Compute Pulse

Craiglist for Compute

[Compute Pulse @compute_pulse]

Want to receive our private GPU deal flow

Subscribe here: https://t.co/9jgwhna1iF

[Compute Pulse @compute_pulse]

@cyfuture has been included on https://t.co/OANYr3dfw5!

See @cyfuture listings here: https://t.co/lWLZ2fsMDg

[Compute Pulse @compute_pulse]

Cheapest H200 right now (observed on-demand):

1. Lambda — $2.29
2. GMI Cloud — $2.60
3. Deep Infra — $2.69
4. TensorPool — $2.99
5. Seeweb — $3.00

Highest managed listing: ~$21.20. That's a 9.3x managed-listing spread.

Live board, 34 providers 👇

— https://t.co/OANYr3dfw5

[Compute Pulse @compute_pulse]

GPU rental prices, 2026-08-10. Managed listing floor to ceiling, $/GPU-hr:

H100 $1.66–$23.33 (14.1×)
H200 $2.29–$21.20 (9.3×)
B200 $3.69–$20.08 (5.4×)
A100 $0.89–$8.19 (9.2×)

On-demand, 1797 live listings. Spread = ceiling / floor.

via https://t.co/OANYr3dfw5

[Compute Pulse @compute_pulse]

Cheapest RTX 4090 right now (observed on-demand):

1. TensorDock — $0.30
2. Runpod — $0.34
3. CloudRift — $0.39
4. Beam — $0.42
5. SwissGPU — $0.47

Highest managed listing: ~$0.78. That's a 2.6x managed-listing spread.

Live board, 9 providers 👇

[Compute Pulse @compute_pulse]

@CPOfficialtx The difficult part is the duration mismatch: long-lived power infrastructure financing GPUs that depreciate within a few years.

[Compute Pulse @compute_pulse]

@krismicinski Exactly. The real measure is economic output per dollar of compute. Revenue can grow while the underlying unit economics worsen.

[Compute Pulse @compute_pulse]

@CryptosandWine Moving up the stack can improve margins—but only if the services create durable demand for the underlying GPU capacity.

[Compute Pulse @compute_pulse]

GPU rental prices, 2026-08-07. Managed listing floor to ceiling, $/GPU-hr:

H100 $1.66–$23.33 (14.1×)
H200 $2.29–$21.20 (9.3×)
B200 $3.69–$20.08 (5.4×)
A100 $0.89–$8.19 (9.2×)

On-demand, 1759 live listings. Spread = ceiling / floor.

[Compute Pulse @compute_pulse]

Cheapest A100 right now (observed on-demand):

1. JarvisLabs — $0.89
2. Deep Infra — $0.89
3. GPU·ai — $0.94
4. Runpod — $1.00
5. Civo — $1.09

Highest managed listing: ~$8.19. That's a 9.2x managed-listing spread.

Live board, 31 providers 👇

[Compute Pulse @compute_pulse]

@dmoses34 @nkagan_official @KalshiResearch @Kalshi The hard part is defining the basis: spot vs. reserved term, region, SLA and utilization can each change “cost of compute.” Which choice was hardest to make certifiable? @compute_pulse maps the underlying rental market.

[Compute Pulse @compute_pulse]

@TMTLongShort H100 already spans $1.66–$23.33/GPU-hr across 38 providers; a sustained doubling near the clearing price, not a few premium quotes, is what would pull new capacity online. @compute_pulse tracks that spread daily.

[Compute Pulse @compute_pulse]

GPU rental, 2026-08-05. Floor and median $/GPU-hr, on-demand:

H100 floor $1.66, median $3.06
H200 floor $2.29, median $4.00
B200 floor $3.50, median $5.99
A100 floor $0.89, median $1.54

1701 live listings across clouds.

— https://t.co/OANYr3dfw5

[Compute Pulse @compute_pulse]

Cheapest H200 right now (observed on-demand):

1. Lambda — $2.29
2. GMI Cloud — $2.60
3. TensorPool — $2.99
4. Seeweb — $2.99
5. Koyeb — $3.00

Highest managed listing: ~$21.20. That's a 9.3x managed-listing spread.

Live board, 30 providers 👇

— https://t.co/OANYr3dfw5

[Compute Pulse @compute_pulse]

@TMTLongShort GPU rental rates can rise while asset returns disappoint. Utilization, financing costs and depreciation still decide the trade

[Compute Pulse @compute_pulse]

@VideoCardz A 20–40% jump points to more than demand alone. FX, distributor inventory and memory costs may all be moving together.

Check latest compute price trend at https://t.co/POl2MukMkz

[Compute Pulse @compute_pulse]

@itsmichaelluu A four-month 10–20x claim needs more than an AI narrative. Incremental returns on all that infrastructure spending are what matter

Check https://t.co/POl2MukMkz for AI trend

[Compute Pulse @compute_pulse]

@teortaxesTex Compute trends are a useful prior, not evidence for one model. Architecture disclosures and observed behavior should carry more weight.

[Compute Pulse @compute_pulse]

@jurad0x Consumer GPUs now compete with AI accelerators for memory, wafers and packaging—not just gamer demand

[Compute Pulse @compute_pulse]

@equities_samjho Model releases may become the new search algorithm updates—except there’s no Search Console showing what changed.