Alibaba Builds an Agent-Specific Chip; OpenAI Kills Sora

Alibaba's chip independence push reaches agent AI — and OpenAI pulled Sora while Disney was still in the meeting.

🖥️ Alibaba Reveals 5nm RISC-V CPU Built for AI Agent Workloads

Decoded: Alibaba unveiled the XuanTie C950 on March 24 at an internal DAMO Academy conference — a 5-nanometer server CPU running at 3.2 GHz, built on open-source RISC-V architecture and designed specifically for agentic AI inference. The chip performs more than 3x faster than its predecessor, the XuanTie C920, delivering over 30% performance gains on agent-specific workloads versus mainstream server CPUs. Alibaba's DAMO Academy built the chip to handle sequential, multi-step processing — the core compute pattern of AI agents — rather than the GPU-dominated parallel training workloads that dominate AI infrastructure investment. By using RISC-V rather than Arm architecture, Alibaba avoids royalty payments and gains full instruction-set customization for specific inference patterns. The company's T-Head semiconductor arm is running two parallel chip tracks: the Zhenwu 810E for AI training and inference, and the XuanTie series for cloud and agentic AI. Alibaba sells AI services through Alibaba Cloud, not the chips themselves. (Reuters, CNBC, March 24, 2026)

Why it matters: U.S. chip export controls have pushed every major Chinese AI firm to accelerate in-house silicon. The XuanTie C950 is architecturally notable: it is the first major Chinese CPU built specifically for agent inference — sequential reasoning chains, task decomposition, and multi-step execution — rather than the parallel compute that GPU investment analysis typically focuses on. Nvidia recognized this compute shift at GTC 2026 when it unveiled the Language Processing Unit from its $17 billion Groq acquisition, also targeting inference. Alibaba is converging on the same answer from the opposite direction: open-source ISA, inside its own cloud, built for Chinese enterprise agent deployments. For Arm (ARM), Alibaba's RISC-V commitment represents a design license that will not be signed. For Nvidia (NVDA), agentic workloads running on CPUs are inference loads not landing on GPUs.

🎨 OpenAI Kills Sora — Disney Was in a Meeting About It When the News Arrived

Decoded: OpenAI discontinued Sora, its AI video generation tool, on March 24 — without advance notice to at least one major creative partner. A Reuters source familiar with the matter said Disney and OpenAI teams were actively working on a Sora-linked project on Monday evening when, 30 minutes into the session, the Disney team received word that OpenAI was dropping the tool altogether. OpenAI did not issue a public statement explaining the decision. Sora launched publicly in December 2024 as OpenAI's flagship AI video product, generating photorealistic video from text prompts. The discontinuation arrives as rival ByteDance had already paused its AI video product Seedance 2.0 globally in early March after copyright pressure from Hollywood studios, including a cease-and-desist from Disney itself. (Reuters, March 24–25, 2026)

Why it matters: Sora was OpenAI's consumer AI video product and its direct entry into the creative AI market against Runway, Kling, and Adobe Firefly. Its abrupt end mid-partnership suggests either a sudden strategic pivot or a legal constraint OpenAI did not disclose publicly. The Disney connection adds complexity: Disney sent a cease-and-desist against ByteDance's AI video product in February over copyright concerns, yet was simultaneously an active Sora collaborator — and still was not warned before the shutdown. Whether the cause is copyright exposure, resource reallocation toward agentic products, or pre-IPO risk reduction is not yet public. For investors, the discontinuation removes a consumer AI revenue category from OpenAI's portfolio and benefits Runway, Adobe (ADBE), and Kling, which now compete for creative studio business without OpenAI's video product in the market.

That's your Wednesday signal. See you tomorrow.

— The AI Decoded Team