Looking ahead: 2025 will be the year of edge AI

Take AI and large-language models, for example. Assumptions about model development are being challenged by Chinese startup DeepSeek, which sent shockwaves through the tech industry and stock markets with the sudden popularity of its chatbot and DeepSeek-R3 large-language model (LLM). 

The quality of responses and performance of DeepSeek approaches or exceeds that of models from companies like Open AI in some benchmarks. However, what is roiling Silicon Valley and Wall Street is the company’s claim of developing its latest model for just $6 million. 

The claim deserves scrutiny given the unclear nature of support from the Chinese government and whether development costs included the cost of all the chips experts suggest are required to train LLM models. Still, DeepSeek’s emergence highlights a crucial shift in how AI models might be developed and deployed, especially at the edge.

The stock market’s dramatic reaction to DeepSeek’s announcement, including Nvidia’s 17% stock price decline on January 27, signals more than just investor jitters. It represents a growing recognition that AI development may not necessarily demand the massive computational resources and budgets that have dominated the narrative. This potential paradigm shift has implications for edge AI, where computational efficiency and cost-effectiveness are paramount. As organizations seek to process more data at the edge rather than in centralized clouds, architectural approaches — including the “mixture of experts” approach used by DeepSeek — could provide a blueprint for developing more resource-conscious generative AI models.

2025 is the year of Edge AI 

Beyond the hype and market volatility, the reality is that a lot of development work on edge AI has been ongoing for years. 

As we navigate through 2025, EdgeIR will keep an eye on what’s possible in edge AI, AI-assisted infrastructure operations, chips and edge hardware that can power new applications (including on-device generative AI). What follows are a few predictions worth tracking.

LF Edge (Linux Foundation):

“AI workloads for vertical industries will drive demand for true AI Edge….Constraints like data privacy, energy and cost will dictate a device-edge-cloud continuum for running AI and data workloads. Edge and Networking open-source projects will be at the forefront of providing the necessary frameworks and connectivity solutions.”

“Easy-to-consume APIs will accelerate the adoption of new services. Projects like CAMARA will enable this new type of monetizing service provider assets. Edge Infrastructure for AI will be another new type of offering, orchestrated and managed by open source technology.”

Vertiv, a provider of digital infrastructure solutions, on collaborating to drive AI Factory development:

“Industry players collaborate to drive AI Factory development: Average rack densities have been increasing steadily over the past few years, but for an industry that supported an average density of 8.2kW in 2020, the predictions of AI Factory racks of 500 to 1000kW or higher soon represent an unprecedented disruption. As a result of the rapid changes, chip developers, customers, power and cooling infrastructure manufacturers, utilities and other industry stakeholders will increasingly partner to develop and support transparent roadmaps to enable AI adoption….In the coming year, chip makers, infrastructure designers and customers will increasingly collaborate and move toward manufacturing partnerships that enable true integration of IT and infrastructure.”

Zella DC, an Australian-based micro data center provider: 

“Edge computing isn’t replacing the cloud — it’s complementing it. In 2025, we’ll see deeper integration between edge and cloud infrastructures, enabling hybrid models that balance centralized and decentralized processing.”

Armada, an edge AI company that provides mobile data centers with compute and internet connectivity:

“Security and IoT breaches are continuing to rapidly evolve and are expected to continue in 2025. If data becomes the ‘new oil,’ companies and countries will want to keep theirs close to the vest. In an increasingly uncertain world, investments in on-prem infrastructure will increasingly serve compliance needs for businesses around the world.”

Code Metal, a startup providing AI-powered development workflows for the edge:

“In 2025, expect a rise in intelligent, edge-centric applications that enhance user productivity. With companies like Intel, AMD, and Qualcomm releasing AI-enabled CPUs for client devices, the term ‘AIPC’ (AI-Powered Client) is gaining traction. While applications like Zoom and Microsoft Office 365 Copilot have started to tap into these capabilities, there’s vast untapped potential across the client ecosystem, including independent software vendors (ISVs). This could lead to a broader range of AI-enabled applications tailored to client devices.”

Conclusion

In the end, the big takeaway from the DeepSeek saga is that the approach of reworking training processes to reduce GPU strain and prioritizing engineering simplicity over computational brute force suggests a future where sophisticated AI capabilities can be delivered through more modest hardware requirements. This efficiency-first mindset, exemplified by the work of companies like Edge Impulse, has already been a hallmark of much of the work in edge AI development, where resource constraints have traditionally limited the complexity of deployable models. What companies stand to gain in 2025 is more focus — and funding — for their efforts.

Article Topics

AI/ML  |  DeepSeek  |  digital infrastructure  |  edge AI  |  edge computing  |  generative AI  |  micro data centers  |  open source