China’s AI Industry Isn’t Overheated – Here’s What UBS Says About the Real Numbers

• Return on Investment (ROI): Ensuring measurable returns on capital expenditure.
• Sustainability: Maintaining a long-term, stable growth trajectory.
• Investment Pace: Gradual increases aligned with actual demand rather than erratic spikes.

There’s a lot of noise right now about whether China’s AI sector is experiencing a bubble.

Spoiler alert: according to UBS Securities (Ruiyin Zhengquan 瑞银证券), it’s not.

On July 24, 2026, Wei Xiong (Xiong Wei 熊玮), a China Internet Industry Analyst at UBS Securities, sat down with reporters and laid out exactly why the narrative of an AI bubble in China doesn’t hold up under scrutiny.

The data tells a different story—one about disciplined capital allocation, structural cost advantages, and genuine technological breakthroughs.

Let’s break down what he actually said, and what it means for anyone investing in or building within China’s AI ecosystem.

—

Chinese Cloud Providers Are Playing It Smart—Not Reckless

Here’s the thing that contradicts the bubble narrative: Chinese cloud service providers are being disciplined about how they deploy capital.

They’re not just throwing money at problems and hoping something sticks.

Instead, they’re maintaining a more stable investment pace with a laser focus on two metrics:

• Return on Investment (ROI): Are we getting measurable returns on what we’re spending?
• Sustainability: Can we maintain this growth trajectory long-term?

This is the opposite of bubble behavior.

Bubbles form when investors stop caring about returns and just chase growth at all costs.

The fact that Chinese cloud providers are emphasizing both ROI and sustainability suggests the market is maturing, not inflating.

As model capabilities continue advancing and inference demand grows, investment in China’s AI industry is expected to increase gradually—not spike erratically.

—

The Training Cost Advantage That Explains Everything

Here’s where things get interesting for investors and builders.

The real competitive advantage in China’s AI industry isn’t mysterious or hidden.

It’s grounded in hard economics.

Training Costs: Chinese Models Cost 10% of US Equivalents

According to Wei Xiong’s analysis, leading Chinese models have training costs approximately 1/10th of major overseas competitors.

Let that sink in.

That’s not a small efficiency gain—that’s a structural competitive advantage.

But here’s the important part: this advantage isn’t coming from cutting corners or subsidies.

It’s coming from three specific factors:

• Algorithmic innovation: Smarter training methods that require fewer resources
• Architecture design: Better model design that’s more compute-efficient
• Model strategy focus: Different prioritization choices (some models optimize for speed vs. size vs. accuracy differently)
• Open-source synergy: Leveraging collaborative ecosystems to avoid duplicate work

This is sustainable competitive advantage, not a race to the bottom.

Inference Costs: Chinese APIs Are 10-20% of Overseas Pricing

Once models are trained, you have to run them (that’s called inference).

And here’s where Chinese providers really show their pricing power:

API pricing for Chinese models is roughly 10% to 20% of comparable overseas model pricing.

Think about what this means for developers and enterprises:

• Building AI applications becomes significantly cheaper
• Startups can experiment with more use cases
• Enterprises can deploy AI at scale without massive budget increases

But here’s what should really catch your attention: despite this aggressive pricing, Chinese firms maintain gross profit margins between 20% and 40%.

That’s not a typo.

They’re selling at 10-20% of US pricing and still pulling in 20-40% gross margins.

While this is slightly lower than US counterparts, it’s still a very healthy level that signals genuine profitability, not unsustainable unit economics.

—

Why This Price Gap Exists (And Why It Matters)

The natural question: how are Chinese companies undercutting US pricing by 80-90% on training and still staying profitable?

Wei Xiong addresses this directly: it’s not aggressive subsidies or reckless price-cutting.

It’s structural.

The cost differences come from:

• Innovations in training algorithms: More efficient ways to teach models
• Architecture design improvements: Better fundamental model design
• Different model strategy focus: Chinese companies may prioritize different performance trade-offs
• Open-source ecosystem synergy: Collaborative development reduces redundant work

None of these are unsustainable or temporary.

They’re durable competitive advantages.

—

The Kimi K3 Case Study: When Chinese Models Command Premium Pricing

Here’s the data point that really challenges the “China can only compete on price” narrative:

Kimi K3 is one of the world’s largest open-source models by parameter count.

Its performance is competitive with global top-tier models.

And its pricing?

It aligns with high-end domestic model standards—not discounted pricing.

This proves something critical: high-quality Chinese models possess significant bargaining power.

Companies aren’t forced to compete on price when they have quality to back it up.

This is even forcing global leading players to adjust their strategies, which is driving further technological innovation across the open-source ecosystem.

Translation: the competitive pressure from high-quality Chinese models is actually accelerating innovation globally.

—

What About Multi-Modal AI? Chinese Companies Are Keeping Up

Multi-modal AI (models that work with text, images, video, audio, etc.) is one of the hottest areas in AI right now.

The good news: Chinese model manufacturers demonstrate stable, competitive output in the multi-modal space.

They’re not lagging behind.

They’re actively competing.

—

The Real Constraint: GPU Bottlenecks Are Still a Thing

But there’s a critical caveat Wei Xiong included in his analysis that shouldn’t be overlooked:

Domestic model training and inference still face significant constraints regarding computing power (GPU) bottlenecks.

This is important.

Despite all the efficiency gains, access to GPUs—especially high-end chips for AI training—remains a real limiting factor for Chinese AI companies.

This affects:

• How fast new models can be trained
• The scale of experiments companies can run
• Competition dynamics with overseas players who have fewer chip restrictions

It’s worth noting if you’re evaluating the long-term competitive landscape.

—

So… Is There Really a Bubble or Not?

Wei Xiong’s conclusion is clear: no, there is not a “bubble” in the Chinese AI sector.

Here’s what the data actually shows:

• Disciplined capital allocation: Focus on ROI and sustainability, not reckless growth
• Structural cost advantages: Based on innovation, not subsidies or shortcuts
• Healthy margins: 20-40% gross margins despite aggressive pricing
• Quality proof points: Models like Kimi K3 command premium pricing based on performance
• Competitive ecosystem: Driving innovation rather than destabilizing the market

What you’re seeing in China isn’t a bubble inflating.

It’s a mature, competitive market where companies are building durable competitive advantages through real innovation.

And as model capabilities advance and inference demand grows, investment is expected to continue—just on a stable, sustainable trajectory.

That’s the opposite of bubble dynamics.

—

References

• UBS: No “Bubble” in China’s AI Industry – People’s Finance (Renmin Caixun 人民财讯)
• UBS Securities Media Briefing on AI Industry Costs – Securities Times (Zhengquan Shibao 证券时报)
• UBS China Investment Outlook and Internet Sector Analysis – UBS China