Skip to content
Frontier Terminal
All theses

Compute

as of

Frontier labs keep pricing power as compute becomes a tradable commodity: token prices hold up relative to GPU rental costs, while demand and capacity buildout continue.

Stance
Bullish
Position
No position
Opened
Horizon
2027-12-31

Compared with CoreWeave, Lambda, Nebius, Crusoe, Together AI, DigitalOcean, Verda, Microsoft Azure, Oracle Cloud.

Read the journal page on GitHub

1. Token prices hold up against GPU costs

Unresolved since

1.73 GPU-h/MTok, 3% below the high of 1.78 GPU-h/MTok on 2026-11-04 (kill at 1.24 GPU-h/MTok); 5 of 203 days of history, needs 203.

Kill condition
The ratio stays 30% or more below its 26-week high for 4 straight weekly readings
Evidence
Median flagship blended token price ÷ median public H100 on-demand price per GPU-hour: a modeled price relationship, not a margin
Latest reading

1.73 GPU-h/MTok

for 2026-11-07

verda.com
Flagship median
$8.00/MTok
Flagships, blended
OpenAI gpt-6-astra $20.00/MTok · Anthropic claude-opus-5-5 $8.00/MTok · Google gemini-4-argon $4.00/MTok · xAI grok-4.7 $3.00/MTok · Meta muse-wonder-1.0 $8.00/MTok
Left out
Mistral, DeepSeek, Alibaba, Moonshot AI
H100 median
$4.62/GPU-h across 8 providers
H100 by provider
CoreWeave $6.16/GPU-h · Crusoe $3.90/GPU-h · DigitalOcean $4.41/GPU-h · Lambda $4.74/GPU-h · Nebius $4.50/GPU-h · Oracle Cloud $10.00/GPU-h · Together AI $3.99/GPU-h · Verda $4.94/GPU-h
No price today
Microsoft Azure
B200 ratio, context only
0.94 GPU-h/MTok (median $8.55/GPU-h across 6 providers)
0.00 GPU-h/MTok0.50 GPU-h/MTok1.00 GPU-h/MTok1.50 GPU-h/MTok2.00 GPU-h/MTokNov 2Nov 3Nov 4Nov 6Nov 7

2. Flagship prices avoid a broad collapse

Unresolved since

$8.00/MTok, 0% below the high of $8.00/MTok on 2026-11-07 (kill at $5.60/MTok); 5 of 201 days of history, needs 201.

Kill condition
It stays 30% or more below its 180-day high for 4 straight weekly readings
Evidence
Median flagship blended token price, (3 × input + output) ÷ 4, across the six labs it opened with that have a mapped flagship
Latest reading

$8.00/MTok

for 2026-11-07

ai.google.dev
Flagships, blended
OpenAI gpt-6-astra $20.00/MTok · Anthropic claude-opus-5-5 $8.00/MTok · Google gemini-4-argon $4.00/MTok · xAI grok-4.7 $3.00/MTok · Meta muse-wonder-1.0 $8.00/MTok
Left out
Mistral, DeepSeek, Alibaba, Moonshot AI
$0.00/MTok$2.50/MTok$5.00/MTok$7.50/MTok$10.00/MTokNov 2Nov 3Nov 4Nov 6Nov 7

3. Frontier labs keep their demand share

Unresolved since

61.0%, 2% below the high of 62.5% on 2026-11-04 (kill at 43.7%); 82 of 111 days of history, needs 111.

Kill condition
It stays 30% or more below its 90-day high for 4 straight weekly readings
Evidence
Combined OpenRouter token share of the nine tracked labs, trailing 7 days (one aggregator’s traffic, not market share)
Latest reading

61%

for 2026-11-06

openrouter.ai
Window
2026-10-31 to 2026-11-06
Each lab
OpenAI 8.3% · Anthropic 6% · Google 2.4% · xAI 0% · Meta 1.3% · Mistral 0% · DeepSeek 39.3% · Alibaba 1% · Moonshot AI 2.7%
0%25%50%75%100%Aug 16Sep 5Sep 26Oct 16Nov 6

4. The physical buildout continues

Unresolved since

198 open roles; 74 of 90 days of history, needs 90.

Kill condition
The count falls 40% or more from its 90-day high
Evidence
Open physical-infrastructure roles, summed across labs with a usable hiring feed
Latest reading

198

for 2026-11-07

api.ashbyhq.com
By lab
OpenAI 0 · Anthropic 32 · Google 3 · xAI 163 · Mistral 0
No public feed
Meta, DeepSeek, Alibaba, Moonshot AI
063125188250Aug 25Sep 12Oct 1Oct 19Nov 7

5. Rented compute demand keeps growing

Holding since

112.5% for the period ending 2026-06-30, against a floor of 30.0%.

Kill condition
Growth is below 30% for 2 straight quarters
Evidence
CoreWeave quarterly revenue against the same quarter a year earlier (SEC filings)
Latest reading

112.5%

for 2026-06-30

data.sec.gov
Value
$2,575M
A year earlier
$1,212M (2025-06-30)
0%50%100%150%200%Dec 31Mar 31Jun 30

6. Data-center power keeps coming online

Unresolved since

15,122 MW, 0% below the high of 15,138 MW on 2026-11-05 (kill at 12,111 MW); 3 of 201 days of history, needs 201.

Kill condition
20% or more below its 180-day high for 4 straight weeks
Evidence
Epoch AI's estimate of current power across the AI data centers it tracks
Latest reading

15,122 MW

for 2026-11-07

epoch.ai
0 MW5,000 MW10,000 MW15,000 MW20,000 MWNov 4Nov 5Nov 6Nov 7

Notes

This is a research journal. It isn’t investment advice. Each status is computed from our stored data against kill conditions written before the data arrived. OpenRouter is one aggregator’s traffic, so treat it as a rough read on demand. A judged claim stays unresolved until the judge rates it.

Not investment advice.

Testing a claim needs a public number. Request a dataset on GitHub.

Data from the last 90 days is CC BY 4.0. Older data is CC BY-NC 4.0, except what we published before 2026-11-04, which stays CC BY 4.0. How you can reuse it.