‘The Biggest Threat to Data Centres Is On-Device AI,’ Says Perplexity CEO Aravind Srinivas

Why It Matters

Shifting AI processing to devices could reshape cloud infrastructure economics and boost user privacy, forcing tech giants to rethink massive data‑centre investments.

Key Takeaways

• On-device AI could reduce data‑centre demand
• $1 trillion data‑centre spend projected by 2030
• Local models improve privacy and cut electricity costs
• Apple, Qualcomm chips advancing on‑device AI feasibility
• Hallucinations expected to be solved within five years

Pulse Analysis

The conversation around AI’s future is increasingly moving from massive cloud farms to the edge of the network. As enterprises pour billions into building ever‑larger data centres, analysts forecast a near‑trillion‑dollar market by the decade’s end. Yet the real strategic pivot may come from chip manufacturers who are shrinking model footprints enough to run sophisticated inference locally. By offloading workloads to smartphones, laptops, and IoT devices, companies can dramatically lower power consumption, reduce cooling demands, and sidestep the logistical complexities of scaling centralized infrastructure.

On‑device AI promises tangible benefits beyond cost savings. Privacy‑first architectures keep user data on the device, mitigating regulatory exposure and building consumer trust. Personalized experiences become more responsive, as models adapt in real time without round‑trip latency to distant servers. This shift also democratizes AI capabilities, allowing smaller firms and emerging markets to deploy advanced services without the capital outlay required for data‑centre access. The competitive edge will belong to firms that integrate efficient, secure AI chips into their product lines, leveraging the rapid advancements seen in Apple’s Neural Engine and Qualcomm’s Snapdragon AI Suite.

However, the transition is not without challenges. Current large‑language models still demand computational resources beyond most consumer hardware, and developers must address model compression, quantization, and on‑device training pipelines. Moreover, the industry must solve the lingering issue of AI hallucinations, which Srinivas believes will be eradicated within five years through better grounding techniques and tighter feedback loops. If these hurdles are cleared, the balance of power could shift dramatically, reducing the relevance of traditional data centres and redefining the economics of AI deployment across the globe.