Perplexity CEO Aravind Srinivas has warned that if capable AI can run locally on personal devices, the economic and environmental case for endlessly expanding large data centres could start to weaken.
Data Packed Locally On A Chip Instead
For most users today, artificial intelligence follows a simple pattern. A request is sent from a phone, laptop, or app to a remote data centre, where a large model processes it before returning a response. This centralised approach has shaped how the AI industry has grown and where investment has flowed.
Srinivas has questioned whether that model will remain dominant over the long term. Speaking on a recent podcast, he argued that the “biggest threat to a data centre” would come if intelligence could be “packed locally on a chip that’s running on the device”, removing the need for much of the inference work to happen in central facilities, i.e., the everyday use of an AI model, such as generating answers, summarising documents, or analysing data after the model has already been trained.
Training, by contrast, is the highly resource intensive phase where models learn from massive datasets, usually using clusters of specialised processors inside data centres.
Srinivas’s argument is not that data centres suddenly disappear. Instead, he suggests that if more inference and personalisation move onto devices, the demand for centralised infrastructure may grow more slowly than expected, raising uncomfortable questions about the scale of current investment plans.
Why This Has Become a Sustainability Issue
The warning comes as the environmental impact of AI infrastructure is drawing increasing attention. Data centres already consume large amounts of electricity, and AI has accelerated that growth. For example, the International Energy Agency estimates that global electricity consumption from data centres could rise from around 460 terawatt hours in 2022 to between 945 and 1,050 terawatt hours by 2030, effectively doubling within a decade as AI workloads expand. The agency also notes that electricity demand from data centres is growing more than four times faster than overall global electricity demand, placing increasing pressure on power grids and decarbonisation efforts. At that scale, data centres would rank among the world’s largest single categories of electricity demand.
However, the pace of growth matters as much as the absolute numbers. For example, the IEA has also highlighted that electricity demand from data centres is increasing several times faster than overall electricity demand, thereby creating pressure on grids, generation capacity, and decarbonisation plans.
Water use has become another point of concern. For example, many large facilities rely on water-based cooling systems, either directly or indirectly through power generation. In water-stressed regions, new data centre projects have faced public opposition and regulatory scrutiny, particularly where local communities see competition for limited resources.
It’s against this backdrop that the idea of moving some AI workloads away from centralised facilities appears to offer a possible route to reducing environmental pressure, or at least slowing its growth.
What On Device AI Really Involves
It’s worth noting that on device AI doesn’t mean abandoning the cloud entirely, as it actually describes running certain AI tasks directly on local hardware, using specialised chips designed for machine learning workloads.
In fact, this is already happening in limited ways. For example, modern smartphones and laptops increasingly include neural processing units, which are optimised for tasks such as image recognition, speech processing, and text summarisation. These chips allow some AI features to run quickly without sending data to remote servers.
Apple, for example, has positioned on device processing as a core part of its approach to AI, emphasising privacy and speed. Microsoft has taken a similar route with its latest generation of Windows laptops, promoting devices capable of handling AI workloads locally through dedicated hardware.
In practice, most current systems are hybrid, e.g., smaller, frequent tasks may run on the device, while larger or more complex requests are still handled in the cloud. The question is whether that balance will shift significantly over time.
Why Local AI May Cut Impact (Or Not)
At first glance, the sustainability case for local AI seems pretty straightforward, e.g., if fewer requests are sent to data centres, fewer servers are needed, and energy and water use could grow more slowly.
However, the reality is more complex, and making AI cheaper and more responsive can increase usage. If people rely on AI more often throughout the day, total energy demand may still rise, even if each individual task becomes more efficient.
There is also the issue of where energy is consumed. For example, a highly optimised data centre running on low-carbon electricity may, in some cases, be more efficient than millions of individual devices drawing power from more carbon-intensive grids. The environmental outcome depends heavily on local energy mixes and usage patterns.
This is why claims that data centres will become obsolete are so controversial, as a shift in where computation actually happens doesn’t automatically translate into lower overall environmental impact.
Smaller Data Centres and Waste Heat
The debate around on device AI is also reshaping how data centre design is being approached. For example, rather than relying solely on vast, remote facilities, some operators are exploring smaller, more distributed models that place computing closer to where it is needed. Known as ‘edge computing’, this approach reduces latency and can improve responsiveness, while also opening up new sustainability opportunities.
In the UK, several projects have demonstrated this approach in practice. For example, at Exmouth Leisure Centre in Devon, a small-scale data processing unit operated by Deep Green uses immersion cooling to capture heat from servers and reuse it to warm swimming pools and hot water systems. The same model has since been applied in other public sector buildings, where computing infrastructure is integrated into heating systems to improve overall energy efficiency.
Facilities with a constant demand for heat are particularly well suited to this model, because the heat generated by local computing can be reused on site rather than being discarded, something a remote hyperscale data centre cannot offer.
These approaches do not remove the energy demands of computing, but they do improve overall efficiency by linking digital infrastructure more closely to real-world energy needs.
Why Large Data Centres Are Still Being Built
Despite growing interest in local and edge computing, investment in large data centres continues at pace and it seems there are practical reasons for this. For example, training the most advanced AI models still requires concentrated computing power, specialist cooling, and robust power infrastructure. Many business services also depend on centralised platforms for reliability, compliance, and security, particularly in regulated industries.
It’s worth noting here that data centres also support far more than AI. For example, streaming, online banking, enterprise software, cloud storage, and collaboration tools all rely on centralised infrastructure and, even if some AI workloads move elsewhere, these services still need to run.
That said, technology companies are aware of the sustainability pressure and are responding with efficiency improvements, renewable energy procurement, and public reporting commitments. These steps suggest preparation for long-term operation rather than an expectation of rapid decline.
The Technical Barriers to a Device First Future
Despite Srinivas’s predictions, he has acknowledged that on device AI faces real technical obstacles. Advanced models place heavy demands on memory, bandwidth, and thermal management. Running them continuously on a phone or laptop can drain batteries quickly and generate heat that hardware struggles to dissipate. Cost is another factor, since more powerful chips raise device prices and limit accessibility.
Progress is being made through smaller, more efficient models designed for specific tasks rather than general purpose use. Researchers and companies are increasingly focusing on models that are “good enough” for everyday work, such as summarising documents or managing routine workflows, without requiring enormous computing resources.
For example, an email assistant that sorts and drafts messages does not need the same scale of model as a system designed to generate long-form creative content across many domains.
What This Means for the Future of Infrastructure
All things considered, it seems the most likely outcome is not a collapse of data centres, but a gradual redistribution of workloads.
Large facilities remain essential for training advanced models and supporting global digital services. At the same time, more inference may shift onto devices and into smaller, local facilities, reducing some traffic and changing where energy is consumed.
From a sustainability perspective, this raises new priorities. Efficient chip design, longer device lifetimes, repairability, and transparent reporting of energy and water use become more important as computing spreads out across billions of devices.
It also sharpens the risk of overbuilding. If assumptions about ever-rising centralised demand prove wrong, the environmental cost is not only operational energy use but also the embodied carbon in construction, equipment manufacturing, and supporting infrastructure.
Srinivas’s warning does not predict the end of data centres. It highlights a growing uncertainty at the heart of the AI boom, where technological change, environmental limits, and investment decisions are becoming increasingly difficult to separate.
What Does This Mean For Your Organisation?
The rapid growth of on device AI is beginning to complicate long-standing assumptions about how and where AI infrastructure should be built. While large facilities remain essential for training advanced models and supporting global digital services, growing interest in on device AI and distributed computing is introducing new constraints on how much centralised capacity is truly needed.
For UK businesses, this has direct implications for how AI is deployed, governed, and paid for. As more AI capabilities move closer to the user, organisations may gain greater control over data handling, latency, and operating costs, while still relying on the cloud for scale, resilience, and compliance. This has direct implications for IT strategy, sustainability reporting, and long-term procurement decisions, particularly as energy prices, carbon targets, and regulatory scrutiny continue to tighten.
For policymakers, infrastructure planners, and local communities, the risk is not simply overbuilding data centres, but committing to energy-intensive infrastructure at a time when the underlying technology is still evolving. Srinivas’s warning does not predict the end of data centres, but it does highlight growing uncertainty around how AI infrastructure should be planned, regulated, and sustained as environmental limits and technological change increasingly intersect.
