There is a fundamental difference between building an AI system that generates text and building one that has to understand the physical world. A language model can learn from an enormous corpus of information that already exists online. Books, websites, code, conversations and documents have effectively created a vast training environment for machines that work with language.
Physical AI has no equivalent. There is no internet-scale archive containing every possible view from a car camera, every rare maritime object seen through an infrared sensor, every agricultural condition, every unusual industrial failure or every way a distant military vehicle might appear through a degraded camera feed.
The data has to be collected deliberately. And the cases that matter most are often precisely the ones that are hardest, most expensive or most dangerous to collect.
That is creating a new infrastructure problem around AI: how do you generate the exact data a machine needs when the real world cannot provide it at sufficient scale? This is the problem that companies such as DiffuseDrive are attempting to solve.
The San Francisco-based startup describes its platform as an end-to-end automated data solution for computer-vision developers, using generative AI to identify gaps in existing datasets and create new data for training, validation and edge-case testing. Its stated goal is to reduce the time required to bring autonomous-machine systems to market by 50%.
But behind that product is a much broader shift: synthetic data is moving from being a way to create more training examples to becoming infrastructure for deciding which examples an AI system needs next. And that requires a very different kind of engineering.
Physical AI has a data problem that language models never had
The first generation of AI infrastructure was largely built around digital information. The next generation has to deal with reality. Autonomous vehicles need to understand roads, pedestrians, weather, lighting and objects at different distances.
Robots need to understand physical environments. Drones need to operate across changing terrain and weather. Industrial systems have to recognize objects and anomalies from imperfect cameras and sensors. Defence systems may need to identify a target that occupies only a few dozen pixels.
In each case, the difficult data is usually not the average case. It is the exception. A perception system might already have thousands of images of a vehicle in normal daylight. What it may lack is that same vehicle partially hidden by foliage, viewed from an unusual angle, at long range, in poor weather or through a particular infrared sensor.
The problem therefore becomes less about collecting millions of additional ordinary images and more about finding the holes in the dataset and deliberately filling them. That distinction is becoming increasingly important as AI moves into physical environments.
Research presented at ICCV 2025, for example, highlighted the importance of diverse, high-quality data for end-to-end autonomous-driving systems and the cost and time involved in collecting large-scale real-world datasets. Gartner has similarly identified edge-case coverage and the operational cost of collecting and labelling real-world driving data as key reasons synthetic data is becoming important for autonomous vehicles.
“The goal isn't to generate more data. It's to generate the data you are missing.” (draft quote)
That sounds like a subtle distinction. Technically, it is a major one.
Synthetic data is becoming part of the training loop
Synthetic data has existed for years. Simulation environments, computer-generated imagery and procedurally generated datasets have all been used to train computer-vision systems.
Generative AI changes the economics and flexibility of the process. Instead of manually constructing every possible scene, a generative model can be used to create variations based on existing data and specific requirements.
But the most useful workflow is not simply: collect data → generate more data → train a model. It is closer to: collect real data → evaluate the model → identify where it fails → generate targeted synthetic data → retrain → evaluate again. The synthetic-data layer becomes part of an iterative development loop.
That is particularly valuable in Physical AI because the most important scenarios may be exactly those that cannot be safely or economically staged in the real world.
A defence company cannot simply manufacture every possible battlefield scenario to create a training dataset. An autonomous vehicle developer cannot realistically wait for every dangerous edge case to occur naturally. A maritime company cannot collect every possible sensor view of every rare object under every combination of weather and distance.
Synthetic data can make those gaps addressable. But there is an important caveat: synthetic data is not automatically useful because it is synthetic. The output has to improve the model.
A beautiful image can still be bad training data
This is where the work of Daniel Bin Schmid becomes particularly relevant. As AI Research Lead at DiffuseDrive, Schmid focuses on generative AI and machine-learning research, with an emphasis on translating new generative-model research into practical capabilities for computer vision.
One of his key areas of work is controllable image generation. Traditional generative AI is often driven primarily by language: describe what you want and the model attempts to create it. That is useful for images.
It is not necessarily enough for machine perception. A computer-vision engineer may not want a tank in a field. They may want a particular object, at a particular location, occupying a particular percentage of the image, under a particular visual condition.
Schmid has worked on bounding-box-controlled generation, allowing spatial coordinates to become part of the generation process. That potentially changes the relationship between image generation and annotation. If the system is told in advance that an object should occupy a particular location and scale, the information defining the object’s position can become part of the generated dataset itself.
“For Physical AI, realism is only one dimension of quality. A generated image has to contain the right object, in the right place, at the right scale and under the right conditions — and ultimately it has to make the perception model better.” (draft quote)
This is an important shift in how generative AI is evaluated. Consumer-facing image generation can be judged by looking at the output. Industrial synthetic data cannot. A human may look at an image and say that it appears completely realistic. A perception model may nevertheless learn the wrong thing from it.
A rare object may disappear from a complex prompt. Geometry may be subtly wrong. Sensor characteristics may not match the deployment environment. An infrared image may look convincing without representing the underlying physical properties of thermal sensing.
For a training pipeline, those differences matter. DiffuseDrive itself emphasizes the distinction between realistic-looking outputs and data designed for deployment conditions, including sensor characteristics and difficult edge scenarios.
“A synthetic image can look completely convincing to a human and still be harmful to a perception model. The question isn't whether we like the image. The question is whether the data teaches the machine something useful.” (draft quote)
The missing piece may be a 30-pixel target
Consider a simple example. A perception model might perform well when an object is large and clearly visible. But when that same object is far away and occupies only around 30 pixels, its performance may collapse.
From a human perspective, the two images represent the same object. From the perspective of the model, they may be completely different problems. That means the useful synthetic dataset is not necessarily the one containing the most beautiful representation of the object.
It is the one deliberately designed around the failure. The same logic applies to camouflage, occlusion, foliage, unusual viewpoints, degraded sensors, bad weather, long-range targets and other edge conditions. This is particularly relevant in defence, where the operational environment is rarely clean.
“In mission-critical computer vision, the metric isn't how beautiful the image is. It's whether the system can recognize the target when conditions are bad enough that it normally fails.” (draft quote)
From research paper to production architecture
This is where Domonkos Haffner‘s role becomes distinct. Haffner joined DiffuseDrive as a founding generative-AI engineer, when the company’s technology was still closer to a vision and a set of prototypes than a finished product.
With a background in physics, machine learning and computer vision, he became responsible for the architecture of the company’s core AI stack: generative models, training and inference pipelines, evaluation systems and the interfaces between data and models.
The distinction is important. His job is not simply to make a generative model work. It is to decide which model architecture makes sense, how it should be trained, how it should be evaluated, how it should be integrated and how the system can survive the next generation of models.
Generative AI is moving quickly enough that a production platform cannot assume today’s model will still be the right foundation in two years. That creates a software-architecture problem. The generative backbone needs to be replaceable without rebuilding the rest of the system around it.
Haffner’s team therefore has to deal with the less glamorous side of AI: memory consumption, training efficiency, inference, caching, evaluation and reliability. In one example, a full-resolution training batch required roughly 72 GB of GPU memory. Changes including optimizer improvements, gradient checkpointing, latent caching and offloading brought that requirement down to approximately 56 GB.
This is the kind of work that rarely appears in an AI demo. It is also the work that determines whether the demo can become a product.
“The interesting problem isn't building a model that can generate an impressive image once. It's building a system that can generate the right data repeatedly, evaluate it, and remain useful as the underlying generative models change.” (draft quote)
The AI researcher’s real job is knowing which papers matter
Generative AI research is moving at extraordinary speed. New papers appear constantly, but not every promising research result survives contact with production.
Academic research is generally optimized for advancing a particular technique under defined experimental conditions. Production systems have different constraints: reliability, compute costs, latency, reproducibility, security, maintainability and integration.
For a startup researcher, the question therefore isn’t simply: “Does this paper work?” It is: “Is there something here that could create a meaningful product advantage, and can we make it robust enough to survive the real world?” That requires technical judgement as much as implementation skill.
“The research literature gives you an enormous number of possible directions. The difficult part is deciding which ones are actually worth turning into engineering projects, and finding the failure modes before you've spent months building around an idea.” (draft quote)
That research-to-product translation may ultimately be one of the most valuable skills in generative AI.
Then comes the part most demos leave out: deployment
A model that works in a laboratory is one thing. A model that works inside a customer’s environment is another. That distinction sits at the centre of Gabor Vecsei‘s role at DiffuseDrive.
As VP of AI Engineering, Vecsei leads engineering across AI, platform and infrastructure; from foundation-model work through to how the product actually gets deployed. His experience spans perception, generative AI, agentic LLM systems and the infrastructure required to train and serve those systems.
He has deployed AI onto constrained edge hardware, consumer platforms handling millions of requests per day and, more recently, environments where systems must operate completely disconnected from the internet. That last requirement is increasingly relevant.
AI is often described through a cloud-first lens: send the data to a large model, process it and receive the result. But some of the most important AI applications cannot work that way.
A hospital may not want sensitive patient information travelling to an external API. A bank may want trading data to remain inside its security perimeter. A government may require sensitive citizen information to remain within its jurisdiction. A defence organization may operate in an environment where the system simply cannot connect to the public internet. The underlying engineering challenge is similar. How do you make advanced AI useful when the internet isn’t available?
Air-gapped AI is not just a defence requirement
The term “air-gapped AI” can sound like a defence-specific concept. It isn’t. Defence is an obvious use case, but the underlying requirement is broader: AI that can operate inside a customer’s own security and data boundaries.
That can mean on-premise infrastructure, sovereign cloud environments, restricted networks or fully disconnected systems. For Vecsei, this changes the engineering problem fundamentally.
You cannot assume an external API will always be available. You cannot assume unlimited compute. You cannot assume that every model can be downloaded, updated or monitored through the internet. And you cannot treat deployment as the final step after the AI has already been built. It becomes part of the architecture from the beginning.
“Air-gapped AI isn't a deployment checkbox. Once the system has to operate without the internet, the architecture changes — from how you package models to how you manage updates, compute, security and observability.” (draft quote)
Vecsei led the transformation of DiffuseDrive’s technology into an enterprise-grade platform capable of operating in highly secure environments.
That reflects a larger trend. AI sovereignty is becoming an increasingly important strategic issue, particularly in Europe.
Europe has a different AI infrastructure problem
The United States still dominates private AI investment. Stanford’s 2026 AI Index estimates that U.S. private AI investment reached $285.9 billion in 2025, more than 23 times China’s $12.4 billion. Global corporate AI investment more than doubled during the year.
Europe is operating from a very different starting point. According to the European Commission, between 2020 and 2025 the U.S. allocated 34% of its €1.33 trillion in venture-capital funding to AI, compared with 18% of Europe’s €252 billion. The Commission also points out that European participation in AI deals falls sharply as deal sizes increase, leaving many late-stage rounds dependent on U.S. and UK capital.
Europe therefore has a funding and frontier-model disadvantage. But that does not mean it has no strategic opportunity. Europe has something else: a huge industrial economy, major automotive manufacturers, sophisticated manufacturing, robotics, aerospace, defence, agriculture and some of the world’s strongest data-protection and regulatory frameworks.
Those industries increasingly need AI that can operate in physical environments. And they often have stronger reasons than consumer software companies to care about where their data is processed, how it is controlled and whether it can leave the organization at all.
Europe’s physical-AI base is already substantial
The industrial robotics numbers illustrate the scale of the physical world Europe is trying to automate. The International Federation of Robotics recorded 542,000 industrial robots installed globally in 2024, more than twice the number installed a decade earlier. Europe accounted for approximately 85,000 installations, including 67,800 in the EU.
Germany alone installed nearly 27,000 industrial robots in 2024. Europe’s robotics market is therefore not hypothetical. There are already hundreds of thousands of machines operating in European industrial environments, creating a growing need for perception, automation and machine intelligence.
At the same time, European companies are rapidly increasing their adoption of AI. Eurostat reported that 20% of EU enterprises with 10 or more employees used AI technologies in 2025, up from 13.5% in 2024. Among large enterprises, adoption reached 55%.
The numbers are still lower than the headline figures often associated with AI adoption globally, but the direction is clear. AI is moving from experimentation into operational infrastructure.
The synthetic-data market is growing with it
The market around synthetic data remains difficult to measure precisely because different research firms define it differently. One estimate from The Business Research Company puts the global synthetic-data market at $920 million in 2026, growing to approximately $3.02 billion by 2030, representing a projected CAGR of 34.5%.
Grand View Research estimates the European synthetic-data-generation market at approximately $58.2 million in 2023, projecting it to reach $506.2 million by 2030, equivalent to a 36.4% CAGR. Its analysis identifies image and video data as the fastest-growing data type in Europe.
These numbers should be treated as market estimates rather than hard facts, but they point in the same direction. Synthetic data is moving from a niche technique toward a significant AI infrastructure category. And Physical AI may become one of its most demanding applications.
Europe is also building the compute layer
The EU is simultaneously investing heavily in AI infrastructure. In 2025, the European Commission announced an expanded network of 19 AI Factories across 16 Member States, backed by more than €500 million in joint EU and national investment. These facilities are intended to provide startups, SMEs and industry with access to AI-optimized supercomputing and technical expertise.
In July 2026, the Commission went further, launching a call for up to seven AI Gigafactories backed by up to €10 billion in EU and national funding, with the expectation that the initiative could unlock at least €20 billion in private investment.
And only days before this article was written, the EU awarded a €387.8 million contract for the LUMI-AI supercomputer in Finland, part of an effort to expand European AI computing capacity. EuroHPC says it has allocated €8.2 billion through 2027 to develop 19 AI Factories across 12 supercomputers.
The message is becoming increasingly explicit: Europe does not want to rely entirely on infrastructure controlled elsewhere. That matters for synthetic data too.
Defence is accelerating the demand for sovereign AI
The defence sector adds another layer. EU member states spent an estimated €418 billion on defence in 2025, equivalent to 2.2% of GDP, according to the Council of the EU. Spending is expected to rise to approximately €454 billion in 2026.
At the 2025 NATO summit, allies also committed to reaching 5% of GDP annually on defence-related spending by 2035, including at least 3.5% for core defence requirements. European Allies and Canada increased combined defence expenditure by nearly 20% in real terms in 2025 compared with 2024.
More defence spending does not automatically translate into AI spending. But it does increase investment in the systems where AI is increasingly embedded: drones, autonomous platforms, surveillance, targeting, logistics, robotics and sensor systems. These are precisely the environments where data is scarce, edge cases matter and cloud connectivity can be restricted.
The EU has another reason to care about data quality
Europe’s regulatory environment adds another dimension. The EU AI Act explicitly requires high-risk AI systems to use training, validation and testing datasets that are relevant, representative and, to the extent possible, free of errors and complete for their intended purpose. It also calls for data governance around collection, preparation, labelling, quality, bias and data gaps.
The significance goes beyond regulation. It reflects a broader principle: data quality is part of AI system quality. For Physical AI, this means that data infrastructure cannot simply be a storage layer sitting behind the model. It becomes part of the engineering system itself.
Small, senior teams may be better suited to this problem
The people building that infrastructure are changing too. Vecsei’s experience has led him to favour small, senior-heavy engineering teams that can operate with considerable autonomy.
The idea is almost closer to a special-operations unit than a traditional engineering army. The problem is not necessarily solved by adding more people. In fast-moving AI development, coordination itself can become the bottleneck.
“The most expensive line item in AI engineering isn't headcount. It's coordination and opportunity cost. You feel it as soon as the team becomes large enough that people stop knowing what everyone else is actually trying to solve.” (draft quote)
His hiring philosophy follows the same principle: find people who can take an ambiguous problem, reason about it, investigate what they don’t know and move forward without requiring every next step to be specified for them.
That is particularly relevant in AI because the underlying technology changes too quickly for every problem to have a predefined playbook.
AI-assisted engineering creates a new dividing line
The same principle applies to the growing use of AI coding tools. AI can dramatically accelerate engineering, but only if the engineer understands what the generated code is supposed to accomplish.
One failure mode is treating AI tools like slot machines: ask for something, run the result, keep generating until something works. The more valuable approach is different.
Use AI to eliminate mechanical work, freeing engineers to spend more time on architecture, experimentation and creative problem-solving. The dividing line isn’t necessarily senior engineers versus junior engineers. It is whether the engineer is using AI to extend what they understand or to avoid understanding what they don’t.
“The best engineers aren't the ones who use AI to avoid learning. They're the ones who use it to spend less time on mechanical work and more time on the parts of engineering that actually require judgement.” (draft quote)
The hidden infrastructure: evaluation
Perhaps the most underappreciated part of synthetic data is not generation. It is evaluation. AI teams often judge generated outputs visually because humans can immediately understand what they see. That creates a trap.
Vibes are useful at the beginning of a project. Researchers need intuition. Engineers need to look at outputs and notice unexpected behaviour. But a production system needs something more rigorous. It needs automated quality checks, heuristics, distribution-level validation and measurements that tell engineers whether the generated data is actually improving the downstream model.
Otherwise, synthetic data becomes an expensive image-generation pipeline. The real objective is model performance.
“Vibes are useful to start. They become a problem when you never build the measurement layer that turns intuition into velocity.” (draft quote)
That measurement layer is particularly important in mission-critical applications, where a visually impressive output can conceal a serious failure.
The EU AI Act’s requirements around accuracy, robustness and cybersecurity similarly emphasize the need to measure and manage system performance rather than treating AI as a black box.
The next generation of generative AI may be defined by control
The first wave of generative AI competed heavily on image quality. The next wave of industrial applications may compete on something else.
Control.
Where should the object appear? How large should it be? What identity should it have? What should be visible and what should be occluded? Which sensor characteristics should be represented? Which specific failure mode is the generated sample supposed to address?
For marketing or creative applications, an aesthetically superior image may be the objective. For robotics, autonomous driving and defence, a slightly less beautiful image that satisfies precise constraints can be much more valuable. That is why technologies such as bounding-box-controlled generation are potentially more significant than they first appear.
They move generative AI from: “Make me something like this.” toward: “Create exactly the example my perception system needs.” And once the generation system can control the image and its spatial annotation simultaneously, another bottleneck can potentially disappear: manual labelling.
Synthetic data isn’t replacing reality
There is a temptation to frame synthetic data as a replacement for real-world data. That is probably the wrong way to think about it. The strongest systems will continue to combine both.
Real data anchors the model in the actual world. Synthetic data expands that world into places where reality is insufficient. The workflow becomes a feedback loop between them.
“Synthetic data isn't a substitute for reality. Reality tells you where the system fails; synthetic data gives you a way to deliberately explore and fix those failures.” (draft quote)
That is why the opportunity is potentially much larger than simply selling generated images. The real opportunity is building the data infrastructure around the AI development loop.
What makes an AI infrastructure company defensible?
There is another important question for the industry. As generative models become increasingly commoditized, what actually creates a durable AI company? A wrapper around somebody else’s model is unlikely to be enough.
If a startup’s product disappears the moment OpenAI, Anthropic, Google or another model provider changes its API, it is probably operating at the commodity layer. The defensible companies are more likely to own something deeper: proprietary workflows, evaluation infrastructure, domain-specific datasets, customer integrations, deployment infrastructure, feedback loops, specialised models, operational knowledge, security and compliance, and, increasingly, the ability to operate independently of any single model provider.
That is particularly important in Physical AI because the customer’s problem is rarely just “give me an image.” It is: Find where my model fails. Generate the right data. Validate it. Train on it. Measure the result. Repeat. And do it inside the environment where my data is allowed to exist. That is infrastructure.
The people building the layer underneath Physical AI
That is ultimately what makes the work of Schmid, Haffner and Vecsei interesting. They occupy three different parts of the same emerging stack. Daniel Bin Schmid is working on the research problem: how to make generative models controllable enough to create useful, precisely targeted training data rather than merely plausible images.
Domonkos Haffner sits between research and architecture: deciding which advances are worth productizing and turning them into a production AI system capable of generating data that improves machine perception. Gabor Vecsei works on the systems layer: taking those models and architectures into environments where AI has to function reliably, securely and sometimes completely disconnected from the internet.
Together, their work reflects a broader transition in AI. The first generation of AI infrastructure was largely about making models bigger and giving them more computing power. The next generation may be about something more difficult: making AI understand the physical world under the exact conditions in which it is most likely to fail. That requires data that does not yet exist.
It requires systems capable of generating it. And it requires infrastructure that can prove whether that data actually made the machine better. For mission-critical Physical AI, that may become one of the most important technology layers of the next decade.
