Since the launch of ChatGPT — much of the investor attention around AI has followed the SaaS playbook: using the technology to try to drive enterprise efficiencies. But the large language models (LLMs) that are powering many of these new B2B tools are designed in a way that means occasional errors in their output are difficult to fully eliminate, making them unsuitable for many business use cases.
Some of Europe’s brightest minds are trying to fix that problem but there’s another class of AI startups emerging that doesn’t suffer from these inaccuracy issues, and has a huge addressable market to target: companies using machine learning to optimise engineering.
Startups around Europe are now using the latest developments in AI model architecture, but feeding them training data in the form of physics equations and maths, rather than terabytes of writing from the internet, in order to build better cars, planes and semiconductors.
If these companies succeed, they say they’ve got the potential to make big inroads into trillion-dollar industries, by using AI to improve the way we design, optimise and invent new products.
50% of the world’s energy
One of these startups is Cambridge-based Monumo. Founded in 2021, it’s developing an AI model to design better electric motors.
“50% of the world's power goes into electric motors today,” founder and CEO Dominic Vergine tells Sifted. “So a 5% improvement in electric motor efficiency would be the equivalent of removing the carbon emissions of Germany and France combined; a huge amount of emissions on a fairly small change.”
Vergine adds that electric motors are an appealing thing to optimise with AI because they represent a “nicely constrained multi-physics problem” built with a handful of components, unlike trying to redesign a whole car, for instance.
Monumo is doing this by developing an AI model that uses some of the machine learning techniques that power LLMs but training them on the maths that explains the physics that affects electric motors. This allows it to design motors for different applications that use energy more efficiently than current systems, which Vergine says are often built with components that are all developed separately from each other.
“With machine learning and the speed of computer processing now, we can start doing something where instead of designing individual components and plugging them together, we design those components as one,” he tells Sifted.
Vergine adds that Monumo’s approach is already creating results that outperform products “that have been developed by the best teams with the best software that was available until now”, by “quite an interesting margin”. Monumo is in conversation with more than 35 companies, he adds.
“We’re now starting to deliver results for one of the top automotive suppliers in the world and are discussing further work together,” he says. “In the next week or so we also expect to begin working with a Japanese car company with the intention of building a long-term relationship with them.”
Big business
Vergine points out that the motor optimisation technology that Monumo is developing can also be applied to electricity generation, as “about 95% of the world's electricity relies on a turbine,” with the other 5% being made up by solar.
This, he says, positions Monumo within wider global efforts to disrupt one of the biggest industries in the world: “We're trying to replace a $5tn a year industry — the oil industry — with something else. That is what decarbonisation is. So the things that start to fill that gap are going to be hugely valuable.”
Another European startup applying AI to valuable engineering sectors is London-based PhysicsX, which raised a $32m round led by General Catalyst. It’s using machine learning to help clients optimise design, across sectors including automotive, aerospace and semiconductors.
Cofounder and CEO Jacomo Corbo tells Sifted that the company is working with five paying customers and that its machine learning technology allows them to test potential designs of components “hundreds of thousands, even a million times faster than numerical simulation”.
With its focus on manufacturing-heavy industries, PhysicsX is also going after a big market. It’s estimated that manufacturing accounts for around 20% of GDP per country, globally.
A better use of AI?
One advantage of using machine learning trained on physics equations, as opposed to LLMs trained on human language, is that these systems don’t “hallucinate” and make mistakes in the same way, according to Dan Garvey, CTO of Cambridge-based startup BeyondMath.
Last year he told Sifted that the company’s AI model is a far cheaper alternative than the previous numerical simulation method of testing component design. It could also allow smaller companies to use precision design processes that were previously too costly.
Corbo adds it’s natural that GenAI models like ChatGPT and image generators are drawing so much public attention, as people can interact with them without technical skills, but questions whether these tools promise the best returns for VCs.
“From an investor standpoint, the question is, ‘Are the returns there? Do they motivate the kind of infrastructure spend and infrastructure build out that we have seen?’” he says.
Corbo is alluding to GenAI tools that are based on LLMs or image generators requiring large amounts of computing power to run, due to the huge amounts of data they are trained on. This also translates into huge energy usage for these systems — a report earlier this year by a group of climate change organisations estimated that AI could lead to an 80% increase in global carbon emissions.
Vergine says that physics-based AI models are far less data and energy-intensive to train and run, and says investors should consider the climate impact of the technology as they decide which AI products to invest in.
“We need to start thinking, is this really justifiable? Is it right to build a chatbot that can be slightly better than a search engine when it's burning that much power?” he adds. “I think we need to be thinking really hard about where we are focusing AI because we can't just burn the planet to generate some funky images.”
