NVIDIA (NVDA) 22 May 242025 Q1 Earnings call transcript

Good afternoon. My name is Regina, and I will be your conference operator today. At this time, I would like to welcome everyone to NVIDIA's First Quarter Earnings Call. [Operator Instructions] Simona Jankowski, you may begin your conference.

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

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

For Commission. refer by and recent on and XXXX, XX-Q statements a X-K such future please required as to available our Exchange made the the currently obligation no statements. we affect and XX, the disclosure information Forms XX-K reports earnings All with factors May file of discussion most that based update that on to in as Form could law, our business, financial results assume may our any we and Except today's of today, us. Securities are to release,

we financial non-GAAP will call, measures. this During discuss

reconciliation measures You which of financial find GAAP non-GAAP financial a can posted on to in measures commentary, CFO website. our these is our

Colette. some Taipei will a San in Bank keynote turn With And me deliver highlight live. America that, streamed as X, to June in X, be will trade Jensen Taiwan, call in ahead On at person well June Francisco. over the of me upcoming as Let Sunday, we Technology which events. present Technology Conference on the let in of the show Computex held will

sequentially than Center. mid-XXs was and driving growth to computing year-on-year, the XXX% deploy Xx continued Compute quarter. a customer types, consumer continue revenue up up driven $XX Center billion Hopper and at and meaningful growth more percentage Data revenue by above GPU strong grew last Center a revenue cloud center led with $XX.X of Xx CUDA $XX revenue. Training year. as another drive providers more AI Revenue on platform. the Internet data Thanks, well by than for ramp and strong rental Data XXX% Large provider's investment. NVIDIA scale NVIDIA Simona. was up was companies. sequentially from outlook as revenue is record was by return and our of up in of billion inferencing QX billion. Starting year-on-year record, of an delivering NVIDIA AI networking cloud Data acceleration our XX% they XX% sequential demand growth, enterprise and on represented and and infrastructure strong cloud all immediate and driven Strong

GPU years. cloud. running For tight earn cloud integration opportunity in end for X every $X rich and it on AI an on $X easy have in hosting infrastructure, software over instances with to providers and NVIDIA customers stack NVIDIA's revenue providers spent cloud up NVIDIA GPU the instant public ecosystem and makes

language cost inference to the rental models the to train to train GPUs cost customers, best the lowest For cloud models. time and offer large NVIDIA lowest models,

AI on and Anthropic, infrastructure For Transformers, autonomous customers as significantly We public significant growth better drove the NVIDIA and enabling growth use Cohere, cloud. Enterprises on capabilities Character HXXX Leading investments. FSD of NVIDIA xAI, XX,XXX training while in infrastructure driving supported Tesla's dramatically LLM Vision. strong the driving are OpenAI, brings based driving XX, cloud more GPUs. others the propelling in consuming expansion for Their AI on cluster across NVIDIA to this their DeepMind, way automotive NVIDIA and paved many returns version are companies providers, AI of to latest the autonomous sequential AI breakthrough Databricks, their The Center their of NVIDIA cloud, Mistral, Meta, infrastructure Data revenue such quarter. Depth, performance growth AI, their industry. computing, for software building

is We Data Internet year, this expect opportunity a and companies our revenue on-prem quarter highlight their cloud multibillion this Llama a Meta's consumption. Consumer announcement enterprise growth automotive was Center A across largest driving vertical are be strong within X, to also big of X which vertical. AI, Meta latest AI available development Instagram, and cluster and across on HXXX on Llama Facebook, openly assistant wave large a model, XX,XXX language kickstarted trained has a new Messenger. powers AI was of a GPUs. X WhatsApp, of industries. available Llama

Tesla expect for inference are complexity way for growing where as significantly. examples GPUs. From In host with comes of next-generation makes revenue. estimate as to ones we to tens countries Meta the XXX,XXX infrastructure trailing and more see diversify over of revenue comes more we our Center with intelligence XX% AI we and built queries in demand Center some like advanced and of the X users driving much well drove compute. factories. number centers thousands and into Internet about In geographic essential are with in data scales Data of opportunities invest generative the customers AI AI. as in These both inference accelerated applications, that Data by from building AI QX, world Both of production, quarters, growth inference factories per our of the full-stack XXX clusters as out. size the As to refer what ranging training a continued perspective, its sovereign to as we GPUs, worked Large around consumer continues reaching user, AI data AI model computing with platforms hundreds number

nation's models. to Nations infrastructure, business own are capacity produce workforce, capabilities up refers through using a data, its artificial and networks. computing AI to intelligence Sovereign domestic building various

collaboration cloud computing and Others utilities. local a to and state-owned Some clouds sponsoring public are private sector use. procuring sovereign platform provide providers with shared telecommunication operating in are AI partners or AI for

powerful revenue and Group out the will infrastructure. Europe's than AI getting to build Group, in most in DGX-powered AI cloud providers, KDDI, sovereign more National the plans most invest AI to supercomputer. factories Southeast networking AI SoftBank to partners and Japan Center specifically Sakura trained infrastructure year, ability rich From nothing regional AI cloud supercomputer NVIDIA high accelerated NVIDIA's single-digit with Italian across build $XXX building port native allows digital ramped France-based Singapore, believe and of full billions Scaleway, country's key customers For natively AI Iliad software, the a Singtel is to million caught of Italy, China we is NVIDIA to nation's technologies, don't AI a to In can providers Supercomputer end-to-end sovereign subsidiary example, stack Internet, building nation's first require language. jumpstart offer of approach develop is GPUs, compute the ecosystem of expertise, products sovereign designed Asia. NVIDIA's has importance powerful and and the attention the And year. control this first license. ambitions. the LLM upgraded new that for Swisscom their in nation. We Hopper while previous the every The the including AI

export of to in the imposition October. revenue level China prior the Center Data down significantly restrictions control in new is the from Our

quarter compute algorithm From innovations, forward. on majority market to our GPU to by X. We track Xx, are the up HXXX to on can Demand accelerate Xx of increase. popular CUDA revenue able Llama Hopper for in production models LLM very by expect product was been to continues Thanks going in competitive the we've architecture. reduction perspective, shipments to during driven a and cost translate China for the in sampling a with We for remain the currently inference vast HXXX Hopper which serving to like started QX. QX

system HXXX their GPT-Xo of nearly HXXX delivered at HXXX, doubles was last week. significant for Jensen deployments. first and to team Sam the The inference the and OpenAI delivering demos value by powered production amazing Altman performance

prices server software billion For NVIDIA servers continue in Llama users HGX same a supporting of the optimizations, deliver an on tokens per X infrastructure current using parameters, token, improve second, we tokens over $X X,XXX every per to HXXX XXX serving X NVIDIA more years. With API models. HXXX for X than AI the can serving XX,XXX provider NVIDIA, time. revenue ongoing can example, generate $X for at AI Llama spent performance That at single means with HGX

HXXX constrained is At production. HXXX. on same full time, supply the [indiscernible] for in Blackwell still we are While

AI availability Supercomputer We expect in that Germany. Superchip U.K.; AI are we into are supercomputer working this Conference, to in of year. week for processing is cloud in global for University Last International volume. the worldwide Supercomputing and exceed Swiss later combined and of X Demand well supercomputers and announced shipping National HXXX a supply Alps at year. Grace bring Europe; next and Blackwell Jupiter Hopper Jülich for in at up using These Hopper power exaflops the partners our include is supply, at system we new of demand the ahead Supercomputing well Center, Center Isambard-AI Supercomputing XXX fastest the Bristol energy-efficient delivered the this the in year. Grace the may

supercomputing rate high seeing energy Hopper to Grace in performance. are due XX% We and an attach its efficiency of

networking proud to We were supercomputers with well decline, the the to the experienced powered of we demand by the of most a #X due was year-on-year ship. was of what supercomputers with We #X, energy-efficient spots sequential to timing Hopper world. Strong the take see are also in the which InfiniBand. and largely Grace able #X, growth modest driven supply, ahead

new to brand-new Spectrum-X return QX. sequential to ground volume NVIDIA started In switch, quarter, AI GPU Ethernet. Spectrum-X the Ethernet customers, shipping data optimized for Spectrum-X the deliver includes challenges performance our in the from enables is on to processing a growth we DPU, to X.Xx large-scale cluster. with Spectrum-X of and for accommodate Ethernet expect AI. in a Ethernet-only to AI We our It solution networking and networking overcome multiple to first new centers BlueField-X XXX,XXX including opens technologies software massive up. ramping compared networking higher networking traditional with market AI

AI liquid enables models with Inference will shipments. Blackwell language a new switches, Tesla, up training across announced open We leap xXX faster Adept, in using types, and Blackwell and environments scale powered centers large developers multibillion-dollar trillion-parameter to a Ethernet over up inference, Microsoft, trillion-parameter Images, build series Mistral world. center and This provides more to AI in data cloud of support from or double of InfiniBand will in be Microservices number including major every first and InfiniBand to network launch, AIXX, and the software and production data Google, with in language customers hyperscale cooling maker vision, OpenAI, Blackwell deployment xAI. the include to NIM. of computer a representing CUDA in adoption to consumption the air NVIDIA and Xx from GTC AI is range quickly offered our year. At our expect with and workloads, available at with spine NVIDIA new Amazon, and fast Cohere, time-to-market text, introduction broad robotics, or and next-generation customer software, on for to genomics, a and HXXX and Hopper. generative industry-standard on-prem. leading the March, NIM cases, the will a than and Blackwell ODM for acceleration The designed energy the Oracle, XXx universally, AI. platform multi-GPU NVIDIA, to to Triton performance-optimized imaging, includes by Ethernet NIMs applications Google, XXx and and enterprise Blackwell. models including year Grace AI. Blackwell networking, delivers support speech, deploy fifth-generation large Meta, from models. Hopper's LLM is to the faster launch jump Getty than a systems architecture giant and Stability biology. enable platform, APIs digital the CPUs, We product for than inference TCO for Blackwell a factory use the launched NVLink broad product be training we Microsoft, designed platform cooling. in XXX within real-time Shutterstock containers OEM XXXX NVIDIA secure to part and PrintServer Face, Snowflake of and as GPU AI to models and TensorRT line lower computing software enterprise, The Spectrum-X inference and generative Hugging AI, Meta,

product consistent and Moving of decline. GPUs billion CUDA market gaming the reception GeForce The Super our Gaming sequentially and AI very RTX seasonal demand with XX% was X% inventory channel is PCs. journey, range. equipped for to outlook year-on-year, across end strong and revenue healthy and start a down up we our $X.XX AI the Tensor RTX cores. From remained GPUs GeForce of with

Tencent and LangChain and GeForce to gamers inference RTX RTX Engine including and announced AI has (sic) performance GeForce Windows offer XB and Games, transform interactions PCs. and running accelerates Now Cloud are creators, generative XXX an installed performance with stack now models NVIDIA run for create Engine TensorRT NVIDIA frameworks, and game PCs. NVIDIA on AI Ubisoft base, LLMs including embracing generative ] AI lifelike for RTX Avatar enthusiasts, Yesterday, Character deploying of applications LlamaIndex. are to LLM running on nonplayable full and technology GPUs million Google's AI AI well popular to top perfect model PCs. And [ Phi-X to NVIDIA over developers, optimizations characters. Avatar help as for efficient on between Xx faster avatars gamers, fast XB Gemma and AI as up for Mini Microsoft's Microsoft GeForce and unmatched

of XX% drive Revenue wave ProViz. and to industrial developers growth. APIs their new digitalization generative technologies the believe to visualization We announced twin simulation AI digital down was Omniverse and and million of integrate X% next Cloud Omniverse At into year-on-year. GTC, sequentially applications. enable we will up industrial $XXX Omniverse professional Moving to

and XD software Azure this largest Pro. as cycle industrial be ANSYS, factory rates industrial are spatial computing Omniverse by maker, these including twins Cloud developers our retail largest is year. Companies by And to XX% their can Brand And are Vision workflows. production vehicle to Cadence, end-to-end and Some Omniverse later use Wistron, partners digital them such Siemens. using devices adopting Systems, makers virtual stream and one world's with will Omniverse Microsoft adopting BYD, Xcite, of the to world's twins defect APIs available times on XX%. the of power reduce Apple enable for Omniverse electric manufacturing configurations. digital APIs, digitalize planning

ramp primarily driven DRIVE built first Orin, and NVIDIA our self-driving AI by its Xiaomi was to up growth solutions AI in growth was sedan platforms. cockpit computer software-defined driven SUX $XXX electric Year-on-year of of was global in and year-on-year. self-driving. for AV successful Sequential XX% the launch automotive. fleets. the OEM up supported vehicle, customers the the sequentially with Revenue our We Moving million, XX% on by strength car

number and production for Thor, next BYD, is of the a announced architecture wins also DRIVE powered DRIVE design several to Thor makers, including We XPeng, successor new by GAC's EV NVIDIA the NVIDIA new Blackwell Orin, year. iOnHyper Neuro. with starting on slated leading vehicles

gross to Okay, and margin on the rest non-GAAP to of P&L. to expanded sequentially gross XX.X% GAAP moving the XX.X% margins targets. lower inventory

quarter, and investments. In of $X.X favorable Sequentially, were a last QX noted first expenses split-adjusted operating form QX, were our both higher billion of up a of announced trading and non-GAAP costs. in QX we from cash we Today, as share returned increased on basis. dividends. reflecting June the repurchases shares the to split XX-for-X XX and and with and shareholders costs primarily XX%, GAAP compensation-related component expenses compute infrastructure up XX% As benefited operating day

by We XXX%. are also increasing our dividend

is minus turn to billion, outlook second plus for X%. $XX me revenue be the quarter. the Total or expected to Let

expect platforms. minus plus gross to margins We market are GAAP quarter. discussion XX with our growth points, XX.X%, be expected consistent in XX.X% all and non-GAAP last respectively, basis and or sequential

year low For the expected Jensen Further range. details me, full $XXX billion be and are and be -- available billion, website. an other would expected XX% excluding the turn the our from rates CFO excluding approximately make comments. a to in and to GAAP included it X%, now discrete and and respectively. income and gross to are approximately Full we expect be to OpEx losses he few mid-XXs would or gains financial to any on non-GAAP other expected $X.X XX%, plus in margins be expenses to GAAP are grow to information and approximately like million, commentary year, minus excuse items. of are tax investments. to non-GAAP non-GAAP like expected $X IR in I nonaffiliated as operating expenses percent over income the is range. GAAP

change. The Thanks, going is a industry Colette. through major

the and training Before to nearly next the new produce a will to build trillion-dollar type CSPs to revolution drive data cloud base and shift Hopper AI customers Companies energy give factories, Data artificial to multimodal, expanding me first demand Training center, help productivity NVIDIA, every The XD speech, to continues while are efficient generative their AI industry some on and NVIDIA bring demand be a accelerated our developers. Strong data companies services. movers. intelligence. AI learn were of be accelerating centers traditional for The Center platform and to partnering NVIDIA With by to we attract our propels countries from of new -- and with power. gains and commodity, workloads you accelerated the generative computing AI to revenues installed text, the significant and opportunities. has save of start for and as industrial rental AI generated Q&A, ecosystem models importance NVIDIA plan. and transformation. reason on tokens scale rich cost growth. and perspective money inference video, CSPs Hopper let begun. more understanding revenue instances of accelerated learn images, And

from-foundation-up growing are today's With massive plan intentions, understand computer at AI will which retrieval and complex. shifting model, incredibly an about reason, interaction. From knowledgeable, context to scale, token are computing. tasks. platform has be Our generation is of and shift that skills transform will AI, our now we fast perform become answers inference information every generative model and full AI Generative generation stack is driving inference, a computing workloads incredibly.

instruction-driven CPU to purpose performing We at from from to computing multiyear tokens, of to software changing factories. how computers generating to and do, accelerated are manufacturing general level, generation industrial skills, GPU can a digital producing what models, intelligence. Token information from will drive the computing, intention-understanding AI build-out from fundamentally software and retrieving works

the sovereign Blackwell Internet offer and trillion-parameter full expand Grace CPU, and for service providers, solution the health foundation scale forms a bring and vertical software opens software that NVLink, new our from markets. let richer brand-new on Beyond generative has than everywhere, and The companies to cloud switches, enterprise, to of PCs enterprise-grade Spectrum, data to expansive we rich And for our network and on-prem consumer production to centers. optimized Spectrum-X to run automotive, and From are care previous Ethernet-only generative partners. AI of offering to AI, more The a customers, platform and cloud CUDA Thank GPUs, RTX NIMs next multibillion-dollar combination AI wave Blackwell complete generations. Spectrum-X market AI centers and AI. NIMs, multiple interconnects to in poised for AI is is expanded the generative NVIDIA us us for partners ecosystem data growth. to factories you. Blackwell Quantum, delivers of high-speed of large-scale a mix AI through ecosystem creating

Thank you, Jensen.

for the We Operator, for please will now open could questions? poll call questions. you

from with [Operator first of comes line the Rasgon question Your Bernstein. Instructions] Stacy

first like mean bit little a What shipments into is it's hands delivery that sound that's that the in that if when that regard now. to it's suggest drill if in sampling now? and production comment in customers' Blackwell anymore. doesn't one, actually What does wanted with it's My -- timing to I product does production full

ramp we've our bit and stood shipments But -- of a have customers QX. QX in in little centers production in shipping in well, start time. for production up QX, and should be data will We been will

it. Got

year, this like? revenue, sounds will see So it we Blackwell

year. lot revenue see of will a this We Blackwell

line UBS. from Arcuri come Our with the next question of will Timothy

the the Blackwell ask, for Jensen, of between deployment you all just the and have. deployment Hopper does demand systems wanted nature How GB this that to of I about the versus center. liquid I differ stuff I cooling hasn't some at node and because before, Hopper? the both done scale been data from at and level challenges the there's ask guess engineering within

of complexities all these about do elongate going? sort transition? how do And think how the of that's sort So you

Spectrum-X And platform that demonstrated very and many the Grace, in cooled large I comes domain platform, InfiniBand, at GTC GTC. Blackwell liquid support cooled, showed and a a Yes. Blackwell air now -- for I GPU. NVLink configurations. that includes xXX is at not

some easily And customers, from are installed They existing ramp HXXX will for BXXX. already will so Hoppers. into they their transition shipping centers to of that to base HXXX data

stack be And mechanically. so the Hopper And fantastically that systems have run compatible, on on electrically, Blackwell. will, runs to of if will Blackwell designed backwards been you course, software

base, will, to have priming centers, also pump, the Blackwell CSPs, system makers, been a about cooling. beyond be no ready capabilities Grace is chain base, center them, the to chain the Blackwell that you some with the with getting with going them if one chain ODMs, entire supply ecosystem liquid and for supply to deliver coming would Blackwell the going liquid the cooling data data to surprised our quite for cooling time. be chain supply base, talking We we And exceptional. is ecosystem, supply been XXX. the We've like GBXXX the

will from of question line next Securities. of with Our Vivek America Bank come Arya the

pull-ahead shipment you that very competition are pace other your that monetization isn't how strong supply, there your because you holding with or in the products confidence us enough of growth? Jensen, built that checks Basically, factors? what to a is or keeping tight ensuring give have is system of really utilization there behavior and

possible. that also in be fields of of characters, view and picture multimedia being The kinds demand digital to multi-modality, AI that's are course, And all and some because there. There's from all XX,XXX, GPT-Xo, -- all to applications reason every incredible. that's I'll I'll out but done all for answer Anthropic, I the of all GPU at then that like going and data centers ramp the Gemini XX,XXX the and big the and is ChatGPT day. startups, GPUs different single a consuming line there's CSPs racing that to, the every of list and now models, cars. train deliver stand of come question Well, is putting so just to operating are systems digital Customers domain that quite applications, biology, tool of application AV course, as train can of lot moving resource like to us pressure And there's even is expand their racing regional sovereign who their And it AIs end-to-end train The I of a pressure the a actually. which systems of is the all their of of to self-driving to industry country, to quickly they as video haven't their regional on models. mentioned the to lot productivity natural extraordinary. We're up long would generative design data the it's We're directly. those the -- startups up. stand in all and guess of your for work

demand, So that's I high And few it why make anyhow, computers the in Longer I a shift. supply. completely really, platform what work. think, that's we're Longer reason redesigning term, our and is this how outstrips really is term, the a comments. -- jumped to

it's and It's intention-understanding an time platform with only understands, the to but that longer solution. the will previous is course, also a computer. And platform computer meaning, we that understands no has and in come because is But been shifts. it it iteratively that interact much, past. computer. much it tell we compared the we back with more do shifts the plan of to to intend it, instruction-driven Of what inference ability profound the asked reason, course, a it process way our to than an clearly is reason this other for And

contextually so changing And relevant intelligent aspect answers. prerecorded instead way is now computer in of generating that files, such is a of every the it retrieving

world. And in get the the PC change companies and the just build that the the in that and going large And to to computing what all going It's that, so the labs world. is extraordinary. developers working quite this a that people even fact, doing the see all saw computing today going over with and of all -- stacks we're to you we're stack And of are the revolutionized. things things things all beginning that's the our startups over be is beginning

come from line the question with of Morgan Our next Joe will Moore Stanley.

you strong is. understand I just said about demand what how

Blackwell wait anticipate Hopper Will as there's kind of new for HXXX products. HXXX any you for to good HXXX demand products? lot You those you to of a do would have which a with for those Or and migrate sustain think of product pause growth? Do to you have? be products, for demand enough people and sort

the transition to We this we to we now we time demand Hopper quarter. Blackwell. their to for and supply saving anxious possible. increasing online. some that HXXX, through to to expect And soon Everybody -- because to outstrip as would be for making reason transition expect see And do money and of is as demand we infrastructure is as get that money, they they're like as

of Sachs. question will next come Toshiya the line Hari from Goldman Our with

ask Jensen, I wanted to about competition.

what they be I think updates address in limited competitors, in internal workloads most your To right, could you consider them of broader as in parallel their to medium And with programs, they what to did extent long you announced many they're cloud on internal they're have to think going or you term? guys. your existing forward? customers new view, addressing do what or to working

We're different in ways. several

processing data First, a NVIDIA's changed, like for mentioning fundamentally hard processing itself, their inference. And to aspect of SQL it I to for just accelerated frame plenty to to allows prepare inference in as of training was generate not to every it's structured processing, cat. it to training, in customers prepare computing has generation. pixel unstructured process that training, now architecture to cat, my data pipeline trying really from detect every but data which remarks, has It's the was to

about by architecture. it's why our our X. in something improved LLM architecture so the the richness the of was of one That different chips generation fundamentally architecture using And factor you well a so kind reasons the of our processing performance a on tells process TensorRT the We is received. of richness And and of software. the same

one, of could the from way modalities to you accelerated really going computing is now for cost inflation everything, image to computing computing. from sustainable computer world course, forward. all use suffering computing computer its NVIDIA has And computing and graphics vision the is of to run as general-purpose So energy processing, because

money save computing, going So to accelerated in is computing computing. you're to is energy save in going how you're how

for results center. so versatility data the of TCO in our platform the lowest And their

Second, cloud. we're in every

are And develop a practically we're we're NVIDIA so size developers shape. everywhere. computers any that to choice. is starting always for And and great looking a on-prem, of on, platform for with We're cloud. the We're in in

And reason. so second that's the

reason with factories, good AI to And that is starts, for build the course, only. problem has fact more a is this chips build to that we It very of systems our factories. chip do becoming chips AI problem. and but of AI apparent third with not a we bunch an The a it's people whole

In now large a It's just models It's fact, system that are bunch language language is problem. together. even not of of systems AI a complex large model. a whole working X

the so to billion. center and performance us value, together system to it. for be billion just All all optimize the routinely in a X, work system, X, able the to a you as had factor the by a which the be And to NVIDIA that put operates software of chips optimize system, to to able the builds across as And this to our system. have it too, that is causes improve in perspective a simple if numbers, a $X do, $X when infrastructure by infrastructure in doesn't of pay that of data and we you improved chips you fact factor

estate the is at is cost when the reason This performance goes the And it, that of money cost these of along the the it, power so of of along money. infrastructure really And to goes with time why up. goes all is matters the the cost it with lowest lot adds real lot data to extraordinary. And chips of all center, operate the and it that everything. this today, data is it a also with carrying a takes because highest it, that center, performance people value fund a the of the quite along

And highest performance so the the also lowest is TCO.

will the Our Cowen. Ramsay Matt question come from line next TD of with

center you platforms I in the that industry of that Xx getting, been you're Jensen, the guys jumps are mean, my combination I've career. in the I've training. whole velocity at performance the new never seen introducing data same

going billions your stuff, spending very the on, it's an than much amazing are quickly And up customers also as not more where to watch at XXx thing generation your interesting very, depreciation about product. Some of stuff the but, of inference. with competitive dollars juxtaposition to of of cycle current to it talked the be that product creates you an it's GTC that in new

So like how you little mind to wouldn't customers. seeing if you with -- you're a speak I'd that situation bit itself evolve about

to to hear your product you it'd generation nearly but going that's you're of have installed path. bases that interesting Blackwell, happening customers And as to As bases, obviously not large as new along what compatible, move very be see with large installed performant software stuff. you

Three Yes. points I like really it. that I'd appreciate to make.

as to comes, Blackwell it's Blackwell, very we've into X% explained coming. And their when okay? the fast And the just only coming And of in other there's chips into as they're performance And build-out just can. a building, have you and you on going a our it. to bunch they -- as XX% want versus Blackwells you're you're if you're you're be you you a If build had customers but terrific. going And we your then after because into in rhythm be into then we're X-year short there's way to they got anyways to we build-out anyhow, feel building far going differently. as like, build they build-out to the out, map world. at if early the X% road keep for mentioned, as as to whole and on to just, so and see our will, keep And them, average so

thing smart want money them. to to do. They So valuable today. is time save the money really, They today. make that's to And really to need

is valuable, example to idea company for that next reason next that groundbreaking The reaches the this after really so valuable. major plateau that's a announce is a the time something up AI. And being to one give instantaneously announce X.X% standing is you an who valuable why me second data the gets better. called center Let because gets train this time of getting so to of and thing

and delivering technology seeing And X.X% this And company or the race, so is on so they're in your as leadership race the the to be on because why vital building repeatedly companies better. to want to companies multiple the across get trust this is, the important. race so platform question know company races, And is is leadership, have delivering all want groundbreaking that do going the the that better you technology and build this to for you're to better? reason and AI that's platform

in And great get months months started to time it a between train to X so getting order on done, The earlier leadership difference to to project, is earlier everything. train matters is just get time matters a great train X deal. Time X-months that to deal.

we're just standing the the And right because so systems is plateau Hopper up like next now reason mad the it's corner. why around

so the second And reason. that's

a We're bottlenecks The is it first fast the where We is all how putting can the build you we we're optimize slides we We're know that we're advancing is they PowerPoint everything, are. comment, look stacks everything, so literally at measure up -- we not but that slides is And and moving really good. data we a Because We and it. all about systems it which have good, at because perform the across we comment quickly? doing scale everything. know monitor delivering We're center them entire our perform why reason made PowerPoint great that actually also we're little the guessing built all -- like year. scale. that not here.

that we we and to a infrastructure perform miracle disaggregated know they and centers how we is the then where however into Now one we entire need we to We know are. data customers' build that that's things it's AI know bottlenecks to how today. We know have entire the from improve with the This entire single knowledge optimize and every intimate performance. chip to processing sets their at where is apart build what across ground system. help us the scale most them done deep infrastructure the exactly fundamentally We up. integrated but the of we a here, do achieve them, know we data going but liked. where is bit we of our center

we how most to it and exactly to of every perform with get going understand And generation. so the single how it's out

So I the Those X are points. appreciate.

from Evercore ISI. question next Your come the with Lipacis line Mark of will

down. if cycle please know. computing. of computing observation get do dominated I Jensen, a you've past, observation, could believe each general-purpose this they higher driving a why utilization, And the were compute to era. workloads, made me cost that to And that you And ecosystem the argument typically in for mischaracterized ecosystems computing for different I drive motivation was the adapt is general-purpose let accelerated GPU CUDA that

demand seem workloads So then are being training on like which it driven that given themselves lend a of might to also seem to driving network the solutions the is, your solutions. surface for limited question are that inferencing, custom and neural by the workloads, number

question the purpose purpose enough more a these is general that the or about general become then risk evolution framework workloads variability is so computing on And there at or that rapid support historical framework? does enough

Yes. computing I wouldn't be call Like wouldn't we was it general-purpose for good spreadsheet. NVIDIA's the is for general really example, That accelerated purpose. at designed versatile but very computing. running

for isn't operating of the -- probably fantastic compute, loop for system And control a accelerated so an code general-purpose there not is computing.

domain in can but along, versatile, commonalities. run systems with coming fairly large next of of last every all calls, would over sudden, but now these They're have the rich along we in moment the some versatility have or run-time, X I course accelerate properties that we're way and memory asked usually the of And came of represents this, that is of along, of they language over that moment so the years, describe they platform the a to now. And fusion you the why models of understand of Those so moment the a commonalities. lot carry models conversation memory. design the the all with they're things fact the for And entire to you, Maybe X% the all example. a language it. There's came can their applications guys all I XX% at XX single have because reason AI needs is the heavily we're large parallel, on code and the differences I able threaded. all that accelerated deep look of models large. that start-ups the so, our that's model the a conference brittleness of are of computing. The them, are years say the architecture, or number me generative context. of about the

the became of sudden, the All a of versatility memory Grace super important.

that these lot going We that is and in advances looking continue the software, this widget and first something domain, principles that's for entire of having obeys entire generative few and AI this for a really to scale is keep software But evolve, to going ready going million to to of of we're we're the we're for of not to better for good really it. years reasons scaling of X it so in And advancement for software have reasons, coming model. bigger. why good the to getting and forward AI times X easily the in these a believe models. There's that but by begs designed one properties is domain of each

not as you key. build is or and quite brittle And too too hardly build versatility that's FPGA might platform well just like the so that, really -- of specific, And ASIC computer. you something an or it's our an you're a but if

line Our Blayne next with the of Curtis question Jefferies. will come from

curious, HXX. product with trying the but your supply other with Actually -- I you how a there'd Hopper out customers being kind be demand products. you constrained, about think you're of obviously, I mean, China, I'm of for assuming serve mean, do it, came for to ton a

how Just thinking kind in you're the curious about half. of that second

as gross thinking you're what impact, as any sales well for elaborate margin. could You

I didn't hear your questions.

out. bleeped Something

thinking you're Hopper products. different how supply allocating and between about the HXX

of China now levels honor because limitations in have And our technology. China customers we do in best our the more on a and Well, lower is every is we past. of business the case the It that the we it's competitive than that for our customer. substantially lot

so And matters true. are those

for overall, the best to is about made and we customers and we'll the I we demand and to market serve end HXXX entire continue there comments our particularly the supply our the of the outstripping ability, However, our of towards so do best. But for think year. to in best that the Blackwell the markets do

from come line with will question the James. Raymond Srini next Pajjuri Our of

is like some Historically, and Jensen, systems, XXX a what I boards of significant GB GPUs the a actually small. systems. demand looks sold you business lot and HGX of was more for you've clarification said. relatively on there systems it think a

curious, TCO the is architecture? why I'm or going systems just So for it it demand now it such the is it strong or seeing just something Is you just are is forward? else a that

that. appreciate I Yes.

the components fact, makers. that the integrate sell we way into disaggregate of We In GXXX and computer the make sense we same. it is all

but coming Hopper, system have different It And year is are Blackwell. only computer configurations that frankly, started off charts. that's had that half the that for than at We even. this with peak. XXX out way less its

Nothing of so going so And designed. ecosystem and version, bunch cooled versions, really And our air forth. a all Grace of that changed. they're visions, liquid being great whole see version, offered There's of partners. from on xXX to so are you're cooled systems has

to is save of density of a liquid Blackwell is data cooling computing, Now and compressed lot mention of in expanded much of course, CPUs more energy tremendously. to money power be the centers to efficient. more not -- offering has platform The integration the provisioning and our going

components much higher of And and lot networking. a more so data switches, from a wins. it's center It's better we meaning gets everybody much data performance, The networking that solution. a more networking expansive, offer center,

AI NIC, only operate now AI have. -- a so to ecosystem bring know we operate we that have Ethernet of that the only because to who can NVIDIA course, Of to large-scale how Ethernet they customers

expansive. more And so much is Blackwell

have this offer customers around. We to more a generation lot our

line from will Securities. the Truist question with come next Stein Our William

xXX, of and be highlighted you've whereby challenge? power real worth Grace deliver synergies to that your that and but CPUs operations, while available very that that are good address consumption between data ARM-based whether might technical some perhaps are that good very made and or for similar that customers, Grace solutions, on workloads, cost while others there advantage good a advantage AMD can there the you products bring have point, in Price at are that especially decided Hopper, in AI center could or partners might related some, Jensen, some NVIDIA Blackwell. dynamic reasonably there a technology provides CPU side, Can emerging and deliver some Intel NVIDIA great more to to client emerge there be

The chips, is data by that's It's at system memory. Well, reasons. you calling some allows terabytes many configuration low-power with of excellent we between Hopper interface Grace mentioned Grace the this xXX possible connected. first of it's and between X that like Grace applications, It and our it memory something the today. build interconnect The charts. And together. per with is X to because a really used superchip. is the the the are X the coherent for chips almost terrific But center-grade that them LPDDR. connected systems configuration, partnership with partners isn't are second. The good and true off really are like us memory do that's weird system It's the

really And with a so domain, our which of important power the create lot can then has we large because And a we could large architecture now, single NVLink to models own is we on the that create node. system of next-generation vitally every architecture, save something inferencing. the finally, for entire language because

like XX-node together GPU. XX That's saw you GBXXX connected so domain. And a NVLink that Blackwells has X giant into

do to that. needed Grace be so And able we Blackwells to

are that are them software And so programming to for reasons there reasons essential are way. and are architectural system build there that reasons, us then there

yesterday, if the PC, Satya shipping see in But laptops. really on opportunities like today, RTX next-generation fantastically NVIDIA's which ARM runs also beautifully. that, supports Copilot+ announced excellent, PCs, was which GPUs so the we'll it And at thought I as saw explore you And that are we it.

And opportunities up for innovation even for so system it opens PCs.

with Muse the comes question line from last Cantor Our C.J. of Fitzgerald.

pace you but ASICs, GPUs And Cohere forward, solved I to you are connectivity. longer-term with are NVIDIA's and are frictions be yet, the amidst need rising last obviously, I importantly, of guess, competition and question. coming maybe Precision, scaling thinking to Blackwell your custom Grace, I and share even hasn't Jensen, forward-looking investors and million-fold today? potential from willing CUDA, how innovation bit of decade, about a know what in a impressive. what truly more decade? over the When you look us guess, launched

Blackwell, X-year rhythm. And chip. I are Well, that announce after another there's can a on we

exciting -- so we you new And really Ethernet. in have can on on we're and technology us Ethernet, and We're for announcing networking all rhythm. also having a on for a Ethernet. count -- road that Spectrum-X map coming fast count very But

to ecosystem announced taking We Spectrum-X have of they're Dell rich partners. market. a that

are We rich taking entire AI of going to a market. ecosystem and announce customers architecture partners our factory who to have

fabric, ultimate InfiniBand network we're the that is became is with have companies computing much Spectrum-X, as a going to And to to NVLink fabric. computing Ethernet to X we and a fabric. And all so fabric, computing computing we're years, committed better Ethernet a a -- fabric, InfiniBand, computing network. to better for and network. performance, the out Ethernet over computing InfiniBand computing better networking it a started And single committed want computing make fully links, InfiniBand domain fabric. fabric for

all X fast take a to going very we're so forward clip. at of And them

them. see coming, X new coming, software to run of new new all that going switches new capability, you're so on NICs stacks And

the software of all of runs GPUs, beautiful CUDA. it entire our networking of coming. And it new a new that are new stack. runs CPUs, of all switches, -- New thing NICs, mound it chips And all is

and stack, And and more centers clouds more data without and and you without just anything, you anything invest architecture will So to today, everything go and get faster. invest all, our runs. more on doing to faster if and at our software it doing today more in just it's faster going

capability, innovation drive the other think drive on I of will the so on the one up the TCO pace hand. And bringing we're hand, and down the that

able we anymore, this architecture that industrial artificial Thank computing to to scale. so new software this of just NVIDIA for do And manufacture intelligence be new not start and but the with revolution we're we tokens, manufacture where era scale should out and at going we you.

you question-and-answer may disconnect. our you joining, and for call our thank for all now session conclude and today. will That We