The technology sector exhibits a recurring pattern every few years. The majority of people are persuaded that the story is resolved by one category that dominates the headlines and attracts all the capital. Then something underneath moves silently. Everyone was never looking for the real money.
That’s precisely what edge computing seems to be experiencing at the moment. While investors have been obsessing over which AI model would “win” over the past two years—GPT-4 or Gemini, Claude or Llama—a more realistic and possibly more resilient competition has been developing in the background. It is taking place on factory floors, in semiconductor design labs, and within the commonplace architecture of facial recognition security cameras and smart utility meters. Not glitzy. Not worthy of a headline. But perhaps the site of the next giant’s construction.
Simply put, edge computing is computing that takes place locally, either on the device itself or very close to it, rather than on your phone or a far-off server farm. Imagine a self-driving car that decides to brake because it cannot afford the milliseconds it would take to ping a cloud server. Or a smart factory camera that doesn’t need to upload raw footage to Amazon’s infrastructure in order to process a thousand images per hour. To solve these real constraints, real silicon is needed. The market is starting to realize that silicon is far more capable than it has ever been thanks to the AI era.
According to Technavio, the global AI edge computing market will expand at a rate of about 32% per year through 2029. That is a real deployment that is currently taking place in logistics, healthcare monitoring, autonomous vehicles, and industrial automation; it is not a speculative forecast based on hype. It’s interesting to note that this growth is independent of the person who creates the best large language model. Whoever creates the best chip will determine the outcome.
Here, it’s important to pay attention to Broadcom’s recent release of its eighth-generation Wi-Fi 8 chipset. It’s the kind of product announcement that has significant impact but seldom appears on the front page. This chipset allows a wireless access point to operate as a machine-learning neural network for the first time, processing and interpreting the data it gathers without sending anything to a distant server. This represents a structural change in the distribution of computation. Qualcomm is operating in a similar manner. It is not necessary for either business to win the model race. They must prevail in the hardware race, which has already begun.
All of this was greatly sharpened by the DeepSeek moment earlier this year. The assumptions supporting the AI investment thesis were rocked when a Chinese lab claimed to have trained a frontier-level model for about $6 million, a small portion of the $100 million or more spent on models like GPT-4. The model itself doesn’t have any defensible value if models can be constructed at a low cost and open-source versions can match proprietary ones in terms of performance. Inference is carried out at the edge by the infrastructure. In short, most startups should stop creating foundation models and start building on top of them, according to Kevin Xu, the head of Interconnected Capital. Applications, services, and the hardware that makes them possible will come after the investment.
The speed at which market valuations will fully reflect this change is still unknown. However, there is a growing perception among the investment community’s more hardware-focused segments that Broadcom and Qualcomm are positioned in ways that Nvidia, despite its dominance in data center GPUs, might not be able to fully replicate at the edge. Nvidia owns the training compute story. We’re still writing the inference-at-the-edge tale.
As this develops, it’s difficult to avoid thinking back to the early 2010s, when Arm Holdings secretly licensed the chip architecture that powered all of the smartphone operating systems while everyone argued over which would prevail. A similar dynamic could result from the edge compute revolution. The business creating the most intelligent model might not be the winner. Without anyone realizing it until the valuation is already in the hundreds of billions, it may be the company producing the chip that powers every smart device, factory, hospital, and driverless car.
