The Internet Trained the First Generation of AI. What Trains the Next One?

There is a useful mental model for understanding where AI systems get their intelligence: they are, in a meaningful sense, compressed versions of the data environments they were trained on.

The first generation of large AI models was trained on the internet — billions of web pages, books, code repositories, forum posts, scientific papers. That data environment had specific characteristics: it was primarily text and images, it was primarily created by knowledge workers, and it reflected a world of information rather than a world of physical operations. The AI systems that emerged from that training are genuinely remarkable at language, reasoning, and pattern recognition within the domain of digitized human knowledge.

But the next generation of AI problems is different. Humanoid robots need to learn how to assemble components. CAD AI systems need to understand engineering design logic. Industrial AI needs to predict production failures. Virtual try-on systems need to model how garments actually behave on real bodies. Autonomous systems need to navigate real-world environments with physical constraints that no amount of internet text can fully describe.

All of these problems require a fundamentally different kind of training data: real-world operational workflow data. And that data does not live on the internet. It lives in factories, engineering organizations, supplier networks, and production systems — most of which have never been exposed to any data collection infrastructure at all.

This is where China becomes interesting. Not in the geopolitical framing that tends to dominate discussions of this topic — but in a purely operational sense. China, particularly the manufacturing ecosystems of Shenzhen and the Pearl River Delta, may be quietly becoming one of the most important environments on earth for generating and operationalizing real-world AI data.

Why Web Data Is Reaching Saturation — And Why It Was Never Enough for Physical AI

The web data problem for AI is real. The most accessible and cleanly formatted text on the internet has largely been incorporated into training datasets already. The marginal value of adding more scraped web content is declining. Synthetic data generation is being explored as a partial solution, but synthetic data trained on synthetic data tends to amplify existing model artifacts rather than introduce new grounding signal.

But even before the saturation issue, web data was never well-suited for physical-world AI. The internet contains descriptions of manufacturing processes, engineering tutorials, product documentation, and factory case studies. What it does not contain is the operational data itself — the actual sensor readings, the real motion sequences, the genuine workflow patterns, the feedback loops between design decisions and production outcomes.

Consider humanoid robotics as an example. Teaching a robot to perform assembly tasks requires training on real human assembly motion — not descriptions of it, not even video of it, but structured operational data about how skilled human operators perform complex physical tasks in real production environments. That data exists in Shenzhen factories where workers assemble consumer electronics at high volume and with extraordinary efficiency. It does not exist anywhere on the internet in a form that is trainable.

Or consider engineering AI for CAD systems. Teaching an AI to generate manufacturable mechanical designs requires training on real engineering workflows — feature tree construction logic, DFM revision histories, the actual back-and-forth between design teams and factory feedback over multiple EVT and DVT cycles. None of that workflow intelligence is published online. It lives in company PDM systems and in the heads of experienced engineers.

The gap between what AI needs and what the internet can provide is widening as AI moves deeper into physical-world applications. Filling that gap requires access to real operational environments. And China has more of those environments, running at higher frequency, than almost anywhere else on earth.

Robotics and Factory AI: Where the Training Data Problem Is Most Acute

I’ve been watching the humanoid robotics space with particular interest from Shenzhen, because the hardware commercialization side of this problem is something I deal with directly. And one thing that strikes me consistently is how underestimated the data infrastructure problem is relative to the hardware engineering problem.

The hardware challenge for humanoid robots is genuinely hard. Actuators, joint torque, thermal management, battery life, structural durability — these are solvable engineering problems, and Shenzhen’s manufacturing ecosystem is actually well-positioned to contribute to solving them. BOM optimization, prototype-to-production iteration, supply chain for specialized components — this is exactly what the Pearl River Delta does well.

But the training data problem may be harder than the hardware problem. Humanoid robots learning to perform useful work in real environments need massive amounts of demonstration data — skilled human motion in real physical contexts. The research community has developed various approaches to this: teleoperation data collection, motion capture, synthetic simulation, transfer learning from existing robotics datasets. All of these approaches have known limitations, and the fundamental constraint is access to high-quality, high-volume demonstration environments.

Shenzhen’s assembly factories are among the most information-dense manual operation environments in the world. Workers performing high-precision consumer electronics assembly — placing components, operating fixtures, inspecting sub-assemblies — are generating operational motion data every working hour. The challenge is that this data is currently invisible to any AI training pipeline. It exists only as human embodied knowledge, distributed across hundreds of thousands of workers across thousands of production lines.

Building the infrastructure to capture, label, and pipe that operational data into robotics training pipelines is a serious multi-year undertaking. The companies working on it are not doing data scraping. They are building deep factory relationships, deploying motion capture infrastructure, negotiating long-term data sharing agreements. It is less like a technology problem and more like a combination of enterprise software deployment and relationship management at industrial scale.

Engineering AI and the CAD Data Problem: What I’ve Seen From the Inside

The engineering workflow data problem is something I have a particular perspective on, having spent fifteen years coordinating between design teams and manufacturing organizations in Shenzhen.

Real engineering data — the kind that would actually make CAD AI systems useful in production environments — looks nothing like what’s publicly available. It’s not STEP files exported from GrabCAD. It’s SolidWorks SLDASM files with full feature trees intact, assembly constraints, tolerance stacks, configuration families, and revision histories that trace design changes back through EVT failures and supplier feedback.

In Shenzhen, most experienced mechanical engineers in consumer electronics use Creo/ProE, not SolidWorks. This reflects the training history of the engineering talent pool in the Pearl River Delta — ProE was dominant during the rapid manufacturing expansion of the early 2000s, and that expertise has persisted. SolidWorks tends to be more common in automation equipment, tooling design, and engineering teams with North American or European influence.

What this means for engineering AI data sourcing is significant fragmentation: the engineering knowledge is distributed across incompatible software ecosystems, informal communication channels — WeChat groups, USB drives, shared folders with no version control — and organizational structures where design ownership is frequently unclear. I’ve been on projects where the “master” design existed in four different versions across three companies’ servers simultaneously, with nobody quite certain which was authoritative.

The engineering intuition problem is equally important and even harder to capture. An experienced Shenzhen mold engineer knows immediately when an ejector pin placement is going to leave witness marks on a cosmetic surface. They know this because they’ve seen it happen eighty times, not because they ran a simulation. That operational pattern recognition — built from years of watching what actually happens between the CAD file and the finished part — exists nowhere in any database. It is embodied knowledge, and capturing it for AI training requires structured expert annotation programs embedded in real engineering workflows.

The DFM feedback loop is perhaps the most valuable engineering data asset for AI systems, and it is almost entirely uncaptured. Every time a Shenzhen factory engineer marks up a design with manufacturability comments, suggests a geometry change to avoid a tooling risk, or flags a tolerance that will cause assembly variance — that is high-quality training signal for engineering AI. Currently it mostly lives in WeChat messages and informal conversations. Systematizing that feedback into structured training data is a significant infrastructure-building challenge.

Virtual Try-On, Apparel AI, and the Physical Fitting Data Problem

The apparel and fashion AI space is a useful case study in why physical-world AI cannot be trained on internet data alone.

Virtual try-on systems — AI that can realistically simulate how a garment will fit and move on a specific body — require training data that combines body measurement distributions, garment construction parameters, material physics, and the real relationship between 2D pattern design and 3D fit outcomes. This data is generated in textile and apparel manufacturing workflows, body scanning operations, and fitting room environments. Very little of it is publicly documented in a form suitable for AI training.

China’s apparel and textile manufacturing sector — centered in Guangzhou, Dongguan, and surrounding areas — runs at a scale that makes it potentially the most data-rich environment on earth for this kind of AI training. The volume of garment sampling, fitting iterations, pattern adjustments, and production runs happening in these ecosystems every day generates an enormous amount of physical-world fitting intelligence. The challenge, again, is that it exists as tacit operational knowledge rather than structured data.

Companies building serious virtual try-on systems are learning that the path to production-quality fitting simulation requires direct operational integration with apparel manufacturing ecosystems — not internet scraping, but relationships with factories and pattern design organizations willing to share workflow data in exchange for the AI tools that result. This is the same pattern that appears across every domain of physical-world AI: the data access problem is a relationship problem, not a technical problem.

Industrial AI: CNC, SMT, and the Manufacturing Execution Data Layer

The industrial AI domain covers a wide range of applications — CNC workflow optimization, SMT process control, production quality inspection, predictive maintenance, supplier coordination intelligence — all of which share the same fundamental data characteristic: the training signal is generated inside production environments that are structurally closed to external data collection.

Consumer electronics manufacturing in Shenzhen generates particularly rich operational data because of the product complexity and iteration speed involved. An SMT line running a high-layer-count PCB with mixed component types at high volume is generating continuous process data — paste deposition, component placement accuracy, reflow profile measurements, AOI inspection results — that is genuinely useful for AI systems trying to understand electronics manufacturing quality optimization. Most of that data currently lives in factory MES systems that are isolated, proprietary, and not connected to any AI infrastructure.

The EVT/DVT/PVT cycle in consumer electronics product development is another underappreciated source of physical-world AI training data. Each validation phase generates structured feedback: what failed, why it failed, what design or process change fixed it. Over hundreds of product development cycles, that accumulated failure-and-fix record represents extraordinary operational intelligence about how hardware products actually go from concept to manufacturable. Companies that have been running EVT/DVT/PVT support operations in Shenzhen for a decade are sitting on institutional knowledge that would be extremely valuable for AI systems trying to understand hardware commercialization risk.

The challenge, as always, is that this knowledge is not in any database. It’s in engineers’ heads, in scattered test reports, in WeChat message histories, in the informal institutional memory of organizations that have never thought of their operational experience as a data asset.

Why Shenzhen Is a Unique Operational Data Generation Zone

From my position inside Shenzhen’s hardware ecosystem, a few characteristics stand out as particularly important for understanding why this geography is uniquely significant for real-world AI data.

The first is iteration speed. Shenzhen’s manufacturing ecosystem can take a hardware concept from sketch to physical prototype in days, and from prototype to first production units in weeks. This iteration speed means that the feedback loops between design and production — the exact loops that generate the most valuable engineering AI training signal — happen at much higher frequency here than in manufacturing ecosystems elsewhere. More iteration cycles means more learning events per unit time means more operational data generated.

The second is supply chain density. The concentration of component suppliers, module vendors, mechanical subcontractors, PCB houses, mold shops, and assembly facilities within a small geographic area means that engineering data flows — however informally — across organizational boundaries at high frequency. A product development cycle in Shenzhen might involve eight to twelve specialized suppliers, each contributing engineering modifications, DFM feedback, and process knowledge. That cross-organizational knowledge flow, if captured, is rich AI training signal.

The third is product diversity. The Pearl River Delta manufactures an extraordinary range of products — smartphones, wearables, audio devices, medical hardware, industrial equipment, toys, home appliances, power tools — across a concentrated geographic area. This means the operational data environment covers a wide spectrum of manufacturing problems, not a narrow vertical. For AI systems training on physical-world data, breadth of operational context matters as much as depth.

None of this is accessible through remote data scraping. Accessing it requires being embedded in the ecosystem — having supplier relationships, factory access, engineering credibility, and the trust of organizations that correctly understand their operational knowledge as proprietary.

The Trust Problem Is Structural, Not Incidental

One thing I keep noticing in conversations about physical-world AI data is that the trust problem gets treated as a minor obstacle — something that can be solved with a good contracts team and a compelling pitch deck. In practice, it is a deep structural challenge that shapes everything about how real-world operational data can be accessed.

Engineering files, production process data, supplier workflow records, and manufacturing execution logs are among the most sensitive intellectual property that any hardware organization holds. An engineering data breach doesn’t just expose product geometry — it exposes cost structure, supplier relationships, manufacturing process know-how, and competitive positioning. For a Shenzhen ODM running production for multiple competing brands, even the metadata around what they’re making and how they’re making it is sensitive.

This means that AI companies trying to build real-world manufacturing data pipelines are not primarily solving a technical problem. They are primarily solving a trust and relationship problem — building the credibility, legal infrastructure, and operational integration that makes it rational for manufacturing organizations to share workflow data as part of an ongoing partnership rather than treating it as a one-time asset extraction event.

The companies that will succeed in building physical-world AI training pipelines are the ones that can offer genuine operational value to manufacturing organizations in exchange for data access. Not just payment. Genuine operational value — tools that make the factory more efficient, quality systems that improve production outcomes, engineering AI that helps the organization’s own engineers do better work. Data sharing embedded in a value-exchange relationship is fundamentally more sustainable than data acquisition as a one-time transaction.

Practical Takeaways for Teams Building Physical-World AI

• Physical-world AI training is an operational integration problem, not a data scraping problem. The data you need cannot be fetched from the internet. It requires sustained presence inside operational environments.
• Manufacturing relationships are data infrastructure. Treat factory partnerships, supplier relationships, and engineering organization access as capital assets, not procurement transactions.
• Shenzhen’s engineering ecosystem is high-value and high-friction. The density of manufacturing intelligence here is extraordinary, but accessing it requires engineering credibility, local presence, and years of relationship-building — not remote deal-making.
• EVT/DVT/PVT cycles are underappreciated training data sources. The feedback loops from hardware validation phases contain concentrated engineering learning that is essentially unavailable anywhere else. Companies with deep access to these cycles have data advantages that cannot be replicated by model architecture improvements alone.
• Tacit manufacturing knowledge is the hardest and most valuable signal. Expert annotation programs embedded in real engineering workflows — capturing the pattern recognition of experienced engineers who know immediately when a design will fail in production — are a serious competitive moat for physical-world AI systems.
• Trust takes longer to build than models. Plan the relationship development timeline realistically. Data access that requires organizational trust cannot be accelerated by throwing engineering resources at it.

Final Thoughts

I want to be careful not to overstate this in ways that shade into the kind of geopolitical framing I find mostly unhelpful. The point is not that China “wins” some AI competition. The point is more structural and more interesting than that.

Physical-world AI is a fundamentally different infrastructure challenge from web-scale AI. The training data it needs is operational, embodied, tacit, and relationship-dependent. It cannot be scraped. It cannot be synthesized from first principles. It can only be accessed through sustained operational integration with the real environments where it is generated.

China’s manufacturing ecosystems — particularly the Pearl River Delta and Shenzhen — happen to be some of the densest, fastest-iterating, most product-diverse operational data environments on earth. That is a fact about manufacturing geography and history, not a political statement. The operational intelligence embedded in these ecosystems is a genuinely significant resource for the next generation of AI development.

How much of that intelligence becomes accessible to AI training pipelines will depend on whether AI companies learn to think like operational partners rather than data acquirers. The companies that embed themselves deeply in manufacturing workflows, deliver genuine value to engineering organizations, and build the trust infrastructure required for long-term data sharing relationships — those companies are building something that model architecture alone cannot replicate.

The internet trained the first generation of AI. Factories, engineering workflows, and physical-world operational systems may train the next generation. And the companies that control real-world operational workflows may become more important to the next generation of AI than the companies that simply control models.

That is not a prediction about who will dominate. It is an observation about where the training signal actually lives — and who is positioned to access it.

I Work FOR YOU, Not Factories. My role is to stand on your side of the table — not the factory’s — and help you navigate the Shenzhen manufacturing ecosystem from the inside.

Need China-Side Hardware Execution Support?

If you’re building a hardware product and need someone who understands the China manufacturing side — supplier evaluation, DFM coordination, EVT/DVT/PVT management, factory communication — this is the kind of work I help with from Shenzhen.

Not a factory broker. Not a sourcing agent. A China-side execution operator aligned with your product goals.

Feel free to reach out directly:

Email: [email protected]

Website: www.easelinktech.com