CLASSIFICATION: CORE STRATEGY — STRICTLY | Peppa Pulse – The Philosophical Archive of Peppa: Thoughts on Time, Value, and Civilization

CLASSIFICATION: CORE STRATEGY — STRICTLY CONFIDENTIAL FOR KAI FOUNDERS & PARTNERS

Breakdown of Production Capacity for 300 Chinese AI Model Manufacturers & 6- Phase Onboarding Timeline 撰写人 / Author: Begger，KAI.com 创始人 / Founder of KAI.com 日期 / Date: 星火纪元 🌍:Sol₂₃:Φ₃:δ₅（2026-06-14） / Spark Era 🌍:Sol₂₃:Φ₃:δ₅ (June 14, 2026)

Methodology Statement

This is not an industry report. This is a battle map.

How many AI model vendors actually exist in China? According to the Ministry of Industry and Information Technology (MIIT) registration data, as of March 2026, over 260 large models have successfully registered under the Interim Measures for the Management of Generative Artificial Intelligence Services. Counting unregistered open-source fine-tuning teams, vertical industry workshops, and inference aggregators operating via APIs, the actual number of accessible model supply nodes far exceeds 300. This document filters exactly 300 providers based on the criteria of “ability to complete technical integration within 90 days and possession of commercial API inference capacity”, categorizing them into four strategic tiers for capacity estimation.

Accessible Daily Inference Capacity (Tokens/day) = GPU Stock × Commercial Inference Utilization × Daily Token Output per Card × KAI Integration Conversion Rate 参数说明 / Parameter Notes:

Based on A800/H800 benchmarks, FP16 inference yields approximately 80 billion Tokens/day/card (weighted average across mixed model sizes). Since the Chinese market primarily uses FP8/INT8 quantized inference, a 1.8x coefficient is applied.

The actual proportion of total GPU time currently allocated to external commercial inference.

The proportion of capacity that vendors are willing to allocate exclusively to KAI channels post-integration.

I. Four-Tier Panorama

Tier 1: Infrastructure Level (15 Vendors)

Possessing GPU clusters exceeding 10,000 cards, proprietary foundation models, and independent training capabilities. These 15 entities control over 60% of China’s total AI computing power. • • • KAI.com Strategic Core Document

Vendor

Est. GPU Scale

Flagship Model

Current Util.

Daily Releasable Capacity

Alibaba (Tongyi/Qwen)

(H800/A800) Qwen3-235B, Qwen-VL 25%

Tokens

ByteDance (Doubao)

(H800/H100) Doubao-Pro-256K 15%

Tokens

Baidu (ERNIE)

ERNIE 4.5, ERNIE-Speed 30% 900亿 / 90B Tokens

Tencent (Hunyuan)

Hunyuan-TurboS, Video 20% 800亿 / 80B Tokens

Huawei (Pangu)

(Ascend 910B) Pangu-5.0-NLP 20% 600亿 / 60B Tokens DeepSeek DeepSeek

(H800) DeepSeek-V3/R1 35%

Tokens 智谱AI（GLM） Zhipu AI (GLM)

GLM-5, CogView-4 30% 500亿 / 50B Tokens

Moonshot AI (Kimi)

Kimi-K2, Moonshot-v1 25% 400亿 / 40B Tokens MiniMax MiniMax

MiniMax-Text-01, Hailuo 30% 450亿 / 45B Tokens

Baichuan Intelligent

Baichuan4-Turbo 20% 300亿 / 30B Tokens

01.AI (Yi)

Yi-Lightning, Yi-Vision 20% 300亿 / 30B Tokens

StepFun (Step)

Step-3-256K 15% 200亿 / 20B Tokens

iFLYTEK (Spark)

(inc. Ascend) Spark-5.0 25% 350亿 / 35B Tokens

SenseTime (SenseNova)

SenseNova-6.0 20% 400亿 / 40B Tokens

Kunlun Tech (Skywork)

Skywork-5.0 15% 180亿 / 18B Tokens

The core contradiction for Tier 1 is not “whether capacity exists,” but “to whom it should be sold.” Domestic price wars have driven margins close to zero, and international export channels are extremely scarce. The global Agent economy pipelines provided by KAI represent a structural arbitrage opportunity for them, not just a value-add.

Tier 2: Elite Strike Level (50 Vendors)

Possessing 1,000 to 10,000 GPUs, proprietary or heavily modified open-source models, with highly differentiated capabilities in at least one vertical dimension. Selected 30 representative vendors from the 50 are listed below: KAI.com Strategic Core Document