In aerospace, meanwhile, drone control systems are incorporating data from a growing number of sensor inputs.
Autonomy can make life easier for human controllers, but it comes at a cost. Advanced AI-based technologies can be incredibly power-hungry – and that growing appetite could soon force engineers to make some difficult decisions, according to Brandon Lucia, professor of electrical and computer engineering at Carnegie Mellon University in Pennsylvania.
“For decades, power has been the invisible enabler of progress – abundant, reliable and taken for granted,” he told Professional Engineering. Now, however, that “illusion of infinite power” is starting to crumble: “Industries that have never considered power a constraint will now have to design around it.”
Speaking in late 2025, he said that electric vehicles (EVs) would soon hit new limits as the growth of ‘AI compute’ – resources required to train and deploy systems – outpaces battery energy density improvements, ultimately threatening journey range and limiting capabilities.
In the air, autonomous flight systems could struggle to balance sensor fusion, navigation and AI decision making within tight thermal and power envelopes.
“Every new advancement will face the question: can we afford the energy it consumes? Power will become the ultimate design constraint – the limiting factor in how we build, deploy and scale intelligence itself,” said Lucia, who is also CEO of general-purpose processor firm Efficient Computer.
One solution will be lightweight ‘edge’ AI chips and dynamic energy allocation to preserve range while maintaining intelligence and autonomy, according to Lucia. But, he said, we are approaching the limits of hardware specialisation.
“The complexity of specialisation will outweigh its benefits, and the next wave of AI performance will hinge less on hardware breakthroughs and more on algorithmic ingenuity,” he said.
Dr Jonathan Lancelot is working on that algorithmic ingenuity. The researcher and senior lecturer at the University of the West of England (UWE) Bristol focuses on making neural networks for computer vision more energy-efficient, particularly in space, defence and remote environment applications.
Weather satellites are a stark example, he said – any onboard AI systems need to be as energy-efficient as possible as there are other vital functions draining the same finite power source. EVs face similar challenges. “You want your remote agent, or your neural network, to function for the car and not use as much energy,” he said.
Edge deployments, with computations carried out on the vehicles themselves, are often preferable to communicating with data centres via the cloud, especially for safety or security. This increases the onboard power demand, however.
AI is so energy-intensive because it relies on complex mathematical calculations known as ‘floating point operations’, Lancelot said. Last year, he developed a new approach that relies on simple logical operations instead, reducing the required compute and cutting energy consumption. “That is the direction we should be going when we’re talking about edge computing,” he said.
It is not just the power demand of vehicles themselves that needs to be considered. Despite the push for edge operations, many processes are still hybrid or cloud-based, including voice assistants, self-learning driver ‘profiles’ and predictive maintenance, according to consulting firm McKinsey & Company. Greater autonomy in cars and other vehicles is therefore adding to the ever-growing power demand from new AI data centres.
“We should not be heavily dependent on data centres,” Lancelot said. “That will eventually come up on energy grid issues. It’ll turn into geopolitical issues… we don’t know what the future is going to look like.”
Thankfully, AI techniques could also bring opportunities for some significant energy-efficiency improvements in automotive and aerospace. A research project led by Imperial College London, for example, explored the use of AI-controlled dynamically adjustable flaps for heavy road vehicles. By reacting to turbulence and other external conditions, the reinforcement learning-based technique could reduce drag by 10% and fuel consumption by 5%. In aviation, AI could help pilots reduce climate change-contributing contrails.
With challenges to overcome and opportunities to take, one of the most important factors will be ensuring there is enough expertise to seize them. This could be a challenge in the UK, however, with a recent report from Coventry University and the Advanced Propulsion Centre identifying a “critical shortage” of AI and digital skills.
Ultimately, Lucia predicted, AI’s influence on engineering is only set to grow and “stick-in-the-muds” will be forced to adapt or fall behind. “The winners of this next phase will be those who design AI systems that can evolve faster than the chips they run on,” he said.
Want the best engineering stories delivered straight to your inbox? The Professional Engineering newsletter gives you vital updates on the most cutting-edge engineering and exciting new job opportunities. To sign up, click here.
Content published by Professional Engineering does not necessarily represent the views of the Institution of Mechanical Engineers.