〔Day Digest〕Multimodal Inference Splits the Stack: DeepSeek Nears Opus-4.8, OpenAI Open-Sources Codex Harness, Morgan Stanley Sees 2027 Cloud Capex +38% — Apple Ponders an AI Bet

0. Daily Arc

The AI story pivoted from parameter counts to the execution stack and multimodal perception. DeepSeek's experimental vision model matched its V4-Flash text baseline while closing on Opus-4.8 in visual-Agent benchmarks ^[1][2]; OpenAI handed the runtime layer to developers by open-sourcing Codex Harness under Apache-2.0 ^[3]. On the capital side, Morgan Stanley sees 2027 global cloud capex up 38% ^[4], Gelanfei shipped a domestic ARM-based SoC ^[5], and Nvidia is simultaneously courting Rebellions ^[6] and pushing Vera-based EDA flows ^[7]. The unresolved tail is Apple, where incoming CEO John Ternus faces the question of whether to break the cash-neutral posture and buy into AI ^[8].

1. The Multimodal & Inference Stack

• **[NEW] DeepSeek-V4-Flash-Vision-Exp** went live on the DeepSeek API platform as an experimental text-and-vision model: text-task scores (Agent, reasoning, world knowledge) match the V4-Flash production release, while visual-understanding Agent benchmarks show a step-change and approach Opus-4.8 ^[1][2]. A free Files API and DeepSeek Harness 0.1.1 were released alongside ^[1].
• **[NEW] SenseTime** open-sourced SenseNova U1.5 Lite, a lightweight native unified multimodal model targeting real visual delivery, native 4K, and image editing ^[9].
• **[NEW] (single-source, unverified):** independent researcher Ben Davis tested an anonymous model "stealth/ox-alpha" on OpenRouter, reporting it beat GPT-5.6 and Claude on some coding benchmarks with a 1.048M-token context window; Davis says he is "99% sure" it is an unannounced Zhipu GLM-5.x flagship, citing video encoder, tokenizer, and audio-refusal signatures as evidence ^[10]. Zhipu has not confirmed.

2. Compute, Chips, and the Open-Source Lever

• **[NEW] OpenAI open-sourced Codex Harness** under Apache-2.0; with only two harness adjustments (inference preservation and context compression), GPT-5.6 Sol's ARC-AGI-3 score rose from 13.3% to 38.3% and output tokens fell roughly sixfold — OpenAI president Greg Brockman wrote that Codex "can drive far more than coding tools" ^[3]. Same model, different harness, ~3x score, ~6x lower cost ^[3].
• **[NEW] Morgan Stanley** sees global cloud capex growing 38% in 2027 ^[4].
• **[NEW] Nvidia** is exploring a deal with Korean AI chip startup Rebellions ^[6]; separately, Nvidia is partnering with mainstream EDA vendors and demonstrating a Vera-MPU-based (not GPU-only) path for chip-design efficiency ^[7].
• **[NEW] Gelanfei** released its first SoC, the GF800, built on an ARM CPU plus a self-developed GPU, targeting endpoint devices, edge computing, and industry AI ^[5].
• **[NEW] Lightelligence H1 2026** results: revenue RMB 79.78M, up 284% YoY, gross margin ~65.8%, adjusted net loss RMB 147M ^[11].
• **[ESCALATED] A China chip denial** landed ahead of Nvidia's big week — the underlying detail in the headline is not in the packet and the feed is a single Google News item, so treat the substance as single-source ^[12].

3. Embodied AI and Robotics — From Demo to Output

• **[NEW] Xinghai Tu CEO Gao Jiyang** said the company completed 100+ "real fulfillment" orders at a JD-operated front warehouse the day before the 2026 World Robot Conference, with zero human intervention ^[13]. Xinghai Tu is delivering "thousand-unit" model-driven productive units this year, and is now measured on labor output, not unit sales ^[13].
• **[NEW] Huixi** released the Huixi Embodied series, a full-stack compute platform for embodied AI ^[14].
• **[NEW] Yuequan Bionic**, with the Jilin Bionic Robotics Innovation Center, released a self-developed product matrix: the Y-Hand dexterous hand line, YQ Motors, and YQ Muscles artificial-muscle modules ^[15].
• **[NEW] Sugon** showcased its cableless box-style supernode scaleX40 — orthogonal wireless-cable Tier-1 interconnect, 40 GPUs per node, 28+ PFLOPS — at the Anhui Computing Industry Ecosystem Conference, claiming deployment cycles shrink from months to hours ^[16].
• **[NEW] Nvidia–Yuejiang Robotics signal:** Nvidia Omniverse and robotics product marketing senior director and Jensen Huang's daughter Huang Minshan visited the Yuejiang Robotics booth at the 2026 World Robot Conference on August 19 to inspect joint physical-AI R&D results ^[17].
• **[NEW] Baidu Apollo Go** began road tests for a driverless service linking Hong Kong West Kowloon Station and Airport Express Kowloon Station under a pilot license from Hong Kong's Transport Department ^[18].

4. Capital, Consolidation, and the Wild Cards

• **[NEW] Deloitte** projects the 2026 global chip market at $975B, with generative-AI-related revenue over half; the AI-driven semiconductor M&A cycle is expected to run at least through 2030 ^[19].
• **[NEW] Kakao** board approved a split into "Kakao AI" (36% of net-asset book value) and "Kakao X" (64%); an EGM is set for December 17, the split is effective January 1, 2027, and Kakao AI's relisting is targeted for January 27, 2027 ^[20]. KakaoTalk's ~50M users will be repositioned as an "agent-style AI interface" ^[20].
• **[NEW] China Huaneng Group** chairman Wen Shugang and Huawei CEO Ren Zhengfei met in Shenzhen on August 20, with Ren highlighting China's twin advantages in power and telecom networks and identifying AI, computing power, and compute-power-electricity coordination as the next collaboration space ^[21].
• **[NEW] Haibo Sichuang** and partners established Haibo Sichuang Intelligent Computing Investment (Hainan) with a scope covering AI application software, AI basic software, AI public data platforms, and AI industry-application system integration ^[22].
• **[NEW] Anthropic** plans later this year a new Claude security system that still retains enterprise customer data for 30 days but allows storage on customer-owned cloud servers ^[23].
• **[NEW] OpenAI** added an enterprise-facing safety layer that detects misuse without retaining enterprise data ^[24].
• **[NEW] Fujitsu** joined Anthropic, Canon, and Microsoft in a royalty-free AI licensing network ^[25].
• **[NEW] Apple wildcard:** incoming CEO John Ternus takes over in September; Bank of America analyst Wamsi Mohan argues Apple is moving away from its long-held net-cash-neutral posture toward higher R&D intensity and possible large-scale M&A — the question is whether Apple finally buys into AI ^[8].

5. What Would Falsify It

• The OpenAI Codex Harness result is the cleanest causal claim in the packet — same model, 13.3% to 38.3% on ARC-AGI-3, ~6x token reduction ^[3] — but it is a single-source report and the benchmark is relatively new; third-party replication on SWE-bench or internal Agent traces would either confirm or fade it ^[3].
• The "ox-alpha = Zhipu GLM-5.x" attribution rests on a single researcher's behavioral and tokenizer tests, with Zhipu not on record ^[10]. If the real provider surfaces, the GLM-5.x launch narrative resets ^[10].
• Morgan Stanley's 38% 2027 cloud capex call ^[4] is a forward projection; hyperscaler capex prints for 2026 (Alibaba, Microsoft, Meta, Google, Amazon) are the first falsification test ^[4].
• The 2026 World Robot Conference claims from Xinghai Tu ^[13] and Yuejiang ^[17] are vendor disclosures at a trade show — independent site audits and unit-economics disclosure are the test.
• Source-quality flag: items ^[24], ^[10], ^[6], ^[25], ^[12], ^[8], ^[7], ^[3] are single-source in this packet and warrant a second feed before any position is sized on them.