Twenty billion dollars is a powerful incentive. That is the amount by which Amazon raised its 2026 capital expenditure estimate, from $200 billion to $220 billion, and the company explicitly identified the rising cost of memory chips as the driver. Not new datacenters. Not more racks. The same buildout, repriced.
Last week we covered TSMC lifting 2nm and 3nm output and an earnings season that had capital markets repricing the buildout. Those pieces were about capacity and profit. This week the story is the number itself — and what it is actually measuring.
TrendForce now projects that the top nine cloud service providers will spend more than $957 billion in 2026, up 106% year over year, with 2027 surpassing $1.5 trillion. The headline reads as an accelerando. Read the line items and it reads as an invoice.
The headline reads as an accelerando. Read the line items and it reads as an invoice.
UBS did the decomposition, and the arithmetic is unforgiving. Memory spending climbs from $71 billion in 2025 to $367 billion in 2026 to $923 billion in 2027. Everything else — compute, networking, power, shells — falls from $631 billion to $525 billion across the same window.
That means roughly 90% of the two-year increase is one component getting more expensive. Memory goes from 14% of the AI capex bill to 64% in two years. The growth is not a quantity signal. It is a price signal wearing a quantity signal's clothes.
TrendForce's mix data says the same thing from a different angle. DRAM and NAND rise from 47% to 68% of major CSP capex between 2026 and 2027, driven by contract pricing: server-DRAM contract prices rose 64% cumulatively in the second half of 2025, with a further cumulative increase of roughly 270% forecast for 2026.
Meanwhile non-memory spending declines. On this arithmetic, the industry is not buying meaningfully more compute in 2027 than it bought in 2025. It is paying more for the memory wrapped around the compute it already planned to buy.
This is where the RBC analysis earns its keep. RBC was first to flag that hyperscaler capex growth was being flattered by runaway memory prices — a single-quarter observation at the time. What UBS's line items and TrendForce's mix forecast now show is that the effect is not a quarterly artifact but the structural shape of the entire 2026–27 capex curve.
The mechanism is straightforward once you see it. Memory sits on the bottleneck. Goldman Sachs expects tight DRAM, NAND, and HBM supply through 2028, with 2027 tighter than 2026, and argues the market still has not fully understood the durability of the cycle — memory stocks trade at mid-single-digit P/Es despite it.
When a critical input is allocated rather than priced, the supplier captures the surplus. Korea's DRAM export price hit $92,183 a kilogram in August, up 401% year on year. That is not a demand statistic. It is a transfer.
And the bill lands on the buyers' income statement. Nebius raised on-demand GPU rates 17%–21% on 1 October, its second hike in three months, with server memory up 41%. CoreWeave disclosed it has kept contracting new compute capacity at higher prices, roughly $40 million per megawatt on short-dated deals.
Dearer chips, dearer money, same buyers. The Fed raised rates on 16 September, its first hike since July 2023, and sees policy near or above 4% through 2027. The largest cloud spenders were already putting 102% of cloud revenue into capex.
So the $957 billion is doing double duty it cannot support. As a demand indicator it is contaminated: a dollar of "AI capex growth" that reflects a memory price increase buys no additional compute, creates no additional capacity, and produces no additional tokens. It is the same rack at a higher sticker.
As a capacity indicator it is worse than contaminated — it is directionally misleading. If memory consumes 47% of the budget in 2026 and 68% in 2027, the residual available for accelerators, optics, and buildings is under pressure even as the headline number climbs 60%.
The profit pool follows the pricing power. UBS's own assessment is that the price-driven share of this spending adds little to real US GDP and shifts income and profits toward memory producers in Asia. SK Hynix holds roughly half the high-bandwidth memory market. Micron added $32.2 billion of quarterly revenue in a year, $31.5 billion of which was gross profit.
That last ratio is the tell UBS left blank: the price-versus-volume split. Micron's fiscal Q3 DRAM average selling price rose by the low-60% range in a single quarter. If revenue steps up again while shipped bits barely move, the boom is price, not compute.
None of this requires demand to be fake. Amazon does not expect to have enough capacity to satisfy anticipated AI demand even after spending $220 billion, expects the same in 2027, and reportedly described visible 2028 demand as "striking." AWS grew 37% year over year, its fastest in 18 quarters, with AI services revenue growing triple digits to a $25 billion run rate.
Demand is real. The capex number is the unreliable narrator.
There is a second-order consequence the headline hides: capital intensity is rising without a matching rise in capacity, which is precisely the condition under which buildout economics deteriorate. Bernstein puts the cost of a 1GW datacenter at $34.6 billion to $39.5 billion depending on architecture, and notes the dominant economic burden is not electricity but capital expenditure and the depreciation it generates.
If memory is absorbing two-thirds of the budget, the depreciation clock starts on a denominator that bought less compute than planned. Bain estimates the industry must generate $6 trillion in annual revenue by 2031 to fund expected infrastructure demand, against roughly $1.8 trillion from existing applications — a $4.2 trillion gap. A rising share of each capex dollar going to memory prices does not close that gap. It widens it.
There is one large uncertainty worth stating plainly. TrendForce's 68% mix figure is a scenario, not a procurement record. The firm does not name the providers in its group or publish underlying methodology. Some long-term agreements signed from Q2 2026 include price ceilings, and a projected second-half 2027 supply ramp could soften the squeeze. The counterweight's coverage is unknown.
Citi stands alone in calling the peak, with DRAM and NAND prices topping out in Q2 2027 and DRAM falling in the second half. Goldman sees the shortfall widening instead. That disagreement is not ours to resolve here.
What matters is that the answer changes what the capex number means, not whether it is large. If Citi's clock is right, buyers holding take-or-pay commitments near cycle-top prices carry the loss — Micron has locked customers into binding volumes through 2030, roughly $100 billion of obligations backed by $22 billion in deposits.
The question for 2027 is therefore not how much cloud providers spend. It is how much of what they spend buys compute rather than pays for the same capacity at a higher price. On UBS's own line items — memory at $923 billion against $525 billion for everything else — that ratio is getting worse, not better, and the industry is still reading the total as a demand signal.