Perplexity CEO warns on-device AI threatens $500 billion data center industry buildout

Aravind Srinivas Challenges the Centralized Data Center Model

Aravind Srinivas, CEO and co-founder of Perplexity AI, has issued a contrarian warning about the future of artificial intelligence infrastructure. Speaking in a podcast interview with Prakhar Gupta, Srinivas argued that the biggest threat to data centres is local intelligence, where AI capabilities are "packed locally on a chip that's running on the device" 1

1

2

2

. This approach eliminates the need for inference on centralized data center infrastructure, fundamentally challenging the prevailing model where companies shell out billions of dollars to acquire GPUs and build hyperscale facilities.

The Perplexity CEO, who previously worked at OpenAI, Google Brain, and DeepMind, called this a "$10 trillion question, hundred trillion dollar question," questioning whether it makes sense to spend $500 billion to $5 trillion on building cloud-based centralized data centers across the world 3

3

. His vision describes a more decentralized AI ecosystem where compute shifts closer to users, reducing reliance on remote servers and potentially disrupting an industry built on massive infrastructure investments.

On-Device AI Gains Momentum Through Small Language Models

The case for on-device AI has strengthened considerably as small language models demonstrate increasingly capable performance on personal devices. Paras Chopra, founder of AI lab Lossfunk, observed while testing a 270-million-parameter variant of Gemma that it was "absolutely wild how coherent and fast it is on my phone" 1

1

. Mobile applications such as PocketPal and Google AI Edge Gallery now allow users to download AI models and experiment directly on smartphones, while Google has shipped on-device features across its Pixel lineup that prioritize speed and privacy without relying on the cloud.

Research institute Epoch AI stated in a recent report that using a top-of-the-line gaming GPU like NVIDIA's RTX 5090 (under $2,500), anyone can locally run AI models matching the absolute frontier of performance from just 6 to 12 months ago 1

1

. This relatively short and consistent lag means advanced AI capabilities are becoming widely accessible for local development and experimentation in under a year. Developers have also experimented with modified versions of powerful open-source models running locally on MacBooks with Apple silicon or on a single consumer GPU, achieving cloud-comparable results for specialized workloads.

Privacy and Personalization Drive the Local Intelligence Advantage

Srinivas has emphasized privacy as a foundational advantage of on-device AI, noting that "all your data lives on your client" and eliminates vulnerabilities inherent in cloud-dependent systems that require ongoing authentication 4

4

. He envisions AI models adapting to users through test-time training, observing repeated tasks, retrieving local data on-the-fly, and automating workflows while keeping everything private. "It adapts to you and over time starts automating a lot of the things you do. That's your intelligence. You own it. It's your brain," Srinivas explained 3

3

.

Gavin Baker, CIO and managing partner at Atreides Capital, echoed this view, imagining a future where smartphones house more memory modules to accommodate pruned versions of frontier AI models 1

1

. He pointed to Apple's strategy, focused heavily on on-device, privacy-first AI rather than relying on powerful cloud-based models. Srinivas noted that Apple has "a massive advantage" due to its M1 chips and power-efficient devices 3

3

. Qualcomm and original equipment manufacturers including Samsung, Lenovo, and HP could also benefit from distributing AI-enabled devices with specialized chips designed for efficient inference.

Technical Barriers and the Hybrid Future of AI Compute Infrastructure

Despite the promise of on-device AI, technical barriers and performance trade-offs remain significant. Srinivas acknowledged that no AI model has yet been released that can run efficiently on a local chip while completing tasks reliably 3

3

. Minh Do, co-founder at Machine Cinema, framed the trade-off succinctly: "You wouldn't expect a poorly performing AI but a cheaper AI if the expensive one can accurately diagnose your grandmother or get all your math problems right" 1

1

.

Sriram Subramanian, cloud computing analyst and founder of market research firm CloudDon, expects a mixed model where inference is split between the cloud and the device to improve performance, with GPUs remaining "the larger pie definitely" for accuracy and high-demand workloads 1

1

. Rajesh C Subramaniam, founder and CEO of edge AI services company embedUR, explained that "what's changing is where inference makes the most sense," noting that many edge hardware workloads are situational and triggered by on-screen context or real-world interactions that benefit from local processing due to latency, privacy, and cost considerations 1

1

.

At the same time, the cloud remains essential for tasks such as large-scale model training, fleet-level analytics, coordination across devices, and continuous improvement of AI models. Hardware economics also present constraints, with DRAM prices rising and power efficiency remaining a critical concern. The debate intensifies as 2026 approaches: will the industry see a hybrid ecosystem balancing cloud and edge capabilities, or will a genuine pivot to edge AI dominance reshape the economics of billions in infrastructure investments? Srinivas's bold stance suggests the latter could materialize sooner than expected, potentially creating what some analysts warn could be an AI bubble if centralized data centers become a "single point of failure" with widespread economic repercussions 3

3

.