Good morning. It's Monday, March 16th, 2026. Today is NVIDIA GTC keynote day — Jensen Huang takes the stage at 11 AM Pacific, and we'll cover actual announcements in tomorrow's briefing. For now, here's what dropped over the weekend and this morning.

**OpenAI GPT-Realtime-1.5 — Voice AI just got meaningfully better** — OpenAI quietly released GPT-Realtime-1.5 and GPT-Audio-1.5 last week, and the numbers are substantial. Audio reasoning is up 5 percent, hitting 82.8 percent on the Big Bench Audio benchmark. Alphanumeric transcription — codes, serial numbers, order IDs — improved by over 10 percent. Instruction-following in live conversation is up 7 percent. The headline feature is tool calling mid-conversation: your voice agent can now query a database or hit an external API while the user is still speaking, with no awkward pause in the audio stream. Pricing is unchanged at $32 and $64 per million audio tokens, and it's a drop-in upgrade — same WebSocket API. Early developer reports describe substantially improved connection reliability, though these figures reflect informal testing rather than OpenAI official benchmarks. One caveat: a subset of users on Reddit have flagged non-English language degradation in GPT-Audio-1.5 specifically, so multilingual deployments should test before upgrading.

**MCP 2026 Roadmap — the protocol grows up** — On March 9th, MCP lead maintainer David Soria Parra published the official 2026 roadmap, and it signals a meaningful maturation of the protocol. The biggest structural change: MCP is moving away from release-milestone-driven development toward Working Groups as the primary vehicle — similar to how the W3C and IETF operate. Three priority areas stand out. First, a "Streamable HTTP" transport that lets MCP servers scale horizontally behind load balancers without holding sticky session state — solving the biggest production deployment headache today. Second, a dot-well-known metadata format so MCP servers can be discovered and described in a standardized way. Third, context portability: context snapshots that can be passed between agents, sessions, or even devices. The spec itself hasn't changed since November 2025, but this roadmap sets the direction for what production MCP infrastructure will look like by end of year.

**AMD officially endorses OpenClaw on Ryzen and Radeon hardware** — AMD published a best-known-configuration guide this week for running OpenClaw locally on Ryzen AI Max Plus processors and Radeon GPUs — the first major chip vendor to officially document and endorse the platform. They're calling the two configurations RyzenClaw, which runs through WSL2 on Windows, and RadeonClaw, for Radeon AI PRO GPU setups. This isn't just a blog post: it's AMD's developer relations team publishing a validated, supported configuration for running local AI agents without any cloud dependency. Combined with the MINISFORUM N5 Max NAS shipping with OpenClaw preinstalled on the same Ryzen AI Max Plus 395 chip, you're seeing the beginning of a hardware-native OpenClaw ecosystem — the same way Android became a default on Android phones. Intel hasn't made a similar move yet.

**Alibaba's enterprise agent push — Qwen-powered, imminent** — Bloomberg reported this morning that Alibaba is preparing to launch an enterprise-focused agentic AI service as early as this week, built on its Qwen large language model. This is separate from the consumer OpenClaw deployment app Alibaba launched on March 13th. The enterprise product is described as an autonomous agent for company operations — think automated workflows, enterprise data access, internal tool orchestration. This makes Alibaba the third major Chinese cloud platform to commercialize in this space, following Baidu's DuClaw service and Moonshot AI's Kimi Claw. The timing is notable: it comes as China's OpenClaw craze is driving government-level attention in Shenzhen and Tencent free-installation booths at their headquarters. We'll cover the actual product when it launches.

That's the briefing for Monday, March 16th. Watch for GTC announcements starting at 11 AM Pacific — Jensen's promised a chip that will surprise the world. We'll have the full breakdown tomorrow.