Moonshot's low prices weren't an attack on capitalism. They were capitalism's opening move.
Moonshot shipped models that matched or beat Western leaders on key benchmarks — then priced API access at a fraction of the cost. K2 cached input at ~$0.15/M tokens vs. OpenAI's higher rates.
The margins were never the point. Those prices were a live demonstration for a different audience: venture capital. By proving world-class performance at disruptive cost, Moonshot built the growth narrative needed to justify a valuation sprint from $4.3B to $31.5B in ~18 months.
This isn't state-subsidized dumping or anti-capitalist warfare. It's hyper-capitalist story-selling. The model is the marketing deck. The real product is the funding trajectory — $3.9B raised in six months, ARR tripling in a quarter. Capital flows to the narrative, not the margin.
If you're benchmarking Chinese AI on unit economics, you're measuring the wrong thing. Their moat isn't pricing power — it's access to patient, strategic capital willing to fund years of losses for positioning. The competitive threat isn't cheap tokens; it's a balance sheet that outlasts yours.
How many European AI startups could survive three years of negative unit economics while scaling? That's the real gap.
One story, unpacked. Now the rest of the week's China signals, in brief:
- Test fleet consists of Apollo RT6 robotaxis with safety operators onboard.
- Initial trials take place in Brent, a London borough with mixed urban-suburban routes.
- Public rides scheduled to begin in 2027.
- Expands right-hand-drive market presence after Hong Kong launch.
- Authorization covers Level 4 testing on urban roads nationwide.
- First Chinese company to gain such comprehensive approval in Germany.
- Uber raised its stake in Momenta.
- WITA-Omni uses a Thinker-Talker-Actor architecture for parallel processing.
- Achieved six first-place scores out of eight, including audio-visual alignment (86.13) and event temporal reasoning (84.64).
- Eliminates disconnect between speech and movement by synchronizing gestures, expressions, and speech rhythm.
- Follows AgiBot's world model Genie Envisioner-Sim 2.0 topping WorldArena.
- HyWorldVLA combines pixel-level and latent-space world modeling with Vision-Language-Action architecture.
- Three-stage training: video VAE, autoregressive token prediction with dual supervision, and action generation.
- Ablation: removing pixel-level prediction drops PDMS to 87.50, removing latent-space processing drops to 89.91.
- Optimal latent feature supervision weight 0.1 achieved best score.
- Series F funding of ~€3.29B ($3.5B) at post-money valuation of ~€32.9B ($35B).
- Round was multiple times oversubscribed and closed early.
- Series G (Pre-IPO) starts earlier than planned with pre-money valuation of ~€47B ($50B).
- Kimi is a Chinese AI chatbot start-up competing with OpenAI.
China Signals analyzes sources from the Chinese-, English- and German-language spheres and contextualizes them for decision-makers with China exposure. All information is always provided with original sources for independent verification.
Imprint · Privacy