Unique Research · Insight / Unique Partner welcome KEEP · October 8, 2023 · Full editorial rewrite of ~2,989 CJK — funding, AIFS/DataPilot, Vector Ocean preserved · inventory type insight
Beijing DataCanvas (九章云极), an AI foundation-software provider, closed a RMB 300 million Series D1. Participants: China Electronics’ Zhongdian Wisdom Fund, Huamin Investment, China Taiping’s Taiping Innovation, Zhejiang Orient’s Orient Jiafu, and AI-focused financial investor Zhuoyuan Capital, among other central/SOE-backed vehicles.
Founded 2013, focused on AI foundation software. 2020: Gartner AutoML key-vendor list; four consecutive years in IDC China ML development-platform Top 3 nationally. Unique Research interviewed Chairman Fang Lei on foundation-model practice and how AI foundation software drives enterprise innovation.
Large enterprises need the ability to build proprietary models
On Jun 30 DataCanvas launched AIFS (AI Foundation Software) and DataPilot as large-model infrastructure for AI/automation journeys.
Fang describes three modes of injecting enterprise data: (1) Prompt mode — OpenAI-style black box; (2) Fine-tune a white-box base (e.g. LLaMA) into “your” model; (3) Train from scratch with proprietary data from day one. AIFS includes DataCanvas’s white-box Yuanshi (元识) large model plus a platform for customers to train proprietary models end-to-end.
“Which path depends on data volume. Only the top ~5% need to train new models; the rest fine-tune. Head enterprises have many shifting scenarios — external models can’t cover everything and vendor quality varies. Only a proprietary model supports long-term development.”
“Cloud-within-cloud” strategy
Fang maps 2017–18 as AI 1.0 — end-to-end turnkey AI vendors (four little dragons grew fast). ~2019 enterprises found turnkey rarely finished real business; lasting capability had to be owned. With rising tech literacy, teams, and policy on controllable tech, China’s LLM market accelerated; DataCanvas’s fast phase began ~2020.
Central-enterprise cloud, compute build-out, and large models are the “three horses” of China’s AI 2.0. Compute demand may even outgrow selling models themselves. Old IDC competition was bandwidth vs Alibaba Cloud; emerging intelligent-compute centers (GPU clouds) leave room for many sizes.
Two years ago DataCanvas bet cloud-within-cloud (AI cloud inside a thousand industry clouds) — partnering BAT clouds plus Tianyi Cloud, Mobile Cloud, and regional GPU clouds; embed DataCanvas software into the cloud SKU so customers train in-house. AIFS upgrades from Foundation Software to Foundation Service.
Fang (ex Microsoft Bing data scientist; early Windows Azure) notes the software + compute + rebate model is mature in the U.S. — Snowflake ~$60B; Databricks ~$40B — win-win for training cost. “Hardware homogenizes; software differentiates; data is the deepest moat. Same hardware — shorter train time, smarter models — is the pain.”
Value is capability step-change, not only cost cut: same headcount/budget from 3 models/year to 30, or from simple to complex models.
“Vector Ocean” as data’s end state
AI and data remain inseparable — now a Data for AI architecture era; large models reverse-empower data (New Data). Warehouses/lakes no longer suffice; after years in databases plus vector directions, DataCanvas proposes Vector Ocean: all data in vector blocks with unified query (lakes still split structured tables vs unstructured files).
ETL and retrieval will be GPT-ized (natural language for once-complex jobs); storage vectorized — hence Data Pilot. DataPilot: multimodal Vector Ocean architecture; on-demand automated integration, codegen, orchestration, analytics; NL data access, analysis, and ML modeling — lowering the full data→ML chain.
Yuanshi’s edge: multimodal including structured data — vector era’s marker. Experimental TableGPT: describe needs in NL; auto-pick stats/ML on private data, explain results, suggest next mining — “what you need is what you get.”
Fang: models raised the ceiling of imagination; what’s missing is not apps but proven business landings and industry-leading models — give mid/large enterprises time to adapt and innovate. AI 2.0 will redefine software: from forward process-automation tools to reverse assembly around human thought — industry from software to thought-ware.
Author: Qiuping · Review: Siky