AI tokens 40% cheaper
One endpoint for frontier AI models, served on SkyPilot Platform’s verified idle GPUs.
$5 free credits per month · zero-data retention
OpenAI-compatible. One line to switch.
One endpoint, one key — your coding agent reaches every model here, plus your own proprietary models through BYOK.
Streaming, tools, JSON mode — no changes to your existing workloads except for the endpoint switch.
Batch, evals, agents — first token in under a second.
Automatic failover — owners reclaim GPUs; traffic automatically falls back to available backends without interruption.
from openai import OpenAI client = OpenAI( base_url="https://tokens.skypilot.ai/v1", api_key="YOUR_SKYPILOT_TOKENS_KEY", ) response = client.responses.create( model="zai-org/GLM-5.3", input="Plan an agent that triages 10,000 support tickets.", )
Everything 40% under list
Buyers save. Owners earn.
40% cheaper. Reserved fleets sit 20–40% idle; serving on those hours costs almost nothing.
Routed to spare capacity. The router picks whichever fleet has free capacity.
Your endpoint never changes. Owners reclaim; traffic shifts.
Opt in, nothing to operate. Through the SkyPilot Platform you already run.
Your jobs always win. Lowest priority, whole idle nodes only, preempted instantly.
Earn on every token served. Those hours earn zero today.
What could your idle GPUs earn?
Show the math
nodes = GPUs ÷ 8 (Kimi-K3 serves on 8-GPU nodes)idle node-hours/mo = nodes × 730 × idle%token revenue = idle node-hours × utilization% × $50–250 per node-houryour earnings = 50% of token revenueUtilization means your average request traffic as a share of a fully loaded node, taken across all idle hours. The $50–250 band is what a fully loaded 8×B300 node earns per hour serving Kimi-K3 at our prices, from throughput measured by SemiAnalysis InferenceX in Aug 2026 (the 8×B200 band scales that by its roughly two-thirds serving throughput), so revenue scales directly with utilization. Actual revenue depends on the traffic mix: agent workloads are usually at the low end, generation-heavy batch at the high end. For comparison, reserved rent on the same node is $30–45 per hour. These are estimates under preview terms, not quotes.
Privacy works both ways
Zero data retention
No prompts, outputs, or logs persist on SkyPilot systems or GPU suppliers.
Caches are partitioned
Prefix caches are salted per user before any customer fleet serves traffic.
Both sides stay anonymous
Buyers can't identify fleets; fleets get no interface to requests.
Signed, tunneled transport
No open ports on fleet clusters; requests arrive over a mutually authenticated TLS tunnel with daily-rotated certificates.
No control over your cluster
No buyer code runs on your machines; the Tokens control plane can't modify your fleet.
SOC 2 Type II
The report is available at SkyPilot trust center on request.
Frequently asked questions
How can it be 40% cheaper?
The GPUs are already paid for. Reserved fleets sit 20–40% idle, and serving on those hours costs almost nothing. We also deploy a highly optimized inference stack, so each GPU serves more tokens per hour. Input, cached input, and output all bill at 40% under list price.
What happens when a fleet owner takes GPUs back?
The owner's own workloads take priority. When an owner reclaims GPUs, traffic automatically falls back to other available backends. Throughput can drop briefly, but the endpoint keeps serving.
Where do my requests run?
Requests run on GPU fleets from SkyPilot Platform customers, and can fall back to our dedicated GPUs or verified inference providers. Fleets can't join anonymously: each one is a verified enterprise customer under contract. Fleet owners do not operate the model server. We deploy it, and if anyone tampers with it the cluster stops serving.
Isn't this just another aggregator?
An aggregator resells the same providers as everyone else and competes on catalog size. Our supply is different: idle reserved GPUs in enterprise fleets that already run on the SkyPilot Platform. Those hours are already paid for, so serving on them costs little. That is why we can price 40% under list. The router is provider-agnostic, and you can bring your own keys.
Which models are available?
GLM-5.3, GLM-5.3-Flash, Kimi-K3, and DeepSeek-V4.1-Flash today, with more open models added during the preview. You can also route your own provider accounts through the same endpoint with your own keys.
What does joining as a fleet owner involve?
You opt in through the SkyPilot Platform you already run, and you can pause or leave the program whenever you want. You earn on every token served from your fleet; the calculator above estimates the amount.
What actually runs on my GPUs?
A model server, deployed through the same SkyPilot Platform control plane that already manages your cluster. It runs unprivileged in a restricted namespace: no service-account tokens, egress limited to the SkyPilot Tokens proxy, and images and weights pulled only from private sources. It writes nothing to your infrastructure, and the exact manifest is visible on your dashboard.
What are the preview's limits?
Seats and duration are limited. You get $5 in free credits each month, and you top up with prepaid credits beyond that. It works best for batch jobs, evals, and agents, where price and throughput matter more than latency.