AWS just raised its 2026 capital spending forecast to $220 billion, blaming soaring memory costs rather than GPU scarcity. For UK buyers, that's a signal the AI hardware crunch is structural — and why AI servers are facing escalating costs due to HBM shortages is now a multi-year planning problem, not a passing blip.
View the data behind this chart
| Feb 2026 estimate | Jul 2026 estimate | |
|---|---|---|
| Amazon annual capex | $bn200 | $bn220 |
What Amazon actually announced
Amazon's cloud division reported $42.2 billion in second-quarter revenue, ahead of the $40.54 billion Wall Street had pencilled in, and a 36.7 percent year-on-year jump — its fastest growth since 2021. Group-wide revenue hit $200.61 billion, up 20 percent, with AWS delivering $16.62 billion in operating income at a 36.8 percent margin. AWS's dedicated AI business and its chips business have each now passed a $25 billion annual revenue run rate.
The headline for infrastructure buyers, though, is the capex line. Amazon has lifted its 2026 cash capital expenditure forecast from roughly $200 billion to approximately $220 billion. CEO Andy Jassy was explicit about the cause: "the higher cost of memory" is pushing the figure up, and he warned the company still "will not have enough capacity to meet all the demand we have in 2026," adding he expects the same to be true in 2027. Shares jumped 14 percent on the news, suggesting investors read tight supply as a sign of pricing power rather than a warning sign.
Why memory, not GPUs, is the new bottleneck
Jassy's language matters: he pointed to "inflated prices right now on some of the components like memory, hard drives, and SSDs" — not chip allocation from Nvidia or AMD. That fits a broader pattern across the hyperscaler market, where manufacturers are diverting DRAM wafer capacity into higher-margin HBM for AI accelerators, squeezing supply of conventional server memory at the same time as demand for AI-grade memory explodes.
Analysts covering the memory market have quantified this: Micron's own HBM total addressable market forecast implies growth from roughly $35 billion in 2025 to around $100 billion by 2028, a near-40 percent compound annual growth rate. SemiAnalysis estimates memory could account for about 30 percent of hyperscaler capex in 2026, up from roughly 8 percent in 2023–2024 — a structural shift in where AI infrastructure money goes. For buyers who haven't yet had reason to understand High Bandwidth Memory (HBM), it is now the single biggest line item driving AI server pricing.
A shortage with no near-term end date
This isn't being framed as a temporary supply hiccup. Market commentary tracked alongside Amazon's results points to DRAM and NAND remaining in short supply until at least 2028, with DRAM prices expected to more than double in 2026 and rise again in 2027.
Component pressure extends beyond memory too — hard drives, some server CPUs, and overall server yields are all cited as tightening, alongside continuing power and energy constraints on new build-out.
The maths UK buyers should borrow from Jassy
Jassy laid out AWS's own investment logic on the earnings call, and it's a useful model for any UK enterprise weighing cloud versus owned infrastructure. Data centre capital is committed roughly two years before servers go live, but once live, that capacity can be monetised for 30-plus years without repeating the start-up spend. Servers and networking, by contrast, are bought only months ahead of deployment, typically break even in under three years, and are contracted against demand — most AI capacity now on five-year terms — with a useful life of five to six years.
The lesson for a UK buyer isn't to mimic AWS's scale, but to recognise that server-level decisions and facility-level decisions sit on very different payback clocks. Locking into long cloud commitments now, while memory prices are inflated, transfers that cost risk onto your own budget for years. Modelling both paths properly — including realistic refresh cycles — is where a Total Cost of Ownership comparison for hybrid on-premise and cloud deployments earns its keep before a contract is signed.
Building a hybrid plan around a 2028 shortage
With Moody's marking up hyperscaler capex forecasts by $85 billion to near $1 trillion by 2027, the direction of travel across the entire cloud market is unmistakable: costs are rising and staying up. Waiting for memory prices to normalise before committing to AI infrastructure is a bet against a market that multiple analysts expect to stay tight into 2028.
Practically, that argues for a deliberate hybrid split rather than an all-or-nothing choice. Workloads with volatile or short-term demand may still suit cloud consumption, where AWS absorbs the capex risk. Predictable, long-running AI workloads — training pipelines, inference at steady volume — often make more sense owned outright, provided the memory-heavy bill of materials is priced and configured correctly from the start. Buyers should plan their AI deployments and estimate GPU requirements against realistic five-year utilisation, not launch-day enthusiasm, before comparing quotes through server configuration tools or vendor-specific paths such as the Dell server configurator, HPE server configurator or Lenovo server configurator.
- 01DataCenterDynamics — AWS reports fastest growth since 2021, Amazon annual capex to hit $220bn on AI memory costs · 31 July 2026
- 02The Register — AWS to spend $200 billion to double capacity by end of 2027 · 6 February 2026
- 03The Register — AWS says server memory shortage pushing customers to cloud · 30 April 2026
- 04Blocks & Files — Memory semiconductor supercycle set to run through 2028 · 21 January 2026
- 05Tom's Hardware — Memory will consume 30 percent of hyperscaler spending this year · 1 May 2026
- 06DataCenterDynamics — Moody's hyperscaler capex forecasts marked up by $85bn to close in on $1trn by 2027 · 1 June 2026