Sponsored by
Arm published a framework for measuring central processor work around AI agents. It published performance results for a new server chip at the same time. An agent's visible answer is only one stage. Many tasks require work before and after it. Software must retrieve information, maintain state, call tools, run code, and check outcomes. Those steps place work on server CPUs. Accelerators may still perform the main model calculations.
10 AI Stocks to Lead the Next Decade
AI isn’t a tech trend – it’s a full-blown, multi-trillion dollar race, and 10 companies are already pulling ahead.
These are the innovators driving real revenue, attracting institutional attention, and positioning for massive growth.
Get all 10 tickers in The 10 Best AI Stocks to Own in 2026, free today.
The framework proposes measuring both defined components and complete tasks. Component tests can isolate work such as orchestration, retrieval, scheduling, or tool execution. End-to-end tests ask whether component gains change task outcomes. That connects processor results with a user-visible measure. Arm says representative use cases and repeatable measurements should anchor the approach. Its white paper groups coding, enterprise operations, research, personal assistants, and robotics into five workload archetypes. The paper calls this a framework for industry agreement, not a finished benchmark suite or single score.
That distinction addresses a measurement problem. Existing processor tests describe specific workloads. Full agent tasks mix model calls, network delays, software choices, and CPU activity. A single elapsed-time result may hide which layer caused a change. An isolated processor test may say little about the task a user sees. Arm proposes combining both views. Replayable task traces are one possible way to reduce variation between runs. The framework also separates CPU component results from model and network effects.
200+ Proven Ways to Make Money With AI in 2026
The next wave of millionaires will be people who figured out how to make AI work for them.
The window to get ahead is still open. But not for long.
Here are 200+ proven ways to make money with AI in 2026.
Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.
Signal65's separate report compares early Arm AGI CPU systems with AMD's EPYC 9755 across seven workloads. It reports an average of 1.99 times the modeled performance per rack. The report says this estimate uses measured server-level results and a rack model. Individual workload results range from 1.43 to 3.19 times. Group averages differ: 1.7 times for tooling, 2.0 times for agent workflows and orchestration, and 2.2 times for reasoning and planning. These are reported test results, not a general ranking of every server system.
The rack figures are modeled using Arm's rated server power and measured per-chip results. Under a 24-kilowatt rack assumption, the report models 21 to 23 Arm servers against 13 EPYC 9755 servers. Signal65 says its systems were early units. It plans follow-up testing on final production hardware. The report states that Signal65 worked with Arm and identifies Arm as the commissioning party. That relationship does not erase the measurements. It is important context when weighing a vendor-linked comparison.
The processor comparisons use different AMD reference chips for different tests. The seven-workload rack average is against EPYC 9755. A separate sandbox-start result compares with EPYC 9965. Combining those benchmarks into one comparison would misstate the report. The sandbox figure concerns startup latency in named runtimes. It does not measure whole-agent quality or task accuracy.
The broader point is architectural. An agent can place repeated demands on CPUs through retrieval, software tools, data movement, and isolated execution environments. These operations can overlap, so concurrency matters alongside isolated speed. Whether a faster CPU improves service depends on the full stack, including models, storage, networks, software, and power limits. Arm's proposal recognizes those boundaries. Its results show one modeled case rather than broad field performance.
For technology teams, the next evidence would be repeatable tests across different agent workloads and independently checked production systems. The proposed framework makes that measurement question explicit. Signal65's early results offer a case study, not a final answer about future data-center design. This research check-in records what has been measured and what remains open.
State of Product reveals what AI still hasn’t solved
80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.
