Confidential Computing for LLMs — TrustedRouter
Run LLM inference behind hardware attestation across every provider. GCP Confidential Space, with remote attestation.
Run LLM inference behind hardware attestation — across every provider.
Confidential computing has been deployable for years. The industry just never put inference behind it. TrustedRouter does: the gateway runs inside GCP Confidential Space and signs the exact binary it's running.
You challenge it with a nonce, get a JWT signed by the CPU's hardware root key, and match the image digest to the open-source build. That's confidential computing applied to the one data path that's becoming the most sensitive: your prompts.
NONCE=$(openssl rand -hex 16)
curl -s "https://api-azure.trustedrouter.com/attestation?nonce=$NONCE" | jq .
# JWT signed by the hardware root key:
# eat_nonce your nonce (replay-protected)
# image_digest SHA-256 of the running container
# pcrs boot-time platform measurementsHardware root of trust.
GCP Confidential Space isolates the gateway from the host. Even the cloud operator can't read what's inside.
One attested production path.
The hosted gateway terminates prompt traffic only inside GCP Confidential Space. There is no non-attested fallback path.
Verify before you trust.
The nonce-bound attestation endpoint lets any client confirm the running image on demand — not just at deploy, but per request.
The overhead is gone. The excuse is gone.
GCP Confidential Space overhead is small compared with model latency. There's no longer a performance reason to run inference outside a TEE — only inertia.
TrustedRouter makes confidential inference a one-line base_url change, across 30+ providers, with the attestation exposed for you to check.
The full argument: “Attestation is All You Need”.
Attestation proves the running binary is the published binary on hardware you can challenge. It does not defeat a nation-state with physical access to the host, and it does not prove the open-source binary is bug-free. The trust anchor is the hardware and cloud attestation chain.
Current routes, prices, privacy, and measured performance.
Catalog facts come from the routes currently configured in TrustedRouter. Performance uses the same cached metadata snapshot as the public leaderboard. Prompts and outputs are not part of these measurements.
| Model | Providers | Context | Input | Output | Privacy | Measured route |
|---|---|---|---|---|---|---|
Anthropic: Claude Opus 4.8anthropic/claude-opus-4.8 |
|
1,000,000 | $5.25/1M | $26.25/1M | varies 5 cited scores | 1390 ms TTFT anthropic · 100.00% available · n=3 |
OpenAI: GPT-5.5openai/gpt-5.5 |
|
1,050,000 | $5.25/1M | $31.5/1M | ZDR 3 cited scores | Warming up |
Google: Gemini 3.5 Flashgoogle/gemini-3.5-flash |
+1
|
1,048,576 | $1.575/1M | $9.45/1M | ZDR | measured google-ai-studio · 0.00% available · n=8 |
MoonshotAI: Kimi K2.7 Codemoonshotai/kimi-k2.7-code |
+8
|
262,144 | $0.735/1M to $0.9975/1M | $3.675/1M to $4.2/1M | ZDR 5 cited scores | measured inceptron · 0.00% available · n=9 |
Z.ai: GLM 5.2z-ai/glm-5.2 |
+22
|
1,048,576 | $0.714/1M to $1.575/1M | $1.575/1M to $5.5125/1M | E2EE 4 cited scores | measured engy · 0.00% available · n=43 |
MiniMax: MiniMax M3minimax/minimax-m3 |
+8
|
1,048,576 | $0.2835/1M to $0.63/1M | $1.155/1M to $2.52/1M | ZDR 4 cited scores | 2045 ms TTFT wafer · 100.00% available · n=13 |
Browse every modelReview provider policiesOpen the full leaderboardSnapshot 2026-08-05T00:30:16Z