NVIDIA vs AMD is all the rage—major CPU architecture showdown: who can set the new standard for AI agents?

Bank of America said the contest between Nvidia and AMD’s CPU roadmap is, in essence, a power struggle over who can define the industry’s yardsticks in the era of AI agents.
(Backgrounder: Intel: Our AI chips will be cheaper than NVIDIA and AMD, targeting inference-first Crescent Island with air cooling)
(Background add-on: “Madam Buffett”’s IPO day 1 smashed $500 million to scoop up SpaceX! ARK Invest sold shares to raise money, but is still buying)

Table of contents

Toggle

• Two technical routes, two design philosophies
• x86 and ARM: a covert battle over software ecosystems
• With AI Day approaching, AMD gets its response window

Nvidia and AMD are currently fighting over standards-setting power in the $170 billion-scale AI-agent CPU market. With the AI wave of agent-based systems, server CPU architecture roadmaps are diverging. Nvidia argues for faster cores, believing single-core performance determines the system ceiling. AMD, meanwhile, pushes for more cores, saying concurrency throughput is the key.

Last week, Nvidia unveiled its Vera CPU architecture, featuring 88 custom ARM cores, 1.2TB/s memory bandwidth, and a single-chip compute die design. Its logic is that AI agents require repeated interactions between the CPU and GPU, with each loop cycle depending on completion of the prior step—so single-core performance directly determines overall response speed. AMD will release its response at this Thursday’s AI Day, and the market expects it to emphasize the “more cores” route. Currently, EPYC 9965’s rack-level throughput in 100kW deployment scenarios is already 2.4 times that of Vera.

Two technical routes, two design philosophies

Bank of America’s view on this roadmap battle is clear: whoever defines the industry’s next set of measurement standards will win.

Nvidia Vera CPUs differ from conventional server processors that simply “stack more cores.” With 88 cores, they are not especially standout in the server CPU space, but Nvidia is betting on “maximum single-thread execution performance.” Vera’s memory bandwidth is 1.2TB/s, and on-die interconnect bandwidth is 3.4TB/s, with all design choices aimed at improving single-core execution efficiency.

Nvidia’s logic is that agentic AI is different from once-and-for-all large-scale parallel computing; it is more like an iterative loop of repeated CPU-GPU interaction. Steps such as tool calls, code execution, retrieval, orchestration, and the like all depend on the previous step being finished. If single-core performance is insufficient, it will delay the entire agent’s response speed, causing the GPU to wait and ultimately lowering the overall utilization of the AI factory.

AMD’s design philosophy is sharply different. AMD believes production-grade AI is closer to a distributed software platform. Components such as databases, APIs, vector storage, orchestration engines, cache, and middleware execute in parallel. In this scenario, the bottleneck is the number of concurrent workflows the system can carry within a fixed power budget. EPYC 9965’s rack-level throughput in 100kW deployments is 2.4 times that of Vera, and the next-gen EPYC 6 is expected to increase that to 3.3 times.

x86 and ARM: a covert battle over software ecosystems

Beyond the argument over “faster and more” core counts, there is another front that is more hidden but equally important: the choice of instruction set—an ecosystem battle between x86 and ARM.

Nvidia Vera is built on the ARM architecture. Nvidia believes that as long as the microarchitecture is strong enough, instruction-set differences don’t matter. Can ARM handle AI workloads? The answer is yes. Can ARM run enterprise software? If performance comprehensively surpasses x86, the software ecosystem will migrate accordingly.

AMD and Intel clearly hold opposing views. Their argument is direct: as agentic AI expands from model inference into enterprise-grade workflows—covering databases, middleware, security platforms, enterprise applications, and other scenarios—over the past decades these have all been optimized around the x86 architecture. The feasibility of ARM replacing x86 cannot be proven by just a few AI benchmark results.

With AI Day approaching, AMD gets a response window

AMD’s AI 2026 Day on Thursday will be the first public response point in this roadmap battle.

Bank of America expects AMD won’t just compete by speed through benchmark data comparisons, because that would fall into Nvidia’s “single-core performance” evaluation framework. AMD needs to redefine the dimensions of competition—shifting from “single-core performance” to “agent-carrying capacity in real production environments.”

Bank of America believes this is the core of the debate. The deciding factor is not a simple comparison of technical superiority; it is which measurement standards the industry ultimately accepts.

The strength of Bank of America’s report is that it reframes a CPU technology roadmap battle into a fight for industry “standard-setting power.”

The compute demands of agentic AI differ from traditional AI training. Traditional training is dominated by large-scale parallelism, while agentic systems are a hybrid of sequential loops and concurrent scheduling. Which architecture is better depends on how the evaluation standards are defined. If the industry makes “agent response latency” the core metric, Nvidia’s approach is more reasonable. If the industry makes “agent capacity per single machine” the core metric, AMD has the advantage.

Thursday’s AMD AI Day will be an important milestone in this debate. But Bank of America’s implied judgment is that the roadmap contest won’t be settled in the short term. The final winner may not necessarily be the side with higher benchmark scores; rather, it will be the company that can drive the industry to accept the measurement standards it defines.

For investors, the investment value of this debate isn’t about deciding whether one roadmap is right or wrong, but about understanding the strategic bets each of the two companies is making. Nvidia is betting on agentic AI’s extreme sensitivity to latency, while AMD is betting on the core needs of production environments for concurrency density. Both routes could prove viable; the outcome ultimately depends on the real evolution direction of AI applications.

Bank of America assigns a Buy rating to both companies: Nvidia’s target price is $350, and AMD’s target price is $620. This shows Bank of America believes the competition is not a zero-sum game; both companies can grow through differentiated roadmaps, though their growth paths differ.