"We expect to grow revenue by approximately 70% in fiscal 2028."
"We commenced production shipments of Vera Rubin earlier this month, having already received purchase orders from every major hyperscaler, AI cloud and system OEM, we expect Vera Rubin to mark the fastest product ramp in NVIDIA's history."
"Further, to support the Frontier Labs infrastructure build-outs, we recently announced partnerships with 6 of the world's leading infrastructure capital providers, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, to establish financing platforms that will raise over $500 billion of third-party capital."
Analyst Q&A 8
70% year-ahead guidance vs demand
Q (Joseph Moore · Morgan Stanley): They asked why the company guided 70% for the year when demand appears much higher, what is constraining the gap to 100% demand, and whether that gap can close over time.
A: Management said AI agents drive enormous compute demand, far above humans, and much of the market (sovereign AIs, NeoClouds, enterprise) is growing fast. They said supply-chain constraints (land, power, cooling, labor) take years to align, but NVIDIA has greater upstream and downstream visibility. They said they have never guided a year out before, and their supply allows them to confidently deliver 70% while they keep working to increase supply.
"And even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%." — Jensen Huang (CEO)
Inference workloads and market share
Q (Christopher Muse · Cantor Fitzgerald): He asked how evolving agentic AI workloads affect NVIDIA's inference market share over time, and how ACIE and Groq 3 LPX fit into that picture.
A: Management described the AI life cycle as multiple phases and said NVIDIA built NVLink 72 rack-scale systems to be a single fungible system that covers data creation through agentic inference. They said this architecture raises revenue opportunity per gigawatt to about $40 billion and preserves customer investment across many model types. On Groq 3 they said it gives record interactivity and low latency for high-speed token generation as a bolt-on for specific high-ASP services, while most data centers will use Vera Rubin and NVLink 72.
"What it allowed us to do was to create one fungible system that allows us to transition from data creation, data preparation, the pre-training the post-training to agentic inference." — Jensen Huang (CEO)
Drivers of the 70% growth
Q (Stacy Rasgon · Bernstein): She asked which products and markets drive the 70% growth, whether pricing plays a role, and what unconstrained demand would look like.
A: Management said unconstrained demand would be much higher and they will work to get more capacity. They split the market roughly in half: hyperscalers are one half, and the other half is ACIE (enterprise, NeoClouds, sovereign AIs) which needs entire factory platforms and is often invisible to outsiders. They also gave per-gigawatt revenue progression: Hopper ~$18 billion, Grace Blackwell ~$25 billion, and Vera Rubin ~$40 billion, noting customers race to newer, more productive generations.
"The other half of the picture is what we call [ ACIE ] and that's all the enterprise, the NeoClouds, the sovereign AIs, that part of the world is invisible to everybody." — Jensen Huang (CEO)
Ecosystem investments and rivals' custom chips
Q (Vivek Arya · BofA): He asked whether the roughly $500 billion in commentary captures ecosystem investments, whether more equity investments are possible, any fiscal '28 cash specifics, and how NVIDIA balances investing while some partners build custom chips.
A: Management said NVIDIA is a full AI factory platform used across the AI life cycle and worldwide, and they expect frontier labs to run on NVIDIA broadly because of its economics and breadth. They called investments in AI labs once-in-a-generation opportunities and said they regret not investing more; they expect these lab partners to be customers and partners for a long time. The CFO added that the biggest parts of the commitments are supply commitments concentrated in the first 3 years and will be used to build needed products, which supports confidence in revenue growth.
"You can see that those commitments, the biggest parts of them are in the first 3 years, and we will use that to build the products that we need." — Colette Kress (CFO)
Impact of open-source models
Q (Timothy Arcuri · UBS): He asked whether the rise of open models is positive or negative for NVIDIA, given frontier model companies drive demand.
A: Management said the world needs both closed and open models, and both are growing rapidly. They said nearly all open models run on NVIDIA because of its global footprint and fungible architecture, and that open models enable startups and enterprises to build proprietary AI. They concluded that success of any model type is good for NVIDIA since their platform runs every frontier model.
"Nearly all open models run on NVIDIA, and the reason for that is because NVIDIA's footprint around the world is the highest, and our architecture is the most fungible." — Jensen Huang (CEO)
Recursive improvement, AGI, and demand
Q (Benjamin Reitzes · Melius): He asked how recursive self-improvement and possible AGI milestones would affect industry demand and NVIDIA's role.
A: Management said demand will inflect further as systems become agentic; he stated most AI are now agentic and future companies will run many continuously running agents. He described agents running 24/7 (for example on DGX stations), improving themselves between runs, and said AI is already doing productive work and generating profitable tokens — more compute would generate more profitable tokens and more profit for services.
"When the world goes to agent, fully agentic systems, you're going to have agents running all the time, working with other agents running all the time." — Jensen Huang (CEO)
Ranking supply constraints
Q (James Schneider · Goldman Sachs): He asked which constraints are most acute (data center power/shell, DRAM, foundries, etc.) and to rank them.
A: Management said the entire supply chain is challenged and running flat out. They said more capacity is coming online continuously, and they will work on yield improvements and with all suppliers to increase capacity. They noted they currently have roughly supply for 70% while demand is much higher and they need the supply chain's help to avoid disappointing customers.
"I think the answer is our entire supply chain is challenged." — Jensen Huang (CEO)
Scaling revenue per gigawatt and capacity
Q (Aaron Rakers · Wells Fargo): He asked whether revenue per gigawatt can scale from 40 to 60 or 80 and whether capacity deployment scales linearly or if something unlocks supply.
A: Management said the goal is to pack more compute per plot of land, ideally much more, and that the trend is toward increasing revenue per gigawatt each generation. They gave a historical progression: general-purpose computing at roughly $3–$5 billion per gigawatt, Hopper ~$18 billion, Grace Blackwell ~$25 billion, Vera Rubin ~$40 billion, and said it will go higher after that. They added that this productivity makes data center investments attractive because returns can be very fast.
"If we could literally get $1 trillion of compute into 1 gigawatt and 1 piece of land power shelf, it would be a fantastic outcome." — Jensen Huang (CEO)
May 20, 2026 earnings call — Blackwell ramp and Vera CPU launch
NVIDIA reported record quarterly revenue of $82 billion, driven by strong data center demand and Blackwell systems.
"Driven by sustained strength in our Blackwell architecture. And demand for GB300 and NVL72 was particularly strong with frontier model builders and hyperscalers each having cumulatively deployed hundreds and thousands of Blackwell GPUs."
"Vera opens a brand new $200 billion TAM for NVIDIA, a market we have never addressed before."
"We are increasing our quarterly dividend from $0.01 to $0.20 per share."
Analyst Q&A 8
Why the company changed segments
Q (Joseph Moore · Morgan Stanley): The analyst asked why management changed the segmentation and the philosophy behind it. He also asked about competitive differences between the two segments and the surprising CPU number across them.
A: Management said AI and computing are very diverse in applications, locations, and governance, so they split the business to make it easier to understand. They described three main segments: hyperscale clouds, AI natives/enterprise/sovereign, and the robotic edge, and said NVIDIA builds full-stack solutions that can be integrated into all these environments.
"So that is the reason why we broke it all apart this way." — Jensen Huang
Grow faster than hyperscaler CapEx?
Q (Ben Reitzes · Melius): The analyst asked if NVIDIA's goal is to grow faster than hyperscaler CapEx and whether hyperscaler CapEx will keep growing rapidly. He asked if management still endorses that view.
A: Management said NVIDIA should grow faster than hyperscaler CapEx and explained the data center business has two large parts: hyperscalers and a diverse second category of AI native and sovereign clouds. He said hyperscaler CapEx is large and growing, and the second category consists of hundreds to thousands of companies and is growing very fast, where NVIDIA is well positioned.
"So first of all, we should be growing faster than hyperscale CapEx." — Jensen Huang
Vera Rubin's impact on inference share
Q (Christopher Muse · Cantor Fitzgerald): The analyst asked how Vera Rubin and extreme co-engineering will affect NVIDIA's inference market share into late 2026 and 2027. He asked about coverage of frontier models and optimization for diverse workloads.
A: Management said they are gaining inference share quickly, helped by more frontier model companies and new partners like Anthropic for whom NVIDIA is securing capacity across multiple cloud providers. He said VeraRubin will be adopted broadly by frontier model companies and expects it to be even more successful than Grace Blackwell, while also noting that NVIDIA already dominates inference in the fragmented AI data-center segment and in physical AI.
"We are growing share in inference very, very quickly." — Jensen Huang
LPX traction and role in platform
Q (Timothy Arcuri · UBS): The analyst asked about traction for custom merchant products like CPX and LPX and how LPX fits NVIDIA's broader platform strategy, including the 20% market comment.
A: Management said LPX is designed for low latency and very high token rate but has low throughput and limited context capacity, so its use cases are narrow and it will be a niche product for some time. He explained LPX can be added by service providers who already offer high token-rate premium services, while Grace Blackwell and VeraRubin cover the full AI lifecycle.
"The LPX is designed for low latency and high token rate." — Jensen Huang
Role of CPUs for agents and $20B
Q (Vivek Arya · BofA): The analyst asked whether CPU workloads for agentic applications are incremental or cannibalistic versus GPUs, and whether the $20 billion figure is for standalone Vera CPUs or part of VeraRubin.
A: Management said the $20 billion number is for standalone CPUs and that Vera is used in multiple ways (VeraRubin, standalone Vera, Vera with CX9 for storage, and Vera with CX9 for security). He explained agents use CPUs for orchestration and tool use while GPUs handle the heavy inference; Vera was designed as an agentic CPU optimized for token-based economics.
"The CPUs of the past were designed to have many cores, so that it could be easily rentable." — Jensen Huang
Where AI-native (neo) clouds belong
Q (Stacy Rasgon · Bernstein): The analyst asked whether AI-native (neo) clouds belong in hyperscale or the AI cloud segment and asked about their magnitude and growth compared with hyperscalers.
A: Management confirmed AI-native clouds are in the second category (AI native/enterprise/sovereign), saying they do not design chips and need a fully integrated, rentable architecture that NVIDIA provides. He said this segment started after hyperscale, but long term it should grow faster and become very large, and he also expects physical AI and robotics to grow quickly.
"Because it represents some you know, $50 trillion to $80 trillion of the world's economy." — Jensen Huang
Vera CPUs as upside above $1 trillion
Q (James Schneider · Goldman Sachs): The analyst asked whether Vera standalone CPUs will be the biggest source of upside above the previously stated $1 trillion platform revenue, and whether other product combinations could increase share.
A: Management said they did not include any standalone Vera CPUs in the prior $1 trillion figure and he expects standalone Vera CPUs to be the second largest incremental opportunity. He added LPX would be another incremental opportunity and that the combination of Vera, VeraRubin, and LPX will address the full AI spectrum from pretraining through agentic inference.
"And so I expect that to be the second largest" — Jensen Huang
VeraRubin ramp timing vs GB-300
Q (TD Cowen): The analyst asked whether the VeraRubin ramp should follow a similar fast slope as the GB-300 ramp or be more gradual given the new silicon and architecture.
A: Colette said they plan to launch VeraRubin in the second half and will start shipments in Q3, with initial assembly and ramp continuing into Q4 and into Q1 of next year. She said they already have demand, purchase orders, and major customers ready, but timing depends on producing and putting together the complex systems, so it's a bit early to say which ramp will be faster.
"We have indicated for a while that we will be launching VeraRubin in the second half." — Colette Kress (CFO)
Feb 25, 2026 earnings call — Record revenue driven by data center Blackwell demand
NVIDIA reported record total revenue of $68 billion and guided Q1 revenue to $78 billion, plus or minus 2%.
"Q4 data center revenue of $62 billion increased 75% year-over-year and 22% sequentially, driven primarily by sustained strength in Blackwell and the Blackwell Ultra ramp."
"We shipped our first Vera Rubin samples to customers earlier this week, and we remain on track to commence production shipments in the second half of the year."
"NVLink 72 has enabled us to deliver generationally 50x more performance per watt."
Analyst Q&A 12
Cloud customers' CapEx growth
Q (Vivek Arya · BofA): Analyst asked how confident NVIDIA is that top cloud customers can keep growing CapEx into next year and whether NVIDIA can still grow if their CapEx stalls.
A: Management said they are confident customers' cash flow will grow because of an inflection in agentic AI driving large compute demand. They argued compute now equals revenues, so more compute translates directly into customer and NVIDIA revenue growth.
"I am certain at this point that we are at the inflection point, we've reached the inflection point and we're generating profitable tokens that are productive for customers and profitable for the cloud service providers." — Jensen Huang (CEO)
Why NVIDIA invests in partners
Q (Joseph Moore · Morgan Stanley): Analyst asked about the role of investments in Anthropic, OpenAI, CoreWeave and others, and how the balance sheet is used to grow NVIDIA's position.
A: Management said NVIDIA's core strength is its ecosystem across clouds, on-prem, edge and robotics, with thousands of AI natives built on their platform. They said investments are strategic and focused on expanding and deepening that ecosystem across the whole stack.
"So our investments are focused very squarely, strategically on expanding and deepening our ecosystem reach." — Jensen Huang (CEO)
Networking and Spectrum-X trends
Q (Harlan Sur · JPMorgan): Analyst asked how the Spectrum run rate is trending now and what management expects as they exit the calendar year, citing Spectrum-XGS and upcoming Spectrum-6.
A: Management said networking is part of NVIDIA's AI infrastructure and that NVLink and their rack-scale approach greatly increased networking needs per rack. They described Spectrum-X as extremely successful and said NVIDIA's networking business is growing fast and becoming the largest in Ethernet.
"The invention of NVLink really turbocharged our networking business." — Jensen Huang (CEO)
Customized chips and architecture roadmap
Q (Christopher Muse · Cantor Fitzgerald): Analyst asked whether NVIDIA will focus more on customized silicon by workload or customer and how the dilate architecture affects that roadmap.
A: Management said they try to avoid extra architecture interfaces unless necessary because each interface adds latency and power. They emphasized architecture compatibility across generations lets software optimizations help older GPUs, and said they will extend the architecture with Grok as an accelerator.
"Architecture compatibility allows us to do that. It allows us to invest enormously in software engineering and optimization, knowing that our entire installed base in the cloud, on-prem everywhere from generations of architectures and GPUs will all benefit." — Jensen Huang (CEO)
Quarterly revenue cadence and gaming supply
Q (Stacy Rasgon · Bernstein): Analyst asked how sequential data center growth should evolve through the year with Rubin ramping, and whether Gaming can grow year-over-year given memory constraints.
A: Management said Blackwell will continue selling while Vera Rubin comes to market and it's too early to know how much Rubin will ramp in the second half, though they expect most customers to buy it. For Gaming, they said supply will be tight for a couple quarters and it is too early to know if year-over-year growth will occur.
"We do expect pretty much every single customer to be purchasing Vera Rubin." — Colette Kress (CFO)
CUDA's role for inference
Q (Atif Malik · Citi): Analyst asked how important CUDA is as more AI investment shifts to inference workloads.
A: Management said CUDA is essential for inference and that NVLink 72 plus their inference stack delivered very large generational gains in performance per watt and per dollar. They argued inference now equals revenues because tokens generated by agents are monetized and data centers are power limited.
"So performance per watt is 50x, performance per dollar, 35x." — Jensen Huang (CEO)
Sustainability of high gross margins
Q (Benjamin Reitzes · Melius): Analyst asked if mid-70s gross margins are sustainable and whether supply visibility into calendar '27 implies sustainability, plus any memory innovations.
A: Management said the key to sustaining gross margins is delivering generational performance per watt and performance per dollar to customers. They said they will keep releasing many new chips and co-designing systems each generation to maintain those advantages and margin levels.
"The single most important lever of our gross margins is actually delivering generational leads to our customers." — Jensen Huang (CEO)
Feasibility of space data centers
Q (Antoine Chkaiban · New Street): Analyst asked about the feasibility, timing and economics of running data centers in space.
A: Management said economics are poor today but should improve. They noted space has abundant energy but different cooling challenges, that Hopper is already in space, and that one strong use case is high-resolution imaging where processing in space avoids sending huge data back to Earth.
"Well, the economics are poor today, but it's going to improve over time." — Jensen Huang (CEO)
Revenue mix: hyperscalers vs others
Q (Mark Lipacis · Evercore): Analyst asked whether non-hyperscaler customers grew faster than hyperscalers and what those customers are doing differently.
A: Management said the top 5 CSPs (hyperscalers) account for about 50% of revenue, while a large and diverse set of other customers — model makers, enterprises, supercomputing, sovereigns — is fast-growing. They said this customer diversity is a major strength and benefits NVIDIA going forward.
"So when you think about our top 5, as we articulated as being our CSPs, our hyperscalers and they have right now sat at about 50% of our total revenue." — Colette Kress (CFO)
Role of Vera CPUs
Q (Aaron Rakers · Wells Fargo): Analyst asked the strategic importance of Vera CPUs and whether they are driven by heterogeneous inference workloads and stand-alone CPU needs.
A: Management said Vera uses different architecture choices, supports LPDDR5, and focuses on very high data processing and single-threaded performance needed in data preprocessing and post-training. They said Vera was specifically designed to be excellent for post-training workloads.
"Vera was designed to be an excellent CPU for post-training." — Jensen Huang (CEO)
Capital allocation and buybacks
Q (Timothy Arcuri · UBS): Analyst asked why NVIDIA does not do a very large share buyback now given strong cash generation.
A: Management said capital return is chosen carefully and supporting the ecosystem — suppliers, capacity, and early developers — is a top priority alongside strategic investments. They also said they continue to repurchase stock and pay a dividend and will look for the right opportunities during the year.
"We will continue to make this a very important part of our process and strategic investments." — Colette Kress (CFO)
Apps driving huge data center CapEx
Q (James Schneider · Goldman Sachs): Analyst asked which application areas (agentic, physical AI, etc.) will drive the long-term data center CapEx inflection and if management still supports the $3 to $4 trillion envelope.
A: Management said future software is token-driven and inference/token generation is central, creating far higher compute demand than before. They identified the current inflection as agentic AI and the next major wave as physical AI — applying agentic systems to manufacturing and robotics — driving large incremental demand.
"The wave that we're seeing now is the agentic AI inflection, and the next inflection beyond that is physical AI, where we take AI and these agentic systems into the physical applications, such as manufacturing, such as robotics." — Jensen Huang (CEO)
Nov 19, 2025 earnings call — Blackwell momentum and Rubin ramp
NVIDIA reported revenue of $57 billion, up 62% year over year, including record sequential revenue growth of $10 billion or 22%.
"We are working on a strategic partnership with OpenAI focused on helping them build and deploy at least 10 gigawatts of AI data centers."
"Looking ahead, to fiscal year twenty twenty-seven, input costs are on the rise but we are working to hold gross margins in the mid-seventies."
Analyst Q&A 8
$500 billion forecast and extra orders
Q (Joseph Moore · Morgan Stanley): He asked if the $500 billion forecast still stands after the quarter and whether the remaining ~$350 billion is still expected, and if there could be upside from more demand.
A: They said they are on track for the $500 billion forecast. They shipped $50 billion this quarter and expect the number to grow because they will likely take more orders. As an example, the KSA agreement is for four to 600,000 more GPUs over three years and Anthropic is also a source of demand.
"We are on track for that as we have finished some of the quarters." — Colette Kress (CFO)
Can supply meet rising AI demand soon?
Q (C.J. Muse · Cantor Fitzgerald): He asked whether supply can realistically catch up with demand in the next 12 to 18 months or if the shortfall could last longer.
A: They said NVIDIA has planned the supply chain closely with partners like TSMC, packagers, memory vendors and system ODMs and were planning for a big year. They emphasized that many AI workloads — pre-training, generative AI, agentic AI — all run on NVIDIA GPUs and are growing together, which drives continued demand. They also noted model quality improvements and many new applications are accelerating adoption.
"NVIDIA Corporation's supply chain basically includes every technology company in the world." — Jensen Huang (CEO)
Revenue per gigawatt and who funds buildouts
Q (Vivek Arya · BofA): He asked what dollars per gigawatt assumptions were used in the $500 billion number and how much of the long-term data center buildout will need vendor financing versus customer or government cash.
A: They said performance per watt improves each generation and gave rough generation efficiency comments: "twenty some odd, twenty to twenty-five" for one generation, "probably 30 to 30, you know, say 30 plus or minus" for Grace Blackwell, and Rubin likely higher. They said performance per watt translates directly to customer revenue, and that hundreds of billions of CapEx for the hyperscalers will be fully cash flow funded, while additional new agentic-AI investments and country- or industry-specific projects will have their own funding sources.
"And in each one of these generations, the speed up is x factors, And therefore their TCO, the customer TCO, improves by x factors, and the most important thing is in the end, you still only have one gigawatt of power." — Jensen Huang (CEO)
Use of large cash flow and ecosystem deals
Q (Ben Reitzes · Melius): He asked how the company plans to use large free cash flow: how much for buybacks versus investing in the ecosystem, and what the criteria are for ecosystem deals like Anthropic or OpenAI.
A: They said a strong balance sheet is needed to secure and plan the supply chain and to support customers and suppliers. They will continue stock buybacks but also invest to expand CUDA and the ecosystem. They described long-term technical partnerships: OpenAI has run on NVIDIA since 2016 and the company invested for deep co-development; Anthropic is being brought onto NVIDIA's architecture and Claude will run on their platform.
"And everything that OpenAI does runs on NVIDIA Corporation today." — Jensen Huang (CEO)
Inference share and Rubin CPX product details
Q (Jim Schneider · Goldman Sachs): He asked where AI inference shipment share might go next year, and asked about Rubin CPX: how big its market could be and what target applications it serves.
A: They said CPX is designed for long-context workloads that must read or absorb lots of data first — PDFs, videos, 3D images — and that CPX has excellent performance per dollar. On inference, they described three scaling laws (pre-training, post-training, and inference) that together make compute needs exponential and said it is hard to know the exact future percentage, but they hope inference becomes a very large part of the market. They also said Grace Blackwell and GB/MB link advances give a multiyear leadership advantage in inference performance.
"CPX is designed for long context type of workload generation." — Jensen Huang (CEO)
Biggest possible bottleneck to growth
Q (Timothy Arcuri · UBS): He asked which single bottleneck — power, financing, memory, foundry, or something else — worries them most as a potential constraint on growth.
A: They said all those items are issues and constraints when growing at this scale, but most are ‘‘attractable’’ and solvable with planning. They noted strong visibility and control over the supply chain and that they have established partners in land, power and financing. They emphasized the architecture delivers the best value per TCO and per watt, and customer demand and platform adoption are increasing.
"Well, these are all issues and they're all constraints." — Jensen Huang (CEO)
Next-year margins and OpEx plans
Q (Stacy Rasgon · Bernstein): He asked what are the biggest cost increases next year, how they will keep gross margin in the mid-seventies (cost cuts vs pre-buys vs pricing), and how to think about OpEx growth.
A: They said they exited the current year in mid-seventies gross margin due to cost improvements and mix, and for next year they are accounting for known input prices and many component complexities. They will work on cost improvements, cycle time, and mix to try to hold gross margins in the mid-seventies, and OpEx will be focused on continued investment in engineering and business teams to innovate software, systems, and hardware. Jensen added they plan and negotiate with the supply chain long in advance and have secured a lot of supply and financial terms with partners.
"So we're taking all of that into account, but we do believe if we look at working again on cost improvements, cycle time, and mix, that we will work to try and hold at our gross margins in the mid-seventies." — Colette Kress (CFO)
Do dedicated AI chips threaten GPUs?
Q (Aaron Rakers · Wells Fargo): He asked whether the role of AI ASICs or dedicated chips has changed and if those designs are now more likely to be deployed versus the GPU architecture.
A: They said modern AI systems are very complex — full racks, multiple switch types, large memory needs and diverse model types — so building a platform involves more than one chip. They argued NVIDIA's strengths are fivefold: they accelerate every phase of AI, are strong at pre-training/post-training/inference, run every AI model, are available across clouds, on-prem and edge, and have diverse offtake to support customers. This description was offered as the reasons the NVIDIA architecture is broadly useful across workloads and clouds.
"You can invest in one architecture, use it across the board, You can use what one architecture, and, not worry about the changes in the workload across those three phases." — Jensen Huang (CEO)
AI-generated notes on the earnings call transcript. Quotes are taken from the transcript. This is not a FolioObs investment view.