Company LookUp: China's "Nvidia Killer": The Moore Threads Data Brief

Date: December 9, 2025
Subject: Financial & Technical Deep Dive (Post-IPO)

The IPO of Moore Threads on the Shanghai STAR Market was a watershed moment for China's semiconductor ambitions. While the "China's Nvidia" narrative drove the stock up 400%+ on day one, the underlying data reveals a company in a violent transition from cash-burning startup to commercial heavyweight.

The most striking data point is not just the revenue growth, but the gross margin swing—indicating the company has successfully moved from selling prototypes at a loss to selling commercial AI hardware at high margins.

• The "J-Curve" in Revenue: Revenue for the first three quarters of 2025 (785M RMB) nearly doubled the full year of 2024. The company projects hitting 1.0–1.2 billion RMB by year-end 2025.
• The Margin Miracle: In 2022, for every $1 of product sold, they lost $0.70 just to make it. By 2024, that flipped to a 71% positive margin, a figure that rivals Nvidia's own legendary margins (~75%). This confirms they are now selling high-demand AI chips (MTT S4000) rather than low-yield consumer cards.
• Valuation Premium: At the IPO issue price, the stock traded at 123x Price-to-Sales (P/S) based on 2024 sales. For context, most mature tech companies trade at 5x-10x, and even high-growth US tech rarely sustains >50x. This is a "scarcity premium"—investors paying for the only viable domestic option.
• Cash Runway: The IPO raised ~8 billion RMB ($1.1B USD). With losses narrowing to ~700M RMB per 9 months, this war chest gives them years of runway to weather sanctions.

Moore Threads has executed a rapid four-generation roadmap in under five years.

Cluster Scale: A critical metric for AI is "how many cards can work together?" Moore Threads has successfully deployed "KUAE" clusters of 10,000 GPUs, a massive milestone that proves their interconnect technology (MTLink) is stable enough for serious model training.

Hardware is useless without software. Nvidia's dominance is built on CUDA. Moore Threads' answer is MUSA (Moore Threads Unified System Architecture).

• The "MUSIFY" Migration Tool: This tool automatically translates CUDA code into MUSA C++.
  • The Promise: "Zero-cost migration."
  • The Reality: Works well for standard PyTorch operators (used in most LLMs). Custom, hand-written CUDA kernels still require manual re-coding by engineers.
• Ecosystem Wins:
  • DeepSeek Integration: Developers have successfully run DeepSeek-R1 (Distilled 7B) on the consumer-grade MTT S80 using Ollama.
  • Frameworks: Native support for PyTorch and DeepSpeed means Chinese AI startups can train Llama 3 or Qwen models on Moore Threads hardware without rewriting their entire codebase.

• Is it profitable? No (Loss of ~724M RMB in Q1-Q3 2025), but losses are shrinking fast.
• Is it growing? Yes, revenue is up 182% YoY.
• Can it replace Nvidia? For inference (running models), yes. For training massive frontier models (like GPT-5 scale), not yet—they are still generations behind on raw bandwidth and energy efficiency.
• The Bottom Line: Investors are not buying Moore Threads for its current P/L statement; they are buying it as a geopolitical hedge. If China is to have an independent AI industry, Moore Threads must succeed.

Would you like to see a direct benchmark comparison between the MTT S80 and the Nvidia RTX 3060 to see exactly where the consumer performance gap lies?