There’s plenty of debate on whether the current AI datacenter buildout is sustainable, but it could also face threats from an unexpected area.
Aravind Srinivas, CEO of AI search startup Perplexity, has outlined an interesting thesis that challenges the fundamental assumptions behind the hundreds of billions being poured into centralized AI infrastructure. Speaking about the future of artificial intelligence deployment, Srinivas suggested that the entire datacenter boom could be undermined not by economics or energy constraints, but by technological advances that make large-scale cloud infrastructure obsolete.
“The biggest threat to a datacenter is if the intelligence can be packed locally on a chip that’s running on the device, and then there’s no need to inference all of it on one centralized datacenter,” Srinivas said on a podcast. “It becomes more decentralized and even better if the models that are coming along with the chip are things that adapt to you.”
The Perplexity CEO explained that personalization wouldn’t necessarily require on-device model training. “You may not even need that for adaptation to you. It could just be data that lives on your computer or your device that can be retrieved on the fly. So retrieval augmented generation, tool calls, all of that can already help personalize things to you,” he noted.
But Srinivas painted an even more radical picture of what truly localized AI could enable. “Imagine we crack something like test time training where there are some tasks that you repeatedly do on your local system and the AI watches it, and you’re fine with AI watching it because AI is living on your computer, it’s not going to the servers, and it adapts to you. And over time it starts automating a lot of the things you do, that way you don’t have to repeat it. That’s your intelligence. You own it, it’s your brain.”
The implications of this vision are stark. “That really disrupts the whole datacenter industry,” Srinivas concluded. “It doesn’t make sense to spend all this money, 500 billion, 5 trillion, whatever, on building all the centralized data centers across the world that do a lot of the intelligence workloads for people.”
Srinivas’s comments arrive at a moment of unprecedented investment in AI infrastructure. Microsoft, Google, Amazon, and Meta are collectively spending over hundreds of billions annually on datacenters and AI chips, while Saudi Arabia has announced plans for a $100 billion AI initiative. The industry has largely operated on the assumption that AI inference and training will remain centralized due to the computational demands of large language models. However, recent developments suggest the on-device AI future Srinivas describes may not be far-fetched. Apple has deployed on-device AI capabilities across its latest iPhone lineup, while companies like Qualcomm and MediaTek are embedding neural processing units into smartphone chips. Google’s Gemma models are designed to run entirely on-device, and Meta and some Chinese labs have released Llama models small enough for local deployment. If the trend toward efficient, personalized, local AI accelerates, the multi-trillion dollar bet on centralized infrastructure could face its most existential challenge yet—not from regulators or energy costs, but from the very technological progress it was built to support.