Unique Research · Interview · August 6, 2024 · Guest: Nie Kaixuan (聂凯旋), Founder & CEO, Songying Tech (松应科技) · Interviewer/editor: Qiuping
Nvidia Omniverse (RTX, physics, OpenUSD) is the reference platform for real-time 3D collaboration, robotics sim, and synthetic data. Songying’s Orca — started months after Omniverse beta in 2021 — aims to be China’s only direct peer: physics-accurate sim + training for embodied robots, aircraft, and autonomous driving, with CAD/CAE, AI, robots, render/physics engines, and sensors under one data fabric.
Interview highlights below.
Q: How does Orca differ from Omniverse?
Nie: Same north star — digitize/intelligize the physical world — but we are the first and still only China vendor that can be lined up against Omniverse. Lineup: OpenUSD real-time 3D data synthesis/render + multi-robot training/sim, with strong progress on embodied training. Advantages: scalable modular architecture (real-time synth, render, physics); gradual open-source ecosystem for embodied sim; distributed multi-GPU beyond single-GPU limits; multi-vendor GPUs (Nvidia + domestic) — not locked to one silicon; born-in-China continuous sim + synthetic-data service (supply continuity). Huang’s physical-AI future needs compute and scenes; Nvidia moves from internet-centric compute to industry-wide infra — Orca wants the same transformation quality in China.
Q: Value in production operations?
Nie: Not a single tool — a real-data sim fabric connecting enterprise systems end-to-end. Agents learn in the virtual yard; results flow back before any physical change. Core: help manufacturers move business logic from human/script execution to AI-driven digital processes — then reshape workflows for market speed.
Q: Historical data vs. real-time consumer demand — the GAP?
Nie: Don’t rip-and-replace the whole plant. Start at the front: AI design concepts → consumer choice; if demand jumps (e.g. 10k → 100k/500k), first simulate capacity strategies virtually, then apply physical change (often via robot reprogramming A→B). Orca supports plant-to-global resource reallocation with lower risk.
Q: Why only you domestically if “simulate before you act” is so valuable?
Nie: Awareness gap. Many still see Nvidia as a GPU seller, missing that CUDA/Omniverse-class software creates the scene value that sells silicon and trains agents/models. Non-specialists underweight the layer; it remains high-potential.
Q: Typical customers?
Nie: AI/robotics cos training embodiments in Orca (sensors: cameras, lidar…); AD partners for safe scenario testing; last year also film/auto-design validations of core modules. (“Qinglong” robot: live work at Lingang Shanghai Electric + Orca virtual factory.)
Q: From buying a mechanism to buying an embodied “employee” / agent?
Nie: Yes — 1:1 physics twin; change virtual first, then physical. Path respects real physics and existing plant infra (e.g. keep Siemens gear) — upgrade efficiency and reduce human-managed steps rather than greenfield replace. Hard parts: perception/decision, integration, data — we provide ongoing support + customer feedback loops.
Q: Multinationals / knowledge transfer; outbound capacity moves?
Nie: Same shared sim world avoids visa/rotation handoff loss when building overseas plants. For China firms moving capacity to MENA/LatAm/South Pacific: virtual industrial env to stress-test staffing, plans, and migration ROI — a day late can cost millions.
Q: Must factories rip out equipment?
Nie: No — intelligize existing workflows; train agents that can run lights-out ops. Integrate design/sim/industrial control/order systems, collect data, train agents as expert operators of CAD/CAE/PLC — not replacements of those tools (aircraft example: agent proposes beyond-A380 designs and orchestrates part plants).
Q: Data mess in Chinese manufacturing?
Nie: Legacy non-standard data is hard. Same bet as Nvidia: unify on OpenUSD for interoperability — skip endless bilateral standard fights.
Q: Why found this; how convince investors; commercialization; roadmap?
Nie: Mid/long-term national & industry gap in high-end physics-accurate sim — create value first, monetize follows. Told angels it’s a long project; grateful for patient angels. Pitch via Huang/Nvidia history (games → crypto → metaverse → AI; ~98% AI training GPU share) and CUDA+Omniverse software moat — if Nvidia is all-in, the layer is valuable. Product launched April; commercialization May; lighthouse customers in marketing/design/manufacturing/ops; expand SI/tech partners. Roadmap: follow fastest physical sectors (factory automation, eldercare, retail, logistics, life services); this year land services; 3–5 years industrialize with SOE/leading manufacturers, then replicate to broader B2B.