The four largest American tech companies will spend about $725 billion building AI infrastructure in 2026, and the spending is usually described as a moat too deep for anyone to cross. The accounting that makes it look like a moat assumes the equipment lasts a long time. The equipment does not. One company has already started admitting it.
Amazon, Alphabet, Microsoft, and Meta plan to spend roughly $725 billion on capital expenditure in 2026, up about 77 percent from the $410 billion they spent in 2025. Amazon alone is guiding to around $200 billion, Alphabet to $180 to $190 billion, Meta to $125 to $145 billion, and Microsoft to roughly $190 billion. About three quarters of the total, on the order of $450 billion, goes to AI infrastructure: chips, servers, and the buildings to hold them. The standard reading of these numbers is that they are a wall. Nobody outside this handful of firms can write checks this large, so the incumbents have bought themselves a position no challenger can reach.
The wall reading borrows its confidence from history. The railroad and fiber buildouts were also ruinous, also funded with more money than seemed sane, and also bankrupted plenty of the people who attempted them. But the survivors ended up owning something that kept paying. Track laid in the 1880s was still carrying freight in the 1920s. Fiber buried in 1999 is still lit today. The capital was spent once and then collected against for decades. That is what makes a buildout a moat: the spending stops and the asset keeps earning.
The asset at the center of this buildout is a graphics chip, and Nvidia now ships a new architecture every year. Hopper arrived in 2022, Blackwell in 2024, Blackwell Ultra in 2025, and Rubin in 2026. A flagship chip bought near the top of the market in 2024 is two generations old before it has finished paying for itself. The toll booth in this story does not get built once. It has to be torn down and rebuilt every couple of years, because the thing collecting the toll keeps becoming the slow option.
Here is where the accounting gets strange. On the books, this fast-aging equipment is treated as more durable than ever. Microsoft stretched the assumed useful life of its servers and network gear from four years to six, effective in fiscal 2023, a change that moved about $3.7 billion of depreciation out of that single year and into later ones. Alphabet, Meta, and Amazon all drifted the same direction. A longer assumed life means a smaller depreciation charge each year, and a smaller charge means a larger reported profit. So in the exact period when the hardware cycle was speeding up, the financial statements declared the hardware longer-lived.
Depreciation is just the mechanism by which the cost of an asset is charged against the earnings it produces. Stretch the assumed life and you spread the cost thinner, which lifts today's profit by borrowing from tomorrow's. One critique circulating among analysts puts the gap at roughly $176 billion of understated depreciation across the industry between 2026 and 2028, on the assumption that the real economic life of these chips is closer to two or three years than to six. Whether that exact figure holds, the direction is not in dispute: if the chips wear out faster than the schedule says, reported profits today are partly profits pulled forward from a bill not yet shown.
The reason I think the treadmill reading is right and the wall reading is wrong is that one of these companies has already blinked. Amazon extended its server life to six years in early 2024, in line with everyone else. Then it reversed part of the move, shortening the assumed life of a subset of its servers and networking equipment back to five years, effective January 2025. That reversal cost it about $677 million of operating income over the first nine months of 2025. A company does not volunteer to lower its own earnings on a whim. It does so when the equipment in the racks is telling it something the depreciation schedule was too generous to admit. Amazon shortening lives is reality forcing its way back onto the books, one footnote at a time.
So the spending looks like a moat and works like a treadmill. A moat is dug once. A treadmill is paid for continuously, and the moment you stop paying you fall off the back. The $725 billion is not a barrier that keeps rivals out. It is a subscription to staying current, and the renewal notice arrives every two years regardless of whether the last round of chips earned back its cost. The incumbents are not protected by the size of the check. They are committed to writing it again.
The framing survives anyway because the gap between this spending and the cash these companies actually generate is increasingly bridged with debt. Borrowing makes the treadmill feel like an asset, because you can finance next year's chips with someone else's money and let the revenue catch up. That works for exactly as long as the cash the chips throw off can cover both the interest on the last batch and the purchase of the next one. The first buildout in this list that cannot do both at once is where the analogy to railroads finally breaks in public.
My conviction is narrow and falsifiable. Over the next two years I expect more useful-life shortenings, not extensions, as Amazon's reversal turns out to be the leading edge rather than an outlier. I expect the share of this capex funded by debt to keep climbing. And I expect the real test to arrive not when AI demand falls, which would be obvious, but when it merely stops accelerating, because a treadmill only feels like a moat while you are still speeding up. The companies running it own the fastest computers in the world for now. What they do not own is a road that keeps earning after they stop laying it.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)