Hyperscaler spending keeps rising while its growth rate collapses, and that is the trade

The single most important number in technology markets right now is not the size of hyperscaler capital spending. It is the second derivative: how fast that spending is growing, and whether the growth rate is holding.

The absolute figures are enormous and still climbing. The five largest hyperscalers are expected to spend somewhere in the region of $775 billion to $800 billion on artificial intelligence infrastructure in 2026, and consensus has that rising to roughly $939 billion next year. Any headline built on those numbers writes itself, and most of them have been written.

The growth rates underneath tell a different story. Capital spending among that group grew about 56 percent in 2024 and about 73 percent in 2025, and is projected to grow around 90 percent in 2026. Consensus for next year is 28 percent. Spending is still increasing, substantially, and the rate at which it increases is set to fall by roughly two thirds.

That distinction matters more than it sounds because of how the companies exposed to this spending are valued. A supplier growing revenue 70 or 90 percent a year is priced on the assumption that the growth continues, and multiples expand to reflect it. A supplier growing 28 percent is a good business, frequently a very good one, and it is not the same security. The transition between those two states is where multiple compression happens, and it happens while revenue is still rising, which is what makes it so consistently surprising to people watching the top line rather than the rate.

It is worth being careful about what a deceleration to 28 percent would actually mean. It is not a bust. It is not a cancellation of the buildout. Nearly a trillion dollars of annual capital spending is an extraordinary level of investment by any historical standard, and it would sustain an entire supply chain comfortably. The question is narrower and more specific: which companies are priced for that number, and which are priced for the number before it.

Canadian investors have more exposure to this than the size of the domestic technology sector suggests, and it is concentrated rather than spread. Celestica is the clearest case, having repositioned itself around artificial intelligence data centre hardware and advanced networking, with guidance revised upward through this year on that demand. When the market re-rates AI infrastructure suppliers, that name moves with it, and its weight in Canadian technology indices means the sector moves too.

The counterweight sits in the enterprise software names, and the contrast is instructive. Constellation Software has been trading around 19 times forward earnings against an earnings growth rate near 57 percent, which is a very different proposition from a hardware supplier priced on hyperscaler order flow. Its growth comes from acquiring and operating vertical market software businesses whose customers are municipalities, clubs, clinics and small operators. Those customers do not adjust their spending because a hyperscaler moved a data centre from 2027 to 2028.

That is the useful frame for reading Canadian technology exposure at the moment. One part of it is a levered call on a capital expenditure cycle that is decelerating from an extraordinary base. Another part is a set of businesses whose revenue has almost nothing to do with that cycle. Both have been rising, and they have been rising for unrelated reasons, which means they will not fall for the same reason either.

There is a further point about where the spending goes that changes who benefits as the cycle matures. Roughly a quarter of hyperscaler capital expenditure flows to Nvidia by one bank's estimate, which is a remarkable share for a single supplier and also means around three quarters goes elsewhere: to networking, to power distribution and cooling, to construction, to land, to the electrical infrastructure that has come to dominate the Alberta conversation. As buildouts move from chip acquisition toward getting facilities actually energised, the mix shifts toward that second group.

The genuine risk in the current setup is not that spending falls. It is that spending continues while returns on it disappoint. Hyperscalers are funding this from operating cash flow, which is why it has been sustainable so far, and each of them is answerable to shareholders who will eventually ask what the return was. If revenue attributable to artificial intelligence products does not begin to justify the capital, the adjustment will not be gradual, and it will happen to suppliers before it happens to the buyers.

The signals worth tracking are more specific than the headline totals. Whether hyperscalers reaffirm or trim capital guidance on their next reporting cycle. Whether supplier order backlogs are lengthening or merely holding. Whether announced facilities are converting into energised megawatts, which is the physical evidence that money is actually being spent rather than committed. And whether the gap between announced projects and connected load keeps widening, because in Alberta that gap is already fifteen to one and it is a better indicator of real activity than any dollar figure.

None of this is an argument to be bearish on a sector that is still growing. It is an argument for knowing which of two very different things you own. A Canadian portfolio holding technology exposure today probably holds both a capital expenditure cycle proxy and a set of businesses insulated from it, and the two have been indistinguishable on the way up. They will not be indistinguishable on the way down, and the moment to understand the difference is while both are still working.