Arm Q&A Session @ COMPUTEX 2025 : Charting the Future of AI

At the heart of COMPUTEX 2025, Arm’s media Q&A offered more than just technical updates—it painted a picture of an AI-powered future where inference, developer ecosystems, and flexible silicon design converge. Without a flashy product reveal this year, Arm instead leaned into substance: the company’s architecture is enabling AI at every layer, from hyperscale data centers to consumer wearables.

Chris Bergey, Senior Vice President and General Manager of Arm’s Client Line of Business, gave his insights and response to all questions.

The Rise of Inference—and the Edge

AI training may dominate today’s infrastructure headlines, but Arm believes the real commercial opportunity lies in inference—especially at the edge. Leonard Lee from NextCurve asked about the shift toward inference workloads and how that impacts Arm’s IP demand. Arm's view is clear: "We’re still in the early innings, but inference is where monetization happens." Whether in smartphones or cloud agents, inference is becoming the engine for scalable AI deployment.

This shift is driving demand for more compute, more memory bandwidth, and more intelligent system balance—a point echoed in Arm's explanation of why hyperscalers like AWS now rely heavily on Arm architecture. Over 50% of compute deployed by AWS in the past two years has been Arm-based, a milestone that reflects deeper architectural shifts across the ecosystem.

GPUs, NPUs, and the Question of Balance

The evolving balance of AI accelerators was another recurring theme. Ian Cutress from More Than Moore raised an incisive question: with GPUs traditionally focused on graphics and NPUs on efficiency, where does Arm see the GPU fitting into future AI workloads?

Arm’s response emphasized pragmatism: while NPUs may deliver exceptional TOPS-per-watt, developers increasingly rely on GPUs for their flexibility and mature software ecosystems. As AI continues to infiltrate graphics workflows—especially in applications like upscaling and super-resolution—the GPU's role as a dual-purpose engine is only growing. “We’re not choosing between accelerating graphics or AI—we’re building GPUs to do both,” Arm explained.

In a nod to this strategy, the company touted the strength of its Mali GPUs and teased upcoming investments in further boosting their capabilities.

Kleidi: Quietly Powering the AI Ecosystem

Few products illustrate Arm’s influence across the AI stack better than Kleidi, the AI acceleration library introduced last year. Since its release, Kleidi has been integrated into Microsoft ONNX, Meta’s ExecuTorch, Google LiteRT, and Tencent’s Hunyuan framework—amassing over 8 billion installs.

Kleidi’s broad adoption speaks not just to performance, but to Arm’s role as an enabler rather than a competitor in the AI space. As Jeff Lin from Omdia queried about Arm’s position in the NVIDIA AI ecosystem, the company made its stance clear: “Our role is to enable through IP. We don’t compete with our partners—we empower them.” From DGX Spark to Grace Blackwell, Arm CPUs underpin some of NVIDIA’s most advanced systems.

From Wearables to Hyperscale: One Architecture to Span Them All

Arm's enduring strength lies in its consistency—from ultra-low power wearables to hyperscale infrastructure. Wen-En from PCADV noted that Arm is perhaps the only architecture offering a complete pipeline from local AI to cloud compute. Arm’s response? “That consistency is exactly what we’ve protected for over 30 years.”

The company emphasized how features like Armv9 security, built-in AI acceleration, and the removal of legacy 32-bit support are shaping a future-proof foundation across devices.

This vision extends to the user experience as well. Arm expects new AI interactions to move beyond touchscreens and keyboards. Think whiteboard conversations, fluid voice interactions, and context-aware interfaces. “We’re not just computing differently—we’re interacting differently,” they said.

Arm in the PC Market: Steady Growth, Not Overnight Disruption

Eric Smith from TechInsights and Kazuki Kasahara from PC Watch turned the conversation toward PCs. In a post-pandemic world, users now expect PCs to deliver mobile-like experiences—silent, all-day performance, seamless AI, and rich multimedia. Arm sees this as a natural advantage for its architecture, rooted in mobile efficiency but now ready for the desktop.

Qualcomm’s claim of 9% market share in the U.S. and EU for Arm-based Copilot+ PCs was cited as a promising start toward Arm's broader 2030 goal: 50% share in client computing. “This is a decades-old industry we’re transforming. It won’t happen overnight—but we’re on track,” said Arm.

CSS: A Platform for Speed, Still in Motion

Some attendees questioned the pace of adoption for Arm’s Compute Subsystems (CSS). Despite claims of accelerated time-to-market, only a subset of partners have reached tape-out so far. Arm acknowledged the variability: different verticals move at different speeds, and more launches are expected by year-end. The feedback loop from CSS development has already helped Arm refine its microarchitecture and better support future partners.

The TOPS Debate: How Much Is Too Much?

As Jeff Lin wrapped up the session with a question on Microsoft's push for 100 TOPS NPUs by 2027, Arm offered a balanced view. Yes, more compute opens the door to more innovation. But it also brings cost and power tradeoffs. “The market will decide where the balance lies. And that’s the beauty of having multiple players,” they said.

Final Thoughts: From Foundations to Flywheels

The session closed on a reflective note. Arm is not just building faster chips—it’s reinforcing an ecosystem that makes AI accessible, flexible, and commercially viable. From silicon IP to developer tools, from edge devices to cloud backbones, Arm’s role is not to dominate but to enable.

And that may just be the most sustainable AI strategy of all.