I told you so. Not that you needed to be any sort of seer to make this particular prediction.
Three or so weeks back, I noted that bottom-up consensus estimates for hyper-scaler capex in 2026 were rising already, where “already” meant that not even a month into the new year, projections for AI spending were up materially.
In the course of documenting those upward revisions, I made a “bold” (note the scare quotes) forecast: Mega-cap capex guesstimates would increase further and in very short order.
Fast forward to mid-February and those estimates are indeed higher. Markedly higher. The projection for 2026 is now $660 billion, up $120 billion compared to consensus as it stood headed into earnings season.
Note from the (staggering) chart above that the forecast for 2027 is now approaching $800 billion. For context, that figure a decade ago was just $33 billion. Try to wrap your mind around that increase.
The chart also shows you the extent to which these companies are approaching the absolute limit in terms of their capacity to fund this spending out of cash flow.
Of course, they’re not funding it all out of cash (they’re borrowing heavily), but the point is just that we’re almost to a place where they couldn’t even if they committed every last spare penny.
The figure below, from Goldman’s Ben Snider, gives you some additional context.
The projected capex/cash flow ratio for the hyper-scalers this year exceeds the same metric for Technology, Media, and Telecommunications stocks at the height of the dot-com bubble (although not the peak for Telecoms specifically).
Should you be worried? Well, that depends on how much you value every dollar of buybacks. “As capex has surged, S&P 500 buybacks have declined, falling by 7% YoY during Q4,” Snider wrote, noting that last quarter marked a third straight quarter of buyback “stagnation.”
“We expect the increasing scarcity of free cash flows and buybacks will strengthen the premium for companies focused on returning cash flows to shareholders,” he added.
But as discussed here last week, corporate America will remain a huge buyer of its own stock. So, alarming as the charts above certainly appear, rest assured the corporate bid isn’t going anywhere.




I think the limitation in rate of growth in electrical power generation, and the limitation in rate of growth of memory chip availability (until the new DRAM fabs start actual production) will be the limiting factor in datacenter growth. The demand for compute will exceed this capacity, particularly from physical AI: robotics and autonomous driving.
Remember “the War on Christmas” which so distressed the GOP?
Well, over the last few days it almost looks like the Whitehouse is launching a new war = the War on AI!
First we had had the newly exhumed Peter Navarro demanding that datacenters provide their own electricity. He even mentioned water usage!
And then today it was reported that Pete Hegseth is going after Anthropic. This may be a negotiating ploy, but it also signals that AI is no longer an untouchable sacred cow in the GOP.
They are following the lead of Ron DeSantis and Josh Hawley in realizing that there is not massive voter support for the continued build out of datacenters if they raise voters electricity rates to help an industry based on the promise of allowing companies to lay off people. (Otherwise why would any large corporation even use it?)
https://www.msn.com/en-us/money/markets/trump-trade-adviser-navarro-says-administration-may-force-data-center-builders-like-meta-to-internalize-costs/ar-AA1WpQUD?ocid=socialshare
https://www.marketwatch.com/story/hegseth-reportedly-considering-severe-penalty-on-anthropic-as-negotiations-stall-16dae92f?st=b7mZ4q&reflink=desktopwebshare_permalink