Phala

Congrats to iExec on bringing dstack into Nox’s Chain of Trust. Nox is iExec’s confidential execution protocol for DeFi and RWA use cases. It lets smart contracts work with encrypted handles while off-chain runners compute over private values. dstack fits at the runner layer. It runs the Nox worker inside a measured CVM, with quote/attest evidence exposed through the dstack SDK path. That lets Nox verify TDX/RTMR evidence, compose hash, and request freshness before trusting the runner. That gives the Chain of Trust a concrete flow: quote → compose hash → freshness → KMS delegation → result acceptance. Links: - dstack docs: https://lnkd.in/e_UZxqh3 - Nox repo: https://lnkd.in/eKTHWEYd - Phala blog: https://lnkd.in/eR-pnhSh

VoxCPM2 turns multilingual TTS and voice cloning into an AI app primitive. Phala keeps voice prompts, audio inputs, cloning config, and app creds inside a TEE CVM. Deploy: https://lnkd.in/gWk79h4h Template code: https://lnkd.in/gX6t8bX3 Upstream: https://lnkd.in/eB9xqFAJ

Blackwell changes the confidential AI performance story. In a new paper by Hang Yin and Kevin Wang, B300 GPU Confidential Computing keeps GPU-local work close to native: BF16 matmul runs at 0.998x of non-confidential performance, and a 96,000-matmul CUDA graph runs at 1.0012x. The bottleneck moves to the confidential VM-to-GPU bridge. Same-context host/device copies serialize. non_blocking copies still block the CPU thread. Small crossings pay roughly 330 us of setup. That changes how confidential inference systems should be built: - schedule bridge crossings deliberately - pool CUDA contexts - make runtime defaults CC-mode-aware - keep weights and KV state resident - expose fabric health as part of the tenant boundary The fixes are concrete: --no-async-scheduling recovered 57% of the dense-decode CC gap, and a CC-aware loader cut GPT-OSS-120B load time from 287.09s to 8.36s. Read the Phala breakdown: https://lnkd.in/ezTeQaMa Paper: https://lnkd.in/evYCQQsQ

Open Notebook brings NotebookLM-style research notebooks to self-hosted AI. Phala keeps sources, notes, prompts, and app state inside a TEE CVM. Deploy: https://lnkd.in/esk5Hdfn Template code: https://lnkd.in/eeKrKhWp Upstream: https://lnkd.in/geyZ7QGu

LegalCiteBench was accepted to the AI4Law Workshop at ICML 2026. The paper studies a concrete legal AI failure mode: whether LLMs can name the right case and cite it correctly without retrieval. Across 21 models, citation generation remains weak. The best scores are below 7/100 on both citation retrieval and citation completion. Verification performs much better, reaching 75.6 on citation error detection and 96.1 on case verification/correction. The paper also introduces Misleading Answer Rate (MAR): how often a model gives a concrete citation or case answer while performing poorly on retrieval-heavy tasks. 20 of 21 models are above 94%. The takeaway for legal AI systems: use retrieval, citation normalization, verification, and calibrated abstention. Exact legal authority needs source checking before use. Blog: https://lnkd.in/eqgsq2S2 Paper: https://lnkd.in/erT6Hhvn Code/data: https://lnkd.in/eHYC48eE

OpenCV turns images/video into AI-ready signals. On Phala Cloud, camera frames, CV pipelines, and extracted vision features stay inside a TEE CVM. Deploy: https://lnkd.in/eZprYDra Template code: https://lnkd.in/eN-QYXcQ Upstream: https://lnkd.in/dc2iZwJQ

OPPO × Phala published a joint paper on Kubernetes Pod-level remote attestation for confidential AI infrastructure. The core idea: before an AI workload receives secrets, data, or service access, the verifier can check the exact Kubernetes Pod: hardware measurement, container image, Pod identity, and cluster context. dstack-capsule brings this into a cloud-native model with Pod-level attestation, privilege fuse, multi-layer sandboxing, and an implementation on Intel TDX + Sysbox. Paper: https://lnkd.in/ebkuqzt3 Blog: https://lnkd.in/ePyWbh_E dstack: https://lnkd.in/eNWYbgKx

Building trust machines for AI 🤖 Phala at #CCSummit

Mattermost is the open-source hub for team chat, incident response, and DevOps workflows. Phala keeps channels, webhook tokens, and app state inside a TEE CVM. Deploy: https://lnkd.in/eJyfMyX9 Template code: https://lnkd.in/eJH44Qbg Upstream: https://lnkd.in/d7a_XExN

PropellerHeads launched Turbine, a trading protocol for large Ethereum trades. Their core trust problem is exactly where TEE matters: private order flow, protected settlement keys, and verifiable production code. The case study shows how Turbine runs the batch solver, Fynd router, and settlement signing path inside a Phala Cloud Intel TDX confidential VM. Orders enter through dstack, routes are computed privately, settlement is simulated and signed in-enclave, then submitted on-chain. The on-chain KMS adds governance: only approved enclave measurements can receive the signing key, so traders can verify which code is authorized. Read the case study: https://lnkd.in/e64QyzpV