The rapid expansion of artificial intelligence (AI) data centers, driven by large language models requiring substantial computing power and energy, is poised to transition into a new phase marked by greater efficiency and decentralization. While the exact timing remains uncertain, experts suggest the era of costly, resource-intensive AI operations will give way to more affordable, compact models running on personal devices.

Currently, AI development relies heavily on powerful GPUs, such as those produced by Nvidia, housed in liquid-cooled data centers that consume gigawatts of electricity. This approach, often dubbed "brute force," echoes earlier phases in computing history, reminiscent of the mainframe era which eventually yielded to smaller, cheaper technologies. Industry observers caution against extrapolating today's costs and performance as permanent, emphasizing that technological advancements typically undergo refinement cycles that lead to significant drops in expense and resource demands.

Venture capitalist Steve Jurvetson highlights a common pattern in technology evolution, referring to a “factor of 10” shift—where new approaches become at least 10 times more efficient, whether through increased performance, reduced cost, or both. Recent market data illustrates this pattern: while major U.S.-based AI providers like OpenAI and Anthropic continue to charge premium rates around $1-$2 per million input tokens, some Chinese models, such as Zai’s GLM and Alibaba’s Qwen, are available for 6 to 12 cents per million tokens—prices nearly 90 percent lower.

This price disparity has contributed to a decline in U.S. AI firms’ market share from 70 percent last year to approximately 30 percent currently. Despite this shift, all parties in the sector are expanding their offerings. Meanwhile, AI architecture itself is evolving toward leaner models. Instead of requiring trillions of parameters, smaller language models with tens of billions of parameters are achieving comparable performance. Research from Stanford University indicates that a model with around 70 billion parameters, capable of running locally on a device such as an Apple Mac, can answer nearly 89 percent of certain chat and reasoning queries effectively.

The deeper transformation lies in the migration of AI capabilities “to the edge,” meaning increasingly powerful AI models running on smartphones and personal computers rather than centralized data centers. This trend is gaining momentum: in September, California-based PrismML unveiled a compressed version of the open-source Qwen3.8 27-billion-parameter model reduced to under 6 gigabytes—making it suitable for high-end smartphones and PCs alike, shrinking the memory requirements by nearly a factor of 10.

Hardware manufacturers are advancing in parallel. Qualcomm recently announced its latest Snapdragon neural processing unit (NPU), incorporating a “mixture-of-experts” design capable of operating 30-billion-parameter AI models within just 3 gigabytes of memory. This is a substantial leap compared to previous mobile AI models, such as the four-billion-parameter model running on Apple’s A20 Pro chip, with Qualcomm offering roughly a 7.5-fold increase in performance. Similar advances come from Google’s Tensor chips in Pixel phones and Apple’s M- and A-series processors, which can execute AI tasks locally—a possible reason why Apple has largely refrained from large-scale investments in AI data-center infrastructure.

Startups are also contributing innovations that reduce the compute and cost demands of AI. Reflection AI, backed by Nvidia, claims its technology performs inference on leading Chinese models with three to four times less computational effort. Meanwhile, companies like Perplexity are working to run AI agents on personal computers, eliminating token-based fees and providing users full control over their data—a significant consideration amid growing privacy concerns linked to cloud-based AI services.

As AI continues to evolve, the balance between centralized data models and edge-based solutions is shifting toward decentralization. Cost efficiencies and hardware advancements suggest that the high-cost data-center boom may gradually end as compact, capable AI becomes broadly accessible on everyday devices.