Memory, Not GPUs, Is the Cost Line That Breaks Your 2027 Cloud Budget

Week of 31 August 2026 | Compiled by Team Paringa

A week where the AI cost story moved off the GPU and onto DRAM, and one vendor published hard numbers on repatriating its own token spend.

What Happened This Week

🔴 Memory to consume 68% of hyperscaler capex by 2027, and the pass-through has started

TrendForce published analysis on 25 August, picked up by The Register on 26 August, showing DRAM and NAND flash will account for 47 percent of total cloud provider hardware spend in 2026, rising to 68 percent in 2027. Memory will cost the hyperscalers more than GPUs.

The price moves behind that are severe. Server DRAM contract prices rose a cumulative 64 percent in the second half of 2025 and are projected to climb roughly 270 percent across 2026. Enterprise SSD prices rose around 35 percent in the second half of 2025 and are on track for a cumulative 235 percent increase this year. HBM contract prices could rise a further 70 to 140 percent in 2027. Total cloud operator capex is set to nearly double in 2026, up 98 percent, then grow another 50 percent in 2027.

Suppliers are already using this as cover. Nvidia is reported to be raising prices 15 percent on products containing its Vera Rubin and Grace Blackwell chips. Amazon has lifted its 2026 cash capex guidance from around $200 billion to approximately $220 billion, explicitly citing higher memory costs. Enterprise spending on cloud infrastructure passed $143 billion in Q2, growing 43 percent year on year.

None of this is absorbed by the providers indefinitely. Memory-heavy instance families, high-memory database tiers and storage-optimised SKUs are the first places this shows up in customer bills, and reserved instance renewals in 2027 are where it lands hardest.

Sources: The Register, TrendForce

Most FinOps plans built this year assume compute unit prices trend flat to down. That assumption is now wrong for anything memory-bound, which is most data platforms, most caching layers and most vector stores. Three actions worth taking before Q4 close. First, model your 2027 renewal at a 20 to 30 percent uplift on memory-heavy families and see what breaks. Second, pull a right-sizing pass specifically on over-provisioned RAM, which is the least policed dimension in almost every estate we see. Third, if you have a three-year commitment window open, the case for locking pricing before the 2027 step-up is stronger this quarter than it will be in six months. Cheap memory was a structural assumption for twenty years. Rebuild the forecast without it.

🟡 Nutanix spends $20m on its own cluster to stop paying per token, expects payback in 12 months

Nutanix CEO Rajiv Ramaswami told The Register on 27 August that the company has spent $20 million standing up its own AI infrastructure, and expects to recoup that inside a year. The driver was internal developer tooling.

"Our software teams have been using AI for coding across the lifecycle," Ramaswami said, initially with tools including Copilot and Claude. "Usage exploded and so did costs." Nutanix has moved to open weight models on an on-premises cluster. "We are no longer paying on a per-token basis." Frontier models and external clusters are still used selectively where the workload justifies it.

The numbers give the decision context. Nutanix reported Q4 revenue of $757 million, up 16 percent year on year, and full year revenue of $2.85 billion, up 12 percent, with net income of $1.5 billion. A $20 million capital commitment against that base is a considered bet, not a stunt. Ramaswami also confirmed Nutanix is moving toward an Arm port, explicitly because customers are balking at current hardware prices.

The same theme surfaced elsewhere. HP executives told The Register on 27 August that rising cloud AI costs are one reason customers are buying higher-specification PCs capable of running AI work locally, framing it around token economics.

Sources: The Register, The Register

A twelve month payback on AI inference repatriation is a genuinely useful public data point, and CFOs should ask their own teams to test it. The qualifier matters though: Nutanix is an infrastructure company with spare capacity, existing datacentre operations and engineers who can run a cluster. Most enterprises have none of those, and a naive copy of this move produces a stranded asset with 20 percent utilisation. The transferable insight is the workload-to-model matching Ramaswami described. Routing routine code completion to a small open weight model and reserving frontier models for the tasks that actually need them typically cuts token spend more than any infrastructure decision, and it costs nothing but engineering discipline. Do the routing work first. Only then price the hardware.

🟡 Nvidia's numbers point at $1.3 trillion of hyperscaler capex next year

Nvidia reported earnings on 26 August, with Jensen Huang forecasting roughly 70 percent fiscal 2028 revenue growth, well above analyst estimates. Alongside it, AWS confirmed it will buy 2 million Nvidia GPUs and adopt Nvidia's new Vera CPU.

The framing analysts took from the call is the capex trajectory. Capex among the top five hyperscalers is expected to rise to $1.3 trillion in 2027, up from around $800 billion in 2026. That is a 63 percent increase in a single year, and it is happening despite growing investor scepticism about return on AI investment.

That scepticism is visible in the market's reaction to individual names. Alibaba's shares fell 5 percent on 20 August after it reported capex up 75 percent and negative free cash flow of 44.7 billion yuan. Alphabet was similarly punished in late July over capex scrutiny. The pattern is that infrastructure spending continues regardless while public markets increasingly refuse to fund it on faith.

Sources: CNBC, CNBC

Hyperscaler capex is not philanthropy. $1.3 trillion of infrastructure has to be recovered through list prices, instance mix and the quiet erosion of discount depth over the next three renewal cycles. If you are negotiating an enterprise agreement in the next two quarters, that is both your negotiating argument and your urgency argument. It also raises the value of anything that reduces lock-in: egress-aware architecture, portable workloads, and a genuine second-source option even if you never exercise it. Buyers who can credibly move keep their discounts. Buyers who cannot will fund the buildout.

What CFOs, CTOs and FinOps Leads Are Dealing With

The discipline has caught up with the problem in scope but not in tooling. The FinOps Foundation's 2026 State of FinOps report, drawn from 1,192 respondents representing more than $83 billion in annual cloud spend, found 98 percent of teams now actively manage AI costs, up from 63 percent in 2025 and 31 percent in 2024. Scope has widened well past public cloud: 90 percent manage SaaS, 64 percent software licensing, 57 percent private cloud and 48 percent datacentre spend. The Foundation formally changed its mission from managing the value of cloud to managing the value of technology.

The gap is measurement. The single most requested capability is granular monitoring of AI spend at the token, LLM request and GPU utilisation level, and commercial tooling has not delivered it at scale. Meanwhile 73 percent of organisations report AI costs exceeding original budget plans. Add memory-driven unit cost inflation on top of an estate nobody can meter properly and you have the shape of 2027: not a spending problem, an attribution problem with a spending consequence.

Three Things Worth Watching Next Week

1. Instance price notices for memory-heavy families — Watch AWS, Azure and Google Cloud pricing pages and reserved instance terms for the first visible pass-through of DRAM costs. It will arrive as SKU mix changes before it arrives as headline price rises.

2. Nvidia's reported 15 percent uplift on Vera Rubin and Grace Blackwell — If confirmed, it resets GPU-hour economics for every cloud provider and every neocloud reselling them.

3. More published repatriation numbers — Nutanix put a real figure on the table. Expect others to follow, and treat any claim without a stated utilisation assumption as marketing rather than evidence.

Questions? hello@paringa.ai

Paringa is an Infrastructure & AI Security, Cost and Governance Specialist. We help organisations see what their AI and cloud estate actually costs, then take the spend out without slowing delivery. | 31 August 2026 | © Paringa Ltd