CCST · Index 1 of 4 · AI Build-Out Family · daily scrape · biweekly composite

Compute Spot Tightness

How tight is AI compute on the spot market right now? Hourly GPU rental rates, availability across 40+ cloud providers, and cross-generation price ratios — aggregated daily into a 0-100 regime score. High score = compute is scarce, cascading demand across the GPU stack. Low score = supply is catching up.

Composite 90 / 100 Tight

Provisional · 1/2 layers live · rental = in development (v2) · 80d history. The composite reflects availability only. The rental-tightness signal is held for v2 — actual $/h versus a per-chip depreciation curve, on median provider-weighted prices — rather than publish a floor-price approximation that can read backwards.

As of Aug 21, 2026
GPU models tracked 95
Source GetDeploying (88 GPUs across 40+ providers)
History · momentum Tightening +40 pts recent

81 points · since 2026-05-26

Layer breakdown

Rental tightness · in development (v2) in dev held · not scored in v1

Held for v2. The honest signal is whether old silicon holds value beyond its expected depreciation path — actual $/h vs a per-chip decline curve, a direct read on the AI-chip depreciation debate. It needs median, provider-weighted prices; the cross-generation ratios below are shown as developing data, not yet scored.

Availability 90 40% base · provider data

Provider count per GPU model + availability flag. Fewer providers + "Limited" tags = tighter supply.

Cross-generation price ratios Leading derivative · 3-12M lead

Ratios between GPU generations reveal demand structure. Persistent legacy-gen pricing post-new-gen-launch is a strong regime-shift signal.

H100 / B200 0.50

Legacy frontier vs current frontier. Higher = cascading demand across generations.

A100 / H100 0.56

Legacy training tier vs frontier training tier. Higher = legacy still economically wanted.

H200 / H100 1.32

Next-gen premium within Hopper family. Spot-market quirks possible.

Per-tier breakdown

Tier GPUs tracked Avg hourly $ Avg providers Weight
Frontier 3 $4.88 8 40%
Legacy training 1 $1.88 2 30%
Agentic 4 $0.96 10 20%
Longtail 3 $1.07 3 10%

Equity Gap · CCST vs Compute Stocks Diverges 1-6M ahead

CCST composite tracks spot-market compute tightness directly. The equity basket (NVDA, AMD, ORCL, VRT) reflects what the market has already priced in. Divergence between the two is the actionable signal: index rising while stocks lag = bullish setup.

Basket 30d return -3.0% NVDA · AMD · ORCL · VRT (avg over available)
CCST composite 30d delta 36 Needs ~30d history (currently accumulating)
Equity Gap 38.99 Positive = stocks lag CCST = bullish setup
Ticker Latest price 30d return
NVDA $216.85 2.3%
AMD $469.45 -15.0%
ORCL $142.07 12.9%
VRT $264.63 -12.1%

AI Build-Out Index Family 4 indices · supply-side AI hardware tracking

CCST · TAISP · MHP · STP — four independent indices, each with its own data source, equity basket, and equity-gap signal. Read them together: when one moves before the others, that's the rotation signal. (Agentic-demand-side tracking lives in the separate Agentic Demand Index.)

93.3 / 100 Tight

AI Build-Out Index — equal-weighted average of the 4 live indices below (4 of 4 reporting). Each index counts once; one with no current reading drops out rather than being imputed.

Methodology

CCST is the first of the Closelook AI Build-Out Index Family — a 4-index suite tracking the AI hardware supply chain (Compute Spot Tightness · Taiwan AI Supply Pulse · Memory/HBM Pulse · Semicap/Test Pulse). Each index has its own data sources, equity basket, and equity-gap signal. Closelook Agentic Usage Pulse (AUP), tracking demand-side AI adoption, is a separate standalone index in its own track.

CCST focuses on the spot market for GPU compute: hourly rental rates, availability across 40+ providers, and cross-generation price ratios. Daily scrape, cron-driven, no manual touch.

How the composite is built. Each layer is scored only when it carries a real signal — a layer we cannot yet measure returns nothing and drops out, and the remaining weights renormalise. We never fill a gap with a neutral 50. Live layers are blended by weight × confidence, so a thin or provisional signal cannot dominate. The reading carries a coverage tag (how many layers are live, and the rental scoring mode) so it is never read as more settled than it is.

Rental signal — v1 and roadmap. Most tracked GPUs are prior-generation, which should depreciate as new gens land — so flat is not neutral. An old GPU holding its rate has beaten the depreciation curve, which is a demand signal, and a direct, data-driven test of the debate over whether AI chips depreciate far faster than hyperscaler accounting assumes. v1 approximates this from cross-generation price ratios (legacy vs frontier). v2 replaces it with an explicit per-chip expected-depreciation curve — actual $/h versus the launch-date-and-age decline path — on median, provider-weighted prices.

Composite 70+ = tight (cascading demand, supply pressure). 40–70 = neutral. Below 40 = loose (supply catching up, depreciation normalising).

Index reading is a diary entry on our process, not investment advice.

FAQ · from the current data

Compute Spot Tightness — CCST — data Q&A

What is the Compute Spot Tightness — CCST?

Daily scrape of hourly GPU rental rates and availability across 40+ cloud providers (88 SKUs), plus a four-stock cross-check — is compute getting tighter or looser? Part of the Closelook Lab (AI build-out).

What does the Compute Spot Tightness — CCST cover?

GPU rental rates and availability across 40+ providers — is compute tightening? 88 GPU SKUs · 40+ providers · Daily.

How often does it update?

Re-baked after the US close on trading days from the Closelook lake, the index workers and vendor series; the page states its own as-of date.

How tight is compute right now?

CCST composite 90/100 — regime Tight, 30-day change 36. Built from daily GPU rental rates and availability across 40+ providers (88 SKUs) plus a four-stock cross-check.

Is this investment advice?

No. Every Lab tool is a data surface in an investment research diary — it describes what the engine measures, never what to do.