NVIDIA Earnings Transcript: VeraCPU and VeraRubin Expand Agentic AI, 2027 Q1

How management expects VeraCPU and VeraRubin to expand NVIDIA’s share in agentic AI and inference without cannibalizing GPU demand

1) The core framework: CPU work is “orchestration/harness/tools,” GPU work is “thinking/inference”

Management’s explicit division of labor is meant to prevent cannibalization by keeping the GPU responsible for the highest-value parts of agentic systems (token generation), while using CPUs for latency-tolerant coordination and tool execution.

• Management describes the agent “harness” and orchestration/memory/tool orchestration as CPU-run, while “thinking” and spawned sub-agents use GPUs for inference. The harness runs on CPU, and tool use (e.g., “search”/“browser use”) is depicted as running on CPU, while the spawned sub-agents “are thinking” and use GPUs. ¹
• They further state that agents have “lower patience tolerance than humans” and want things “quickly,” prompting NVIDIA to accelerate tools/data-processing engines so they “run on CUDA,” but the overall economics still push toward more tokens being generated fast—Vera is positioned for token-per-dollar economics. ²

Implication for cannibalization: NVIDIA is positioning Vera (CPU) to capture incremental CPU demand for orchestration/tooling across billions of agents, while still keeping inference (token generation) on GPUs—so the CPU enables more inference calls rather than replacing GPU inference. This is reinforced by management’s view that “every time they spin these off, you are gonna need to do inference… that is where the thinking happens… on GPUs.” ³

2) VeraRubin is positioned as a GPU-centric inference expansion lever, not a substitute

Management’s discussion frames VeraRubin as an advanced inference platform that improves throughput and economics, which should increase GPU demand by making AI factories more attractive (and more profitable to run).

• They attribute “agentic AI” arrival to tokens becoming profitable and compute capacity becoming revenue/profits, with NVIDIA described as the “platform of this era.” ⁴⁵
• For VeraRubin, management highlights performance/economics that would deepen GPU usage: “up to 35x higher inference throughput” and “up to 10x greater AI factory revenue compared with Blackwell.” ⁶
• They also claim they are gaining share in inference very quickly, with the reason tied to broader frontier model coverage (e.g., more frontier model companies and adding Anthropic) and then asserting VeraRubin will outperform prior platforms in that inference mix. ⁷

Implication for cannibalization: If VeraRubin increases inference throughput and factory revenue, it should expand total inference deployments and GPU-backed demand; it is not presented as displacing GPUs, but rather as strengthening NVIDIA’s inference “share of wallet.” ⁶⁷

3) VeraCPU captures a new CPU TAM while still complementing (and pairing with) Rubin GPUs

Management explicitly states VeraCPU’s TAM expansion is largely standalone CPU revenue—separately tracked from Rubin—while VeraRubin systems are sold in combination arrangements. That structure is intended to broaden NVIDIA’s monetization across agentic workloads without eliminating GPU requirements.

• Management says the “world’s first CPU purpose built for agentic AI” opens a new “$200 billion TAM” and that it’s paired with major hyperscaler/system maker deployment partnerships. ⁴
• They state visibility to “nearly $20 billion in total CPU revenue” this year, and specify production timing/ramp (“commence production shipments of VeraRubin… in the second half… starting in Q3”). ⁸
• When asked how much of the $20B relates to stand-alone VeraCPU vs being included in VeraRubin, management answers: “The 20 billion is for standalone CPU.” ¹
• They also define four ways Vera is used, starting with VeraRubin (Rubin GPUs connected to Vera), plus other Vera configurations (standalone CPU, Vera with CX9 and software stack for storage, and Vera with CX9 and software stack for security/confidential computing). ¹

Implication for cannibalization: By carving out stand-alone CPU revenue as incremental (not replacing GPU revenue already embedded in Rubin systems), management is effectively arguing Vera monetizes CPU-heavy portions of agentic systems while Rubin continues to monetize inference via GPUs. ¹

4) Why management believes GPUs still win “thinking” (inference) even as CPU orchestration grows

Management provides a behavioral model of agentic systems that naturally increases inference volume—creating a tailwind for GPUs rather than a substitution.

• They forecast “billions of agents” (growing from current levels) where agents spin off sub-agents, and “every time… you are gonna need to do inference”; they explicitly locate “thinking” on GPUs. ³
• They say orchestration runs on CPUs and that sub-agents use GPUs when spawned. ³
• They also indicate work like “simulators” can run on CPUs or GPUs, which suggests flexibility, but the agent “thinking” path is still framed as GPU inference. ³

Implication for cannibalization: Even if CPU load rises (orchestration/tooling), the agent architecture management describes creates more inference events that increase GPU utilization. ³

5) Additional structural support: NVIDIA’s full-stack approach is designed to “optimize around diverse AI workloads”

Management argues that the platform-level approach helps prevent displacement because NVIDIA is not selling CPU-only or GPU-only value; it’s selling an integrated stack that can address diverse workloads with CUDA-based tool acceleration.

• They describe accelerating “all of the world’s tools and data processing engines” so agents’ tool usage “runs on CUDA,” because tools run faster with GPUs already present. ²
• They state Vera “was designed to be an agentic CPU,” motivated by agent economics (“tokens per dollar… tokens per dollar”) rather than classic “dollars per core,” i.e., CPU is being optimized for agentic workload economics rather than competing with GPUs for token generation. ²
• They also emphasize Rubin’s confidential computing platform (VeraRubin as “end to end confidential computing”), which aligns to secure data center environments rather than replacing inference compute. ²

Implication for cannibalization: Integrated acceleration and economics are positioned to increase total “tokens per dollar” performance across CPU+GPU components, maintaining GPU relevance in the token generation path. ²

6) Direct cannibalization question: management’s answer ties to workload roles, not “CPU replacing GPU”

In the Q&A, an investor directly asks whether CPU workloads are “incremental” or “cannibalizing what the GPU would have done otherwise.” Management’s response does not indicate substitution; instead it explains CPU’s role in agent harness/tool execution and provides the standalone-TAM framing.

• Management explicitly frames CPU demand as harness + tool execution (IO/orchestration/memory management/tool use on CPU; tool use such as search/browser running on CPU), while (per their earlier agent model) the thinking/inference uses GPUs. ¹³
• They clarify the $20B figure is standalone CPU, further indicating the CPU opportunity is not merely a re-labelling of GPU revenue. ¹

Bottom line interpretation: Based on these excerpts, management expects VeraCPU and VeraRubin to expand NVIDIA’s share by capturing incremental CPU-side workload for agents (orchestration/tool execution) while maintaining (and even increasing) GPU-side inference demand driven by more agent/sub-agent inference calls and by Rubin’s inference throughput improvements. ¹³⁶

7) What must be true for this non-cannibalization thesis to hold (as implied by their claims)

While the excerpts don’t provide quantitative cannibalization estimates, they do imply the following conditions:

• GPU inference remains the “thinking” bottleneck for spawned sub-agents (“thinking happens on GPUs”). ³
• VeraCPU improvements map to orchestration speed/economics (agentic harness/tooling on CPU; tokens per dollar as the economics target), not to replacing token generation on GPUs. ²¹
• Rubin’s inference economics and throughput improvements increase deployment demand, reinforcing GPU utilization rather than substituting away from it. ⁶
• Demand growth from frontier model proliferation and cloud/partner compute capacity supports incremental inference capacity creation (share growing in inference “very quickly,” with partnership-driven capacity expansion for Anthropic). ⁷

Overall, management’s expectation is fundamentally architectural: CPU enables agent orchestration and tooling; GPUs handle inference; VeraRubin amplifies inference throughput—together expanding NVIDIA’s share in agentic AI/inference without cannibalizing GPU demand. ³¹⁶