40% cheaper AI tokens
Frontier AI models, served on SkyPilot’s managed and verified idle GPUs.
$5 credits to start · Zero Data Retention
Cut token costs by 40%.
One endpoint, one key
Use frontier open models, or use proprietary ones through BYOK.
Streaming, tools, JSON mode
One URL switch, no changes to your existing agentic workflows.
Agents, evals, batch
Scale from development to production AI workloads.
Automatic failover
SkyPilot Tokens automatically routes traffic from idle GPUs to dedicated ones when reclamation happens.
from openai import OpenAI client = OpenAI( base_url="https://tokens.skypilot.ai/v1", api_key="YOUR_SKYPILOT_TOKENS_KEY", ) response = client.responses.create( model="meta-models/Muse-Spark-Open-816B-A42B", input="Plan an agent that triages 10,000 support tickets.", )
Every model, 40% cheaper
Muse Spark Open
GLM-5.3
Kimi-K3
DeepSeek-V4.1-Flash-1017
Bring your own keysUsing idle GPUs from verified customers.
Token buyers save. Fleet owners earn.
40% cheaper
Tokens are securely computed on SkyPilot-managed idle GPUs.
Routed to spare capacity
SkyPilot Tokens route requests to idle GPUs, with automatic fallback.
No interruption
If an idle GPU node is reclaimed, traffic falls back to dedicated nodes.
Earn from the idle GPUs in your fleet
Earn paybacks when not in use, reclaim automatically on new workloads.
Opt in, nothing to operate
SkyPilot Platform already manages the scheduling of your GPU fleet.
Your jobs take priority
Token pods are automatically removed when you submit new workloads.
Calculate the token revenue from your idle GPUs
Kimi-K3 on 8×B300 nodes| How revenue grows | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Token revenue | $15M | $19M | $25M |
Show the math
- Nodes512 GPUs ÷ 8 per node64
- Idle node-hours a month64 × 730 h × 30% idle14,016
- Serving node-hours14,016 × 80% serving11,213
- Revenue per node-hour32 requests at once, measured$110
- Token revenue a month11,213 × revenue per node-hour$1.2M
- Token revenue a yeartoken revenue a month × 12$15M
- Our prices
- Kimi-K3: input $1.80, cached input $0.18, output $9.00 per M tokens.
- Serving load
- 32 requests at once on one 8-GPU node: 79,778 input and 565 output tokens per second, 91% of input from cache, 42.5 ms per output token. Earns $110 per node-hour. This is the measured point with the highest throughput that keeps latency under 50 ms per output token and under 2 seconds to the first token.
- Measured by
- SemiAnalysis InferenceX's agentic-trace benchmark of Kimi-K3 on one 8×B300 node, read 2026-10-28. About the AgentX traces
- Traffic
- Real multi-turn coding-agent sessions from InferenceX's AgentX traces, about 125,400 input and 890 output tokens per request with 91% of input served from cache, so output is under 1% of tokens. Shorter contexts or longer answers earn less per node-hour at our prices, because an output token takes far longer to produce than a cached input token.
We take privacy seriously
SkyPilot Tokens: No prompts, outputs, or logs persist here or on the fleet
SkyPilot-verified fleet: We verify the model served; tampering stops serving
Your fleet: Operated by the SkyPilot control plane already managing your fleet
SkyPilot Tokens: Public-facing gateway, outside your VPC
Zero Data Retention (ZDR)
No prompts, outputs, or logs persist on SkyPilot systems or GPU suppliers.
Caches are salted per account
On the model servers we deploy, each account gets its own cache salt, so one customer's cached prompts are never reused for another.
Both sides stay anonymous
Token buyers can't identify fleets, and fleet owners can't identify buyers. Fleets store no prompts or responses.
Signed, tunneled transport
No open ports on fleet clusters; requests arrive over an mTLS tunnel with daily-rotated certificates.
Owners keep control of their clusters
LLM calls are made from SkyPilot verified and hardened images. No user tool calls or code execution on the fleet.
SOC 2 Type II
SkyPilot is SOC 2 Type II compliant. Report available on our Trust Center.
FAQ
How can SkyPilot Tokens be 40% cheaper?
SkyPilot Tokens is built on a novel approach: serving tokens on SkyPilot-managed GPUs from verified customers. These customers already use SkyPilot Platform to schedule their GPU fleets, and they can opt in to allow SkyPilot Tokens serving on idle nodes.
Serving on idle GPUs (with dedicated fallback capacity) means we can offer 40% cheaper cost than most inference providers.
What happens when a fleet owner takes GPUs back?
For fleet owners: The owner's workloads take priority and will transparently preempt any Tokens jobs. The workloads are scheduled as usual.
For token users: In-flight requests automatically get re-routed to other available GPU nodes, including other idle GPU nodes or dedicated fallback capacity. Latency can increase briefly, but the endpoint keeps serving.
Is SkyPilot Tokens secure?
Yes. We take security and privacy design very seriously. Zero Data Retention (ZDR) is always on. Your requests and responses are not stored by us or any third party. Refer to the privacy section above for technical details. Reach out to our team for in-depth discussions.
Where do my requests run?
Requests run on GPU fleets from SkyPilot Platform’s verified customers, and can fall back to our dedicated GPUs or top-tier partner inference providers.
| Backend | What it is | Data handling | Traffic |
|---|---|---|---|
| Verified SkyPilot Platform customer fleets | Idle GPUs in verified customers' fleets, serving a model server we deploy and operate | Zero Data Retention; per-account cache salt; requests arrive over an mTLS tunnel | Most |
| SkyPilot dedicated nodes | Our own single-tenant nodes on AWS, OCI, CoreWeave, Together AI, and Nebius, running the same model server | Zero Data Retention; per-account cache salt | Most |
| Nebius Token Factory | Partner inference API | Zero Data Retention enabled for our account | Spill-over only |
| Together AI | Partner inference API | Zero Data Retention enabled for our account | First for some models |
| Baseten | Partner inference API | Zero Data Retention enabled for our account | Spill-over only |
Every enterprise customer that joins the fleet goes through technical verification and a legal contract. Model servers are deployed by SkyPilot Tokens and use a limited set of specific hardened images.
Which models are available?
Muse Spark Open, GLM-5.3, GLM-5.3-Flash, Kimi-K3, and DeepSeek-V4.1-Flash-1017 today, with more open models continuously added. You can also bring your own keys (BYOK) to access OpenAI GPT or Anthropic Claude models through the same endpoint.
What does joining as a fleet owner involve?
You opt in through the SkyPilot Platform you already run, and you can pause or leave SkyPilot Tokens at any time. You earn on every token served from your fleet: use the calculator to estimate the token revenue.
New to SkyPilot? Sign up for SkyPilot Platform and become a verified provider.
What actually runs on my GPUs?
A model server that is (1) deployed through the same SkyPilot Platform control plane that already manages your fleet; (2) using a limited set of hardened container images maintained by SkyPilot.
The model server runs unprivileged in a restricted namespace: no service-account tokens, egress limited to the SkyPilot Tokens proxy and the cloud storage model weights are pulled from, no access to your data volumes, and images and weights pulled only from SkyPilot-managed sources. It writes nothing to your infrastructure. Exact manifest is available for audit.