AI server prices are moving in weeks rather than quarters, with some components swinging as much as 40% week to week. The volatility is making cost forecasting difficult for GPU cloud providers.
AI Compute · Rack Economics
Nvidia's Next AI Rack Nearly Doubles in Cost — Driven by Memory
The build cost of the Vera Rubin "VR200 NVL72" rack hits ~$7.8M, almost double the previous generation. With prices now swinging as much as 40% week to week, AI compute is starting to look like a commodity market.
GB300 NVL72
~$4M
prev. generation
VR200 NVL72
~$7.8M
next generation
Nearly 2× the build cost — and a new memory tier is the biggest reason why
+435%
memory cost increase
~25%
of total system cost is
~$50K
a single Rubin GPU —
40%
week-to-week price swings
The same GPU, wildly different prices
On-demand hourly rates for the H100 (80GB) span a huge range across providers.
→
$14.90
priciest — over 30× higher
Spot / reserved plans cut 60–90%, but carry preemption risk.
COMMODITIZATION — the upside
More flexible procurement. Developers route experiments and inference to low-cost providers to dodge big-player markups; some workloads shift to TPUs.
UNPREDICTABILITY — the cost
Weekly price moves make budgets hard to forecast. Long-term contracts get revised; AWS has raised GPU reservation prices multiple times as memory shortages worsen.
What's next: Blackwell (GB300) is rolling out now; the Rubin generation ramps from late 2026 into 2027. Nvidia is said to be exploring a model where operators buy the infrastructure and share revenue on top of hardware sales.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…