TEE-Verified Compute on Sui
Mainnet Launching End of May
Sui Basecamp Dubai 2026 Sponsor
Early Access · For Teams
TRAIN YOUR
LLM ON A NETWORK
NOBODY CONTROLS.
Most AI teams rent GPUs from AWS, Lambda, or Vast.ai — then trust the provider to not log their data, not throttle their jobs, not disappear. VRAM removes the trust requirement entirely. Every compute job runs across a decentralized GPU network, scored inside hardware enclaves, and anchored on-chain. Your model. Your data. Your provenance record.
Vast.ai / Lambda
Centralized provider
No on-chain audit trail
Trust-based scoring
Bittensor / Templar
Validators see gradients
Score can be gamed
No provenance record
VRAM Network
Decentralized GPUs
On-chain checkpoint hashes
TEE-verified · zero trust
Hardware-attested scoring
Every gradient is evaluated inside a Nautilus TEE (AWS Nitro Enclave). No validator ever sees your weights, parameters, or training data. The scoring binary's PCR fingerprint is registered on-chain — modified code fails verification and the score is rejected.
On-chain provenance, forever
Every checkpoint hash is anchored on Sui. Your training run produces a cryptographic audit trail — useful for compliance, for model registries, and for anyone who needs to verify a model's origin without trusting your word.
Distributed training, open protocol
Supports the federated training loop today. HuggingFace-compatible job intake and full SFT support ships with the mainnet training marketplace. Apache 2.0 at TGE — inspect every line.
Pay per window, not per month
No 12-month reserved instances. No hourly minimums. You submit a job, the network prices it, you pay in $VRAM per 10-minute window. Early access teams receive $VRAM credits at TGE.
10-minute reward cycles, not monthly billing
The network settles every window — roughly 10 minutes. Real-time visibility into loss convergence, miner performance, and compute utilization. No waiting for an invoice to know if your job ran.
0
Validators see your data
How It Works
FROM JOB SUBMISSION
TO ON-CHAIN PROOF.
You don't manage infrastructure. You define your job, the network handles the rest — and you get a verifiable record at the end.
01
Submit your training job
Define your model config (HuggingFace-compatible), dataset endpoint, and window budget. We provision the network allocation and return a job ID.
02
Network mines your job
Miners pull your training task each window. They train a slice, upload compressed gradients to your encrypted R2 bucket — credentials sealed via Sui Seal IBE, readable only by the enclave. (Storage migrates to Walrus at v0.7.)
03
TEE validates every gradient
Nautilus enclaves score each gradient against held-out batches, measuring loss improvement. Bad gradients get zero weight automatically, on-chain. The scoring binary's PCR is verified before any result is accepted.
04
Checkpoint anchored on Sui
After each aggregation, a checkpoint hash is written to Sui. Immutable, timestamped, publicly verifiable. Your model's provenance record exists independently of anything we control.
05
Retrieve your model
When training completes, export the final checkpoint. You own the weights, the Sui provenance record, and the full audit trail — all in your own storage bucket.
Comparison
HOW WE COMPARE
TO THE ALTERNATIVES.
Designed for teams who can't afford to trust their compute provider.
VRAM Network
Vast.ai / Lambda
Bittensor / Templar
Decentralized compute
✓ Yes
✗ No
✓ Yes
On-chain audit trail
✓ Every checkpoint
✗ None
✗ None
Gradient privacy (TEE)
✓ Hardware enforced
✗ Provider can see all
✗ Validators see gradients
Tamper-proof scoring
✓ TEE deterministic
Partial (trust provider)
✗ Can be gamed
Fully open source
✓ Apache 2.0
✗ Closed
Partial
No reserved instance / contract
✓ Pay per window
✗ Hourly minimums
✓ Yes
Early access credits
✓ For approved teams
✗ No
✗ No
Ready to Start?
YOUR MODEL.
YOUR PROOF.
Apply for early access. We review every application and reach out personally. Business emails only.
Request Early Access →