About & methodology

What this is

An AI reads each provider's public terms of service, SLAs, acceptable use policies, and AI-specific terms against a fixed 10-term schema, and records what it finds with a citation and a source link for every value. These are the AI's summaries of what the documents say, not the documents themselves and not legal advice. They can be wrong, incomplete, or out of date, which is why every value links to its source: so you can verify it yourself before relying on it. It does not advise, recommend, or rate.

Methodology

• Archival. Every fetched document version is preserved with a timestamp and content hash; nothing is overwritten.
• Extraction. A structured pass with Claude (Opus) reads each provider's documents against a fixed 10-term schema, returning a value, a confidence level, and a citation for every field. Values it cannot support are recorded as “not specified” or “unclear”, never guessed.
• Human verification. Extractions can be corrected; corrected fields are marked human-verified and survive re-extraction.
• Provenance. Every value links to its source document, with the fetch date and version hash behind it.

Coverage

Providers & documents

• Amazon Web Services: Amazon Compute Service Level Agreement (EC2/ECS/Fargate), AWS Customer Agreement, AWS Service Terms, AWS Responsible AI Policy, AWS Acceptable Use Policy
• Microsoft Azure: SLA for Microsoft Online Services (consolidated; incl. Azure compute/GPU), Microsoft Customer Agreement (published), Microsoft Product Terms for Online Services, Microsoft Acceptable Use Policy (Online Services)
• Google Cloud: Compute Engine Service Level Agreement, Google Cloud Platform Terms of Service, Google Cloud Service Specific Terms (AI/ML data use), Google Cloud Platform Acceptable Use Policy
• CoreWeave: CoreWeave Terms of Use, CoreWeave Acceptable Use Policy
• Lambda: Lambda Terms of Service
• Crusoe: Crusoe Legal Center (Cloud Platform Terms of Service)
• Runpod: Runpod Terms of Service
• Vast.ai: Vast.ai Terms of Service

Term dimensions

• Capacity & reservation terms: Existence and nature of reserved or committed-capacity terms in the PUBLIC documents: reserved instances, committed-use/savings plans, take-or-pay, use-it-or-lose-it, cancellation rights. Where the meaningful commitment terms are negotiated privately (enterprise commitment vehicles), say so explicitly with 'negotiated, not published'. Do not infer terms that are not in the documents.
• Data use & AI training: What the terms say about the provider's rights to use customer data, and specifically any language about training or improving the provider's AI/ML models on customer content. State the default position (e.g. 'does not train on customer data' vs. 'may use to improve services') and any opt-in/opt-out.
• Suspension rights: The grounds on which the provider may suspend the service (e.g. AUP breach, non-payment, security risk, legal process) and whether prior notice is required or suspension can be immediate.
• Termination: Termination for convenience rights (which side, and any notice period), and the effect on data retrieval after termination - the post-termination window during which the customer can still retrieve data.
• Unilateral modification: Whether and how the provider may unilaterally change the terms, and any notice commitment before changes take effect (e.g. 30/90 days' notice, or 'effective on posting').
• Governing law & disputes: Governing law, the forum/venue, whether disputes go to binding arbitration (yes/no), and whether there is a class-action waiver (yes/no). Keep each element short.
• Availability definition: How uptime/availability is defined and the commitment percentage(s). Categorize the basis as one of: instance-level, region-level, cluster/throughput-based, API-error-rate-based, or other. Include the commitment percentage(s) (e.g. 99.99% region, 99.5% instance). If the provider publishes no uptime commitment, value = 'no SLA / uptime not committed'.
• Credit regime: Service-credit structure: the tier mapping (uptime threshold -> credit percentage), the maximum credit, and whether credits are stated to be the sole and exclusive remedy. Give the tiers compactly.
• Claim mechanics: Whether the customer must file a claim to receive credits, the claim window and how it is measured (e.g. within 30 days / end of billing cycle), and what evidence the customer must supply.
• SLA exclusions: Notable exclusions from the SLA as a short categorical checklist plus a brief note. Look for: force majeure, customer fault/misuse, scheduled or emergency maintenance, beta/preview services, single-AZ deployments, throttling, external network/ISP issues. List the ones present.

Disclaimer

These values are an AI's reading of each provider's public terms, not the terms themselves and not legal advice. They can be wrong, incomplete, or out of date. Every value links to its source document, so verify anything yourself before you rely on it. Nothing here creates an attorney-client relationship.