The Omniscient Bulletin · 2026-06-04
The Omniscient Bulletin — June 4, 2026
Microsoft ships its own frontier coding model the same week it pulls Claude Code from its Windows-and-Office division; Sam Altman names cost as the customer's second-biggest theme; DeepSeek finally takes outside capital as US Commerce closes the offshore-subsidiary loophole on chip exports; Google ships a frontier multimodal model that runs on a 16-gigabyte laptop.
The bill came due. Microsoft is winding down Claude Code across its Windows-and-Office division and shipping its own coding model the same week. Uber tells investors it burned through its entire 2026 AI budget in four months. Sam Altman tells customers, on stage, that cost is now the second-biggest theme he hears, behind only capability. DeepSeek finally takes outside capital after a year of insisting it didn't need any; Commerce closes the back door it would have used to buy frontier chips. Google ships a frontier-class multimodal model that runs on a 16-gigabyte laptop. The buyer side started saying no this week. The supplier side started building the answer.
Frontier
Microsoft Build ships seven from-scratch MAI models, beats Claude on coding benchmarks, and routes the result into Copilot
At Microsoft Build, Mustafa Suleyman presented seven models built from scratch on commercially licensed data, no third-party distillation. The flagship, MAI-Thinking-1 (35B active parameters, 256K context), posts 97.0% on AIME 2025 and matches Claude Opus 4.6 on SWE-Bench Pro. MAI-Code-1-Flash beats Claude Haiku 4.5 by 16 points on SWE-Bench Pro (51.2% vs. 35.2%) and is rolling to Copilot's model picker. The same week, Microsoft is winding down Claude Code across its Experiences and Devices group. The supplier and the customer are the same company.
Labor
The buyer side starts saying no: Walmart caps tokens, Uber burns its 2026 budget in four months, Microsoft retreats from Claude Code
Sam Altman, at OpenAI's Intelligence at Work event, named cost as the second-biggest theme from customers, behind only capability, and read back the client meme: 'My company spent my entire 2026 budget in Q1.' Walmart capped per-employee tokens on its Code Puppy agent after demand outran budget. Uber burned its entire 2026 AI allocation in four months on Claude Code and Cursor; per-engineer costs ran $500 to $2,000 a month, and the COO said the link to consumer innovation 'is not there yet.' Microsoft is pulling Claude Code from its Experiences and Devices division by June 30.
Fortune / Bloomberg / The Next Web→
Industry
DeepSeek takes outside capital for the first time at $52 to $59 billion, with Tencent and CATL as the largest external checks
Reuters reported DeepSeek is closing its first outside fundraise at roughly $7.4 billion (50 billion yuan), valuing the lab at $52 to $59 billion post-money. Founder Liang Wenfeng is contributing 20 billion yuan; Tencent is in for roughly 10 billion yuan, CATL for 5 billion yuan. Final talks are underway with China's national AI fund, NetEase, and JD.com; the roster is expected to stay under ten names. The lab that bootstrapped through 2025 on founder capital has concluded the next leg of frontier compute cannot be absorbed internally.
Policy
Commerce closes the offshore-subsidiary loophole on AI chip exports the day before the executive order lands
On May 31, the Bureau of Industry and Security affirmed that US export-license requirements for advanced AI chips apply to all firms with a Chinese parent or headquarters, including their offshore subsidiaries. BIS framed its own answer simply: 'The answer is yes.' Nvidia confirmed licences are required to ship controlled products to PRC-headquartered companies. The guidance, which targets Blackwell and Rubin GPUs, landed 24 hours before Tuesday's executive order on frontier models and on top of the rescinded Biden AI Diffusion Framework.
Compute
Google DeepMind ships Gemma 4 12B as a frontier-class multimodal model that runs on a 16-gigabyte laptop
Google DeepMind released Gemma 4 12B on Wednesday: the first mid-sized Gemma with native audio inputs, the first to use an encoder-free multimodal architecture, and small enough to run on a consumer laptop with 16GB of VRAM or unified memory. Google says it 'nears our larger 26B MoE model on standard benchmarks.' Licensed under Apache 2.0, it fills the gap between the edge-friendly E4B and the 26B mixture-of-experts variant. Open weights, native multimodal, laptop-resident.