Acceptable Use Policy

FlatCompute offers flat-rate pricing: you pay a fixed monthly price and get unlimited token generation within your plan's technical limits. "Unlimited tokens" means there is no per-token billing and no monthly token cap. It does not mean unlimited capacity or unlimited simultaneous requests.

Each plan defines two hard limits enforced by our LiteLLM gateway:

• RPM (Requests Per Minute): How many inference requests you can send per minute. Enforced at the gateway level — requests exceeding this rate receive a 429 response.
• Max Concurrency: How many inference requests can be in-flight simultaneously. Requests exceeding this are queued, not dropped.

Plan limits by tier:

Long-context requests: The context window specified by your plan (64K–1M depending on zone) is a per-request limit, not reserved capacity. Long-context requests are isolated in separate queues to prevent one user's large request from degrading others. Builder, Pro, and Enterprise plans may have additional constraints on simultaneous long-context requests.

What is NOT fair use:

• Creating multiple accounts to aggregate RPM or concurrency beyond a single plan's limits.
• Sharing a single API key across multiple users or organizations to effectively multiply capacity.
• Using automated scripts to maximize every available request slot 24/7 with the intent of reselling or commercially exploiting the compute capacity beyond your own use.
• Attempting to bypass rate limits, concurrency caps, or queue mechanisms.

FlatCompute's core promise is that we do not oversell GPU capacity. Here is what that means in practice:

• Seat caps are real: Each Zone has a maximum number of seats (e.g., 70 for Qwen 35B-A3B, 50–150 for DeepSeek V4). Once a Zone is full, no new subscriptions are accepted. There is no waitlist bypass or secret tier.
• Seats map to capacity: We calculate seat limits based on the GPU's realistic throughput at the model's operating point, not theoretical maximums. We publish the math: hardware specs, cost per month, and break-even analysis for each Zone.
• Concurrency is enforced: Each plan's concurrency limit (1–12 simultaneous requests) is a hard cap. If all users in a Zone hit their max concurrency simultaneously, the GPU can still serve them without degradation. This is by design.
• Capacity monitoring: We monitor GPU utilization, KV cache usage, queue wait times, and TPOT (time-per-output-token) in real time. If p95 TPOT degrades beyond acceptable thresholds, we stop accepting new subscriptions for that Zone.
• Automatic closure: Zones automatically close at 80–85% seat utilization or when performance metrics degrade. This threshold is conservative — it leaves headroom for burst traffic.

This commitment is the foundation of our business model. If we oversell, performance degrades for everyone, and the flat-rate model breaks. We would rather close a Zone and turn away revenue than compromise the experience for existing subscribers.

You may not use FlatCompute to generate, store, or distribute the following types of content:

• Content that is illegal under Spanish law, EU law, or the law of your jurisdiction, including but not limited to: child sexual abuse material (CSAM), non-consensual intimate imagery, content that promotes terrorism, or content that incites violence.
• Content that infringes the intellectual property rights, privacy, or other rights of third parties.
• Malware, ransomware, or other malicious code, or content designed to facilitate cyberattacks, phishing, or social engineering.
• Personal data of third parties processed without a lawful basis under the GDPR.
• Content that defames, harasses, or threatens specific individuals or groups.
• Content designed to manipulate elections, spread disinformation at scale, or impersonate real persons without consent.

We do not pre-screen or monitor inference requests in real time. However, we investigate reports of abuse and may log request content during investigations (see our Privacy Policy).

You may not:

• Attempt to gain unauthorized access to the Platform, its infrastructure, or other users' accounts.
• Attempt to reverse-engineer, decompile, or extract model weights from the Platform.
• Use the Platform to scan for vulnerabilities in third-party systems.
• Introduce malware, viruses, or harmful code via the API.
• Interfere with or disrupt the Platform's infrastructure, including the LiteLLM gateway, vLLM backends, or rate-limiting mechanisms.

If we become aware that you have violated this AUP, we may take the following actions, depending on severity:

• Warning: For first-time, non-severe violations, we will send a warning and request corrective action.
• Rate-limit reduction: We may temporarily reduce your RPM or concurrency limits.
• API key revocation: We may revoke one or all of your API keys.
• Account suspension: We may suspend your account and all subscriptions.
• Account termination: For severe or repeated violations, we may terminate your account permanently.
• Legal action: For violations involving illegal content or harm to third parties, we may report to authorities and cooperate with investigations.

We reserve the right to take immediate action without prior notice for violations involving CSAM, terrorism, or imminent harm.

If you believe your account was suspended in error, contact us at abuse@flatcompute.com.

To report a violation of this AUP by another user, email abuse@flatcompute.com with the following information:

• The API key prefix (sk-fc-xxxx...) or account email of the alleged violator, if known.
• A description of the violation.
• Evidence (timestamps, request IDs, output samples).

We investigate all reports within 48 hours and take action as appropriate.

We may update this AUP from time to time. We will notify users of material changes via email at least 14 days before they take effect. The "Last updated" date at the top of this page reflects the most recent revision.