Take a typical case: a bank's innovation team pilots ten different models in a year, then realizes none of them share data, governance, or access controls. The bill for integration rivals the bill for the models. China's enterprises are hitting that wall, and it is reshaping where generative-AI money goes.
Key takeaways
- IDC projects China's generative-AI software market will reach about US$3.54 billion (35.4亿美元) by 2025.
- The same analysis argues enterprises are shifting from experimenting with many models (大模型) to demanding a single unified development platform for data, models, and apps.
- IDC China's AI research director Lu Yanxia (卢言霞) described the local market as still in an early stage despite fast model progress.
- The thrust is enterprise spend on platforms — RAG, fine-tuning, deployment, security — not just consumer apps.
The number
International Data Corporation (IDC) projected in late 2024 that China's generative-AI software market would reach roughly US$3.54 billion (35.4亿美元) by 2025. The figure covers software — not the servers or services wrapped around it — and reflects both accelerating model capability and a wave of applied deployments.
For context, that is software spend only. Hardware and services push the broader generative-AI total much higher, but mixing those scopes would confuse the picture, so the cleaner software figure is the one worth tracking.
From models to platforms
IDC's accompanying report, on China's generative-AI application-development platforms, makes the more interesting point: as companies scale past pilots, they want one platform that handles data preparation, model fine-tuning, RAG and prompt support, model deployment, and security. The analyst view is that "large-model platforms" are splitting into two kinds — foundational platforms that merge with enterprise data and analytics, and agent-development platforms heading toward low-code and no-code.
That is a structural shift. In 2023–2024 many Chinese firms raced to adopt the newest model; in 2025 the bottleneck is operational — wiring models into existing systems with governed data and auditable output.
Still early, says IDC
Lu Yanxia, IDC China's research director for artificial intelligence, cautioned that although foundation-model capability kept improving through 2024 and generative applications began landing, the sector is still at an early stage. Her note is a useful counterweight to headline growth rates: platform definitions are not yet settled, talent is short, and regulatory uncertainty lingers.
The caution matters for anyone reading the US$3.5B figure as destiny. A fast-growing market can still be immature, with vendors churning and buyers unsure what they actually bought.
What it means for China's AIGC field
The money story is moving up the stack. Consumer image and video apps grab attention, but IDC's lens puts the durable spend in enterprise platforms that tie models to business data. For Chinese software vendors, that is both opportunity and pressure: they must package generation as governed, integrable infrastructure rather than a flashy single feature.
The platform thesis also reshapes competition. A firm that owns the integration layer — connectors to databases, identity systems, and compliance logs — can swap the underlying model as prices fall, insulating itself from the model churn that is burning through venture capital. Buyers, for their part, are learning to evaluate vendors on governance and deployment, not leaderboard scores.
The open questions
Two things will decide whether IDC's trajectory holds. First, whether Chinese enterprises actually consolidate on a few platforms or keep a sprawl of point tools. Second, whether the security and compliance features platforms promise mature fast enough to satisfy regulated industries like finance and healthcare. Both are unresolved, which is exactly why IDC still calls the market early.
A modest number in context
The US$3.5B software figure looks large in isolation, but it is still a sliver of China's total enterprise-software spend, and far smaller than the headline numbers attached to foundation-model training. That gap is the point: IDC is measuring the layer that enterprises actually pay for today — integrations, governance, and deployments — not the capital burned on frontier research. If the platform thesis is right, this line item grows as model costs fall and the savings get reinvested into workflow tooling. If buyers stay stuck piloting, the number plateaus well short of the projection.
Honest limitations
This piece rests on IDC's published forecast and secondary Chinese coverage of the report; we did not access the full paid IDC document. The US$3.54B figure is a single-vendor projection and carries the usual forecasting uncertainty; different IDC scopes (adding hardware and services) produce larger totals, which we deliberately did not mix in. "Early stage" is the analyst's qualitative judgment, not a measured threshold. We did not independently verify IDC's segment breakdowns or the precise wording of the platform taxonomy.
What readers can do now
- Enterprise buyers should shortlist platforms on data governance, deployment, and security, not benchmark scores alone.
- Vendors should package generation as integrable infrastructure with RAG, fine-tuning, and audit trails.
- Observers should track whether 2025 spend confirms IDC's platform thesis or stays stuck in model hopping.