NVIDIA and AMD are competing over the core metrics for agentic AI server CPUs

Overview

This report discusses the competition between NVIDIA and AMD over server CPU architectures and evaluation metrics in the era of agentic AI. With the Vera CPU, NVIDIA emphasizes single-thread performance, low latency, memory responsiveness, and GPU utilization; AMD emphasizes EPYC's ability to support more concurrent workflows within a fixed-power rack. BofA summarizes this debate as a contest between "faster cores" and "more cores" and describes the server CPU market opportunity as potentially expanding to about USD 170 billion by 2030.

Core views

The core view is that the key KPI for future AI infrastructure CPUs has not yet been fully determined: if agentic AI is primarily constrained by the completion time of a single agent task, NVIDIA's low latency, single-thread performance, and system-level co-design are more advantageous; if it is primarily constrained by the number of agents that can run concurrently per rack, AMD's framework of more cores, throughput, and service density is more compelling. The report expects AMD's upcoming AI 2026 Day may not be just a benchmark comparison, but more likely an attempt to define new metrics preferred by the industry.

Analysis framework

The report uses an approach combining architectural comparison and workload decomposition, breaking agentic AI into CPU and GPU loops such as tool calling, code execution, retrieval, and orchestration, and comparing NVIDIA Vera's monolithic compute die and ARM ecosystem path with AMD EPYC's chiplet architecture, x86 software compatibility, and concurrent throughput thesis. The analysis also extends to the software ecosystem debate between x86 and ARM, arguing that long-term optimization of enterprise databases, middleware, security platforms, and application stacks may influence CPU selection.

Methodology notes

• AI Infrastructure EvaluationSingle-agent completion time vs. agents per rack

  agentic AI CPU KPI

  NVIDIA emphasizes low latency, single-thread progress, and GPU utilization, while AMD emphasizes concurrency, throughput, and service density; the two correspond to different definitions of bottlenecks in AI agent workloads.

• Semiconductor Architecture ComparisonMonolithic die vs. chiplet architecture

  CPU architecture trade-offs

  NVIDIA Vera uses a monolithic compute die and emphasizes scalable coherence, while AMD relies on its proven chiplet architecture to improve concurrent throughput and power efficiency.

• Software Ecosystem AnalysisARM vs x86 ecosystem

  Instruction set and enterprise software compatibility

  The report argues that if microarchitecture delivers better agent performance, ISA may become secondary; however, AMD and Intel can emphasize x86's long-term optimization and validation in enterprise software, databases, middleware, and security platforms.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• NVIDIA Corporation (NVDA.OQ)

  Core covered name; the report maintains a BUY rating and uses Vera CPU as the focus of its agentic AI infrastructure thesis.

  Strengths

  Leader in AI compute and networking markets; Vera emphasizes single-thread low latency, memory responsiveness, and GPU utilization, and can work in concert with system-level components such as Rubin GPU, Spectrum, and BlueField.

  Weaknesses

  Faces factors such as global AI project volatility, gaming cyclicality, power constraints, China's compute export restrictions, and regulatory scrutiny.

  Comparison

  Relative to AMD, NVIDIA emphasizes faster cores and end-to-end AI system co-design rather than simply core count or x86 compatibility.

  Risks

  Weak consumer gaming, intensifying competition in AI and accelerated computing, China shipment restrictions, volatility in enterprise/data center/automotive sales, slowing capital returns, and government scrutiny of its dominant position in AI chips.

• Advanced Micro Devices (AMD)

  Primary comparison name; represents the argument for more cores, concurrent throughput, and the x86 ecosystem.

  Strengths

  The EPYC platform emphasizes higher concurrent workflows under fixed power, with AMD estimating that EPYC 9965 and EPYC 6 can achieve about 2.4x and 3.3x Vera benchmark throughput, respectively, in 100kW deployments.

  Weaknesses

  Gains in AI CPU/GPU share still need to be delivered, while growth in cyclical markets such as PCs, embedded, and game consoles remains slower.

  Comparison

  Relative to NVIDIA, AMD places greater emphasis on throughput, concurrency, and service density when production-grade AI is viewed as a distributed software platform.

  Risks

  Execution of the first rack-scale products, timing and scale of Middle East AI projects, fluctuations in consumer and enterprise spending, outsourced manufacturing dependence, and a maturing game console cycle.

• Intel (INTC)

  Relevant competitor; together with AMD represents the x86 ecosystem and enterprise software compatibility argument.

  Strengths

  Server CPU and external foundry/packaging opportunities may extend over a longer cycle, and the x86 ecosystem benefits from years of enterprise software optimization and validation.

  Weaknesses

  Manufacturing ramp and yields remain uncertain, the PC market is mature, and it may continue to face CPU share losses.

  Comparison

  Relative to NVIDIA's ARM route, Intel can emphasize x86's software inertia in databases, middleware, security, and enterprise applications.

  Risks

  Intel Foundry 18A/14A yields or ramp may fall short of expectations, lack of major external foundry customers, weaker-than-expected PC trends, and CPU competitors accelerating share gains.

• Arm Holdings (ARM)

  Relevant ecosystem name; NVIDIA Vera uses ARM-based cores, linking the ARM ecosystem to AI data center CPU opportunities.

  Strengths

  Long-term data center content, silicon/chiplet, and AGI CPU opportunities may improve valuation support.

  Weaknesses

  Highly dependent on SoftBank, rising operating expenses, and exposure to the mature smartphone market.

  Comparison

  Relative to the x86 camp, the ARM route benefits from customized AI CPU design, but must contend with shorter validation history and enterprise software stack compatibility issues.

  Risks

  Semiconductor unit cyclicality, exposure to the mature smartphone market, competition from data center x86 and other custom Arm CPUs, RISC-V competition in low-end consumer markets, geopolitics and the Arm China relationship, Qualcomm/Nuvia litigation, and a limited free float.

Key data

• NVIDIA ratingBUYThe report cover page shows the BUY rating is maintained.
• NVIDIA target price350.00 USDThe report gives a Price Objective of 350.00 USD.
• NVIDIA current price207.29 USDThe report cover page and stock data table list the price as 207.29 USD.
• Implied upsideApproximately 68.9%Calculated based on the 350.00 USD target price and the 207.29 USD current price.
• Vera CPU core count88个Olympus ARM-based coresNVIDIA's disclosed Vera CPU architecture parameter.
• Vera memory bandwidth1.2TB/sVera CPU memory bandwidth disclosed in the report.
• Vera on-die interconnect bandwidth3.4TB/sThe report's disclosed on-die fabric bandwidth.
• Server CPU TAMApproximately USD 170 billion, by 2030EThe report states that server CPU TAM could expand to about USD 170bn by 2030E.
• AMD EPYC 9965 throughput estimateAbout 2.4x Vera benchmarkBased on AMD's rack-scale throughput model for 100kW deployments.
• AMD EPYC 6 Venice throughput estimateAbout 3.3x Vera benchmarkThe report cites AMD's forecast for next-generation EPYC 6 Venice.

Impact & implications

For investors, the key to the debate is not simply comparing core counts or instruction sets, but judging whether real agentic AI production environments are truly constrained by latency or by concurrency. If the market accepts NVIDIA's "faster cores" framework, the co-design of Vera with Rubin GPU, Spectrum switches, BlueField storage/NIC, and other components will reinforce NVIDIA's AI platform premium; if the market shifts toward AMD's "more agents/per-rack throughput" framework, AMD's competitiveness in server CPUs and the x86 ecosystem may be re-rated.

Risks

• NVIDIA faces competition in AI and accelerated computing markets, volatility in global AI project sales, power constraints, and China's compute export restrictions.
• There is uncertainty around AMD's execution of rack-scale AI products, customer adoption pace, and outsourced manufacturing dependence.
• If the actual bottleneck in agentic AI is not single-thread latency, NVIDIA Vera's core thesis may be weaker than AMD's concurrent throughput framework.
• If enterprise software compatibility and the importance of the x86 ecosystem rise, adoption of ARM-based CPUs in enterprise AI workflows may be constrained.
• The semiconductor industry remains affected by cycles in PCs, gaming, data center capex, and geopolitics.

What to watch

• Whether AMD AI 2026 Day proposes agentic AI CPU KPIs accepted by the market.
• Which better explains customer purchasing in real production environments: single AI agent completion time or per-rack agent concurrency.
• The system-level synergy of Vera CPU with Rubin GPU, Spectrum switches, BlueField, and other components.
• The actual throughput, power consumption, and software stack performance of EPYC 9965 and EPYC 6 Venice in 100kW deployments.
• How enterprise customers trade off database, middleware, security platform, and application compatibility between ARM and x86.
• Whether the valuation implied by NVIDIA's 350.00 USD target price can continue to be supported by growth in AI compute and networking markets.