Design Partner Program

We're taking a small number of design partners (target: one to three) before opening anything self-serve. The trade is explicit:

You get

• 90 days free, up to ~25 robots' worth of inference.
• A dedicated endpoint (your VPC peering or IP allowlist), your checkpoints deployed, and a measured latency report for your actual network path.
• A latency SLO in writing: p95 chunk latency < 75 ms end-to-end (H100 tier; we currently measure 22.8 ms p95 server-side + ~1 ms network in-region, so this is a conservative ceiling, not an aspiration).
• Direct line to the person who built the serving stack. Feature requests land in days.

We get

• A real fleet workload instead of synthetic benchmarks.
• Permission to reference the results (anonymized until you approve otherwise).
• Blunt feedback on the protocol, onboarding, and pricing.

After 90 days

• Pricing on the order of $1 per robot-hour (final pricing set with partners; our published cost basis is $0.04–0.07/robot-hr, so the number is set by value, not cost).
• No lock-in: the protocol is documented, your checkpoints are yours, export anytime.

Good fit

• 5–100 robots doing learned manipulation (warehouse, kitchen, lab, light industrial).
• Already using or fine-tuning open VLAs (SmolVLA, π0, GR00T, OpenVLA) — or wanting to without buying a GPU per robot.
• An engineer who can integrate a ~30-line client against a documented TCP protocol.

Not a fit (yet)

• Safety-of-life control loops (surgical, automotive). Off-robot inference needs a qualified on-robot fallback we don't provide today.
• Sites with no reliable network path (we'll measure; if physics says no, we say no).

Start: the program is invite-only while we onboard the first partners. A public request form will open here when slots do. Have your robot count, policy/checkpoint, control rate, and site location ready — the first step is a latency measurement on your network path.