Apple Vision Pro Chief Leaves for OpenAI Hardware Role
On June 27, sources close to the matter told NUPIAO that Paul Midey, Apple’s Vice President overseeing both the Vision Pro headset and the company’s smart glasses R&D division, is stepping down next week. He’s set to join OpenAI’s hardware team immediately after, where he’ll lead the development of their next generation of full-suite AI smart devices. For nearly seven years, Midey has been at the helm of the Vision Pro hardware engineering team, a device that was once hailed as Apple’s next core computing platform.
Guo Mingji: iOS 27 Will Double Down on System-Level Apple Intelligence Integration
On June 27, Mingji Guo, a long-time analyst tracking Apple’s supply chain from TrendForce International Securities, released a report highlighting that iOS 27 will take system-level integration with Apple Intelligence to the next level. His latest industry survey reveals a key detail: to ensure smooth performance under heavy AI workloads, the upcoming mid-range iPhones powered by the A20 processor in the first half of 2027 will bump up their DRAM specs to 9GB (1.5GB × 6-die). That’s a noticeable jump from the current A19 models which sit at 8GB (2GB × 4-die). Meanwhile, the three high-end models expected in late 2026 featuring the A20 Pro chip—including the foldable and two 18Pro variants—will stick with 12GB (1.5GB × 8-die).
DeepSeek Teams Up with Peking University to Launch DSpark: Boosting Inference Speed by 60–85%
On June 27, the DeepSeek team joined forces with Peking University to publish a research paper titled “DSpark.” Based on speculative decoding techniques, this new framework proposes a fresh approach to accelerating large language model (LLM) inference.
The paper points out a major flaw in existing parallel “draft generation” methods: while they can generate longer tokens at once, the weak correlation between tokens often leads to higher rejection rates and wasted verification compute power. To solve this, DSpark introduces a semi-autoregressive structure. It adds a lightweight sequential module onto the parallel generation backbone, significantly strengthening dependencies between tokens and boosting draft quality.
Additionally, DSpark features a “confidence-based dynamic verification mechanism.” This allows the system to adaptively adjust verification lengths based on request success probabilities and current system loads, effectively cutting down on useless computation. Offline tests showed a massive improvement in acceptable generation lengths. When deployed in the live DeepSeek-V4 system, the method boosted inference speeds by roughly 60% to 85% compared to baseline models, while also reducing throughput loss during high-concurrency scenarios.
Qualcomm Plans to Bring Datacenter Chip Tech to Smartphones for Better On-Device AI
News broke on June 27 that Durga Malladi, Executive Vice President at Qualcomm, announced plans to bring the datacenter chip technology unveiled this week directly into smartphones. The goal? To supercharge local AI capabilities on mobile devices.
Qualcomm’s newly launched High Bandwidth Computing (HBC) architecture uses a vertical stacking design, tightly integrating memory with compute units. This setup dramatically improves data transfer speed and efficiency. The first generation of this architecture is scheduled for launch in datacenters next year, with commercial availability expected by 2028. “Datacenter technology won’t stop there,” Malladi stated, noting that talks are already underway with smartphone, PC, and automotive manufacturers to bring these advancements to their ecosystems.
Agibot Rolls Out Its 15,000th General-Purpose Embodied Robot
On June 28, NUPIAO learned that Agibot has officially announced the mass production exit of its 15,000th general-purpose embodied robot, the Elf G2.
This follows a rapid pace of production milestones: Agibot celebrated the rollout of its 10,000th unit, the Expedition A3, back on March 30 this year. Just before that, on December 8 last year, they marked the delivery of their 5,000th general-purpose embodied robot.
Embodied World Model “WoW” Passes Regulatory Filing
On June 26, the dual-brain models Pelican-VL and WoW (“WoW”) from the Huishi Kaiwu platform at the Beijing Humanoid Robot Innovation Center successfully completed the latest batch of generative AI service filings with the Beijing Cyberspace Administration.Beijing Humanoid is now set to officially launch token services for its entire Huishi Kaiwu model series, planning to roll out API access capabilities in phases for industrial clients, research institutions, and developers.
