AI's hidden storage bill: an 8-second video leaves 500MB behind
MIT Tech Review AI · rss · 2026-10-22
An MIT Tech Review piece (sponsored by WD) argues AI infra planning fixates on GPUs while ignoring the storage cost inference leaves behind:
- WD's internal measurement: an 8-second 720p AI video (3.97MB) leaves over 500MB of back-end data — logs, cached context, intermediate renders, backups, safety records — a 130x footprint vs. the delivered file
- Training is a budgetable, finite project; inference runs continuously and its data never retires. McKinsey projects inference demand at 35% CAGR through 2030 (vs 22% for training), exceeding 40% of total data center demand
- An IDC survey of 763 decision-makers: 95% report data growth from generative AI (61% seeing 25%+ growth); 96% anticipate needing faster archive retrieval for inference/RAG
- Storage bottlenecks surface in OpEx and retrieval delays, not CapEx lines; budgeting storage as "a rounding error" can backfire within 12-18 months
The article closes with three questions for CFOs/CIOs about retention policy, storage procurement vs. data lifecycle, and what happens to storage costs if AI usage doubles.
More from Infra
- Dev shows off self-built infra stack — HTTP clients, attention kernels, model routers — to quit dependencies — wavefnx · 2026-10-22
- Can a $100 MacBook Pro M2 24GB serve local LLMs for Home Assistant? — Ejo2001 · 2026-10-22
- Tune your threshold before adding an LLM judge: optimal escalation rate was 0% — Reasonable_Royal_621 · 2026-10-22
- Where your $20 Claude subscription goes: an interactive AI value-chain explainer — aishashok14 · 2026-10-22
- llama.cpp WebGPU adds Q2_0 support, Ternary Bonsai prefill jumps ~50x to 2000 t/s — pmttyji · 2026-10-22
- CVE-2026-111943: Compromised intermediate certificate authority discovered — keks24 · 2026-10-22