Where AI Becomes Tangible: A Journey Through Humanoids, Chips, and Sovereign AI in Shenzhen and Taipei

Those who want to know where artificial intelligence is heading should look beyond software to screws, server racks, and factory floors. This is what we took away from our journey through Shenzhen and Taipei in November 2026. We traveled with a group of entrepreneurs from the Netherlands, Germany, Sweden, Austria, and the United States, as part of the AI Experience trip. We visited the Huaqiangbei electronics market, robotics manufacturer UBTech, NVIDIA, chip designer MediaTek, and sensor company Ainos.

What we saw can be summarized in one major shift. As AI rapidly turns software into a commodity, the strategic advantage is shifting to what you can physically grasp: integrated hardware, proprietary data from the physical world, and control over the technology that runs a company. China is building the factories and robots for this. Taiwan is building the chips and models, explicitly keeping control in its own hands.

Key Takeaways

• Hardware is the new differentiator. As software becomes a commodity, the advantage shifts to tangible, integrated hardware. Shenzhen builds this just-in-time.
• Humanoid robots operate independently. At UBTech, robots execute tasks without a remote controller. Over 100 factories are running batches of dozens to about 100 units, compared to 20 to 30 companies with a single test unit just a year earlier.
• Factory data is scarce. UBTech has captured 800,000 hours of factory data for world models; broad generalization requires an estimated one billion hours.
• AI agents rely heavily on CPUs. This is why NVIDIA is building its own CPU rack with 256 CPUs and an AI factory consisting of five types of racks.
• Taiwan opts for sovereign AI. Banks and hospitals collaboratively train models via federated learning, while the data remains local.
• The core question for any AI investment: How much control do you retain over your data, your model, and your business continuity?

Why is Huaqiangbei in Shenzhen so important for hardware and AI?

Huaqiangbei demonstrates that China can go from a rough sketch to a working product without waiting on suppliers. Our trip started in this district of Shenzhen. It is an electronics mecca of market halls where we saw literally every screw, cable, and advanced chip available. Anyone walking in here with an idea has everything within reach to build it.

Walking among the stalls, we witnessed China's true strength: lightning-fast adaptability. Export restrictions on Western technology, such as NVIDIA chips, have fueled this strength rather than slowed it down. The country is moving increasingly convincingly toward technological self-sufficiency, and Huaqiangbei is the visible proof of that.

The entire system operates just-in-time, and we observed two decisive advantages in this. The first is maximum autonomy: no dependence on complex, vulnerable supply chains on the other side of the world. The second is an unparalleled time-to-market: prototyping and building happen at top speed, without costly time lost to delayed deliveries.

A conclusion emerged here that kept resurfacing throughout the rest of the trip. Precisely because AI is rapidly turning software into a commodity, the strategic advantage is shifting. Tangible, integrated hardware is once again the ultimate differentiator, and whoever can build it the fastest has a head start.

How advanced are humanoid robots in China? Our second visit to UBTech

Humanoid robots in China have gone from demonstration to practice in a single year. We saw this at UBTech, the robotics market leader in China. Humanoids are booming, in the United States just as much as in China, and at every demonstration we wonder if it is more than a moving machine with a flashy exterior. At UBTech this year, we received a more concrete answer than ever.

It was our second visit in two years, and precisely this repetition made the progression visible to us. The robots now genuinely execute tasks independently, without anyone operating them via a remote controller.

The scale has also tipped. Whereas last year 20 to 30 companies were experimenting with a single test unit, over 100 factories are now testing or running batches of dozens up to about 100 units simultaneously. If we compare this to the American competitor Figure AI, our conclusion is that the development is accelerating massively.

The second shift we saw is under the hood. The world models, the "language models for the physical world," are becoming more reliable and powerful. UBTech has now recorded 800,000 hours of factory data.

That is still far removed from the estimated one billion hours required for broad generalization. But it is directly applicable data, and this is exactly the data that software companies without factory floors are missing. Whoever has the factory, has the dataset.

What is "the Chinese mindset" behind the rise of robots?

Chinese tech companies are led by builders and steer for the long term. To understand how humanoids and robots will continue to develop, we consider it important to understand that Chinese mindset. Two characteristics continually stood out to us during the visits.

Chinese companies are run by engineers: true builders, and far less by marketers and investors. Building products and engineering are the driving forces, not the presentation.

In addition, the Chinese think much further ahead. In Europe, we tend to evaluate companies based on how well they perform in the short term. In China, we saw a focus on the long term, and that determines which investments make sense.

The building blocks for humanoids are in the making: the world models, the factory data, and the Shenzhen supply chain. We expect there will come a moment when all this converges into a valuable mass product. Exactly when, time will tell.

Why is NVIDIA building its own CPU? What we saw in Taipei

NVIDIA is building its own CPU because AI agents primarily do CPU work, leaving the expensive GPU idle otherwise. In Taiwan, the tone was different from Shenzhen: less scale, more precision. In Taipei, at NVIDIA—the company that grew big on GPUs—we stood in front of a rack full of CPUs: 256 of them in a single two-meter-high server cabinet. The obvious question was: why would a GPU company do that?

The answer lies in how AI agents operate. Agents do relatively little GPU work, and the number of GPUs an average company needs is often overestimated. An agent thinks briefly (GPU) and then works for a long time (CPU): looking up an order, reading a document, waiting on another system. Then it thinks briefly again. This happens hundreds of times, and the expensive GPU sits idle in the meantime.

NVIDIA's answer is therefore not one type of rack, but five. One for thinking (GPU), one for doing (CPU), one for speed, one for memory, and one for networking. Together, they form a single AI factory.

The agent switches hundreds of times per task between short thinking on the GPU and long working on the CPU; the five racks at the base are NVIDIA's answer to that pattern.

For companies deploying agents, we see a sobering lesson here. Their computational needs do not resemble those of model training: they primarily require CPU capacity, memory, and fast connections, and fewer GPUs than is often assumed.

What is federated learning and how does Taiwan use it?

With federated learning, Taiwanese companies collaboratively train AI models while their data remains within their own organizations. The need for control ran like a common thread through all our visits in Taiwan. The island is building its own local, and thus sovereign, language models, combining them with federated learning. We find this unique, and in our view, an example for Europe.

We saw how this works at the banks. Taiwanese banks are jointly training AI to detect fraud. Each bank uses its own transaction data and only shares the model updates. In this way, banks benefit from each other's insights while the original customer data remains local.

Hospitals are doing the same to train AI that helps detect tumors on brain scans. Learning together without centrally collecting patient data.

Behind this lies a pragmatic observation: for many applications, the best model is not necessary at all. A smaller, specialized model can be cheaper and faster, and you can deploy it under your own management. By training collectively while retaining IP and corporate data, powerful alternatives to closed frontier LLMs emerge.

How do MediaTek and Ainos maintain control over the AI hardware stack?

MediaTek and Ainos each occupy a distinct position in the AI hardware stack and maintain control there. At MediaTek, we saw that the chip designer develops custom chips where clients can incorporate their own designs. This allows them to retain more influence over the technology on which their AI runs.

At Ainos, AI was literally given a nose. Sensors in chip factories measure airborne chemicals to detect anomalies. This creates a proprietary data source from the physical world, essentially an LLM for smells.

The pattern we consistently observed: Taiwan is a master at finding unique sweet spots in the AI hardware stack. Rather than competing with the largest models, they occupy an indispensable part of the supply chain and maintain control over it.

What does this mean for European entrepreneurs?

Our most important lesson: every AI investment should come with the question of how much control you retain over your data, over modifications to the model, and over your business continuity. This question recurred at every visit, and it is less technical than it sounds.

Shenzhen answers that question with its own just-in-time supply chain and factories that produce their own data. Taiwan answers it with sovereign models, custom chips, and physical world data sources. In both cases, we saw that control is not an afterthought, but the foundation of the strategy.

For us as European entrepreneurs, this trip holds up a mirror. Do not evaluate AI vendors solely on price and performance, but on who controls the data, the model, and business continuity. Consider smaller, specialized models under your own management, and look to industry peers to train collaboratively, just as the Taiwanese banks do. And recalculate how many GPUs you actually need, because agents run predominantly on CPUs.

What lies beneath is harder to copy: companies led by builders who steer for the long term. Humanoid robots are the most visible example of this. The building blocks are coming together, and as far as we are concerned, the question is not if, but when. Time will tell.

Frequently Asked Questions about the AI Experience trip

What is the AI Experience trip?

The AI Experience trip is an educational study trip for entrepreneurs visiting AI and hardware companies in Asia. The 2026 edition went to Shenzhen (China) and Taipei (Taiwan) with a group of entrepreneurs from the Netherlands, Germany, Sweden, Austria, and the United States.

Which companies and locations did we visit in China and Taiwan?

In Shenzhen, the Huaqiangbei electronics market and robotics manufacturer UBTech. In Taiwan, NVIDIA in Taipei, chip designer MediaTek, and sensor company Ainos, complemented by sessions on sovereign language models and federated learning at Taiwanese banks and hospitals.

Why is Huaqiangbei in Shenzhen important for AI and hardware?

Huaqiangbei is the largest electronics market in the world. Every component, from a single screw to an advanced chip, is immediately available. This makes just-in-time prototyping and production possible, with maximum autonomy and an unprecedentedly short time-to-market.

How advanced are humanoid robots in China?

At UBTech, humanoid robots independently perform tasks without a remote controller. Over 100 factories are testing or running batches of dozens up to about 100 units, compared to 20 to 30 companies with a single test unit a year earlier. UBTech has recorded 800,000 hours of factory data to train world models.

Why is NVIDIA building its own CPU?

Because AI agents think briefly (GPU) and work for a long time (CPU): looking up an order, reading a document, waiting on another system. The expensive GPU sits idle in the meantime. NVIDIA is therefore building five types of racks (GPU, CPU, speed, memory, network) that together form a single AI factory. In Taipei, there was a rack with 256 CPUs.

What is federated learning and how does Taiwan use it?

Federated learning is a method of collaboratively training an AI model without sharing data. Each participant trains on their own data and only shares the model updates. Taiwanese banks use this for fraud detection, and hospitals use it to identify tumors on brain scans.

What is the most important lesson from the trip for European companies?

Every AI investment should include the question of how much control you retain: over your data, over modifications to the model, and over your business continuity. Smaller, specialized models under proprietary management and collaboration via federated learning are powerful alternatives to closed frontier LLMs.

Glossary

Term	Meaning
Humanoid robot	Robot with a human-like form (torso, arms, legs) that can perform tasks in a human environment, such as in a factory
World model	"Language model for the physical world": an AI model that learns how objects, forces, and actions behave, trained on practical data like factory hours
Federated learning	Collaboratively training an AI model where each participant trains on their own data and only shares model updates; the source data remains local
Sovereign AI	AI models and infrastructure managed independently by a country or organization, without reliance on foreign or closed providers
Frontier LLM	The largest and most advanced type of language model, typically closed and offered by a limited number of parties
AI agent	AI application that autonomously executes multi-step tasks: short reasoning on a GPU, long execution on a CPU
AI factory	NVIDIA's term for a data center built from specialized racks for GPU, CPU, speed, memory, and networking
GPU and CPU	GPU: processor for parallel computing tasks like AI models; CPU: processor for general tasks like querying systems and processing documents
Just-in-time	Production philosophy where components are available exactly at the time of use, without large inventories or long delivery times
Huaqiangbei	Electronics district in Shenzhen, the largest electronics market in the world
Time-to-market	The time between the idea and the market introduction of a product