CONTACT SALES
SECURE INFERENCE FOR ENTERPRISE

End-to-end encrypted inferencefor every model.

Prompts are encrypted before they leave your machine and decrypted only where the model runs. Nothing readable crosses the wire, and zero data retention means nothing is kept after the answer.

Sealed in the client. The proxy, the host, and the operator never hold the key.

454 t/s
NEMOTRON 3 ULTRA · #1 ON ARTIFICIAL ANALYSIS
0
DATA RETENTION · ENFORCED AT THE GATEWAY
2.7×
LOWER COST THAN THE #2 PROVIDER
300+
MODELS · ONE ENDPOINT, ONE BILL
01 // SPEED AND SECURITY
MODEL:
GPT-OSS-120B
HARDWARE:
1× NVIDIA B200
ENCRYPTION:
END TO END

Unmatched speed & security, never a compromise.

Agent-tuned, one NVIDIA B200 serves gpt-oss-120b faster than SambaNova and Groq accelerator chips. Every request can stay encrypted end to end.

Output tokens per second, one request

Single stream, 10k-token prompts. Higher is better.

PROVIDEROUTPUT TOKENS/SECEND-TO-END ENCRYPTION
CerebrasWafer-scale engine
1,729.0
NOT OFFERED
BLACKBOX AI1× NVIDIA B200
723.2
YES
SambaNovaRDU
706.8
NOT OFFERED
GroqLPU
473.5
NOT OFFERED
AzureGPU
361.0
NOT OFFERED
Nebius (Base)GPU
357.7
NOT OFFERED

Providers: Artificial Analysis, P50 over the past 72 hours, October 16, 2026. BLACKBOX AI: best verified result from our agent run on one B200, measured on the same methodology by a fixed harness.

READ THE BENCHMARKS
02 // SECURE GPUs

Deploy GPUs at scale, encrypted out of the box.

Rent through Blackbox and each machine comes with access controls, spend limits and an audit trail that bare metal and other GPU clouds leave to you.

ADDED BY BLACKBOX
ONE VERIFIED SIGN-IN
Your team signs in once for every cloud, with a verified email.
SCOPED API KEYS
Each script gets its own key. Revoke it and it stops on the next request.
SPEND LIMITS
Caps on machines, hourly spend and launches are checked before any provider is called.
SSH KEY CHECKS
Weak keys and pasted private keys are rejected. We store fingerprints only.
ONE AUDIT LOG
Launches, terminations, refusals and key changes from every cloud, in one record.
NO PROVIDER KEYS
Your team never holds a cloud provider credential, so there is none to leak.
TALK TO ENTERPRISE
NVIDIA B300
BLACKWELL ULTRA / SXM
NVIDIA B300 Blackwell Ultra SXM accelerator: architectural assemblyA single SXM accelerator module with a cooling plate, chip package, circuit substrate, and socket. The exploded view separates these layers.COOLING PLATECHIP PACKAGESUBSTRATESOCKET MOUNT
PART
B300 SXM
VIEW
EXPLODED ISOMETRIC
LAYERS
4
SCALE
ILLUSTRATIVE
SEPARATION
100%
SHEET
02 / 04
An illustrative B300 assembly: cooling plate, chip package, substrate, and socket. The exploded view separates these layers.
SIGN-IN:
VERIFIED
API KEYS:
SCOPED
PROVIDER KEYS:
NONE
Available to Enterprise customers
03 // FORWARD-DEPLOYED ENGINEERING
ENGINEERS:
FORWARD-DEPLOYED
SCORING:
FIXED HARNESS
BASELINE:
LAST BEST

Recursive improvement, run by agents and our engineers.

A forward-deployed engineer sets up Blackbox inside your environment and points our agents at your workload. The agents tune the engine on your live traffic, and every improvement becomes the baseline for the next cycle.

THE LOOP

  1. YOUR ENGINEER DEPLOYS

    A private endpoint, an air-gapped instance or GPUs you own, set up with your team.

  2. AGENTS PROPOSE

    Agents try changes to the engine against the shape of your real traffic.

  3. THE HARNESS DECIDES

    A fixed harness scores every change. Only a change that beats the current best is kept.

  4. THE BEST IS THE NEW BASELINE

    Changes that hold up across workloads become engine defaults for every deployment.

REPEAT FROM 02, STARTING FROM THE LAST BEST
AGENTS A1-A3NEW CHANGESPEC DECODELIVE TRAFFICNEVER PAUSEDFIXED HARNESSKEPT ONLY IF FASTER
ONE AGENT RUN, GPT-OSS-120B ON ONE B200
Best verified output speed, run start to latest result

Single-stream tokens/sec. Each step up is a change the harness kept.

Best verified output speed, run start to latest result
Agent run, start to latest verified resultValue
Start≈480
Latest723.2
FASTER IN ONE AGENT RUN
1.5×
TOKENS/SEC ON ONE B200
723.2
EXPERIMENTS, SCORED AUTOMATICALLY
83
HUMAN EDITS TO THE ENGINE
0
WORK WITH OUR ENGINEERSHOW THE LOOP WORKS
04 // TRAINING
METHODS:
3
DATA:
YOURS
STEERING:
YOUR ENGINEERS

The job ends. The improvement doesn't.

Other platforms run training as one scheduled job on the dataset you upload. On Blackbox, agents keep working from your data and evaluations, and your engineers steer them.

Teach the model what good work looks like.

Agents fine-tune an open-weight model on examples of the answers your team expects, then keep optimizing it between runs.

FROM YOUR EXAMPLES
Training starts from examples of good answers, drawn from your own data.
TESTED ON NEW CASES
The tuned model is tested on cases it has not seen before it is deployed.
YOUR ENGINEERS DIRECT
Engineers set the goals and review the results. Agents do the optimizing.

Round after round, kept only if better.

Agents start a new round of post-training as soon as the last one is scored, so the model keeps improving after the first job ends.

SCORED ON YOUR EVALUATIONS
Every round is measured against the evaluations your team defines.
KEPT ONLY IF BETTER
A new checkpoint replaces the current model only when it scores higher.
YOUR TEAM STEERS
Engineers set the preferences and policies the agents work toward.
ROUND AFTER ROUNDKEPT ONLY IF BETTERREJECTEDYOUR TEAM STEERS

Reward the outcomes your work needs.

Agents run rollouts continuously against the environment and reward you define, and every round of rewards updates the policy.

CONTINUOUS ROLLOUTS
Agents run the policy through your environment again and again, not once on a schedule.
REWARD YOU DEFINE
You decide what a good outcome is, and every rollout is scored against it.
ENGINEERS STEER EVERY RUN
Your engineers adjust the environment and the reward as the runs go.
POLICYAGENTS RUN ITROLLOUTSUPDATEREWARD YOU DEFINE
TALK TO ENTERPRISEAvailable to Enterprise customers
The most secure inference for enterprise

Any model. Full speed.
Your data stays yours.