and debt, while Google execs pledged to double the company's capital spending to compete in the AI race. These efforts are part of an estimated $5 trillion in investment J.P. Morgan research expects to be invested globally in AI infrastructure, data centers and power systems by 2030.
Meanwhile, leaders at enterprise IT vendors and frontier AI companies are wondering aloud whether even that will be enough.
"People are underestimating the capacity that's going to be needed, even now," said Jeetu Patel, president and chief product officer at Cisco, during an onstage discussion with OpenAI CEO Sam Altman at Cisco's AI Summit Feb 3. Patel cited the $5 trillion JP Morgan estimate.
“Maybe, if that really gets spent quickly, that will be enough," Altman replied.
Patel expressed doubt about that during his introductory remarks at the summit.
"We just don't have enough power, compute and network bandwidth now, memory [and] data center shells to go out and satiate the needs of AI," he said.
The current pace of AI infrastructure development also isn't enough for Amin Vahdat, senior vice president and chief technologist for AI infrastructure at Google. He said during a separate onstage discussion with Patel that enterprise hardware refresh cycles are too slow to keep up with AI's true capabilities -- and are likely to remain that way.
"If we could only cut down the lead time of hardware design to delivery by a factor of 10 right now, from the time we say, 'Hey, here's an amazing new piece of hardware,' to the time where we have it at scale in the data center, " Vahdat said. "If we could, in a 10-year period, get that down to, let's say, three months … I don't know how to do it. I don't know if anybody does, but if we could get down to three months, from an efficiency, capability and change-the-world perspective, it would be a radically different place."
Instead, cloud providers such as AWS are not only adding more data center space, power and cooling to accommodate the latest AI infrastructure hardware, but also maintaining previous generations of hardware for long periods.
“We actually are completely sold out of and have never retired an A100 server," said Matt Garman, CEO of AWS, during another onstage discussion at the summit. "Part of where the [AI chip] industry today is getting so much efficiency gain … is [by] reducing the floating-point accuracy."
Thus, some customers with high-performance computing workloads can't move to newer generations of Nvidia chips, such as Blackwell, because their calculations aren't as precise, Garman said.
"A lot of the world, actually, and some of the old applications rely on that level of precision," he said.
Garman also predicted that investment in lower-cost AWS Trainium chips will help boost AWS margins and lead to lower customer prices long-term.
'All that spend absolutely puts pressure on margins'
For some industry observers, however, it's impossible for the scale of vendors' investments not to affect how much they charge for services, at least in the short term.
That trend we've seen over the last few years of declining prices for cloud services could well be at an end.
Jim FreyAnalyst, Omdia
"Demand is outstripping supply for data center and compute capacity, which is driving massive investments into all supporting technology and systems, and all that spend absolutely puts pressure on margins," said Jim Frey, an analyst at Omdia, a division of Informa TechTarget. "When demand outstrips supply, prices naturally rise. … We are seeing basic economics at play here."
The takeaway for IT buyers is to prepare for potential cost increases in cloud infrastructure, Frey said. Enterprises also face rising costs for flash data storage due to overwhelming AI infrastructure demand in 2026.
"That trend we've seen over the last few years of declining prices for cloud services could well be at an end, at least while this big wave of buildouts to support AI takes place," Frey said. "Will it ever return? Probably yes, but not for at least a couple more years."
Another analyst predicted this could accelerate ongoing cloud repatriation, and boost the adoption of FinOps and observability tools, along with cloud auto-scaling features, to rein in costs.
"Can IT buyers fight or resist the price increase? No. Can they better prepare? Yes," said Naveen Chhabra, an analyst at Forrester Research. "Over the past several years, we have seen a lot of waste in cloud spend … because they lack the maturity and controls to make the necessary changes to their infrastructure in a timely manner to control the cloud waste."
Cloud pricing shifts sway IT decisions
In the last two months, AWS and Google pricing fluctuations have prompted IT pros to consider how they might offset increased costs.
In mid-2025, AWS slashed prices for some of its EC2 instances with Nvidia GPUs by up to 45%, but in early January, it also upped the price for its EC2 Capacity Blocks for machine learning (ML), which enable users to pay an upfront fee to reserve compute capacity, by about 15%.
For one AWS customer, the cost of high-end GPU compute was already too high, prompting it to swap out the default Linux memory allocator on smaller, cheaper EC2 instances to keep audio and video encoding workloads there.
"For anyone using Linux, the memory allocator by default is [The GNU C Library], but we optimized memory usage on those applications using jemalloc," said Anuj Tyagi, a senior site reliability engineer at a communications company he requested not be named because of policies prohibiting him from representing it in the press. "That reduced the memory usage of the application by half."
However, that kind of reconfiguration isn't for everyone, and not all Linux operating systems support jemalloc, Tyagi said.
Later in January, Google Cloud Platform (GCP) doubled the price per gibibyte (GiB) for data transfer out of its content data network (CDN) Interconnect, Direct Peering and Carrier Peering network services in North America, from $0.04 per GiB to $0.08 per GB. The pricing change does not apply to general GCP internet rates for Premium and Standard licensing tiers, according to a Google spokesperson.
"We continue to offer free data transfer for customers exiting Google Cloud, as well as for those performing 'in-parallel' processing, in response to the principles of cloud interoperability and choice outlined in the EU Data Act," the spokesperson said.
The spokesperson attributed the pricing change to "significant investments in our global infrastructure to provide high-throughput, low-latency connectivity for our customers."
The CDN Interconnect and network peering pricing changes will affect customers in specific scenarios, such as those who wish to combine Google's CDN services with third-party providers or establish secure high-speed connections between on-premises and cloud infrastructure.
"I don't expect a great deal of short-term impact -- because we're all-in on Google Cloud, we don't export much data," said a chief architect at a large technology company in the Southwest, who spoke on condition of anonymity. "But keeping network traffic from a company's internal network from traveling over the public internet on the way to and from Google's network, for customers of Google Workspace, is a common security desire in large corporations, especially more conservative ones."
Beth Pariseau, a senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism. Have a tip? Email her.
Dig Deeper on Cloud infrastructure design and management
