Good morning. It's Sunday, March 29th, 2026. Here's your tech briefing.

**Artemis II — three days to launch.** The crew has arrived at Kennedy Space Center, and it is real now. Commander Reid Wiseman and Pilot Victor Glover, along with Christina Koch and Canadian astronaut Jeremy Hansen, flew in on supersonic jets and met the press yesterday. Liftoff is targeting April 1st at six twenty-four PM Eastern — and for the first time since Apollo 17, humans will venture beyond low Earth orbit. The crew is now in quarantine, the SLS rocket is on Pad 39B, and NASA says there are no outstanding issues. This one is actually happening. Three days.

**Huawei's new AI chip is finding its market.** Reuters broke an exclusive overnight — ByteDance and Alibaba are both planning to place orders for Huawei's new 950PR AI chip after customer testing delivered strong results. Huawei plans to ship around 750,000 of these chips this year, and the news lands as a meaningful milestone: two of China's biggest AI spenders are now pivoting toward domestic silicon rather than waiting for Nvidia hardware they can no longer get under export controls. The 950PR is purpose-built to challenge Nvidia in the China market, and while it won't match Blackwell on raw performance, it's apparently good enough for production workloads. This is the most concrete sign yet that China's AI infrastructure stack is becoming genuinely self-sufficient.

**Mistral enters the voice race with Voxtral TTS.** Mistral AI has released Voxtral TTS, an open-weight text-to-speech model — and it is notably good. The 4-billion parameter hybrid model achieves 70-millisecond streaming latency, making it competitive with proprietary realtime voice APIs. The architecture is clever: a transformer backbone handles semantic meaning, a flow-matching module handles acoustic texture, and a neural codec handles final audio output. The result is natural-sounding, low-latency speech that can run locally. It drops under a CC BY-NC license, which means developers can self-host it without the data privacy concerns or per-character pricing of cloud APIs. For anyone building voice agents or offline assistants, this changes the calculus. Notably, Voxtral landed the same day as two other open-source TTS models — from Cohere and Tencent — making last Thursday something of a voice AI triple-drop.

**RSAC 2026: agentic AI is real, adoption is not.** The RSA Security Conference wrapped in San Francisco this week, and the headline theme was the gap between where vendors are and where enterprises actually are. SiliconAngle's breaking analysis put it bluntly: vendors are in the third inning, enterprises are still in the first. AI ROI at scale remains elusive for most organizations — only mid-to-low teens of firms are seeing measurable returns. Meanwhile, the security angle is darkening fast. Cisco warned that one wrong agentic action can cause irreversible enterprise damage, and the conference floor was full of new products addressing the MCP security crisis we've been tracking all month. The agentic wave is real, but the gap between demo and deployment has never felt wider — or more consequential.

That's your Sunday morning briefing. Artemis II launches Tuesday. The hardware wars are getting interesting. And the voice stack just got a lot more open. I'll be back tomorrow.