---
title: "NVDA (Trans): Vera CPU alone could deliver $20bn in revenue this year"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/topics/40971574.md"
description: "Below is Dolphin Research's Trans of $NVIDIA(NVDA.US) FY27 Q1 earnings call. For the earnings analysis, please see 'NVIDIA: Rising competition, shifting AI bottlenecks — can even a market darling stumble a bit?'I. Core takeaways from the earnings report 1) Shareholder returns: the company returned a record $20bn to shareholders this quarter. The quarterly dividend was raised to $0.25 per share from $0.01, and an additional $80bn share repurchase authorization was approved."
datetime: "2026-05-21T03:07:09.000Z"
locales:
  - [en](https://longbridge.com/en/topics/40971574.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/40971574.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/40971574.md)
author: "[Dolphin Research](https://longbridge.com/en/news/dolphin.md)"
---

# NVDA (Trans): Vera CPU alone could deliver $20bn in revenue this year

**The following is Dolphin Research's transcript of**$NVIDIA(NVDA.US) **FY27 Q1 earnings call. For our take on the print, see '**[**NVIDIA: Rising rivals, shifting AI bottlenecks — can even a universe stock stumble?**](https://longbridge.cn/en/topics/40969925?channel=SH000001&invite-code=294324&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=4ecec065-48e2-47d0-bee5-f12cdc3e7b41)**'**

**I. Core takeaways from the results**

1) Shareholder returns: A record $20bn was returned this quarter. The quarterly dividend was raised from $0.01 to $0.25 per share. A new $80bn buyback authorization was added, on top of $39bn remaining under the prior plan. The company plans to return Approx. 50% of FY free cash flow to shareholders.

2) Q2 guide: Total revenue is guided to $91bn (+/−2%), with QoQ growth driven mainly by Data Center. GAAP and non-GAAP GPM are 74.9% and 75% (+/−50bps). GAAP and non-GAAP OpEx are Approx. $8.5bn and $8.3bn, respectively.

3) FY guide: FY27 GPM is still expected to stay in the mid-70% range (Approx. 75%). OpEx YoY growth is raised to 45%+, driven by higher R&D and accelerated use of AI tools. GAAP/non-GAAP effective tax rate is guided to 16%–18% (down from 17%–19%) on mix shift by region.

4) Key metrics: Total revenue was $82bn (+85% YoY, +20% QoQ), with a QoQ absolute increase of $13.5bn, a record high. This marks three consecutive quarters of YoY acceleration and 14 straight quarters of QoQ growth. GAAP GPM was 74.9% and non-GAAP GPM 75%, both roughly flat QoQ. Non-GAAP OpEx rose 12% QoQ, and the non-GAAP effective tax rate was 16%, below prior guidance. DSO was 45 days (Q2 expected to normalize to ~55 days). FCF was $49bn (vs. $35bn in Q4), a record.

5) Supply chain and visibility: Supply commitments, including inventory purchase obligations and prepayments, rose to $145bn. **The company reiterated revenue visibility of $1tn for Blackwell + Rubin over CY2025–CY2027 (unchanged)**. **It also added visibility for Approx. $20bn of annual revenue from standalone Vera CPU, opening a new $200bn TAM**. China Data Center compute revenue remains excluded from guidance (H200 export licenses were approved but no revenue yet, with entry into China still uncertain).

**II. Earnings call details**

**2.1 Management highlights**

1) Reporting structure and new segment disclosures. a) The company adopted a new reporting framework with two market platforms: Data Center and Edge. b) **Data Center is split into two sub-markets: Hyperscale (public cloud and the largest consumer internet companies) and ACIE (AI cloud, industrial and enterprise, and sovereign AI)**, **the latter capturing growth in AI-specialized DCs and AI factories**.

c) Edge covers Agent and Physical AI endpoints including PCs, game consoles, workstations, AI RAN base stations, robots, and autos. d) The website now provides a 9-quarter revenue recast under the new framework.

2) Data Center. a) Data Center revenue was $75bn (+92% YoY, +21% QoQ), driven by continuing Blackwell ramp. Demand for GB300 NVL72 is especially strong at frontier model developers and hyperscalers. Both have deployed Blackwell GPUs in the hundreds of thousands, making it the fastest ramp in company history.

b) **By split: Data Center Compute revenue was $60bn (+77% YoY), while Networking was $15bn (nearly 3x YoY)**. c) Hyperscale sub-segment revenue was $38bn (~50% of DC, +12% QoQ). ACIE was $37bn (+31% QoQ), with **AI cloud up over 3x YoY and sovereign AI up 80%+ YoY**. NVIDIA AI infrastructure is now deployed in nearly 40 countries, covering $50tn of GDP, and sites above 10MW nearly doubled in a year to 80+.

d) Networking: Spectrum-X now exceeds all other Ethernet vendors combined. InfiniBand grew over 4x YoY, driven by next-gen XDR. e) Pricing and lease yields: H100 lease rates are up 20% YTD and A100 cloud prices up 15%. Customers can earn attractive returns even past GPU depreciation, reinforcing the AI infra financing ecosystem.

3) AI infrastructure and ecosystem. a) Industry CapEx outlook: Analysts see hyperscaler CapEx topping $1tn by 2027, with annual AI infra spend reaching $3–4tn by decade-end. b) Two drivers of AI penetration: hyperscale workloads (search/ads/recs/content understanding) are migrating from CPU to accelerated GPU architectures, and AI-native products/services (inference, reasoning, Agentic AI) are scaling up.

c) **Blackwell has been adopted by all major hyperscalers, cloud providers, and model developers**: **OpenAI**'s GPT 5.5 is co-designed and trained/deployed with Blackwell and tops the artificial analysis charts. **Microsoft**'s 'Fairwater', the world's most powerful AI DC, went live early with hundreds of thousands of Blackwell GPUs. **AWS** will add over 1mn Blackwell and Rubin GPUs starting this year and is partnering on Spectrum networking. **Google** will offer Blackwell via cloud with confidential computing support. NVIDIA is deepening work with **Anthropic**, expanding capacity via AWS, Azure, CoreWeave, and SpaceX/X.AI. Share in frontier AI continues to rise, with partners including Anthropic, OpenAI, Gemini, xAI, Meta MSL, Microsoft AI, TML, Reflection, Perplexity, Cursor, and others.

d) Performance leadership: On MLPerf Inference, Blackwell Ultra swept all tasks. GB300 improved throughput 2.7x in six months, cutting per-token cost 60%. e) **Vera Rubin roadmap: mass production shipments start in Q3, integrating seven dedicated chips and five accelerated racks**. Versus Blackwell, inference throughput can be up to 35x higher, and AI factory revenue potential up to 10x. Google's A5X bare metal instances can support up to 960k Rubin GPUs across sites.

4) Vera CPU and new TAM. a) **Vera CPU is the company's first CPU purpose-built for Agentic AI**, based on custom Arm cores and co-designed end-to-end with Rubin GPUs and NVLink. It delivers up to 1.5x higher single-thread vs. x86, 2x better energy efficiency, and 4x rack density. b) It opens a new $200bn TAM; **visibility for Approx. $20bn of standalone CPU revenue this year**. The goal is to become one of the world's largest CPU suppliers.

c) **Four Vera modes: paired with Rubin GPUs (1 Vera per 2 Rubin), standalone CPU, Vera+CX-9 for storage stacks, and Vera+CX-9 for security and confidential computing stacks**. Tight supply is expected through Vera Rubin's lifecycle.

5) Edge and Physical AI. a) Edge revenue was $6.4bn (+10% QoQ, +29% YoY). Blackwell workstation demand is strong, while consumer softened slightly on higher memory and system prices. b) Physical AI generated over $9bn in the past 12 months. With Uber, compute support for Robotaxi fleets will extend to nearly 30 cities across four continents by 2028. Leaders in industrial, surgical, and humanoid robotics are developing on NVIDIA's platform.

6) Capital allocation and outlook. a) Priority order: R&D and strategic ecosystem investments, then shareholder returns. NVIDIA must keep delivering the 'lowest cost per token' and 'highest tokens per second' to power customer and ecosystem expansion. b) CEO closed with five points: (1) NVIDIA is the only platform that runs all frontier AI models; (2) the company underpins core data processing, ML, and AI workloads across all hyperscale clouds; (3) with a full-stack AI factory and global ecosystem, it uniquely serves AI-native clouds, sovereign AI, and on-prem enterprise/industrial deployments; (4) CUDA extends to the edge — robotics, autonomous driving, embedded healthcare, AI RAN — and Physical AI will spawn billions of autonomous and robotic systems; (5) Vera is a new growth engine for Agentic AI, opening a $200bn TAM.

**2.2 Q&A**

**Q: What drove the segment realignment this quarter? How do the two DC sub-segments differ competitively, and how do the surprisingly large CPU numbers split across them?**

A: First, a quick correction to Colette's earlier slip — the quarterly dividend was raised from $0.01 to $0.25, not $0.20. That extra $0.05 matters to large holders. On the segments, the goal is to clarify our business. AI and computing are highly diverse: by domain (language, 3D graphics for manufacturing and robotics, proteins in life sciences, small-molecule chemistry, materials science, physics for energy, science labs, higher education), by application (enterprise, energy, manufacturing), by location (hyperscale cloud, AI-native cloud, on-prem enterprise, factory floors, supercomputers, and the edge — autonomous vehicles, robots, and a growing number of compute nodes inside fabs and assembly plants). In future, every base station and wireless network will be AI-driven. Governance can be public cloud or dedicated DCs for regulatory, confidential computing, or national security reasons.

NVIDIA is unique in designing end-to-end, tightly co-optimized, full-stack platforms while staying open. Some customer groups, such as enterprises, want a vendor that integrates everything to buy and use, not to build. So the DC market has many segments, each addressed by an open, full-stack NVIDIA solution.

In simplified form, our business breaks into three: **(1) hyperscale clouds**, where we help accelerate data processing and ML, support internal AI, and bring the NVIDIA ecosystem to their public clouds; **(2) AI-native, on-prem enterprise and industrial, and sovereign AI**, which are growing fast because every industry, country, and company needs AI, and our complete solutions make 'build-it-yourself' viable; **(3) robots/edge** — computing was personal, and will become personal AI, with autonomous cars as a prime example and many other robotic systems, including base-station RANs becoming robotic. The tech stack, OS, operating model, and GTM differ across the three. Hyperscale GTM is simplest (5–6 players), while the other two touch 250k+ companies and require deep vertical AI understanding. NVIDIA's vast accelerated libraries (computational lithography, CFD, particle physics, molecular dynamics, etc.) are crucial to serving the latter two, and the scale now warrants this segmentation to help outsiders see how we operate.

**Q: DC ex-China grew ~120% YoY this quarter and is guided to be near +100% next quarter, while the market sees hyperscaler CapEx up 90%–100% this year. Can you sustainably outgrow hyperscaler CapEx, and will hyperscaler CapEx remain high beyond this year?**

A: Yes, we should outgrow hyperscaler CapEx, for the reasons tied to our segment structure. Conceptually, DC breaks into two big buckets, though it is more nuanced. First are the hyperscalers — the CapEx you cited — roughly $1tn this year, and I expect it to keep rising. Computing is now the revenue engine; without compute, you have no revenue. AI is compute-intensive versus legacy SaaS. That is why frontier AI firms like Anthropic and OpenAI are scaling in months what took SaaS a decade. Hyperscaler CapEx of ~$1tn is tracking toward $3–4tn.

The second bucket includes AI-native clouds (regional and global), AI startups worldwide, 250k enterprises (many building AI factories), and industrials that must colocate compute where data is generated — a chip fab cannot rely on distant cloud response. This also covers sovereign AI clouds. These customers do not want or cannot design chips or assemble systems; they need turnkey systems to buy and operate. We have 5–7 hyperscale customers in the first bucket, but in the second bucket there are hundreds to thousands today and tens of thousands in the future, with rapid growth. Physical AI and the $100tn of industries that IT has barely touched in 30 years fall into this second category and are being transformed by AI.

**Our share in the first bucket is increasing with Anthropic coming on board**, where we will materially expand their capacity in the years ahead. **In the second bucket, few can deliver a platform-level solution like ours** — vertically integrated by design yet modular to customer needs — so our share there is very high.

**Q: As Vera Rubin nears launch and investors focus on inference share, how will Vera Rubin and tight co-design affect your inference share in H2 CY2026–CY2027?**

A: Our inference share is rising quickly. Frontier model developers are increasing in number this year — Cursor, Perplexity, TML, Reflection, and others — and **we added Anthropic this year**, which is expanding extremely fast. We have secured capacity with them on Azure, AWS, CoreWeave, and others, and the capacity coming online for Anthropic across this year and next is substantial. **We had near-zero coverage at Anthropic before; our share there is now climbing sharply**.

Vera Rubin will outperform Grace Blackwell — every frontier model company will start with Vera Rubin from day one, which was not the case in the Blackwell era. So Vera Rubin's starting point is excellent and should surpass Grace Blackwell. The inference share discussion above is largely within hyperscale. In the second category of AI DCs, where we are almost the sole provider, nearly all inference runs on NVIDIA, and in Physical AI we are virtually the only supplier with years of execution. Overall, our inference share is increasing very rapidly.

**Q: Can you share customer feedback on custom merchant parts like CPX and LPX? You once said this could be ~20% of the market. How is LPX progressing, and how does it fit the broader platform strategy?**

A: LPX is designed for low latency and high token rates, but with lower throughput, smaller model capacity, and weaker context handling. Tasks like coding or Agentic workloads require swallowing large context, where LPX struggles. So LPX use cases are not broad; it targets service providers with mixed token services, where high token-rate services are niche but command high premiums. My prior view stands.

I expect LPX and other SRAM-based, decoder-class, high token-rate accelerators to remain 'niche' for quite some time. Grace Blackwell and Vera Rubin support the full AI lifecycle — data processing, pretrain, post-train, RL, and inference — and Grace Blackwell is the best platform today for these jobs. If customers already have high token-rate services, LPX can be layered on to enhance that slice. Whether it is 20% or 10% depends on AI's evolution. Today I see it well below 20%, though one day premium tokens could reach 20%, and we are ready to build this capability with providers. I am confident here.

**Q: There is debate that Agentic apps need lots of CPUs, potentially more CPUs than GPUs. Is this incremental or cannibalizing GPUs? Also, does the $20bn figure for Vera CPU refer to standalone Vera or include Vera Rubin?**

A: **The $20bn refers to standalone CPU**. Vera has four modes: first, Vera Rubin — we will sell millions of Rubins with 1 Vera per 2 Rubins, priced accordingly; second, standalone Vera CPUs; third, Vera with CX-9 and storage software stacks; and fourth, Vera with CX-9 and security, isolation, and confidential computing stacks. I expect Vera to be supply-constrained across the Vera Rubin lifecycle, so the $20bn is for standalone CPUs.

On CPU vs. GPU: an Agent is essentially a harness — e.g., Claude Code wraps the Opus model, Codex wraps GPT 5.5. The harness provides I/O, orchestration, memory management, and tool invocation (browser, compiler, Python interpreter, etc.). Harnesses run on CPUs, and tools are called on CPUs — if AI asks the Agent to search or use a browser, that runs on CPU. There are roughly 1bn human users today; I expect billions of Agents in the future. Agents will use tools — like people using PCs — spawning sub-Agents, where each spawn requires inference; 'thinking' happens on GPUs, while orchestration runs on CPUs. Sub-Agents think on GPUs, and when Agents use simulators, databases, or EDA tools, those run on CPUs or with GPU acceleration.

That is why we work deeply with Cadence, Synopsys, Siemens, Adobe, and others — moving tools, data engines, and DB engines onto CUDA, because Agents tolerate even less latency than humans and demand faster execution. We will need more CPUs, and Vera is designed as an Agentic CPU. Legacy CPUs targeted many cores for renting cores out — cloud economics were dollars per core. Future AI economics are tokens per dollar or dollars per token, needing fast token generation and processing — Vera's sweet spot. Ultimately, we build AI infrastructure: top-tier storage (hence STX), networking (Spectrum-X), GPU and inference (NVLink 72), security and confidential computing (Vera Rubin is the first end-to-end confidential computing platform), and top-tier CPUs. We cover them all.

**Q: For the new segments, should Neo Cloud (AI new clouds) sit in hyperscale or ACIE? With both segments now similar in size, do you expect ACIE to outgrow hyperscale or grow in line?**

A: You are right — **AI-native clouds belong in the second category (ACIE)**. They do not design chips and cannot stitch components into an AI factory. They require ultra-low first-token latency and an architecture that runs all models for all customers with high offtake. NVIDIA's architecture fits perfectly: we supply the full bill of materials, with any gaps filled by ecosystem partners, fully integrated.

They can rent capacity to nearly every AI startup, SaaS, enterprise, and industrial, making it the easiest-to-rent, best-TCO, and most financeable platform. These traits match AI-native cloud needs. They resemble OEMs and large enterprises, so we put them in the second bucket. **The second bucket lagged the first** because hyperscalers have the strongest CS and DC capabilities and focus on consumer apps — where accuracy tolerance is higher and UX gains suffice. Industrial and enterprise apps require AI to be capable, safe, and economically accretive before mass adoption, so the second bucket grew more slowly — which the data show.

Over the long run, **industrial and enterprise are the main battlegrounds for the real economy** — they represent $50–80tn of GDP and will grow with AI. Thus, ACIE will become larger than hyperscale over a long enough horizon. Near term (the next several years), both will grow rapidly, but ACIE likely grows faster. I also hope Physical AI and robotics — the third bucket — enter a high-growth phase within five years.

**Q: At GTC you gave $1tn revenue visibility for Blackwell + Rubin, seemingly excluding LPX, Rubin CPX, and Vera CPU. Will Vera CPU be the largest upside beyond the $1tn, and what else could expand the TAM?**

A: Beyond the $1tn, in order of priority: first, continued share gains with frontier AI customers — I expect further increases. Second, **the $1tn does not include standalone Vera CPU**, which I see as the second-largest upside. Agentic systems have a very large TAM, customers are keen on Vera, and we will ship a lot of Veras. Third is LPX — as noted, LPX has ultra-low latency and great interactivity on SRAM but limited throughput and context. Together with Vera Rubin and LPX, we cover the full AI workload spectrum from pretrain and post-train to inference and Agentic systems.

**Q: Colette said in prepared remarks that GB300 is the fastest product ramp in company history. Vera Rubin is rack-scale with new silicon — will its ramp mirror GB300 or be more gradual?**

A: We have said multiple times that Vera Rubin will debut in H2, with first integrations starting in Q3 and a continuing ramp in Q4. It is hard to say now if it will be faster or slower than GB300, but demand exists and orders are in place, and nearly all major customers are ready. These are highly complex systems, with coordination across sub-system mass production timelines. So it is early to conclude, but **we begin in Q3, continue in Q4, and the scale in Q1 next year will also be very large**.

<End of text\>

**Risk disclosure and statements:**[**Dolphin Research Disclaimer and General Disclosures**](https://support.longbridge.global/topics/misc/dolphin-disclaimer)

### Related Stocks

- [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md)

## Comments (3)

- **空空郎君 · 2026-05-22T02:33:48.000Z**: NVIDIA acquired Groq a while ago, spending $200 billion, and made it all back in just half a year? 🤓🤓🤓
  - **北海狮子** (2026-05-22T14:15:15.000Z): The stock price is beyond saving, this stock really shouldn't be touched.
  - **空空郎君** (2026-05-22T14:15:59.000Z): What stock are you talking about?