TradingKey - The AI infrastructure's center of gravity is shifting, and orchestration between inference calls has become the primary bottleneck in agentic AI systems that link tools, APIs, memory, and multiple agents to accomplish tasks with multiple steps on CPUs.
As a result of this, Arm (ARM) will report its fiscal Q4 2026 FY results on May 6, 2026. Arm investors will want to see evidence that demand for CPUs is structurally higher than it was before Q4 2026 FY and will continue to be into 2026–2028, and that Arm can turn this cycle into revenue, increase its market share, and sustain its margin profile.
During the rise of chatbots from 2023–2025, most of the processing power came from GPUs with very little coming from CPUs. Agentic AI changes the way that load will be distributed. While GPUs still do inference, the majority of time between inference steps will be spent performing orchestration on the CPU side: sending calls to tools, sending calls to APIs, updating memory, synchronizing threads, and coordinating the activities of multiple agents. This orchestration will account for 50%–90% of the end-to-end latency, meaning if the CPUs can’t keep up, that reduces how much the GPUs can be utilized. Therefore, CPU-to-GPU ratios per AI cluster will increase, as will the number of CPU cores per megawatt of power consumed, meaning that CPU racks will be tightly co-located with accelerator racks.
Multi-agent systems (MASs) are expected to further increase the number of tokens generated by each user as APIs and tools are called through complex chains of activity. According to Arm, the increase could be as high as 15 times the number of tokens per user, which serves as an indicator of increased intensity of orchestration and higher CPU cycles being consumed. The signs of demand pressure are already showing up in the x86 supply chain. So far in 2026, average CPU prices have increased between 10 and 15 percent, and a number of press stories have mentioned long lead times (e.g., Intel has a reported backlog of six months, and AMD ships in eight to ten weeks). As such, Arm's performance per watt and its dense and modular nature of compute capabilities have transitioned from being limited to mobility to also being applicable to the new constraints being experienced in the AI data centers.
Arm is the architecture behind the world’s most popular CPU and it has shipped a staggering 350 billion chips through licensees, dominating 99% of the mobile CPU market. That came from heterogeneous designs with RISC-based philosophy that divided work across high-performance cores and efficiency cores to minimize power draw while maintaining throughput. The same playbook has now been implemented in the data center.
Arm cores already reside in noteworthy data center CPUs—from Nvidia’s (NVDA) Grace and Vera to custom CPUs for Amazon (AMZN), Alphabet (GOOGL) (GOOG), and Microsoft (MSFT). Alphabet has claimed its Axion CPUs provide around 65% better price-performance and 60% more energy efficiency when compared to similar x86 chips, and Microsoft’s Cobalt 100 and Amazon’s Graviton4 demonstrate similar efficiency and cost benefits.
With AI clusters tipping to more CPUs, that efficiency premium is going to become a first-order design constraint, rather than a nice-to-have. At Arm’s latest “Arm Everywhere” event, CEO Rene Haas described a road from 30 million CPU cores per gigawatt in today’s AI data centers to as high as 120 million cores per gigawatt to alleviate agentic bottlenecks—an explicit fourfold step-up in core density.
The competition is changing too. Nvidia is marketing Vera CPUs as GPU hosts and as independent agent processing platforms. But Arm is no longer simply an IP licensor in this race. The company unveiled its own rack-scale CPU systems, essentially bringing up the rear in the standalone CPU platform space with air- and liquid-cooled models. This opens a new procurement route to hyperscalers so that they can independently adjust CPU-to-GPU ratios while also reducing ecosystem lock-in.
For the current quarter and forecasted quarters, the outcome should provide concrete milestones regarding Arm's designed-in-house AGI CPU. The signals which drive that conclusion will be through production qualifications/updates with Meta as Customer #1, visibility as to additional hyperscaler wins, and any comment as to order intake pertaining to air- and liquid-cooled racks. The company provided guidance of ~$1 billion in AGI CPU revenue for FY2027–2028 and investors will want initial deployments beyond just forming lighthouse wins, and if supply chains are ready for scale by the latter half of CY26.
There will be much importance as to providing color on the 2nd Generation Tape Out and software ecosystem readiness for agentic orchestration workloads supporting the long-term $15 billion goal.
Arm-based CPUs are now increasingly found alongside custom AI accelerators from hyperscalers, etc. Analysts believe Arm-licensed CPUs will account for between the majority to an overwhelming 90% of custom AI ASIC servers by 2029, with the second half of 2026 being a major inflection point.
Updates on the ratio of Armv9 and CSS-based designs in these custom platforms will be relevant not just for units but the royalty rate uplift. Because CSS can increase effective royalty to a level that is closer to 10% of the chip value, a confirmation of a few more CSS signings results in a disproportionate upside earnings impact relative to just unit growth.
There’s the Windows ecosystem’s turn to Arm-based processors, which is a structural theme separate from the data center. AI PCs are still ramping up to achieve longer battery life, enhanced NPU throughput, and improved collaboration between CPUs, NPUs, and GPUs—all of which align well to Arm’s efficiency-first design philosophy. In the near term, the greatest challenge is how quickly Arm can gain market share with its designs in PC from the x86 incumbent players without sacrificing software compatibility or performance on legacy workloads. Investors should watch for developments in OEM wins, readiness for Windows on Arm, and whether the adoption of Armv9 and continued support for CSS is resulting in higher royalty revenue per device as a means to mitigate smartphone cyclicality.
Arm has a major opportunity for growth; however, it faces an uphill battle from serious competition. Nvidia is pushing its Vera processor as both a host CPU and a standalone computing platform. In addition, Arm is up against already-established x86-based systems in some enterprise and High-Performance Computing (HPC) markets. In addition, several hyperscalers have developed their own in-house Arm-based processors; Amazon has its Graviton, Alphabet has its Axion, and Microsoft has its Cobalt. The large success of these in-house CPU designs adds tremendous validation to the Arm architecture but also impacts Arm's ability to capitalize on a significant increase in adoption of Arm (CSS) and sales growth of high-value products. Also, while the current x86 environment should ease supply constraints in the near term, this will also reduce overall market prices for CPUs. Finally, the most significant risk facing Arm's valuation is that the execution and market opportunity assumed in its valuation will be delayed due to the slow ramp of AGI or the slow adoption of Armv9 CSS or the slow transition of PCs to new technology, which will likely negatively impact Arm's stock price due to lack of investor support for its multiple.