01enterprise · live
OpenRouter
enterprise tier · drop-in
“Drop-in OpenAI-compatible endpoint with verifiable, no-log routing. The receipt is the audit trail.”
18B+ tokens
no-log · verified routing
Private LLM catalog
Frontier models with private runtime.
OpenAI-compatible models with hardware-backed privacy and ACI verification. Keep your SDK flow, change the endpoint, and copy the real call when you need it.
Phala: Qwen3.8 27B Uncensored (Aggressive)
262K 上下文
$0.30/M input
$1.50/M output
$0.05/M cache read
Model requests are routed through confidential AI providers with TEE support.
Check allPrivate inference, by construction.
What you say to the model stays between your client and an attested CVM. Three primitives — encryption, TEE, no-logs — make that a property of the build, not a promise.
Encrypted into the TEE
- Encrypted on both hops, each channel key bound to attestation
- RA-TLS terminates inside the CVM, not at a load balancer
- No plaintext intermediary on the host
工作原理
逐步查看单个请求的端到端流程。
关闭 dstack，看看究竟失去了哪项保障。
dstack 上的隐私推理
ACI gateway attestation, signed receipts, and verified upstream model sessions
活跃边缘未激活步骤中新增
步骤 1 of 5 · 随滚动响应1
步骤 1 / 5发送任何字节前先验证构建
Client SDK fetches the gateway’s TDX quote with a fresh nonce and runs dcap-qvl locally — when the caller pins accepted compose hashes, it confirms the measured release is one of them. The trust decision is client-side; Phala is not asked to vouch for itself.
With dstack: 用户持有信任根，锚定在 Intel 的 TDX 硬件签名之上。
同一 SDK。同一端点。默认机密。
cURL · 即插即用
Hit inference.phala.com/v1/chat/completions with the OpenAI request shape. Receipt headers come back on every response — even from curl.
cURL
$ curl https://inference.phala.com \/v1/chat/completions \-H "Authorization: …" \-d '{"model":…}'x-receipt-id: rcpt-e0ee..x-aci-keyset-digest: sha256:3eff..
PYTHON
from openai import OpenAIc = OpenAI(base_url="…phala.com/v1",api_key=os.environ["PHALA_AI_API_KEY"])r = c.chat.completions.create(…)
OpenAI Python SDK
`base_url="https://inference.phala.com/v1"` and you’re done. Existing code keeps working; capture x-receipt-id from the raw response when you need proof.
一个统一的验证器
Every route returns the same ACI artifacts: a gateway attestation report, a signed receipt, and an attested-session reference for the serving upstream.
UNIFIED PROOF
unified verifierall match
TinfoilDeepSeek V4.1 Flash 1017PhalaKev 4B
0G Private ComputerGLM 5.3
Near AIGLM 5.3 Flash
ChutesQwen3.8 27Bone format · any provider
OPENROUTER
openrouter · phala2026-10-26
41.2Btokens / day
models with traffic10
sourceopenrouter.ai/provider/phala
Live on OpenRouter
Phala serves confidential models as a provider on OpenRouter. The daily token count comes from OpenRouter’s public Phala provider page.
two-hop RA-TLS · TLS key bound to TDX quote
tunneled · no plaintext intermediary
hop 01 · client → gateway
inference.phala.comSPKI in gateway report
hop 02 · gateway → model CVM
Phala model CVMSPKI in model quote
RA-TLSSPKI pinX.509tunneled
两跳 RA-TLS，直达模型
The first TLS hop terminates inside the gateway CVM, whose TLS key is bound to its TDX quote. For Phala-hosted models, the second terminates inside the model CVM, whose TLS key is bound to its own quote. Other providers bind the second hop to keys from their own evidence. There is no plaintext intermediary on the host.
response · /v1/chat/completions
200x-receipt-idrcpt-e0eefe…x-aci-identitysha256:3def…x-aci-keyset-digestsha256:3eff…sessionupstream.verified
verify receiptmatches attestation
Signed receipt + attested session, every response
Every response carries x-receipt-id. Fetch the ACI receipt, verify its signature and response hash against a fresh gateway attestation, then follow upstream.verified.session_id when you need deeper audit evidence.
in production today · 3 live partners
Confidential inference, in production.
OpenRouter routes its enterprise tier through Phala. NEAR AI ships verifiable agent inference. OODA AI runs decentralized GPU TEE.
02web3 · live
NEAR AI
verifiable agent inference
“Verifiable agent inference for autonomous, on-chain workflows. Every model call lands on-chain with proof.”
100% receipts
on-chain verified · zk inference
03public-co · live
OODA AI
NASDAQ-listed · decentralized GPUs
“Decentralized GPUs with hardware attestation guarantees. No host root, no off-band access, no policy promises.”
12M tokens / day
TDX + H100 · hardware-attested
OpenAI-compatible
drop-in /v1 surface
TDX + H100/H200/B300
CPU + GPU TEE
5–15% overhead
vs bare-metal
No host root
compose-hash IS the policy
AI 解决方案路径
在 AI 触及密钥时使用隐私模型。
隐私模型端点是第一个入口点。同样的隐私原语也适用于代理、数据工作流和训练。
Agents
隐私 AI 代理
在可验证的运行时中运行代理的密钥、工具、记忆和操作，而不是放在可见的自动化云中。
Training
隐私模型训练
在保持数据集、梯度、检查点和评估轨迹处于边界内的同时，基于专有数据调整模型。
private training run
Observe without exposing weights.
01
dataset
sealed
02
fine-tune
running
03
eval
private
04
checkpoint
verified
loss curve
proof attached
attestation.json
Data
隐私 AI 数据
将模型移动到敏感记录旁，在不向模型运营方暴露原始数据的情况下返回已批准的输出。
source
EHR data
source
Customer records
source
Internal docs
TEE clean room
query without raw access
approved output
aggregate only
no row exportproof linked
Deploy private inference
OpenAI-compatible. Attested. Receipt-backed.
Drop in with the OpenAI SDK you already use. Point at inference.phala.com and capture x-receipt-id for the signed ACI receipt.
- 01OpenAI-compatible base URL
- 02TDX + H100 / H200 / Blackwell
- 03Signed ACI receipt per response
- 04Gateway attestation + upstream sessions
- 055–15% TEE overhead vs bare-metal
