We’ve been here before: why the edge will break AI’s compute monopoly
AI compute is concentrating in a handful of mega-clusters. History, and the physics of the real world, suggest the pendulum will swing back to the edge.
By 2028, just two AI labs could control half of the world’s newly deployed computing power. That is the striking projection from Dylan Patel, Chief Analyst at SemiAnalysis, in a recent conversation on the Dwarkesh Podcast.
Add a projected $10 trillion of global infrastructure investment by 2030, severe constraints on power grids, and a deep hardware asymmetry between the United States and China, and it feels as if we are entering entirely new territory. The prevailing narrative is clear: centralised mega-clusters are the only way forward.
But is this concentration of computing power actually unprecedented?
We have been here before
In the 1960s and 1970s, IBM dominated the world’s computing through the mainframe. If you wanted serious computing power, you went through a single, centralised titan. Its position felt unassailable.
That dominance wasn’t broken by a bigger mainframe from a competitor. It was broken by a change of paradigm: personal computing. Computing power moved out of the data centre and onto the desk, and the economics of the whole industry changed with it.
We believe the same cycle is going to play out with AI.
The case for centralisation
It is worth being fair to the current trajectory. Frontier labs such as OpenAI and Anthropic are building formidable efficiency advantages. They can turn a single watt of energy into more capable output, and more revenue, than almost anyone else. That lets them outbid the rest of the world for hardware, and reinvest the margins straight back into ever-larger data centres in pursuit of models that help improve the next generation of models.
For training frontier models, that logic is hard to argue with. Training will likely stay concentrated for some time.
The hard wall: the physical world
But extreme centralisation runs into a limit that capital alone can’t move: physics.
As AI moves from chat interfaces into the physical world (autonomous vehicles, robotics, defence systems and real-time industrial automation), the architecture has to change. Three constraints make that inevitable.
The latency trap. Machines operating in the real world can’t wait for data to travel to a data centre thousands of kilometres away and back. In a moving vehicle or on a production line, a split-second delay can be the difference between a successful manoeuvre and a serious failure.
The bandwidth bottleneck. Billions of physical devices generate enormous volumes of raw sensor and video data. Shipping all of it to the cloud for processing is a logistical and financial burden that grows with every device deployed.
The privacy and reliability imperative. Critical infrastructure and local robotics have to keep working when the network drops or a cloud region goes down. Many of these systems must be able to run entirely offline, and many handle data that should never leave the site at all.
Train in the centre, act at the edge
None of this means the mega-clusters disappear. The likely end state is a division of labour: the largest clusters train the models, and the edge is where those models are put to work in the real world. Models get distilled, compressed and specialised so they can run on vehicles, machines, cameras and gateways close to where decisions are made.
We see this in our own engineering. One of our computer-vision projects runs real-time object detection and tracking entirely on a Raspberry Pi 5, controlling a pan-tilt camera with no round trip to the cloud. It is a small system, but it illustrates the point: when a machine has to react to the physical world, the intelligence has to live close to it.
The pendulum swings back
Technology tends to move like a pendulum between centralised and distributed computing: mainframes, then personal computers; then the cloud, and now, increasingly, the edge. Right now the momentum is swinging hard towards centralised compute. As AI becomes physical, we expect it to swing back towards local, distributed intelligence.
The mega-clusters will train the models. The edge is where real-world execution will happen. And that may be how the frontier’s grip on computing power finally loosens.
Will the laws of physics break the data-centre advantage, or is this time truly different? We’d be glad to hear your view.