Published · Sources reviewed . Source publication dates are listed separately below.
Start with the service behind the rate
The recent B300 debate mixes private buyer experiences, public rate cards, screenshots and forecasts. Those are useful leads for checking a particular offer. Without a matching configuration and contract, they cannot establish that the whole market has become cheaper or more scarce.
Compare the exact accelerator and system, GPU count, location, included CPU and memory, network arrangement, billing unit, minimum duration and interruption policy. Keep a whole-instance bill beside a normalized GPU-hour rate. A small interruptible instance and a large assured cluster solve different problems.
Keep observation and effective dates separate
A notice announcing a future price change describes a future effective date. A screenshot without a reproducible source remains a screenshot. Neither should silently replace the recorded current offer, and re-reading an old notice should not change the date of its underlying observation.
A lower quote may reflect a different term, region or service guarantee. A higher aggregate median may reflect a changed provider sample. Use the rental comparison's matched-offer history to investigate rate changes within a stable population, then inspect the terms that population records.
Why duration can matter
Anthropic's November 2 filing says its third-quarter compute included short-term commitments costing substantially more than comparable multi-year compute. That company disclosure supports checking duration. It does not establish the terms or quality of any private B300 quote, or a universal premium for all buyers.
Ask whether an upfront payment is credited against the contract, refundable, or additional to the stated charge. Count the payment once in total cost and separately show when the cash is due. Unknown cancellation or interruption terms should remain unknown.
Evidence: Anthropic Form S-1/A, November 2, 2026
Recorded offers behind the comparison
This excerpt uses the rental comparison's current observations. Each offer keeps its own date and billing terms; different services are not pooled into a market price. Posted rates do not establish launch availability.
Modal · Serverless / per-second
$7.0992/GPU·hrOlder quote · Oct 20, 2026
B300 GPU task; one GPU; billed per second
Region: Provider regions vary
Contract terms and source
- Minimum commitment
- Not recorded
- Upfront cash
- Not recorded
- Cancellation / refund
- Cancellation not recorded · Refund terms not recorded
- Interruption
- Interruptible · restart loss not recorded
- Additional resources
- Billed separately · extra hourly total not recorded
AI Compute Tracker (aicomputetracker.com), CC BY 4.0; third-party terms apply to data from GPU cloud pricing pages
Modal published pricingHyperstack · On-demand
$7.40/GPU·hrObserved Nov 3, 2026
NVIDIA B300 288GB; maximum per GPU: 28 pCPUs and 240 GB RAM; actual allocation not specified; published per GPU-hour
Region: Provider regions vary
Contract terms and source
- Minimum commitment
- Not recorded
- Upfront cash
- Not recorded
- Cancellation / refund
- Cancellation not recorded · Refund terms not recorded
- Interruption
- Guarantee not recorded · restart loss not recorded
- Additional resources
- Inclusions not recorded · extra hourly total not recorded
AI Compute Tracker (aicomputetracker.com), CC BY 4.0; third-party terms apply to data from GPU cloud pricing pages
Hyperstack published pricingNebius · On-demand
$9.50/GPU·hrOlder quote · Oct 29, 2026
NVIDIA B300 NVLink; published per GPU-hour; billed per second
Region: uk-south1, eu-west2, us-north1
Contract terms and source
- Minimum commitment
- Not recorded
- Upfront cash
- Not recorded
- Cancellation / refund
- Cancellation not recorded · Refund terms not recorded
- Interruption
- Guarantee not recorded · restart loss not recorded
- Additional resources
- Inclusions not recorded · extra hourly total not recorded
AI Compute Tracker (aicomputetracker.com), CC BY 4.0; third-party terms apply to data from GPU cloud pricing pages
Nebius compute pricing documentationTogether AI · On-demand
$9.99/GPU·hrObserved Nov 3, 2026
NVIDIA HGX B300 GPU Cluster; published per GPU-hour
Region: Provider regions vary
Contract terms and source
- Minimum commitment
- Not recorded
- Upfront cash
- Not recorded
- Cancellation / refund
- Cancellation not recorded · Refund terms not recorded
- Interruption
- Guarantee not recorded · restart loss not recorded
- Additional resources
- Inclusions not recorded · extra hourly total not recorded
AI Compute Tracker (aicomputetracker.com), CC BY 4.0; third-party terms apply to data from GPU cloud pricing pages
Together AI published pricingClaims and evidence limits
One low quote proves a market-wide price collapse.
Not established by this evidence
The quote's configuration, service class, date and contract must be comparable before it can support a price change. A single buyer experience does not define the whole market.
All GPU-hours have the same economics.
Depends on scope and assumptions
Anthropic reports different economics for short-term and multi-year capacity. Scale, utilization, workload and service terms still need separate evidence.
A posted rate confirms that capacity can be launched.
Not established by this evidence
This site's posted offers do not verify launch availability. Confirm the configuration and capacity with the provider.
Primary sources
- Anthropic Form S-1/A, November 2, 2026 · published . Reviewed 2026-11-03.
Company disclosures describe what the company reported. A review date does not turn a forecast into an observation or establish a private quote's terms.
Apply the explanation
Background reading: Compute Scarcity on AI Wiki. AI Wiki is run by the same team as AI Compute Tracker.