Unique Research · Insight / study-tour dispatch · January 3, 2024 · Hosts: Unique Research lead Wu Wei (Delta) and the tour organizers · Contributors: Luo Hui (Yunxigu), Cheng Weizhong (Zhongke Shenzhi), Wang Ling (DREAM OASIS), Dr. Xue Chong (Quanzhentong)
From 26 November to 5 December 2023, Unique Research lead Wu Wei and the organizers took a cohort of China-based AI founders through Silicon Valley and New York. The ten-day itinerary mixed sessions with NVIDIA, Meta, and Google; founder conversations at U.S. startups FlashIntel, Collov, Boding International, and Gloco; campus visits to Stanford, Princeton, and Rutgers (whose AI/robotics labs rank among the U.S. top three per CN); and a closing discussion with McKinsey AI-application specialists on enterprise landing. The most colorful threads were insider chatter on the OpenAI leadership drama and stories from crypto circle veterans.
For many participants it felt like a compressed “mini study-abroad”: each founder mapped frontier practice back to their own vertical and left charged by the density of U.S. AI activity.
NVIDIA’s next GTC runs 18–21 March 2024. Unique Capital plans a follow-on U.S. camp covering Seattle and the Bay Area, GTC attendance, and China–U.S. AI project roadshows in Silicon Valley. Interested founders may scan the CN QR for consults.
Stanford visit & a founder-led market
Domestic joke of the season: many founders no longer “to VC,” and many VCs no longer “to LP.” The U.S. trip underscored an AI investment market still driven by founders — and by conviction. Roughly 40% of U.S. VC dollars sit in the Bay Area; without an AI narrative, fundraising is hard. Stanford innovation-circle members each pushing their own AI project rekindled mid-career founders’ appetite to build. Live WeChat shares kept remote friends in the loop; the notes below are what returned founders chose to publish.
Luo Hui, CEO, Yunxigu Tech
1. Campus and early-career founders everywhere. Often 3–5 classmates or colleagues spin up a side project part-time — low burn. Microsoft and Google supply compute. Teams ship, get metrics or revenue, then raise. Vertical LLM bets span mental-health companions, virtual partners, support bots, marketing, AI tutors, smart home décor, and more. Dense app experimentation feeds foundation-model progress; China needs the same app layer for a healthy ecosystem.
2. NVIDIA Q&A on cost/performance paths. Luo floated four paths: (A) more/better data and leaner params for cheaper inference; (B) algorithm work that cuts GPU burn; (C) chip-architecture iteration; (D) process-node shrinks. NVIDIA engineers downplayed D: 7nm is “enough,” process cost share for GPUs is limited, and 2–3nm may raise cost without clear power wins. A/B/C all viable; NVIDIA leans on C — e.g. GH200 near-memory (CPU/GPU shared memory) claiming ~7× performance. China SoC shops chase similar in-memory compute to break von Neumann “memory/power walls.” AI will remake chip architecture and OSes; PC/phone/embedded storage stacks and OS layers may see major change in 1–5 years.
3. AGI in 1–3 years was a common hallway view — after physical strength, general intelligence may also surpass humans. Prepare accordingly. Luo thanked Unique Capital for immersing the group in a leading AI environment.
Cheng Weizhong, CEO, Zhongke Shenzhi
1. GenAI remakes the startup template. Small teams win: ship core algorithms yourself, rent the rest. Capital piles for user growth matter less early; young talent can try. 2. Shorter fail cycles. Ship fast, read feedback, quit in 2–3 months if needed, return to Big Tech — and share failure notes openly. 3. China leads some GenAI niches (e.g. digital humans, where quality bars are higher). Should those teams attack the U.S. market?
Wang Ling, CEO, DREAM OASIS
AI saturates every conversation — projects, LPs, even Web3 chats rebranded around GPT. Big-Tech ICs often R&D across 2–3 domains. Side projects of 3–5 people iterate in 2–3 months. Sharing culture and data-first salon talk feel more research-bottomed than China community BD talk. Valley founders run deep tech + talent/capital/compute; China founders often win on business-model clarity, cost control, and execution. Wang’s AI-startup notes: teams of 3–5 can reach meaningful scale; earliest wins may sit in to-professional / SMB efficiency and some large-B point tools; consumer UX still needs better audio-visual / persona fidelity before durable paid intent; large-B AI is slowed by process complexity and internal politics.
Dr. Xue Chong, Founder & CEO, Quanzhentong
~10 AI founders on the trip. Xue argues LLM “hallucination” and forgetting are features for human-like behavior, not only bugs — his firm’s AI-assisted EMR already drafts notes and decision support for doctors. Company styles differ: Jensen Huang stays close to new NVIDIA projects; Meta’s campus feels like a free utopia with biweekly all-hands from Zuckerberg; Google staff seemed calm about GPT and confident in Gemini. Xue’s takeaway map: Meta stays open-source; OpenAI and Google monetize; NVIDIA stays ahead; China chip firms catch up. Chinese AI founders abroad often share U.S. elite-school pedigrees and early USD funding, pricing ~7–8× and R&D cost ~3–5× vs China peers at similar stage — a software analogue of early manufacturing export. Rutgers (founded 1882) PI sessions felt like pure research days at JHMI. Talent density in U.S. universities remains a core U.S. edge.
Next camp preview — 10–18 March 2024
With GTC 18–21 March (Huang keynote, 600+ sessions, 200+ exhibitors), Unique Capital’s next camp covers Seattle & Silicon Valley, GTC, a Valley investor roadshow for China founders, visits to Microsoft / Amazon / Google AI orgs, startup founder sessions, and top-university stops.
15 CN images (campus, GenAI meetup, NVIDIA, Meta, Google, Rutgers, McKinsey, Fisherman’s Wharf, Collov, FlashIntel, Boding, etc.); attach on ingest.
— END —