For the past two years, the spotlight in the tech industry has almost entirely been on GPUs. NVIDIA's soaring stock price and the shortage of CoWoS production capacity led many to believe that the future of computing was limited to parallel acceleration. However, with the evolution of AI technology in 2026, the underlying logic of the entire market is undergoing a dramatic reversal—GPUs are responsible for accelerating "thinking," but all other miscellaneous and coordinating tasks are now being performed entirely by CPUs.
When inference applications began to favor domain-specific architectures such as TPUs and ASICs to avoid the high power consumption and deployment costs of GPUs, another, even bigger turning point emerged: the explosion of AI agent applications and the full arrival of the era of "heterogeneous computing".
The AI Agent wave is revolutionizing hardware demands, with low-power, multi-core CPUs becoming the new favorite.
In pure large-scale language model dialogues, the GPU is the absolute protagonist. However, the workflow of Agentic AI is quite different. It consists of a large group of AI agents that need to coordinate with each other, call external tools, transmit information, and allocate and manage memory. This means that in the operation of AI Agent, as much as 50% to 90% of the latency falls on the CPU workload.
The recent AI Agent craze sparked by OpenClaw (the "lobster" CPU) has directly ignited a market demand for low-power, multi-core CPUs. This not only explains why the Mac mini, which boasts high energy efficiency and uses an Arm architecture, is sold out, but also reflects the current supply shortage of server CPUs. Even facing challenges in production scheduling flexibility, Intel is focusing its resources on...New Intel 18A process productsOn the other hand, they are trying to turn the tide on the server and edge.
In response to this sharp shift in demand, AMD CEO Lisa Su addressed the recent Morgan Stanley Technology, Media and Telecom Conference.franklyThe market is growing much faster than predicted 3 to 6 months ago. She pointed out that even top hyperscalers admit they “seriously underestimated” the massive computing demands of AI infrastructure on CPUs.
Saying goodbye to the "universal chip," AMD sets its tone for heterogeneous computing and deep bonding strategies.
Lisa Su further clarified the current development of AI chips: As the AI arms race enters deeper waters, tech companies' thirst for computing power has shifted from "just having it" to "precision and efficiency." Future AI infrastructure will not have a "one-size-fits-all" chip; heterogeneous computing will be the ultimate solution.
This explains why AMD's Ryzen AI series was able to seize the initiative and actively deploy in the data center sector. Lisa Su revealed that AMD and OpenAI are actively planning their first gigawatt-level computing power deployment, and the next-generation MI450 chip is being "jointly verified" by both companies. Through signing performance-based warrants with Meta and OpenAI, this deep collaboration model, binding the company's equity all the way from the underlying silicon chip architecture, not only secures huge future orders but also builds an ecosystem moat for AMD that is extremely difficult to replace.
However, Lisa Su also warned that this infrastructure boom is accompanied by a substantial increase in memory prices. This will drive up systemic pricing and put greater cost pressure on the personal computer (PC) market, and the market must closely monitor developments in the second half of the year.
The Inevitability of Arm's Journey from Silicon Intellectual Property to "In-house Chips"
With the consensus that heterogeneous computing is becoming mainstream, Arm CEO Rene Haas publicly confirmed in July 2025 that he was considering "developing his own chips," which is no longer just a simple test of the waters, but a precise market strategy.
For a long time, Arm's business model has been to sell architecture and IP licenses without manufacturing the actual chips. However, as demand has changed, simply providing IP can no longer meet system manufacturers' urgent needs for "time to market" and "high-level hardware and software optimization." Arm's strategy of investing in in-house chip manufacturing is not about competing with companies like Qualcomm or MediaTek in the consumer mobile phone market; its core objectives are more focused on the following three dimensions:
• Define distributed computing and chiplet standards:As ArmPredictions for technology trends in 2026The computing architecture is moving entirely towards a "distributed" approach. Arm's self-made chips are most likely reference silicon based on the Chiplet architecture, using physical chips to demonstrate the energy efficiency limits of its architecture in the collaborative operation of AI agent arrays or data centers, allowing ecosystem partners to catch up more quickly.
• Accelerating the integration of physical AI and edge AI:With the intelligent innovations from autonomous driving and industrial robots to wearable devices, Physical AI (physical AI) is emerging.Requires decision-making capabilities with extremely low latency.Arm, through its self-developed chips, can directly demonstrate its hardware and software integration acceleration solutions for these power-sensitive edge scenarios.
• Tailor-made underlying hardware for SLM (Small Language Model):Compared to the massive LLM (Liquid Logic) machines with hundreds of billions of parameters in the cloud, the mainstream in the future will be SLM (Simplified Logic) running on edge devices. Arm's self-made chips will inevitably be deeply optimized for SLM inference and memory bandwidth, seizing the discourse power of edge AI.
The delicate balance after NVIDIA's exit: the starting point for ecosystem reshaping.
It is worth noting that although NVIDIA has confirmed...Sell all Arm shares.However, the 20-year-long technical partnership between the two companies remains strong. This also suggests that Arm's self-developed chip plan will cleverly avoid the massive cloud AI training market dominated by NVIDIA.
Arm's strategy is very shrewd: NVIDIA's GPUs are responsible for heavy training and massive inference in the cloud, while Arm's self-made chips (or ecosystem chips based on this design) are responsible for handling massive AI agent coordination in data centers and real-time responses from SLM and physical AI on end devices.
There is demand and a vision, but Arm also faces the reality that TSMC's advanced process capacity has already been monopolized by major players. In this transition from "single-chip" to "heterogeneous computing," Arm's move is less about transforming into a pure IC design company and more about setting a new performance benchmark for its architecture with physical chips in the new era dominated by AI agents. This ensures that in the golden age of CPU resurgence and heterogeneous computing, Arm's architecture will become the preferred design for supporting millions of agentic workloads.
