Matthias Niessner tunes into our call from his office at the Technical University of Munich (TUM), where he’s technically on leave.
“I’m currently on entrepreneurial sabbatical,” the professor says, adjusting his chunky black headphones. “I still make sure my graduate students have some stuff to do, otherwise I have no obligations to the university right now.”
Considered one of Europe’s top AI researchers, Niessner has led the visual computing and AI lab at TUM — a well-trodden training ground for unicorn founders like Celonis’s Bastian Nominacher and Personio’s Hanno Renner — for the past eight years.
His research also laid the groundwork for AI video startup Synthesia, which he cofounded in 2017 and is now valued at $2.1 billion.
Last year, Niessner stepped back from both roles to take on a new challenge. While huge leaps forward have been made in AI — from ChatGPT to image, audio and video generation — there’s yet another frontier to explore: immersable 3D worlds that you can interact with and walk around in.
His new company, Spaitial, is building foundational models to generate these 3D worlds from inputs like text or images, which could be used for everything from virtual reality gaming to urban planning — and has already raised a $13 million seed round. But building models for interactive 3D environments presents unique challenges.
“We want to create worlds that don’t just look like virtual worlds, they should also behave like virtual worlds,” Niessner says. “This is an order of magnitude harder than static visualisations.”
When I ask how long the models might take to build, Niessner turns coy: “I can’t share the full timeline, but it’s not going to be years, and there will also be double steps in between,” he says. “We know how to do it. We just have to build it.”
The difficulties of go-to-market
Spaitial’s first task is building 3D foundational models; next is figuring out who will use them — and for what.
Niessner says this is the “biggest challenge” any AI company faces today.
He encountered the same issue early at Synthesia, which uses generative AI to create realistic avatars for marketing videos and communications.
“We were doing movie dubbing and very quickly realised it’s a tough market,” he says. “If a Hollywood production pays $50–100k for a movie to be dubbed properly, there were only about 20 movies a year that could afford that. The whole market was just too small.”
The breakthrough came when Synthesia pivoted to corporate clients, helping global firms like McDonald’s create multilingual training videos. That’s when the company really took off.
Niessner has “no illusions” that Spaitial’s go‑to‑market path will be equally tricky — but he already has some idea of the technology’s potential use cases.
One use case Niessner can imagine is 10 year olds creating their own video games from a text prompt, and placing themselves inside them. The technology could also have applications in design and architecture. Construction firms or architects, for instance, might use it to generate immersive 3D models of buildings — which real estate agents could then show to clients in virtual walkthroughs.
Niessner also sees potential in entertainment and, further down the line, in robotic training — where rich 3D simulation environments are critical.
Spaitial is still weighing up whether to build these use cases in-house — by partnering with other companies — or get developers to license the model to build downstream applications themselves.
Andre Retterath, general partner at Earlybird, which led the seed round, says he doesn’t expect the commercial side to be a problem for Spaitial: when it opened its waitlist in May, thousands of people signed up in the first few days explaining how they’d use the tech.
“The question is not if the demand is there or if the market is big enough. It’s really a question of, what’s the sequence? Which market has the strongest pull dynamic, which market is, from a value chain perspective, the most advanced?” he says.
There are other companies working on similar projects. Among them is World Labs, which is building large world models (LWMs) to enable AI systems to “perceive, generate and interact” with the 3D world as it says on its website. The company was founded by renowned Stanford University AI professor Dei-Dei Li in 2023 and is already valued at over $1bn.
A ‘conservative’ approach
Niessner knows there are other challenges awaiting Spaitial. Not least, accessing enough computer power.
Training large foundational AI models (like GPT or DALLE, for example) requires massive amounts of compute — often thousands of GPUs running for weeks or months on end, which naturally incurs huge costs. It’s partly why companies like OpenAI, Anthropic, and Google DeepMind have raised billions in funding.
Niessner says he “doesn’t believe” in raising billions upfront like some of the big players in the US AI scene and that the $13m his company raised “isn’t nothing.”
Nevertheless, he acknowledges Spaitial will need to raise more money eventually — maybe tens or even hundreds of millions — but insists it won’t be in the billions.
“OpenAI is doing great, but there’s a huge risk involved,” he explains. “How much more can you raise before you hit a ceiling? And when will you become profitable? We didn’t want to raise hundreds of millions without having much product yet. There has to be a clear value inflection point before you can justify those massive rounds.”
And for startups, “over-raising too quickly can backfire quite dramatically,” he adds. “We’ve seen it happen. We wanted to avoid that.”
Spaitial plans to be “conservative” in how it spends its resources. For example, by only doing training runs, when a model is trained on new data, that are necessary, as these can quickly become expensive.
Retterath says Deepseek showed how a combination of architectural innovations and training strategies delivered significant efficiency gains that can be adapted for 3D models too.
Compared to text, audio, or image-generation, where these recent more efficient approaches did not yet exist, he expects the overall investment required to achieve a commercially viable 3D foundation model today to be lower.
“Methodologically, 3D models stand on the shoulders of previous modalities and thus the initial quality will be very high, leading to fewer training runs and less incremental releases.”
The opportunity for startups
Would all this be in vain if Sam Altman or his peers at Google with far deeper pockets started building what Spaitial is trying to?
Niessner is pretty confident they won’t.
“The top people who were building this at those companies are now my cofounders,” he says, grinning, “and the reason why they actually were interested in doing a startup is because big companies inherently move very slowly.”
Spaitial’s other cofounders are David Novotny, who led Meta’s text-to-3D asset generation project, Ricardo Martin-Brualla, formerly of Google’s 3D teleconferencing platform Beam and Luke Rogers, former senior director of pricing and purchasing at Cazoo. Niessner rented a room from Rogers in Palo Alto while doing a visiting professorship at Stanford and the two have been friends since.
“It’s very difficult for an existing company to change their DNA, right?,” Niessner continues. “Microsoft dominated the software market, but they didn’t get search right. Google dominated the internet market, but they didn’t get social media right. It’s just difficult for big companies to capture the new technologies in new markets.”
And that, he argues, leaves a big opportunity for startups like his to plug the gaps.
Finding the right people
Some argue Europe isn’t capable of building big AI champions, as most of the experienced talent and capital remains in Silicon Valley.
But Niessner isn’t in a hurry to move to the US anytime soon, especially as he enjoys his European liberties such as clean streets and relative safety, he says.
But if building the best company possible meant moving to the US, he’d do it.
“I think the key is: do you get the right people together? And if (the employees) are happy with the place where they live, then I think it doesn’t matter. AI companies will operate globally. There’s no, “oh we would just be in Europe for the sake of Europe,” he says.
“I know there’s a lot going on in AI sovereignty, which I am not a big fan of at all — because I believe I want to be the best company on earth, and I don’t care if I’m in Europe or in the US,” he adds. “I think we have to think bigger.”
He cites companies like Freiburg‑based Black Forest Labs — which has big US backers like General Catalyst, Andreessen Horowitz, and Y Combinator CEO Garry Tan — as proof that European AI can thrive. Though he admits European companies often run into challenges when raising big later stage rounds.
What Europe does have is talent, adds Niessner. Spaitial is in the process of building up its ten person team — and is currently hiring for five positions to be onsite at its Munich office, including an infrastructure engineer and a research scientist to design and develop 3D foundational machine learning models.
The company is not about to hire another ten or twenty people though, says Niessner. “AI companies are typically tight. It’s always quality over quantity.”
While Spaitial’s roadmap is still taking shape, Niessner remains laser‑focused on solving a problem he believes no one else has cracked — and if he succeeds, we might all soon step into 3D worlds we don’t just observe, but truly experience.
“In Europe, we have great people, we have great engineers,” he says. “We just have to have ambition in order to get there.”
Read the orginal article: https://sifted.eu/articles/big-interview-matthias-niessner/
