Bank of America says the split between Nvidia and Advanced Micro Devices on CPU design is really a fight over who gets to define the yardstick for the AI agent era. The bank said the two companies are competing for standard-setting power in what it described as a $170 billion agent CPU market.
As agentic AI gains traction, server CPU architecture is breaking into two distinct camps. Nvidia is arguing for faster cores and says single-core performance sets the ceiling for the whole system. AMD is taking the opposite side, making the case that higher core counts and concurrent throughput matter more.
Two architectures, two ways to measure the market
Nvidia last week introduced its Vera CPU architecture, built with 88 custom ARM cores, 1.2TB/s of memory bandwidth and a single compute-die design. Its thesis is that agentic AI requires repeated back-and-forth interaction between CPUs and GPUs, and each loop depends on the previous step finishing first. In that setup, single-core performance has a direct effect on end-to-end response times.
AMD is due to respond at its Thursday AI 2026 Day event. Bank of America expects the company to lean into a “more cores” strategy. In a 100kW deployment scenario, EPYC 9965 already delivers rack-level throughput that is 2.4 times that of Vera, and the next-generation EPYC 6 is expected to extend that to 3.3 times.
BofA’s view is blunt: the winner may be the company that persuades the industry to adopt its benchmark, not simply the one with the stronger headline specs.
Nvidia is betting on single-thread speed
Vera stands apart from the traditional server CPU approach of stacking up more cores. In server terms, 88 cores is not especially high, but Nvidia is centering the product around maximum single-thread performance. Vera also carries 3.4TB/s of on-die interconnect bandwidth, and BofA said the design choices point in the same direction: improving execution efficiency at the core level.
The logic behind that strategy is tied to how Nvidia sees agentic AI workloads. Unlike one-off massively parallel training tasks, agent systems involve repeated loops between CPUs and GPUs. Tool calls, code execution, retrieval and orchestration each depend on the prior step. If single-core performance lags, overall agent response slows, GPUs sit idle waiting for the next instruction, and utilization across the AI factory drops.
AMD is framing the problem around concurrency
AMD’s design philosophy is very different. The company sees production AI as closer to a distributed software platform, where databases, APIs, vector stores, orchestration engines, cache layers and middleware run in parallel. In that environment, the bottleneck is how many concurrent workflows a system can carry under a fixed power budget.
That is where BofA says AMD’s numbers stand out. EPYC 9965 delivers 2.4 times Vera’s rack-scale throughput in a 100kW deployment, and next-generation EPYC 6 is expected to raise that figure to 3.3 times. The point of comparison, then, is not how quickly one task finishes in isolation, but how many agent workloads the system can sustain at the rack level under real operating constraints.
The x86 versus ARM question is running underneath the debate
Beyond the “faster versus more” core-count argument, the contest also touches instruction sets and software ecosystems, with x86 and ARM emerging as a second front in the battle.
Nvidia’s Vera is based on ARM. Nvidia’s position is that if the microarchitecture is strong enough, the instruction set matters less. By that logic, ARM can support AI workloads, and if performance clearly surpasses x86, enterprise software will follow.
AMD and Intel take a different line. Their argument is that agentic AI is moving beyond model inference and into enterprise workflows, where databases, middleware, security platforms and business applications have been optimized for x86 over decades. On that view, ARM’s ability to replace x86 cannot be established through a handful of AI benchmarks alone.
AMD AI 2026 Day is the first public reply point
AMD’s AI 2026 Day on Thursday is shaping up as the first major public response in this debate.
Bank of America expects AMD to avoid a simple benchmark race on speed, since that would accept Nvidia’s single-core framing. Instead, the bank thinks AMD needs to redefine the contest around real production capacity for agentic workloads.
That is the crux of the report. BofA argues that the outcome will depend less on a pure technical comparison and more on which company gets the market to accept its preferred way of measuring performance.
BofA keeps buy ratings on both companies
The report treats the CPU dispute as a battle over standard-setting power rather than a straightforward hardware showdown. Agentic AI places demands on compute infrastructure that differ from traditional AI training. Training has mainly been about large-scale parallelism. Agentic systems combine sequential loops with concurrent scheduling, so the better architecture depends on the benchmark being used.
If the industry decides that single-agent response latency is the core metric, Nvidia’s case looks stronger. If the benchmark becomes agent capacity per rack, AMD has the clearer edge.
BofA said Thursday’s AI 2026 Day will be an important moment in the argument, but it does not expect the issue to be settled quickly. The eventual winner may not be the company with the better benchmark score, but the one that gets the industry to accept its definition of what should be measured.
For investors, BofA said the key is not choosing one route as right and the other as wrong, but understanding the strategic bets each company is making. Nvidia is betting that agentic AI will be highly sensitive to latency. AMD is betting that production demand will center on concurrency density. Both paths could work, depending on how AI applications develop in practice.
Bank of America rates both stocks a buy. It set a $350 price target on Nvidia and a $620 target on AMD, signaling that it does not see the rivalry as a zero-sum contest and that both companies can still grow through different product strategies.