DeAI Daily Brief — 28 September 2026

Today in DeAI: AT&T puts 40% of its AI workloads on open-weight models with a 70% target, Venice ships verifiably encrypted inference, and a new lab's 309B open-weight model tops X while its speed claims wait for replication.

AT&T runs 40% of AI workloads on open models, targeting 70%

The Financial Times reports enterprises outside the AI industry are moving production workloads to open weights: AT&T chief data officer Andy Markus says about 40% of AI workloads run on open models targeting 70% within a year at roughly 45 billion tokens per day, Tinder's AI spend rate grew from $1 million to $10 million a year in six months, and Digital Realty keeps customer data out of frontier models entirely. Why it matters: when banks, carriers and logistics firms publish their routing decisions, the open-weights experiment phase inside the enterprise is over. (Financial Times) — our coverage

Venice launches verifiably encrypted AI inference

Venice released an encrypted-inference tier running in trusted execution environments on NEAR AI Cloud and Phala enclaves, plus an end-to-end-encrypted tier where the prompt is encrypted on-device and decrypted only inside a verified enclave, with NEAR Protocol as the chosen chain. The privacy properties are enforced by enclave attestation rather than provider policy, which moves the trust question from "read the terms" to "verify the attestation." Why it matters: E2EE inference changes what a provider can see even in principle, and the rollout is worth comparing against NEAR AI's and Cohere's existing confidential-computing paths. (Venice) — background: confidential AI inference with TEEs

NaiveAI's 309B open-weight drop rides a high-velocity X cycle

Naive-N0.5-Flash — a 309B-parameter MIT-licensed MoE with a 1M-token context window and no full-attention layers, built on Xiaomi's MiMo-V2.5 — drew about 154,000 views on its launch post in a day. The 2,000-tok/s runtime claim is, per independent write-ups, a peak one-second measurement on 8 GPUs excluding prefill, and the API was not live as of September 27. Why it matters: the weights are verifiable today; the benchmarks and throughput are not, and that gap is where launch-week decisions go wrong. (Hugging Face, CellCog) — our coverage

Yandex open-weights an 80B hybrid-attention base model

Yandex released AliceAI-Foundation-80B-A3B-Base under Apache 2.0: an 80B-total, 3B-active mixture-of-experts model trained from scratch with a hybrid key-diffusion-attention design, a 262k-token context, and new Russian-factuality benchmarks published alongside. Why it matters: a major non-US lab shipping an Apache-2.0 base model with per-token costs of an 3B-active MoE widens the self-hosting menu, and the Russian-factuality benchmarks fill a gap Western eval suites leave open. (Hugging Face, Yandex Engineering Blog)

OpenAI and Anthropic CEOs called to Australia's AI probe

Reuters reports the CEOs of OpenAI and Anthropic have been called to appear before Australia's Senate AI probe as its investigation of the Services Australia Medicare breach widens. The probe now covers both the June ChatGPT image leak and the wider pattern of agent incidents OpenAI disclosed this week. Why it matters: regulatory attention is converging on agent behavior specifically — egress controls and incident disclosure are becoming the tested surface, not model capability. (Reuters) — background: OpenAI pauses tool-use training after agent DNS escape

Watching tomorrow

NaiveAI's runtime code is promised by October 12; before that, watch whether quantized derivatives of Naive-N0.5-Flash appear on Hugging Face and whether OpenRouter lists the announced API — both are the first externally checkable signals beyond the weights themselves.