Ask how much companies are spending on AI data centers in 2026 and the honest answer is: more than almost anyone predicted a year ago. Microsoft, Alphabet, Amazon, and Meta are together on track to spend around $725 billion in capital expenditures this year, up roughly 77% from 2025, according to analyst tracking of the four companies' own guidance. Layer in OpenAI's separate infrastructure commitments and the number gets even harder to process. This is not a story about whether AI is useful. It is a story about whether the spending behind it is sustainable, and that question does not have a clean answer yet.
How much are companies actually spending on AI data centers
The four largest cloud and AI companies have each given specific 2026 capex guidance. Amazon is projecting about $200 billion. Microsoft set its 2026 capex at roughly $190 billion, well above what Wall Street analysts had expected. Alphabet raised its own guidance to as much as $190 billion. Meta increased its full-year range to top $145 billion. Add them up and you get close to $725 billion from four companies in a single year, more than the entire annual GDP of most countries.
That figure does not even include OpenAI, which does not build its own data centers at that scale but has signed compute and infrastructure agreements that analysts estimate exceed $1 trillion in total nameplate value through the mid-2030s, spread across deals with Oracle, Microsoft, Nvidia, AMD, Amazon, and others. The single largest piece is the Stargate project with Oracle and SoftBank, aimed at roughly 10 gigawatts of computing capacity and around $500 billion in spending.
Where the money is actually going
Most of this spending is not abstract. It is land, concrete, cooling systems, electrical substations, and, above all, chips. Nvidia's GPUs remain the single most expensive line item in almost every one of these builds, which is part of why companies have been willing to sign multi-year, multi-billion-dollar supply agreements just to guarantee they get enough of them.
Companies have also cited rising costs beyond the chips themselves. Memory pricing has gone up. So has competition for the land, electrical grid capacity, and skilled construction and engineering labor needed to physically build these facilities fast enough to keep up with demand. In some regions, the bottleneck is no longer money at all. It is how quickly a new substation or transmission line can be built.
The power problem behind the spending
The electricity these data centers need is growing almost as fast as the spending itself. Gartner estimates worldwide data center power demand will rise about 27% in 2026, reaching 132 gigawatts, up from 104 gigawatts in 2025. In terms of actual electricity consumed, that translates to roughly 565 terawatt-hours in 2026, up from 447 terawatt-hours the year before. AI-optimized servers are expected to account for close to a third of all data center power consumption this year.
The International Energy Agency projects this keeps climbing well past 2026, with global data center electricity consumption potentially reaching 945 terawatt-hours by 2030. For comparison, that would put data centers in the range of consuming as much electricity annually as a mid-sized industrial country. That is the physical reality behind the spending numbers: this is not just a financial bet, it is a bet that enough electricity can actually be generated and delivered fast enough to run all the hardware being bought.
Is there an AI bubble, and who thinks so
This is where the real disagreement starts. CNBC surveyed 40 tech leaders and analysts on whether current AI spending qualifies as a bubble, and the results split along fairly predictable lines. On the skeptical side, GMO co-founder Jeremy Grantham co-authored a paper arguing the situation looks like an extreme bubble more than a new golden era. Bridgewater founder Ray Dalio and JPMorgan CEO Jamie Dimon have both said publicly that something about the pace of spending looks off.
On the other side, Goldman Sachs and J.P. Morgan analysts have argued the spending is justified because it is being funded by highly profitable companies and matched by real, fast-growing cloud AI revenue. They point to infrastructure suppliers like Nvidia, Cisco, and HPE posting results that look less like a market running out of steam and more like one that cannot build fast enough to meet demand. Both camps are looking at the same $725 billion number and drawing opposite conclusions, which is a large part of why the debate has not settled.
The circular financing question nobody can fully untangle
One specific concern keeps coming up among more skeptical analysts: circular financing. The pattern looks something like this: Nvidia invests billions into OpenAI. OpenAI uses some of that money to buy cloud computing capacity from Oracle. Oracle, in turn, uses that revenue to buy more chips from Nvidia. Money that leaves Nvidia's balance sheet as an investment eventually comes back around as revenue, and by the time it completes the loop, it becomes genuinely difficult to say who is actually paying for what.
Bernstein Research analyst Stacy Rasgon flagged this directly after reports that Nvidia could guarantee as much as $250 billion tied to one such arrangement, noting it would "clearly fuel circular concerns." Nvidia shares dropped roughly 4.5% intraday on that news. The worry among critics is not that these deals are illegal or even unusual, it is that they can make demand and revenue look stronger than they actually are, making it harder for investors to tell what is organic growth versus companies effectively financing their own customers.
The accounting fight over how long a chip actually lasts
A separate but related concern involves depreciation. Investor Michael Burry, known for his early bet against the 2008 housing market, has argued that companies including Meta, Oracle, Microsoft, Google, and Amazon are stretching out how long they claim their AI chips will remain useful, which lowers the depreciation expense they report each year and inflates earnings without changing any actual cash flow. Burry estimated this accounting choice could understate depreciation industry-wide by roughly $176 billion between 2026 and 2028, and specifically suggested Oracle's and Meta's profits could be overstated by around 27% and 21% respectively by 2028.
Microsoft and Google reportedly tell auditors their chips will remain productive for six years, while Meta uses five and a half. The companies defend these as audited, carefully modeled projections. Goldman Sachs pushed back on Burry's framing too, pointing out in an April 2026 report that older Nvidia A100 and H100 chips are still commanding rental prices in the secondary market consistent with five-to-six-year useful lives, which suggests the market itself is not pricing in the rapid obsolescence Burry is describing. Neither side has definitively won this argument yet, and it will not be settled until enough of these chips actually reach the end of their claimed lifespan.
The takeaway
What makes this moment hard to summarize honestly is that the bulls and the bears are not disagreeing about the numbers, they are disagreeing about what the numbers mean. Nobody credible disputes that roughly $725 billion is being spent this year, or that data center power demand is climbing by double digits annually. The disagreement is over whether that spending is being matched by real, durable revenue, or whether some of it is being recycled between the same handful of companies to make growth look more organic than it is. Both things being at least partly true at once, real infrastructure need and real financial engineering, is exactly why this is not a story with a tidy ending yet. Anyone with money in AI-adjacent stocks, or a business that depends on AI tools staying cheap, has a real reason to keep watching how this resolves over the next year or two, not just take either side's word for it.
