AI Infra Framework
AI Infrastructure Bubble Tracker
Utilization, leverage and shortage pricing signals for the post-Nvidia-capex cycle.
24 May 2026 · YK Research
Contents
The Mispricing
AI demand is real. The bubble risk is not “AI is fake.” The risk is that investors extrapolate peak scarcity into permanent economics. In every capacity bubble, the market first pays for access, then realizes the scarce asset was financed at the wrong price, with the wrong leverage, against utilization that only existed during shortage.
The tracker asks one question: is AI capex producing durable ROIC and useful GPU-hours, or are lower-quality capacity wrappers being bid up because everything with “AI infrastructure” in the label temporarily looks sold out?
Real Pricing Snapshot
This chart is actual public pricing scraped from RunPod and Lambda pricing pages, not a made-up bubble score. It tracks posted on-demand GPU-hour prices; it does not claim true utilization.
Normalized Memory Cost
Same source rows normalized by advertised VRAM. Useful for spotting scarcity premium, but still imperfect because instance bundles include CPU, RAM, storage and networking.
What We Can Track Right Now
GPU rental $/H100-equivalent-hour
trackableSource: cloud list prices, marketplaces, spot feeds
Read: falling price while capacity expands = shortage easing
Provider availability / lead time
trackableSource: AWS/GCP/Azure/OCI pages, neocloud quote screens
Read: more regions available = weaker scarcity premium
Cloud capex productivity
trackableSource: AMZN/MSFT/GOOGL/META filings
Read: capex up + ROIC stable = healthy; capex up + ROIC down = air pocket
Operating margin under AI load
trackableSource: AWS/Azure/GCP segment disclosures
Read: margin compression says suppliers/power capture value
Leverage in infra wrappers
trackableSource: public debt, private-credit reports, lease disclosures
Read: debt-funded GPUs are fragile if rental rates fall
Quality minus junk spread
trackableSource: public equity baskets
Read: junk outperforming quality = shortage beta / mania
HBM/DRAM pricing and supply
trackableSource: Micron/SK Hynix/Samsung commentary, contract-price reports
Read: memory still tight = bottleneck real; supply surge = margin risk
Social / options mania
trackableSource: call volume, short interest, X/Reddit mentions
Read: retail flows moving low-float AI labels = bubble behavior
Proxy-only / hard data
- True global GPU utilization outside owned clusters
- Model FLOPs utilization by frontier lab
- Private neocloud occupancy, cancellation rights and customer concentration
- GPU loan-to-value, covenants and refinancing terms in private credit
- Cost per useful token by model after uptime, batching, software efficiency and power
Source Rows
Raw rows behind the charts. Every value comes from the committed public-pricing snapshot; no placeholder GPU coverage scores remain.
Provider Universe
First live collector should expand from the current RunPod/Lambda snapshot into public cloud quotes and marketplace listings, then layer availability status and contract term. Do not clone proprietary SemiAnalysis data; build independent snapshots from public/provider data.
Price metric
Availability metric
Utilization proxy
Owned-cluster metric
Leverage: Where Bubbles Break
Hyperscalers
Neoclouds
Power / data center
The telecom-fiber lesson is brutal: demand can be real and investors can still lose money if the asset is financed at shortage prices and bandwidth prices collapse before the debt amortizes. GPU wrappers have the same shape if H100/H200 rents normalize faster than lenders expect.
Prior Bubble Analogues
No numeric overlap score here. These are qualitative analogues only; turning them into numbers without a source would be fake precision.
Telecom fiber
Shale
Solar / EV
Dot-com
Analogue Reads
- Dot-com: real user adoption, bad monetization, new metrics hiding weak unit economics.
- Telecom fiber: real bandwidth demand, debt-funded overbuild, price per unit collapses.
- Shale: high-cost producers outperform during shortage, then die when marginal price normalizes.
- Solar / EV: real adoption curve, too many entrants, capex boom, margin collapse, leaders survive.
Live Tracker Roadmap
1. Pricing index
Normalize all quotes to $/GPU-hour, $/GB VRAM-hour and $/FP8 PFLOP-hour across Nvidia and AMD chips.
2. Availability index
Capture available, limited, quote-only, sold-out, unknown by provider, region, SKU and contract term.
3. Effective-cost index
Adjust list price by uptime penalty, utilization penalty, memory size, interconnect quality and software maturity.
4. Financial overlay
Join capex, depreciation, operating margin, ROIC, debt and backlog for hyperscalers, neoclouds, power and memory names.
5. Bubble score
Score whether shortage pricing is being validated by utilization and ROIC or just financed by leverage and narrative.
Data Audit Notes
- Pricing charts are backed by committed JSON from public RunPod and Lambda pricing pages, refreshed by scripts/fetch-ai-infra-gpu-pricing.py.
- The current tracker uses posted on-demand GPU-hour prices and advertised VRAM. It does not infer utilization from invented scores.
- Global utilization, private neocloud occupancy, cancellation rights and private-credit LTV remain explicit data gaps until source feeds exist.
- Provider prices are bundled instance economics; CPU/RAM/storage/network quality and software maturity can make effective token cost differ materially.
How To Read It
Not bubble
Early warning
Bubble
Opportunity
Gavin-style bottom line: do not ask whether AI is real. Ask who owns scarce supply, who financed it, what utilization is assumed, and what happens when scarcity ends.