NVDA - Earnings call transcript Q4 2025

Due to licensing restrictions, you must log in to view earnings call transcripts.

Operator

Good afternoon. My name is Krista, and I will be your conference operator today. At this time, I would like to welcome everyone to NVIDIA's fourth quarter earnings call. [Operator Instructions]. Thank you, Stewart Stecker, you may begin your conference.

Stewart Stecker

you. Thank and fiscal conference XXXX. everyone, fourth for to of quarter welcome afternoon, call NVIDIA's the Good Executive Chief With Huang, Executive Kress, President today and President me from Chief Officer; NVIDIA and Officer. Financial are Jensen Vice and Colette

of the call XXXX. call I'd live call being be financial to today's The will discuss on that replay The for for website. webcast be Relations our is available without property. like first Investor until NVIDIA's you fiscal It to or webcast reproduced remind content prior quarter our of NVIDIA's consent. the written is results conference can't transcribed

results During we differ call, number on to materially. statements based a current actual expectations. of significant this uncertainties, may make are and subject may These risks forward-looking our and

could and future release, factors reports and results disclosure For financial business, to and Exchange XX-K of earnings that in may with please on today's Commission. a forms XX-Q discussion affect recent our Form refer X-K the file the the Securities we most that our and

our any of assume XX, obligation information update based on law, statements. us. today, available required no are as to All XXXX, currently we Except to February as made such statements by

reconciliation will financial discuss website. During GAAP this call, we non-GAAP measures. of Confine financial measures our non-GAAP posted CFO our is on a in these financial to measures commentary, which

let call the over turn With me Colette. that, to

Colette Kress

billion of XX% QX of up Stewart. billion. year-on-year outlook Thanks, our $XX.X sequentially was above $XX.X and another Revenue quarter. and and record XX% -- up was

billion, year. up For XXX% fiscal from $XXX.X XXXX prior revenue was the

year. was revenue from $XXX.X center quarter, Data Center. doubling for billion In and billion, prior a year-on-year. of was up $XX.X the than more Center record, XX% fourth XX% Data XXXX fiscal Data the with start sequentially Let's revenue

Hopper and sequential ramp continued the Blackwell growth. commenced XXX As

and in revenue ramp customer sales its our strong our company's We is Blackwell to quickly product unprecedented fastest In history, full delivered in scale. year are is Blackwell adoption. $XX of production QX, Blackwell multiple increasing the expectations. meet in we speed supply configurations, This and billion exceeded demand. across expanding

compute next unlock and over racing the XXX,XXX capabilities. more. cutting-edge XX% Our Xx generation models QX train to to jumped will year-on-year. next clusters already for are started infrastructures of With infrastructure with size. scale level to of or start be and Data Center of for Customers it the Blackwell, this common GPUs Shipments have AI these multiple sequentially revenue

such developers learning to cases. model derivatives for fueling XX,XXX for leverage enterprises customization techniques fine-tuning Llama the training over and software foundation Post model. are as and demand reinforcement hosts Face alone as domain-specific and infrastructure from distillation Hugging use NVIDIA and freighted models tailor

than scale magnitude, orders compute of post-training customization is collectively demand more The massive of can and and model pretraining.

throughput test driven compute Our shot and inference DeepSeek-RX thinking and token inference. experts higher reasoning transformer Hopper up inference. revolutionary It XXx cost XXX. for built and is Blackwell scaling more time task Postgre], X supercharges is models AI like XXXx new LLM to reasoning by AI with accelerating, architected Grok inferences. Long reasoning was of mixture can demand for require lower and reasoning [ XXx AI to Blackwell models versus is compared per OpenAI, engine X.

complexity scale. efficiency ensuring the the response growing the of its X, of infants needed And of Gen tackle delivers throughput NVLink XXx to Domain PCIe cost time, throughput and

STAG across costs. into Companies whole and industries to NVIDIA's tapping platform performance boost are slash inference

LLM. with NVIDIA Server by reduced TensorRT costs inference screenshot and Infant NVIDIA inference XX%, Perplexity XXX using throughput and for cut sees TensorRT feature. its its and X million monthly tripled Now queries costs, Triton

libraries. up at with acceleration NVIDIA, across search billions TensorRT of visual TCO achieved Bing major Xx for and Microsoft a images speed savings

GBXXX across for evoluting AI deployments to early ensuring to obsolete Blackwell the inference inference, a of training from over architecture. first programmable market has for a entire market. in model AI and accelerates against post infrastructure great demand applications, are large cloud, X,XXX new addresses investments pretraining, for to earmarked on-premise, the Many every Blackwell architecture CUDA's inference. enterprise. rapidly

developers Our performance of and lowest And a pace economics. We to is inference our ecosystem, customers' unmatched. NVIDIA in We're full continuously and driven large highest the our and years. delivered for costs last X.X ROI. optimizations the XXXx in just including stack innovation TCO reduction X improve million the

of Large half In GCP, OCI QX, around data to Blackwell first our about these world customer GBXXX the nearly were Azure, for bringing to to center some meet the stand large CSPs up certain and systems and cloud AI. CSPs year-on-year. of represented with surging increased Xx demand regions revenue. sales AWS

cloud Quantum GBXXX and percentage hosting NVLink cluster-based demand rapidly with and rising NVIDIA factory incidents XXX,XXX launched increased Regional as reasoning data continued Switch of revenue, and X models reflecting GPUs We've agents. a globally build-outs a center AI InfiniBand. AI for

revenue grew expanding driven generative Xx cases. language AI. deep systems, include data use AI recommender set and Internet Consumer understanding, of Agentic by These vision, year-on-year, synthetic learning an and generation search

on is the inference serving vast to ads Meta's NVIDIA's of Grace across applications. GBXXX xAI train models. adopting AI Facebook its cutting edge advertising and Grok Andromeda of engine quantities Hopper example, next For generation Instagram, runs Superchip

ad memory Andromeda to Hopper's by large boost throughput and monetization meaningful and enhanced Xx, inference harnesses deliver and personalization in interconnect ROI. fast Grace jumps

on and product AI accelerating applications support, Agentic revenue RAG data fraud including accelerated model and developers increased customer for NIMs NVIDIA GPU AI and to create inventory across We management. introduced Llama year chain and deploy of detection demand Nemotron agents range model supply family fine-tuning, processing. Xx a nearly and help workflows Enterprise

and and generative ServiceNow Agentic with speed and among agent using enhanced Health Illumina, NVIDIA Institute are AI models. Arc platform to first leaders including discovery AI. Pioneer and to care SAP AI new Clinic providers, Mayo Leading use genomic the health drug IQVIA, care are advanced services research

increasingly world, platforms and robotics development. As and AI power to the being AI digital expands beyond NVIDIA adopted are physical software infrastructure

early center, both. data NVIDIA the the virtually largest is applications where or and of autonomous company developing One in robotics car AV the vehicles and every on

automotive vertical grow Group is AV and to it this $X development NVIDIA factory year. fiscal Motor to billion and At adopting revenue robotics initiatives. announced expected accelerate smart NVIDIA's to CES, technologies is Hyundai approximately

platform. transformers, NVIDIA Vision sensor the self high in fidelity simulation supervised World learning, multimodal require development breakthroughs will we Foundation At fusion driving CES, more are and compute. and announced AV model COSMO XXx

ridesharing Just platform. the as robotics. The first to language, Language are Uber, AI is Cosmos companies, physical revolutionized revolutionize AI, have and automotive a models foundation giant among including the adopt to robotics

perspective, demand our infrastructure Data was Countries the across globe by revenue compute growth ramp billion investment their the redefined in Center build-out U.S., the in to the strongest the infrastructure is initial the AI in geographic into AI building From as AI for set XXX XXX EUR sequential surging. a EUR ecosystem invest billion years. and glimpse a EU's global driven AI Blackwell. up coming are offer France's initiatives

solutions of market in roughly China competitive. percentage Center levels at very seen China Data sales the percentage. total change a The export in believe revenue, shipments onset Absent we in center data on any center will that remains remained controls. of regulations, remain as for Now data the current China well below

continue export our to comply with will controls serving while customers. We

Networking sequentially. revenue declined X%

systems robust XX%. GPU at compute Our networking is to attached over

Spectrum-X NVLink major with to revenue represents and growth from InfiniBand, small NVLink and large with NVLink transitioning XX X Switch are vector. increased new We Spectrum-X. a

We in to expect growth return networking to QX.

a new Switch AI systems offers NVIDIA requires compute. class networking. of NVLink for scale-up

huge Ethernet X quantum out, computing are large incentive for environments. Spectrum Spectrum-X Spectrum-X. Microsoft HPC success. CoreWeave For and building for been OCI, we with Ethernet for a and AI AI has factories supercomputers Azure, scale enhances the offer others and

to into networking AI announced Yesterday, infrastructure. enterprise portfolio to will use reach, The industry. every Cisco its footprint Spectrum-X will data and first With their build help integrating Stargate NVIDIA centers bring global large Spectrum-X. Cisco enterprises Ethernet

revenue and demand $X.X $XX.X gaming moving decreased strong year-on-year, the XX% year-on-year. throughout remains and sequentially billion Now XX% to and revenue of Full of increased X% Gaming year holiday. billion ARPCs.

were supply shipments QX by However, impacted constraints.

QX strong expect as increases. supply sequential growth We in

powered desktop Series here. developers fuse GPUs and creators for gamers, laptop fifth computing, The are and the architecture, RTX AI new visual by Build GeForce they redefining graphics XX and Blackwell RT a tops. geometry including and shaders, and performance lighting. cores Xx technologies, leap human new rendering Tensor UQ'sXXXAI cores neuro deliver fourth-generation featuring and GPUs and digital These AI-driven generation

The new generation, unprecedented to frame more to AI-driven frame Xx frame It also up fidelity. visual with models DLSS real-time X X. for packing X and industry's parameters the application of Xx turning first rendered compute into boost rates features Xx transformer

March We also -- that XX%. available technology extends starting with GeForce these top a incredible from life to announced Blackwell world's new GPUs up And be NVIDIA in laptop an battery Max-Q manufacturers. the by will wave laptops of

$X.X include Revenue industry XX% of business. Full sequentially visualization care. year-on-year. increased billion year-on-year. up Key XX% health automotive revenue and Moving year million to professional demand driving X% our of $XXX and was verticals

NVIDIA generative leading are NVIDIA for Technologies technologies workloads. workstations. ANSYS, leveraged and platforms AI design, reshaping simulation in fueling Increasingly, software demand engineering being and are RTX Cadence these Siemens and from

and certified record Strong scale Automotive. including its XX% of the cars up announced autonomous year-on-year. Continental up announced Orin and million, driven Toyota, ramp X% moving NVIDIA deploy powered by growth NVIDIA increased $X.X Thor. At on continued we to build trucks safety vehicles by Now at a maker was Revenue will the world's We $XXX next-generation auto XXX% and Full year-on-year. was vehicles, largest NVIDIA CES, Aurora revenue sequentially driverless billion the year robotaxis. will running DriveOS. Drive in

autonomous of safety platform Rheinland, the authorities has our safety and set the of TUV Hyperion TÜV that Finally, third-party platform for like industry vehicle AV comprehensive passed is and NVIDIA cybersecurity. SUD NVIDIA grade assessments. automotive end-to-end a received first industry's Drive X assessments foremost

Okay.

of architecture. Moving and margin to non-GAAP with our sequentially Blackwell was gross margin XX% XX.X% of P&L. as the first down rest the and GAAP deliveries was the expected gross

for As is with of NVIDIA discussed networking build air several AI last types quarter, multiple chips and infrastructure liquid-cooled different a center. options data customizable and Blackwell

availability, our configurations customers. ramping We Blackwell, exceeded in to QX our expectations providing several increasing system in

low to As gross ramps, we expect Blackwell margins XXs. in be the

Blackwell focused cost the demand of will manufacturing are strong they When gross Blackwell as we and initially, ramped, -- the year. fully have infrastructure. improve this margin to opportunities return improve We meet late race expediting customer to to out to and the mid-XXs, many systems fiscal build we on

X% QX, higher and up to reflecting were billion compute the we operating for new operating dividends. development expenses form product introductions. and XX%, returned expenses repurchases were infrastructure engineering higher costs non-GAAP and shareholders costs share in GAAP and $X.X cash of Sequentially, In

Let first a ramp significant revenue in of with quarter. is plus to to be its -- outlook or we in And $XX Blackwell minus the Total expect billion, Continuing me X%. the demand, turn expected strong QX.

from growth both Center growth and Data expect We Data networking. Center, and we compute sequential in sequential Within both Gaming. expect

gross non-GAAP approximately and and be XX $X.X expected GAAP to to billion, respectively. billion respectively, and points. and plus operating expenses XX%, minus are non-GAAP or basis XX.X% be margins are GAAP expected $X.X

fiscal expect year We in 'XX operating year grow expenses full mid-XXs. to to the be

approximately other any discrete from losses non-GAAP non-GAAP and gains are items. an incoming plus be nonmarketable and rates GAAP equity and GAAP are income be $XXX or tax securities. minus million. and excluding expected XX%, expected held expenses excluding to to of publicly X%,

are financial new in included financial IR information our and information details the other agent. website, AI CFO a available commentary Further including on

upcoming for events closing, community. the me highlight In financial let

Francisco XX. Jose, XX Cowen Health San be on Stanley XX, the in forward starting news-packed join X and Conference us for to look a the financial on X. the our in We events. GTC Media Boston California. these Please Telecom March session next and Conference Q&A TD analysts Jensen Annual Morgan March will will a our March keynote at we for Technology, Monday, on conference San at in seeing March and We for deliver host you will at day, March Care

first discuss XXXX our to of for quarter fiscal the earnings scheduled XXXX. is call May Our results XX, for

questions. I -- operator, we the up are open to call, going to

you start that, If that could would great. be

Operator

[Operator with CJ Fitzgerald. your Muse Cantor first And from question comes Instructions].

Christopher Muse

overall and mean of the potential Jensen, such clusters? this we're NVIDIA [indiscernible] for seeing as between an how and reinforcement and -- to your training increasing customers? think you the for guess the of does do And what inference, potentially lines me, blurring clearly shows learning future compute inference about impact I dedicated promise,

Jensen Huang

are scaling the appreciate scale from that's law, we because now I to multiple C.J. loss. do that data scaling now are and we There's have that continue multimodality, to have pre-training that Yes, to used came reasoning going There pretraining.

And training is train for second post-training than tokens. it's post sensible feedback, models. synthetic enormous or learning, learning amount generated the in reinforcement then kind could, human tokens to And is reinforcement reinforcement feedback, use you using skilling, AI the sense while basically you're data And verifiable learning, of of of generating computation models post-trade. rewards. higher AI synthetically learning are pretraining. that an actually using you The amount that's generate AI reinforcement

part, And search. the of basically that's This is you one-shot the the you've The there chain is needed reasoning, inference beginning. tokens this you one-shot is test beginning. large the that examples beginning. [indiscernible] is already of models the just of more thinking, new the time long the amount and or They're mentioned have And just third all the a same language compute amount part thought, of generated the in XXXx capabilities And ideas. than compute scaling.

simulation-based of search-based compute and that generation even hundreds than today The thousands thousands, in the more our times could thoughtful future. extremely millions is and next have could that models be idea of times and hopefully,

is -- regressive. And architecture? question such it models an the some of design how so you are of do Some the auto

Some based. are models of the diffusion

it of some of want Some the data -- inference. center have to your disaggregated you times

Sometimes is compacted. it

And NVIDIA's data a popular. is reason which of We center, out the every hard architecture why the model. to so run it's what so best configuration is figure is

majority great mind. level. models training, a with The are idea more of all takes you look compute when new many We We reasoning in the Blackwell and vast of performing. at at Blackwell it's designed is that of today actually to times And inference training. our

models tens of faster, AI times thinking But scaling, throughput. test reasoning amazing time is XXx higher for were what's long really

those and And in are your more architecture use allows pretraining so to on center now out board. is you that be to our post have and when or a you inference, going configure now, And fungible across the ways. use your all incredible different to data scaling center Blackwell training is you of doing based easy data

in ever than before. so fact, And concentration of more we're much seeing, much, architecture unified a

from with the Your next comes Joe question JPMorgan. Moore line of

Joseph Moore

Stanley, lot could you the where the And talk prepared we've of you ramp? I complexity you about wonder if then and remarks, said as CES, sort the seen systems you Morgan actually. of -- GBXXX general of are challenges you that rack a the at talked about of in have. terms in level availability

just and the NVLXX you beyond there level the maintained Are bottlenecks to enthusiasm still consider chip level? systems And your at a for above have platforms?

Jensen Huang

that today than CES. a Well, And the because lot I'm was is CES. more for we've more shipped enthusiastic at reason I since

that racks, We plants Blackwell have into some XXX manufacturing the the X.X of each Blackwell racks. components go Grace million one

impatient to complicated. Yes, Grace some $XX of as delivering quite their high, incredibly to customers it's extremely going anxious Blackwell continue quarter. Blackwell, to we last is We're revenues billion and successfully and scale are get have demand And and to ramped up systems.

online number You've probably and web, systems coming of Blackwell seen on we fair the about a them, celebrations of course. have Grace

engineering and for of teams. a our own installation our goes teams and Grace software have own fairly Blackwell We large design

of bring Microsoft open online. the has, many course, now starting has quite successful come about been see theirs. you're and up AI public of has, to CoreWeave

about And is so is the doing, great. I and are our we're we're answer question think of nothing what doing your to great, partners but all easy doing

Operator

next the Bank with Your Securities. Vivek line from question of of Arya comes America

Vivek Arya

they you for in is . is innovations next your wouldn't my has with, confidence that strong if demand QX any DeepSeek gross changed the to sustain And has you And whatever up that mind margins? dashboard then the year? you. if question Jensen, Colette came on view is the can bottom and confirming for way? that What into give

Colette Kress

take the gross the regarding first first question there the part Let of margin. me

soon will point, we At our be margins make manufacturing this gross provide ramp, manufacturing, we XXs. expediting focusing During low that are the in as on expediting our possible. to can our our Blackwell to sure as customers

our is once Blackwell we round. Blackwell cost -- sorry, improve fully I'm does gross rounds, And and it Our margin. once our our fully can

this be probably to expect later we the year. So mid-XXs in

you cooled. customizable water they and in cases. the about have networking systems what speak They They've cool and options. Jensen through their got Walking complexity, are liquid heard multiple some

there and now, focus forward. as opportunity us going So know margins we are on possible. these the an manufacturing we our customers to to to gross going right complete soon as But improve is getting for

Jensen Huang

we fairly investment be several the machine capital out We of based of amount learning. the that line good to that of forward, know things that centers have going of a software building are know We is on majority sight data towards. going vast

and generative your . reasoning AI are you going type computing to so architecture And in center. accelerated be AI, the data want of

AIs, from there we of still that have, that vibrant AI, really AI online in of for it's a And still or know opportunities start-ups need top exciting of number breakthroughs also them whether and course, next quite coming are are innovative, physical amount The are partners. developing each start-ups X of infrastructure. and We reasoning Agentic AIs. new forecast as fair plans our the computing many

GPUs the whether I near signals the to fundamentally and level has fact the be AI-based -- And term the CPUs, it's Midterm the previous to runs and learning with that signals, that. signals on know POs CapEx years. software software compared midterm that from coding like to or near-term has are hand the systems. machine things and so scale-out forecasts course, then would infrastructure that we long-term do changed accelerated of runs signals of computing and signals, on think and And

sovereign some that kind The is it enterprise, out, really as own software. amount wave their to we way tapped about out is AI good AI have early generative that future the AI next of physical roll of and and build AI coming, so another think only we've their sense And AI, and maybe as is Agentic ecosystems. this you fairly a different for consumer then search for software. for of consumer regions robotics the AI of days advertising, recommenders, And

one barely we're we these see them. off We in can much the see are can them different see development ground, happen. of center of places these these And each of activity and will the in obviously, so and because, we can and happening all this great

midterm, So long term. near term,

Operator

line of JPMorgan. with question Your comes from Sur next the Harlan

Harlan Sur

your understand for and -- half Ultra the . supply can next-generation line manage the launch the How half the also the Blackwell simultaneous Your you'll this team's still products? track demand Blackwell current ramping generation customers cadence. to dynamics help in with of be year? Blackwell set Jensen, And this these product annual of that second solutions? still year, us the execute chain do in on the is of Ultra given the to second Ultra in X ramps team is you

Jensen Huang

Yes. is half. second Blackwell Ultra

As an people first had and partners chain us of that of recovering probably a recovered, the light. you speed a fully was did Blackwell amazing The know, we so supply months. just the helped of at all our recover of and us course. hiccup team job We're cost couple many

now we've so successfully of Blackwell. ramped And production

all information, with we'll They laying do everybody our We've all been that and the this have necessary next Blackwell partners all on new that and and is to annual an of processors, and have and the course, online. train of customers, with working coming next The of Ultra But with stop new is work rhythm doesn't of transition. train. memories new proper the networking, out.

It's and lot a Blackwell went going Blackwell between an NVLink to to exactly Blackwell the harder NVLink XX-based the time from system a system because is Ultra, we Hopper same. from system. This X architecture

the had of system, change. was transition. challenging the architecture that So the the the power This a chassis, all to delivery, quite hardware, of

Blackwell to closely We've after been next Ultra partners And that. slot on Vera big, already slot after for and we're called of and the Rubin the the will all again, going will revealed speed the that on is But are partners a our in. all also provide of that very with click right step-up. our working in of And transition and right click huge so that up getting to preparing transition. transition

show to about that. you talk products, to come Really Rubin what so place Ultra, GTC to you to piece. exciting and I'll Vera Blackwell we GTC, so and And after come then new

Operator

question next UBS. with Timothy the Your comes from Arcuri line of

Timothy Arcuri

GPU you these are custom ASIC? building? heterogeneous about balance Jensen, hear to superclusters about a we heard Or distinct infrastructures merchant . speak that customer Can We ASICs. customers some to and on between and something these of use kind ASIC Is will remain of planning GPU. fairly lot the both

Jensen Huang

ways, different areas completely ways. We're models unaggressive whether Well, you've general optimized we some built for we all in different models or text is vision-based several of We're or One, you're or intercept. things models very -- than models great NVIDIA'S in . some models. in or architecture in different diffusion-based ASICs, multimodal it.

-- algorithms. rich our most our sensible, We're because all great target were initial exciting so is is is stack architecture on of it the innovations stack software our that ecosystem so and of software

And much, of curation we're post data really inference all training also way end-to-end data, the in by time scaling. to from the reinforcement tough much to narrow. the definition, to with from processing, We're so the the training, used of good training of learning data, course, more general than the

in not and We a every on-prem. be we're we're cloud, we we're robot. we're could could in general, one because And So in just we're everywhere. be cloud, end-to-end,

great Our starting a anybody up a and architecture much accessible more new initial who's target is target company. for

everywhere. we're so And

fixed not say thing our they're size is is They're these Xx third fast. always per that is anywhere is translates And fixed Remember rhythm or are And Xx, so in the from in that directly fixing to our unusual, which in size. it to performance performance incredibly power. to I our if centers would revenues. Xx data watt

data performance higher. or that for center your throughput gigawatt gigawatt higher, or Xx is that the or you Xx XXX-megawatt center, if the is center Xx a data so XXX-megawatt data And the have in if revenues

past its is than of the data reason tokens the monetizable that And generated. through centers AI are because so directly different is factories

to ROI. that incredibly incredibly revenue capturing the valuable And all for architecture reasons building companies of so the these our fast are so the being of and is generation throughput things just fast token

think And so I third is reason performance. the

that's XXx no the would sits obvious Building amount than software software the top on stack it And complex more an new ASIC architecture is AI then thing growing is I on our that is that the today And architecture the of is different do. the we a say because X of is We world fairly build And that last ecosystem what incredibly hard. advancing was and than ago. top years very is exponentially of building is quickly. architecture.

So bringing that whole ecosystem on top of multiple chips is hard.

built, the into is power decision gets new a business decision in a be comes, a I AI limited reasons. mean and would size, that about gets those chips say say when finally, so the made, There factory again. because I a designed fine. And engine, deploying that is just X and this this, And and deployed. seen has a will are to then you've business lot in processor chip of it over but doesn't new over And time that in

has performance, much our importantly, And technology not deploy lightning advanced, our more to only better is fast. very ability is more it much, and capability software

everybody the enough are of heart, things these so as And now. faint for knows

a And different well, we lot win. we do reasons so there's of why why

Operator

Reitzes comes line question of the Research. with Melius next from Ben Your

Benjamin Reitzes

about U.S. Jensen, regulations towards whether U.S. so the was I $X question. great billion here And can And explaining sequentially. a shift geography-related this geographies. factors there's concern you Your guidance wondering, the fast job up sequentially. can U.S.? or up as U.S. the the slack in pick rate, But other growth it's okay. growing strength. probably some a I we how demand like surge whether up if about keep is think there continues mix and a on underlying if if did underlies -- is with just was go kind throughout your of that's looks be year, it's you the And to China so going the this towards of that

just if maybe dynamic collect could weigh So go and can you wondering that in. . through

Jensen Huang

export was But same approximately is previous what It's percentage. same QX of the before and about it's it control. percentage China the quarters. as approximately half as in the

incredible to modern services services respect takeaway that delivery and is mainstream. With gone it's is shopping in the is everywhere, AI modern AI It's but It's software. geographies, has software modern software, AI everywhere. used software.

If from was you buy a were AI is to milk to you, delivered involved. quarter

almost health Every provides use everything service not student AIs the of so Climate company company care will consumer No a will services as use will. financial tutor, fintech use AI. AI. services AI at use And that Every core tech it. being a company Fintech use and Mineral to higher that AI. university every discovery uses and mainstream AI of -- safe into application. it AI. AI now number say think has gone uses fairly education, I that every The integrated so is every AI, it's

the And hope this to the course, that, in way technology advance new of at safely transition. with advance is -- and and of our helpful that, society. we're I beginning And to a continues we're -- do that believe

software going going the by All is future hand software be And data computing AI. AI. is that will that be decades all be all based to world And decades and world improving be on computers be accelerated, say is, based on I almost to and and infused will learning, forward, it's computers be coding of think future fairly of to to remember, a with that mean so us based of computers forth. the and services that data flywheel for behind been have safe centers has in they've so I and and built purpose built. beginning of what and And CPUs been and part machine general services software on the ultimately, going on will -- the been

And that have taken modernizing decades years in that And computers into out. we're to really journey. X build

beginning of that fairly in the we're And sure so I'm new this era.

opportunity lastly, AI. software address a then technology the world's no has tool has. had ever larger to GDP part of the No And than ever

part software tool the can of larger world's in any time GDP much now is address a that so And this than more a history.

a And step in And of perspective, we be beginning. to we . the is back it has way big about so really just from look and way and the the at small, whether think we're that that. take when in or about the context think you something growth

Operator

with comes of question next from the Your Rakers Aaron line Wells Fargo.

with question ISI. Lipacis next from comes Your Mark Evercore

Mark Lipacis

clarification. I question. Colette a -- Xx faster than the growing does had if up make that year-on-year grew the that so, hyperscalers? a and that And say Did for the would quarter? within you the it the January fast for clarification data enterprise center

for cloud workloads, external question, biggest equipment used. Jensen, being and solutions, your hyperscalers external services is the they are buy both purchasers then, And workflows you, of for that but the internal enterprise

give spend is, a workflows and question enterprises a consumption how splits develop of new expect And of how does as ecosystem. that that that between your to mix? internal? up, So AI sense that your -- and service, applications come you become us and And external can part you these impact the would workload you larger hyperscaler

Colette Kress

Enterprise understand work. where in for with CSPs their very our inference we on and enterprises are working our But similar to were also own seeing areas are you Thanks language can what these mind, in working the that the large on working in Keep keep models, mind, business. be Yes, the Sure. is question large both servicing. with regarding grew CSPs. it to important and to

well both their as are as your own. growing quite well. enterprises building on of with both in They're Your CSPs are terms

Jensen Huang

external our of them used with gear internal because closely them, of And The the We have that to business. infrastructure -- that internal half you consumption. about have CSPs large they work and we're could they optimize consumption say. for to using of. of all advantage NVIDIA workloads a CSPs course, take as are very are of And internal consumption,

the fungible. Spark, be And like one the hand, hand, data used processing processing the fact could we're that we on on other AI for video

the of If also is the then infrastructure longer, the is lower. useful useful so And much much is better. TCO life life our

the for not industry -- term, is part the we industry at is so the And how answer believe, growth the computer by today industrial. you and going if computer forward? you or the reason is is And second far look that if do served largely by is will, CSPs, and it see I because what larger is, of not enterprise long the

we both enterprise, things. example they say use hard When and things soft give me make the because let company let's So example. car an and as you an

car for so we call AI to will and systems exciting at and Agentic make some you more and That's planning things to case the to employees employees software what systems GTC, the guys design to market employees plan tools, agenetic company. we really are And a be and in of build enterprise company, AI. share have productive to Agenetic their with operate

cars. On today, manufacture the also there's that cars, They other that the they AI. fleet entire AI of cars X need cars giant the on billion the trains need this road. treats And hand, system an

data, cars have Someday, and collecting factory. we'll there will billion of be and on AI cars they factory. Whereas road, single be future AI today the a factory X be a all the they'll will and robotic factory our cars, them be and they'll car an every one those an using in improving have

system. And then a the inside car is itself robotic

human computer the the of be there's computers computer a be it it are be mower. build X a And It course, so be and that the you a today. being can that's could the there a involved There's It could lawn could the robot for warehouse. or that could could be people. machineries building, developed tractor, helps be, that AI as a see, could

things common are can't you effects. friction go new they and effect. and to meaning to inertia, AI those These physical and of of those the AI. just cause to understand I, and the all have physical that words we permanence meaning systems physical AIs object require the And languages, They of sense of understand but have of type and but type learn call world,

we So that AI. physical call

of work starting. using lot we on. great the the of then is and now going that. part companies, that's This there's agent and really a hear some way that, people are after systems talking it physical then and And the there AI Agentic we got robotic about era, things revolutionize after inside just to of beginning That the whole you AI AI

And far that providing The computers are which of be my X And represented GDP are representing heavy the will by so -- these that for industrials brand or is long of industries those. and term, by a kind sense makes mall, all is world's companies either this larger new. sense.

Operator

Your next Wells the Aaron question comes Fargo. line Rakers with from of

Aaron Rakers

think how approach some about if now maybe I'm the do of in general. look AI we opportunity. refresh that front of as you GBXXX a you just XXXX And X-year cycle that. see infrastructure the Gen cycle that's think from have replacement when or to Rubin curious where we road map I'm us, the in you deployed been curious how in the really at Jensen, perspective? saw whether anniversary Hopper in it's about we And it's . inflection the start if you

Jensen Huang

it. appreciate I

First are that it reason processing use and the now right and that because And and still so of Pascal are cases data amperes. use one Voltus you things is major CUDA people curation. always is all, of using Blackwell, programmable because could the data is there for

a that not at. an is circumstance model find good You very AI

to model, a circumstance You it's car. let's say, a present language that vision

vision language You present model. that circumstance a to

looks very is actually happened and it. in good the I model this was at language vision what The circumstances, said,

randomization the take find model whole do then an an was. then domain circumstances that go circumstance and You prompt bunch that to AI of and examples. response generate to lake data And whole your whatever and AI go to like use of you a other that, other in prompt you

go then from And the train you bottle. can that,

Hopper data the and to machine And to then training an And so for create go which then you and use you search. could present do you curation data your data peers systems and learning-based processing set, training.

past. less architectures the if in the you -- can these are everything. wants of place, you then onto on they're one the put infrastructure each completely have all But so installed intensive so base CUDA-compatible And and of workloads everything

very of employed. All [indiscernible] well

Operator

for that question, We with Atif question one comes have time and more from Citi. Malik

Atif Malik

Colette. follow-up margins a for have gross question on I

So the understand the to quarter April like earlier per giving would XX about year. of mix. broader you're range parts I quarter kind we to points get there much the the but mid-XXs and have the NVLink XXX And you ramp impact fiscal is still don't know to that And are to to many semiconductor. half Blackwell basis for moving of tariff yields, question, the second tipped bottom,; the Ethernet end

kind gives half back of you trajectory year? the in that this the what of confidence in So

Colette Kress

Yes. Thanks for the question.

a we terms Our how Blackwell complex system, put gross a opportunity pieces the in in of look we . gross margins, quite tremendous improve that over material can on our they're that at lot margins time. together amount everything better to and of a of of different

configurations many that. we as to will well that help us able Blackwell on have be do Remember, different

we that. if short really of our a probably If work. going that customers, will for possible improve ramping term, we're get begin also can these strong So together, after soon not, can. do working we in the to some completed as lot of can we as we we it If start

Tariff until what plan at it's it's this an of the U.S. is, much. its how both where bit understand a government's point, unknown little an it's timing, unknown further we and

So would, awaiting, at controls of export follow are always tariffs this manner. but we again, time, in we course, and/or that

Operator

and Ladies our session. sorry. I'm conclude does question-and-answer gentlemen, that

Jensen Huang

you. Thank

Colette Kress

couple of going are a to Jensen [indiscernible] things. We open up to

Jensen Huang

for time Thank law, you, beyond thank time to AI Colette. or scaling is reasoning. is AI, evolving The computation. into perception another demand resenting you. Blackwell inference test scaling, wanted just more With generative observing I we're AI and extraordinary.

like ignited models DeepSeek-RX apply an AI open time the The world-class compute. more it's inference innovation. consume excellent more global source even Models XXXx Grok consume it X, a smarter Future compute. importantly, has more model. OpenAI, much scaling. But DeepSeek-RX reasoning models model that can the reasoning answer. can has more the Reasoning models -- thinks enthusiast are reasoning

learning Nearly applying thought scale of model's and developer or every to AI performance. their reinforcement like RX is RX chain techniques

intact. scaling But loss Foundation computing. I scaling is being models of are laws, it's no enhanced still for enough. earlier, X mentioned multimodality, longer have The with pretraining now driving traditional and as growing. demand We the AI AI remains

We have scaling X dimensions. additional

skilling, time Post-training a of pre-trading, can model We Inference scaling. Blackwell this and more single single alone. post where demand easily where XXXx orders distillation that magnitude fine-tuning, require a test more reasoning than query from learning, training platform time and for transition pretraining defined compute. compute reinforcement moment, scaling and

than production. Blackwell's Each XXx XXX,XXX marvel. fabric models, AI software produced in nearly faster all operators. reasoning and by engine NVLink manufacturing configuration across this X.X million sites engineering is an Blackwell rack full technologies scale-up process Blackwell components in and FPX factory Blackwell transformer NVLink XX XX Grace XXX Hopper. new of led is

start sovereign physical AI of AI are AIs, we're corner. AI age the of the is scaling. at AI of AI advancing AI, beginning and at and time We're around life reasoning speed. the right But at inference just enterprise the multimodal

XXXX. We will in grow strongly

have and or of increasingly AI. forward, will every become computing dedicate self-operated. Going Data CapEx most will will and centers data either accelerated to factories company centers renting AI

to and about us of you. I couple going want a all new weeks. us other AI, for in a today. I'm GTC We're more. whole talking thank reasoning bunch joining Thank and of Rubin products. to you joining physical networking, AI Blackwell at Ultra, be computing,

Operator

conference concludes This call. today's

now may You disconnect.

NVIDIA (NVDA) 26 Feb 252025 Q4 Earnings call transcript