Ex-Oracle expert: AI datacenter racks depreciate in 5-6 years, Vera Rubin racks top $8M
SumitGup · x · 2026-10-21
An expert interview with a former Oracle employee released by AlphaSense argues demand is the only thing holding the AI infrastructure story together:
- Cost structure: infrastructure is 40-50% of a 1GW datacenter buildout, with GPUs 70% of rack spend; NVIDIA Vera Rubin racks now cost well above $8M each.
- Depreciation: rack infrastructure is depreciated over 5-6 years (Microsoft at the high end, AWS 5), buildings 20-30, networking 10-15 — but economic life extends beyond accounting treatment. CoreWeave and other neoclouds are re-upping 4-6 year contracts on older A100/H100 GPUs, keeping them monetizable via sustained demand, lagging supply, and vendor software support.
- Scale economics: power/cooling/networking are fixed costs; smaller on-prem deployments (retail, banks) won't disappear — utilization is what matters.
- Custom silicon: pays off via volume × cost delta × perf TCO by skipping NVIDIA/TSMC margin layers; OpenAI's Jalapeno and hyperscalers win via system-level co-optimization, but neoclouds can't shoulder the upfront investment.
More from Infra
- Corporate token spending growth collapses to 0.2%, Apollo warns compute demand may not be unlimited — SumitGup · 2026-10-21
- Kyutai shrinks Pocket TTS to 1.9M params, running 103x real-time on one CPU core — alexcovo_eth · 2026-10-21
- Neoclouds now gross ~$38 per GPU-hour at 155 trillion tokens/day per cluster — basedjensen · 2026-10-21
- DeepSeek reportedly earns $38 per GPU-hour on tokens — what does it mean for GPU pricing? — andrew_n_carr · 2026-10-21
- Hyperscaler insider: GPU order-to-install lead time has always been 1–2 quarters — BenBajarin · 2026-10-21
- Efficient Computer's Electron E1 chip claims 100x energy efficiency, named a Time Best Invention — rebeccakaden · 2026-10-21