Since hitting a 52-week high late last year, ARM shares have fallen almost 40%. This dramatic decline prompted Susquehanna to reassess the company’s valuation in early January, upgrading the stock from neutral to positive and raising its 12-month price target by 15% to reflect renewed confidence in ARM’s ability to capitalize on AI-driven demand for its architecture. The firm cited recent initiative changes—including tighter partnerships with major OEMs and revised licensing terms—as catalysts for upside potential, despite the recent pullback in broader tech sentiment.

At CES 2026, ARM officially unveiled its new Physical AI division, marking the company’s most significant restructuring since its public listing. The division will focus on chip designs optimized for robotics, autonomous systems and edge-based inference, with ARM allocating 20% of its R&D budget—approximately $250 million this fiscal year—toward physical AI projects. Management projects the new unit could contribute up to 10% of total revenue by 2028, assuming industry adoption in sectors such as manufacturing automation and smart logistics accelerates as expected.

ARM remains partially owned by its former parent, and SoftBank’s substantial margin loan backed by ARM shares introduces a risk of forced selling if share prices continue to weaken. SoftBank disclosed in its latest earnings report that roughly $4 billion of debt is secured against ARM stock; a further 20% decline from current levels would trigger covenants requiring additional collateral or partial loan repayment. Analysts warn that a rapid downturn could exacerbate volatility, especially if broader AI or semiconductor sentiment turns negative.

Speaking at the World Economic Forum in Davos, ARM CEO René Haas emphasized that the United States has sufficient power generation to support AI growth, but the real obstacle lies in getting electricity to data centers. He cited permitting delays, aging transmission lines and localized bottlenecks as key constraints, noting that some regions face lead times of 24–36 months for new line approvals. Haas projected that addressing these grid limitations is critical to enable the next wave of AI model training and deployment, forecasting that infrastructure improvements could unlock up to 30% more effective compute capacity by 2027.