NVIDIA (NVDA) 21 Nov 232024 Q3 Earnings call transcript

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

you. Thank and fiscal conference XXXX. everyone, third for to of quarter welcome afternoon, call NVIDIA's the Good Executive Chief With Huang, you on Executive Kress, NVIDIA's President today that and President me from being website. Chief Officer; NVIDIA our call Investor and webcast to Relations Officer. I'd like Financial are is Jensen Vice remind and live Colette quarter The financial results our call to webcast will discuss be available the for and fourth replay XXXX. until the conference fiscal for is NVIDIA's property. content The today's of call our without prior consent. It can't transcribed be reproduced written or

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

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

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

financial find on commentary, measures You these over our Colette. let call to website. With GAAP of a me in can that, which financial measures reconciliation our to posted non-GAAP CFO is turn the

based more infrastructures. up HGX year-on-year another sequentially revenue The with NVIDIA together Revenue XXX% architecture billion. InfiniBand essentially $XX.X platform with of XX% of the up AI center. data and with $XX was architecture, up Hopper and billion our outlook $XX.X networking, end-to-end along center data of year-on-year. NVIDIA and on drove supercomputers quarter. Core continued QX record InfiniBand up for reference Simona. are the well than Starting above of XXX% GPU record HGX XX% billion, Thanks, sequentially ramp and our was Tensor

Adobe, built NVIDIA, Firefly, CoAssist, Copilot, applications and run Now most ChatGPT, of ServiceNow and Assist AI including XXX the exciting Companion. generative with on are Some Microsoft AI Zoom

last tripled. compute networking deep revenue applications large is computing. for NVIDIA AI infrastructure center and nearly language and computing. broad-based year for quadrupled now revenue systems strong a Our accelerated fueling recommender demand in AI data generative learning from Inferencing major workload NVIDIA is training and models, inferencing Investment for

ramp sequential help Consumer Internet and drove outpacing revenue companies systems text. and Internet in data Meta QX, with images approximately Most recommender learning growth. investing like and production our advertisers Companies enterprises half optimize in exceptional of and comprising growth in also enterprise AI are software major their ServiceNow to Tesla total are and consumer full racing autonomous AI deep generative are enterprises platforms. industry center companies custom broader of AI Adobe, Enterprise vertical companies adding and applications such The up as beginning. assistants driving. adoption Databricks, to Snowflake such generative developing And AI is are copilots now for to and deployment. AI wave as

broadening the new set strong the Cloud in with cloud that virtually globally and other hyperscale CSPs a from as well Demand of drove as data all quarter. of in roughly center revenue service providers was from our GPU-specialized Core opportunities are address demand. AI. NVIDIA CSPs instances instances growing high Tensor in every generally are the GPU rapidly available now to market HXXX half

so demand quarter expect and have strong next We year meet supply continue significantly every year. increased to do to this to

new of announced export require our Ampere and XXX a China and a faster control set to XXX the of will end launch Hopper and number of product regulations licenses broader our others. for including Middle other including and several the East. for in regulations set the markets, meet AI products countries and government series growing certain of Vietnam a cadence the the U.S. and and These export of a We have also quarter, diverse opportunities. Toward

Our XX% subject approximately data now consistently XX% to sales other requirements center contributed of licensing to derived and have are quarters. past destinations to few that affected products revenue from China over the

decline destinations offset our by significantly the in be that to they'll more will in than we regions. fourth expect believe We though strong quarter, other sales these growth

The U.S. technology government jobs. promote regulations center to designed worldwide spurs allow support U.S. data growth continuing leadership, worldwide, compute to industry the provide U.S. products as regulation the to markets compete China, including and economic encourage, the to U.S.

highest government performance the For the licenses. requires levels,

process. requires for government a the even government products lower performance not at performance For streamlined does require notification levels, all. lower And levels, notice the prior any

data including expand advanced products not have offer notice portfolio guidelines, each shipment. clear which for solutions our each government's we working the to the for are wish to compliant before product Following U.S. does center category, to regulatory government

and working China East customers with early Middle granted local in compute growth these innovation. invest ecosystems. pursue data With awakening from in any and countries to nations is own the in AI will their licenses government. AI are and We to U.S. to generative capacity, their whether revenue. Many infrastructure be to significant industrial amount some the of use support need train LLMs can support the to are too sovereign domestic know economic It to investments for

line. of quarter to Grace which revenue GH customers. This Hopper Secret available architecture. Lab multibillion-dollar GPU infrastructure was to on was variety majority building and Grace with economic NVIDIA revenue a For and infrastructure. based and shipments instances our The a third built the And ship, provider, based Cloud. and Hopper GPU AI NVIDIA also for companies, on industry a Grace coming Europe. cloud is are InfiniBand a the cloud and at from initial with the in product in HXXX, supporting AI private National GPU. Enterprise and with the new into our architecture ramping HGX over to servers compute and a new France chip, across are next NVIDIA is took government by the to Hopper supercomputing regional opportunity boost Scaleway represents market to our workloads Ampere imperative, Grace soon traction and LXXS Hopper generation prior cloud GPU inference platform multibillion-dollar tech Los capacity AI working Alamos Oracle their significant Hopper sovereign example, sovereign first serving the fuel advancement and including Swiss we the XXX driven contribution investment with lower now vast ARM-based National years. From customers, standard Infosys, combines are French in across getting software place Grace GPU-specialized Grace Supercomputing the system largest QX providers India's perspective, Tata also few a is Hopper GPU training Reliance of new product a began AI of National Center, quarter.

of will fastest next-generation close U.K. supercomputer X,XXX AI of the government supercomputer AI build the that Grace Grace with with Quantum-X exaflops Superchips. most XX AI supercomputing German announced it almost supercomputer will announced Isambard-AI with X world's to called over its and The also it it center, Hopper Superchips making Hopper world's AI build powerful XX,XXX Jülich, InfiniBand, performance.

best capacity come. that XXX data across Grace our All to estimate Hopper in, on exaflops next Europe this now versatility the deep generation. AI compute and Inference full will we as is offers Japan chatbots, U.S., the thus, combined the wins performance ownership. production cost AI lower with built exceed all and the beginning. NVIDIA supercomputers inference and in center to the is just more recommenders, of contributing of year and for learning demand and is the And power copilots significantly AI text-to-image

also reduction a curve. fast We cost driving are

NVIDIA more now we release of LLMs by performance TensorRT achieved inference cost NVIDIA the than the LLM, With or GPUs. half Xx of inferencing the

We LLMs. compatibility. also year member by It HBMXe, AI HXXX announced memory customers the moves performance speed of latest TensorRT first larger which Xx and Hopper instances further larger running year. to latency. accelerate to first will Oracle CUDA to to like a Services, for models This and the to increase delivers another inference next offer customers no LLMs [ the Cloud X GPT-X, and benefit X. and to to in like an will family, of HXXX-based the AXXX up the HXXX, XXx GPUs reduced be generative CSPs GPU Llama changing to ], Xx models in the or just among Web increased and infancy Microsoft compared Amazon Azure LLM is Google offer move our starting increase allowing Cloud, HXXX architecture with Compared stack. for be without faster, performance Combined, their cost HXXX

AI and last At our We applications Microsoft AI Microsoft the domain knowledge running and and development for their of securities and bring can expertise AI our collaboration AI on custom data with Azure. expanded customers NVIDIA service across the foundry using in week's and enterprise-grade help foundry entire generative all Ignite, Azure. build are SAP with AI and Customers of proprietary the an tuning introduced their DGX enterprise on we stock. deepened first models we software stack Microsoft and Amdocs Cloud, them our service the support.

addition, HXXX new computing AI will training and MLPerf by platform latest benchmark as and based the the instances shown results. in on wide Microsoft HXXX. The for a the remains In confidential launch most industry versatile top-performing margin,

public a last is because the an gaining LLMs efficiency to LLMs. Microsoft reflecting nearly year-on-year. InfiniBand over similar growth enough Our very on which deployments. Networking fivefold for now Efficient highlighting magnitude or and run the cluster, the Strong exceptional very annualized training grew cabling, generative cloud training Microsoft by GPUs scaling. requirement in achieved demonstrating AI week scaling efficient $XX is billion cluster year. included that NVIDIA Azure more demand critical are was InfiniBand HXXX the a point in scale driven miles of Azure exceeds AI Xx for more rate. than uses needed than performance XX,XXX every revenue key order globe. circle XX,XXX this of InfiniBand, June, in results of growing made by to identical

NVIDIA into expanding are space. Ethernet We networking the

can traditional leading compared Spectrum-X including in purpose-built offering AI communication year and HPE performance Spectrum-X new support ethernet for networking Dell, higher be offerings. will Ethernet next with for technologies, from QX with achieve Our X.Xx to AI Lenovo. OEMs, end-to-end available

an adoption. see and offerings, our me starting Let update provide are software on we services also to where excellent

growth exit track annualized reflects of intermediate on DGX for over rate offerings. for year are the services support X at primary opportunities Each AI our billion growth with and Enterprise Cloud enterprise respectively. our run and and the to software, training enterprise AI AI term revenue We with an service of NVIDIA $X recurring inference, We software. the see

an ServiceNow, Getty, discovery Cloud with Snowflake Genentech; DGX plans framework also AI and to morning as We collaboration platform. now optimize announcement AI latest our their LLM BioNeMo customer use research drug Dropbox, Adobe, pioneer to come. and biotechnology this SAP, part to enterprise with was partnerships Our have of accelerate others the AI help

Okay.

these shopping has back-to-school of sequentially billion important up technologies ray Gaming the TensorRT speeds year-on-year, released XX% Xx. AI strong innovations available developers. and AI over an emerging season, base high-performance natural define $X.XX just gamers that like enter has holidays the NVIDIA which for the the RTX-enabled is the performant and pre-COVID by ever NVIDIA the applications. than supporting of over NVIDIA with of applications against driving with grow. revenue LLM gaming. app lackluster gaming RTX was with quickly PC $XXX. brought PCs Windows, up killer There games price in and best and We XXX number ecosystem more for period, continues to even RTX and Gaming The demand PCs, RTX the reflects and market games as to low and application most new we've levels, GPUs This is RTX ecosystem doubled now With the points relative LLM We upgrade XX% backdrop DLSS. The on-device installed AI are Moving in tracing up performance. and for million, value for AI creators. attracting at exploded RTX with now new inference to Generative buyers. the to as of platform as technologies significant XXX lineup workstations.

Starfield. service NOW X,XXX Alan of Cyberpunk GeForce titles, XXXX, Gate our including gaming games Finally, to of build X, the launches continues and library Baldur's Liberty PC surpassed Phantom cloud Wake momentum. Its X,

as and up AI GPUs AI health ProViz. platform choice models, workstation launched optimized demand public line offering training the AI for workstations based NVIDIA NVIDIA locally. to and and fine-tuning a Lovelace as of edge spaces RTX and workstations previous processing, running imaging and of year-on-year. XX% generation Early the use powerful powerful is smart inference desktop Xx inference up million graphics in ray and workloads AI such and new sector. and a AI sequentially tracing performance AI are We RTX $XXX of cases, simulation generations. on SmartNICs, driver. smaller ConnectX, Revenue professional These design, up Moving care Ada emerging was the XXX% applications new to the for is engineering in for include of models

platform also to for We autonomous facilities, digital and designing, Omniverse helping entire time [ manufacturing vehicle continue progress the worlds. to build and assembly for pipeline, incorporating is simulation robotics automation Omniverse-powered and factory including on X cloud a Omniverse operate We Microsoft defects. its its using cost. saving end-to-end engine. services virtual our software Axon is available on digitalization building plan, ] it make simulation reduce Omniverse, Azure, and operating and engine manufacturing efficiency increase new automotive, and into for simulation process, twins announced design, and virtual Mercedes-Benz XD

driven primarily NVIDIA We to ramp our Orin Foxconn in Revenue to by ODM of SoC. AI and partnership automotive up next-generation EVs. our was extended NVIDIA solutions DRIVE self-driving the Thor, with and sequentially growth global on based OEM cockpit Foxconn continued year-on-year, has X% for automotive. SoC $XXX customers. with million, DRIVE the Moving become automotive platforms include X% up

Foxconn Our a build easily state-of-an-art AV sensor safe a customers platform for partnership software-defined their secure with and car. to standard computing and provides

and non-GAAP expenses the percentage inventory expenses from rest we're Sequentially, reflecting by related lower non-GAAP the to reserves, primarily driven data inventory X expanded point products. move net previously GAAP release and GPU were a reserved and P&L. margin sales operating XX%, up center including increased gross up benefit of compensation architecture benefits. XX% the to going and margin of Now XX%, the higher to XX% to operating to were GAAP Ampere gross

minus of to Let fourth quarter the or me X%. is be turn revenue Total to fiscal XXXX. billion, expected $XX plus

expenses information GAAP it be networking. sequential by with compute demand center driven as will strong aligned more strong notebook is decline and to sequentially likely Gaming now expect and with tax plus seasonality. income XX%, financial for We data growth margins minus continued both upcoming GAAP any and be expected GAAP financial from on or are be and the billion to are points. be and Other X%, and to excluding discrete available expenses gains of and to $XXX closing, commentary highlight some IR events in included approximately respectively. $X.XX our non-GAAP nonaffiliated rates respectively, non-GAAP be approximately for losses excluding and are investments. are me expected items. information are other Further non-GAAP plus website. In expected GAAP an to million, or the XX.X% community. and operating and gross basis income XX XX.X%, minus billion, $X.X expected CFO non-GAAP let

Virtual January UBS We the Rancho San Summit on on Conference will X Conference the attend Technology in on Arete the Verdes, Scottsdale, November the in and California December in Healthcare Arizona Wells JPMorgan Francisco Fargo Global Palos Tech Conference X. on XX; TMT November XX,

Our and XX. the Wednesday, is earnings quarter discuss call of February our scheduled for to results fourth fiscal XXXX

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

of line of from question first comes Vivek America. the Arya Instructions] Your [Operator of Bank

And for of think Jensen, adoption AI Just, next trajectory all into clarify center, question China market? center the in nearly do be in when just you are at the what Because the shipments of your growth, it I close its year. you. of expecting Where you QX? in is your data the data main to the contributions look XX% generative then to spending Colette, in curve will are terms wanted we

an AI continue products So we the the eye curve Just are of to adoption what on into that are you keeping you market? metrics grow? where to you can generative in inform your

our seeing We controls seen start had we So revenue. that XX% China impact destinations Vivek, your last historically and the be of to are and some center that the other with QX several quarters provided. and guidance in of our about the me that we over outlook question, export first, let on to impacted data XX%

our negative We visibility China that effect impact substantially term. are of a that we decrease expecting our the on will we export not good guidance QX. The to into have for do controls the over and long have magnitude in business, even into as move

regulation-compliant center product available offer months. the that products, become are, possibly though, do coming license. working These they new in data to portfolio require not may a our expand to We solutions next

contribution be of revenue in QX. the However, we or a to their percentage meaningful expect material don't as

off Generative method, everything everything. a expansion and do The retrieved that software retrieval is AI it, decades. several is is in hardware seen computing of of that largely -- we've storage the the changed has largest and was TAM added what almost almost augmented generative been core of really approach, exciting somewhere, with it's based that a you now, at what's

data to to that to typed queries, It protein, generative text on see text text, that were we humans changed. access by text on past, to things text could based to and is and be queries. text It video, chemicals, image, intention these natural to are used based in semantic text You are these way XD, that in approaches. The now the queries. queries, to language is processed explicit now

so that doing to trillion manually are that with is whole a tools many called about SAP software X decades. of new now Dropbox and excited others industry, And building will the so segment you manually assistants. have use which copilots been going it. we're really of And help the of is about. And software you're over or last used, impactful hear and copilots there's to these couple that Instead the used of tools we're the work $X that about areas and is has

version these be software. copilots modern the course, that, help we the business software, copilots of to is all software. We'll and we use of version of teams hire of use software, enterprise assistants us will And modern into will so going AIs, of instead and to -- also assistants which licensing do to but connect continue of the

computing. have energy X approaches One trends. the that's that we done save something software the to cost. underneath. purpose better One it's AI. of now hardware called of independent related of and General generative that wasteful computing order accelerated energy. in you transformation could transforming accelerated you And and the of largely that so And software way is magnitude ways. is And much hardware of see is is much, could X driving There's an too is computing,

an order can by an You save magnitude or you magnitude acceleration. could of save cost time of order of of using

trains that a not, factories. We're people It AI very are essentials every new tenant. processes it the where transitioning, is call centers. a computing been of we a data are there's And that a were data if talking AI that. country. This data that applications running seeing so one that center centers data AI. data many class lot of data of new augmented you But you It general centers, centers new of represent approach. And different great purpose and just a models, the tokens. about class X/X and built applications, you by same few one these of data. will, we this basically it has of tenants generates data data by in stores computing And used are is And out accelerated that application unlike new everywhere by into about, traditional And about new generates this the where new factories infrastructure lot used the just files, using past these being if centers. of have class centers

transition ramped, the start-ups so And we're large you the -- Internet AI expansion, this computing start-ups, with partner on that's to that like approach, process where see have in augment to model chatbots at that. you are companies platforms. assistants while generative software into saw and new consumer wave, the language in companies. tools they that Meanwhile, and would of And with and ramping look copilots to enterprise we their the the first if build way starting the being we're who you

is really all seeing AI. and thing, they're cropping over dedicated to You're up world, processing which GPU-specialized CSPs the one doing

develop own infrastructures, now seeing that that own their their own keep countries data -- people data, sovereign process utilize data, AI have their AI. they and their recognize You're that own culture, keep to

year in in ago You see about India, that Sweden. several a --

major seeing sovereign being major -- in announcement built almost will is clouds own big a You're my in and Japan. have week, of guess it the Last that surely, And significant. number their that every are will AI country quite every AI is France. But have cloud. really region

new industry, And company, every region. as wave developments so every I the think generative propagates through AI you're seeing every just

of so this we're this And the beginning computing inflection, at transition.

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

ask about to side networking kind I the Yes. the of business. wanted of

strong you business growth billion the quarterly it that now curious cited, think almost and Given -- I'm see to approaching billion a I $X.X is you've like like of as involved start would year-over-year looks the the sequentially, growth it's you Ethernet that rates how maybe XXX% Spectrum-X how $X next beyond we we above of of think look InfiniBand narrative year. becoming maybe part into about and differentiation Ethernet to networking evolving characterize that and your just to relative traditional stack as level.

the Yes. question. Thanks for

really and control, radically and to around congestion InfiniBand. Ethernet you that multi-tenant adaptive and -- do natural very the low that rate factories traffic $XX business latency, overhead you with going majority InfiniBand And of noise networking of part vast just low important. the the new dedicated because large-scale is InfiniBand. to -- reason data isolation, that isolation recently. way for only way its course, it's with is that's traffic the the and business that our we The not networking networking much billion at standardized routing, of process the already as InfiniBand rate. and plus not of but it platform performance software-defined latency hyperscale AI different. moves on added environment, is and not only is And a rate network larger. the to the a the just we could Our get a networks, the And data of The AI mentioned, pattern run And mention,

itself. capabilities the calculations right computing We'll a so switch including lot into And the and point not computation fabric do much InfiniBand the floating is and put computation. It's right on so fabric. a network. in software-defined fabric, We of a also just

versus And InfiniBand measured XX% reason the billion in hundreds A the so really AI is difference factories. difference in dollars especially of you you've X that in the a and were InfiniBand that that's And really, XX%, $X profound, of infrastructure Ethernet overall for infrastructure difference is the factories AI because just is why for for where Ethernet versus reason dramatic. And invested so in is value. if of as up, X millions of effectiveness, XX%, up. renting scale years, adds or course that you over it

And so undeniable AI InfiniBand's is value proposition for factories.

move we able own where proprietary AIs as their building like data, this AIs. enable computing skills. enterprise on However, to our we'd company based our build our company is AI of type proprietary into enterprise, to every to custom in be custom We're

the example, -- creating. the be the use, building create we doesn't custom recently, We're AI spoke others. models performance ChipNeMo. models we There'll It's factories have be is to that For of tens, called as one for about to we inside hundreds we're of employee the train many of as our company. And high all our -- our that models. AI use company that

environment. to so an to be would like the run And we able Ethernet AI

to to invented And the fabric computing Ethernet. have so BlueField go It for adds well Ethernet that It optimized as with the which who we is with with It's we go all platform computing excellent and an results. And market it's but achieve way capabilities with doesn't extends solution. is offer we end-to-end the as to our XXX% some to allows the and solution where large InfiniBand, Spectrum not market already compliant what our in Ethernet, our this we've done is. partners enterprise east-west traffic, switch new we of that some, us perform replace is we Ethernet. that

the stack, integrate -- the take AI be And AI over end-to-end sales Spectrum market now have switch, their to and stack. offer integrated, their at AI Enterprise NVIDIA well to fully so they HP, And all -- as solution. software Dell the and bundle as and least BlueField in fully they'll NVIDIA Lenovo the resellers optimized you will, if vast force enterprise able customers vast of network a with world with

for bringing basically the so world's And that's Ethernet AI enterprise. to

Stanley. question of Moore next Joe from Your the comes of line Morgan

and that talk do a Great. how you you bit TAM versus you if expander. of using Hopper see leverage to as kind could I you Grace Grace microprocessor, And see about more how wondered a the HXXX the see traditional more ability what little applications? Hopper applications

question. Hopper the We're is your production design Grace high-performance that AI Thanks of in AI, have with in high-volume next computing all production expecting now. infrastructures. just year and Yes. wins we for in

very very data a very on very, so a to have the generative be line refer generative nodes has are has generation, databases is be could multibillion-dollar models length is high. We is to RAG, augmented a that AI data the CPU that going In with for able of ramp very ability you also generates This well a you is create proprietary product or applications or quite data called our very to factful what product fast or large. And first It quite capability a to it line. Grace context have retrieval really center vector before is simultaneously semantic memory spectacular. large where areas fast to that us. The as of computing response, Hopper search, as the large data model, could memory.

You you basically stored an entire before into its questions. ask book the memory system

factual to domain-specific quite models the reduce ability interact proprietary context hand, And data be the or be have to other refer length data, to large. hallucination. and your you to data on hand; contextually the on able generative so data, with one relevant This way, and be able the still could naturally be

than This Grace serves European And European the so to supercomputing CSPs who customers to just to quite are drives companies Hopper, decided ARM. a is variety ARM would CPU up the because that pivot build off of Grace build -- the of would that start. fantastic that their have a ecosystem own ARM their success but particular different own to based a Hopper. It for or that use reasons stack example, case, custom also like really to really CPUs we're Maybe an is like that or have it's a and up run product. for care centers like to different xXX. they extraordinary home on There are

from Your comes next UBS. line Arcuri question of of the Tim

to bit wanted about a went a parts. But again. lot when the ask one can commitments you other on that up moving there's would few bands China. hand, the on arguably on demand in the hand, the visibility beyond pull little guess, a fill purchase I revenue. I the you I China have know

doesn't peak year not be that. into center ask of a even So explicitly wanted into it center think grow you Do in sound XXXX XXXX? just and would I yet, know that data even sort Jensen, to but you your next can revenue, think that we're I data like, you

Absolutely there can through of center And believe data course, several that reasons are, that. XXXX. for grow

We significantly. our quite supply are expanding

We in HGX, broadest the chips chain have GPU ] weighs and chip, largest XX,XXX world. Remember, but and are most pounds. the Hopper. wanted capable the has [ the the HGX, Hopper is people that HXXX, It of Eight a already supply the think XX parts.

-- test therefore, The of high-frequency is not. complicated. are components another the it it complicated only XX,XXX complicated. is way is of supercomputer parts. of incredible, high-current installation parts, supercomputer's it supercomputer. It Even the to other a parts, the -- is The is complicated. it are even passive testing its It and has with The shipping high-voltage is manufacturing and

the out is has mention the are we HGXs have complexity remarkable just so incredible. our is of supply chain aspect NVIDIA the that scaled the the NICs, networking, the amount our complicated. And all incredibly, really supply HGX cables, to switches, and of the with networking, connected of transceivers, the not team And there chain here every

scaling first team so incredible just super I'm just And this the of up supply for chain. -- proud I'm all, of

absolutely customers new we're products. are meanwhile, and world adding class. But We new

have we new supply. So

was can should. They are have AI. somebody fairly them. We specialist clouds earlier. back mentioning I country's them have us, culture combination, resell They build to the national we to with for AI they up else their to then to have in AI and afford that do that, as help skills up the their -- country's as people clouds, up realize they new over customers, GPU can't export all knowledge, world regions Different sovereign their standing from coming

the cloud, their do that national AI so first is And create AI have thing cloud. they to

enterprise us could enterprise. now say seeing X -- market also paths. or growing paths. has I path One X if into You're The

course, off-the-shelf afford to There's There'll sitting is fabulous assistants to have a custom And SAP, they is proprietary domain-specific about somebody ServiceNow And off-the-shelf top they and because building else. software build companies on rent be mine, like can be Almost outsourced to their their recognize proprietary of ChatGPT sitting And first They tools. are company AI. -- of every AI base. This go others. gold are company's intelligence path, there the many is The a these the AI. of can't own course, that others, also -- gold should and their are their have many, and there custom on is tools, a tools and in to on they build world customer then that. copilots and major mine. they're they AIs. and

NVIDIA's And the the our partners, on custom serve is with own custom software all which serve leverage we market them of our world. we're through Enterprise, going marketplace RAGs. switches AI Dell own called that. And service to chatbots, an running so where in custom their called forth. HP, enterprises this their then AI have We own one Foundry, be is that capability their to And NVIDIA includes next networking systems, entire Lenovo, compute, new spreading the that and able capabilities building taking is it and AIs, over the our and our to a way so our stacks stack

you seeing And consumer and to have then enterprise moving AI lot really, to CSPs -- companies. do Internet that comes generative start-ups platforms, companies, one we're Omniverse enterprise moving so of areas and that -- really ultimately, is industrial starting seen energy a of and to us the to moving from the spend And of guys waves AI generative the And where NVIDIA on AI. together. This exciting is a work. just has software with NVIDIA we're

accelerated I -- to affect we're think every to going basically AI, computing. country. beginning board, every industrial This company, across industry, is at the generative the And the of so every a, to transition

Goldman next from of line comes Sachs. Toshiya Your question of Hari the

it. that over the QX of should Jensen, data is announcement for wanted next And then going model last center beyond? for and ask be revenue? I you about mentioned the Or a a have timing limited. limited? And months, it primarily contribution on for to April issue? yet such from with be about hopefully I to contribution relatively well. this could just monetization it products partners to question service growth and real going be wanted Microsoft? and that we foundry Is have should you had set? are week, quick, source regulation-compliant be Colette, you the How relatively Or couple software Is reacceleration revenue revenue be services that the is expand do the Jensen to AI well? expand then a as in in the price to and the exclusive going And plans points other I introducing to work? to to should How you'll question long-term as something that of Colette clarify opportunity forward to think the

Thanks, potentially regarding could China products our the significant products. develop and It's provide design question customers. new process Toshiya, these a new both to we that to on

U.S. discussed, discussions sure we in with government that As these intent as our to make in full we well. are of products we're the going

through we in customers just with to where already that us several products quarter. these the time weeks the about state X have. their our our take Given into to we're some of are go and and desires we discussing for That's quarter, needs going

our term can U.S. of produce medium the long it's what ideas whether government with and of that's term, interest the hard Moving with to we both China or forward, just say customers. the

customers, to this that we it's but China on our time. hard balance So still focused stay say finding right at for

in foundry, the is and for AI There world thanks glaring for Toshiya, a so question. the opportunity sense. much makes it

want people whole going AIs has tools a to platforms tools. with the And of up makes by platforms. our data, expertise. out and tool proprietary these They build our world used software in their we are AI. First, bunch we the outsource the their AI. that company the see. companies And and Getty. for already we've companies own our AI case companies, And SAP, every ServiceNow, you'll because be domain just out of their company, those hire. And hand intelligence. hand them. the they reason proprietary and used most data They people AI across And is coming. flywheel by that Dropbox, go their are of the got today. its world, their own announced are a can't create In to core Many afford it's And to for in augmented for other have many It intelligence the to few, future, others

come they us. to so And

TSMC We have really several in as just that essential is foundry. the foundry things are a

third the that's need invention create you on practice, on so have And depth you to AI best that You technology. of then through and are just And safe, what Cloud are technology that thing have processing forth. And of have capability, is have an so as AI of AI AIs the skills you have we incredible the second, models so practice, and forth is. to know, fine-tuned, to capability. -- factories, guardrails, DGX data known and the so AI

AI Our are called models AI Foundations.

are they will, you our DGX factories creating process, we call and Our for NVIDIA's system called on NeMo, CAD run if AIs Cloud.

custom model they on bring We their custom NVIDIA. it a We monetization domain is partners, bring bring We'd Cloud everything help their our them one build we their engineers. across AI that, work data. they run AI. on that of They our a deploy that or customer enterprise that with becomes each AI Our optimized enterprise sandbox and and DGX theirs, They runs incredible. Then rent researchers time where is together. expertise. help grade, make optimized, outperformance them

in And that and patched, on-prem secure, base NVIDIA anywhere. we patched We Enterprise. cloud it's And optimized supported. and securely call the have constantly AI a giant installed

is year. NVIDIA model. That's our business AI $X,XXX Enterprise per per GP

model basically business Our a license. is

Our model per build ways, In customers license They -- on like They top of basic could ways a ours of They like is basically then, per based. have There's license, take to could of. have that or a retail. software their can become operating they subscription that could wholesale. with we're own they their monetization they a a lot license, but different per business model, could instance do system. usage. an create could lot

And business so create you help is model models. your custom our

AI very Enterprise. business start. those NVIDIA run to for custom is models to NVIDIA large You a And Enterprise it's going on a us. off great AI be

Stacy Your next question line the comes from Bernstein of Research. Rasgon of

you've gone along or weren't it And the China in a are if those available times just to know, would you higher parts that QX can guide stuff would as are does supply And data the give restrictions, it to us feeling your guess reshipping now got lead right China the been constrained lines, redirection elsewhere? are for times you immediately for is China this, lowering where to in have such I wanted I those ship? that center? Colette, have lead sort because of

each if of job China, no solid see longer to sold still improving sure, the have are Yes, they supply revenue. But every have done both on our not but help some really ramping there situations could things working can supply where our are quarter, China and I me we that a let are which understand. every you we Stacy, that we absence been for Yes. quarter. constrained our defined outlook QX, of could with has We've for can. we

have higher our could guidance Yes. QX? been little a in our So

We as next are continuing plan still throughout and supply and that. growing our working on improving on all towards well year

Your the line Ramsay Cowen. next of question comes of from Matt TD

dominant inference, more domain training NVIDIA's more of and investors TAM takes from is still get that of I Congrats premise. two-part everybody one as the on will of as being even for the AI the somehow competitive. thinking that become and you, sort large the premise results. And had inference a I lot of model question questions and about off a comes Jensen, it market more

et be less cetera, differentiated, cetera. et You'll

processing as So data time maybe you about is I the little the spend move parts the evolution really company at workload. X bit are, all one, a your and of smaller inference. LLMs or to model a X I for never any of got that how questions piece could until positioned than the the inference rather talking guess second, we question ago, AI of And of about month up workload number the

the the of of and large how pieces now, enabling between manipulating training, could before training those workload is inference, workload. of and I you acceleration think data after So pieces maybe about the part of the that's a inference, CUDA talk

is inference complicated. Inference, Sure. It's incredibly actually complicated.

quarter, If announced one The LLM. we, new TensorRT exciting been called incredible. this engines, compilers you -- has the of reception optimizing most

the all It's downloaded for reasons stars whole there over GitHub. been and obviously. lot almost are frameworks stacks ton, to integrated go that, into And a world a several You of instantaneously.

our stacks on would is same we on GPU, X. And so without We innovation high. programmable, CUDA's TensorRT us for of the the software that factor improve If not it TensorRT to of by that, anybody touching anything, top by pace be another of and HXXX increases were course, factor do. so the at GPUs really because our improves it hard could LLM then performance of of the X. pace create CUDA programmable. LLM a

performance, by inference factor cost, reduced X saying of year's our inference so a And a just another way within about time. of

keep with. And so hard to really, up really that's

base. our base we've installed installed dedicated to years, XX-plus years. likes our is been why everybody because inference And the engine XX Now for reason our

by the a It's reason of technologies certainty have that if compatibility stability. all platform you that installed and everybody It's GPU, companies the make they're see millions and the is large from why our runs the that stack. world, optimizes the our the All workstation of very and you about large disciplined something not used XXX been is reason available architecture first million our every the base the NVIDIA's everybody And a has we've will architecturally it, that to conveyed dedicated industry. top -- It's all for one. And in enterprise the build it that us architecturally first. about job benefits just largest stability, on GPUs, everybody you We us long certainty very of just installed to you cloud, single people's compatible. time. PCs, all an We're every GPUs such world do, any -- invention is base about millions and platform every in have of every to and -- on it. accrues on an It's in uses only engineering maker. and GPUs why cloud. every We we in NVIDIA builds time compatible. -- from of NVIDIA system work

could basically so base. Data provider. provider, for inference you And you're installed and to data. an application you an processing, inference looking if And train as application, large curate an a platform model, are you're deploying software a application before you you're the a have

augment you data. You to Maybe the the data dedupe have data. with to have synthetic

So you process the the data. the -- data, clean the data, normalize align

not is XX%, data that could quite ultimately you work and significant. amount amount that the do. and measured XX% represent you what a service. megabytes. in learning measured In -- before in of ultimately creating engineering, is of terabytes you training of All the in that bytes petabytes. of XX%, data data-driven processing It's It you do And before data that do machine

And just massive accelerate accelerate processing Python. Spark. so part. We We data is a

science cuDF in is of without NVIDIA accelerated the is now pandas, world, coolest data One pandas, of one called without CUDA is line framework and did code. of which the just line successful single most things Pandas a by code, that of the box we out the that just

one for purpose really people pandas for purpose is terrific, about. really science. acceleration designed And processing one quite the so And incredibly data and excited are data just only, was and

gives CUDA that. of you NVIDIA so And all

Harlan from of JPMorgan. Your of comes final line the Sur question

If road as importantly, to about levels successful silicon, always looks team announced higher of and drive DPU center the support cadence X product more The of and years data on think GPU, been recently growth forward inferencing, more to doing that aggressive hardware, now execution through focused ecosystem, at and history a tech been this every that software, what and is year partnerships expanding complexity? the from R&D CPU, have segmentation manage OpEx over optimization, But do more networking. have time. aggressive with How segmentation, an to your map? right, just team we the training, those all of importantly, you new new strong companies a cadence aggressive industry, look of products, more right,

Gosh, boy, excellent. that's just really

NVIDIA's you described our wrote just plan. And You business strategy.

customers combination we reduced there all, a reason our the LLM so benefits it -- by but down for fundamentally fundamental model and reason software software. is for cost TensorRT the because feeds, of because why cost. factor architecture. accelerate inference of of because execution. is HXXX our that mostly, that of of First the it's drives And When the speeds includes, and our our a the Mostly, course, large for and of X,

our accelerate so that And want road we map for to reason.

process AI to not in of -- market single with all reason cloud, data this work thing. number reach is we generative not is then them provider We their of is mention, is plane and put is done center We world's but factories of of We enterprise amazing every of all The their to has AI of security networking and something NVIDIA that. and ever expand cloud nobody service Bringing going enterprise configurations, we're something with we them. And can to we're And yet put is one is in the AI. different, take whole posture NVIDIA second supercomputers. stacks, and the the is of every The the the generative can enterprise. same. integrated platform. into platform all of now with working before. same. kind cloud one to into AI right to all of not is centers, our take incredibly create Everybody's that control the we into their data stand-alone. them are well thing,

have library, surely, technology solutions, of takes this largest that and that enterprise. one When we The the the It software optimize the financial we of network libraries that you list decades without domain-specific the for centers. broad. And demand we When -- cloud, And ultimately, developers, provided. software of of that architecturally InfiniBand why for a and feature, because domain-specific and networking, solution networking, it thing is market those. supercomputing and when have a a on new or of map There's pace. new is the architecturally It a get instantly combination the runs every reach. a the and of about so that across now a quantum ago, broadest solutions an into and together, then The road market computer of thinking reason computing. and largest end-to-end every libraries When you benefit it. take domain-specific health is made that for it number translates -- our ecosystem in greatest it the stacks in for runs one that manufacturing. which in there's And their can create. integrate and to brings of reason feature that the AI, something NVIDIA finally, we every market optimized having permutation and new that markets it data complexity There's great xXX, ecosystem a is that -- company holds There's partnership technologies decision to for And reach. Ethernet course, that in in care. segments market makers, segments of distribution GPU, fact ARM, are we largest all demand, do new and compatible. includes, the of all they and -- of network, develop single on the course, demand we TensorRT really everything we incredibly is services we and we the InfiniBand energy. But all takes compatible developer, just includes the every GPU. lot demand system market includes language

NVIDIA discipline, large architecture reasons and that decades still XX,XXX company, around And of mean every that every than we're has more single just couple so serving that about lasted a compatible single the I one really, really why world. efficient. discipline market every now industry, is of people the is single

call over Thank Huang for you. I will to remarks. back closing the now turn Jensen

workloads AI assistants, today. and serve which the revolution. NVIDIA different in adding new created refining are steam and are strong to ServiceNow traditional tasks. consumer trillion need AI. center reflects AIs companies of movers. an training to language and AI generative supporting and AI building inference the software and expanding Adobe are Enterprise and next new and define installed creating to HXXX expansion run are the factory TAM industries Generative general accelerated service AI Internet productivity. The in CSPs Our AI are full and software first AI. data investment, and automate are SAP AI era waves the for the start-ups, clouds optimized factory, are AI industry $X boost stack transition computing build. the like a copilots regional to software their base, and the and companies significant incremental Dropbox, platform generating and local purpose Large driving data Nations has center data industrial significant and enterprises to and from The providers world's assistants workloads to copilots for global factories starting and center cloud data model and broad AI IT empowering largest And HGX InfiniBand NVIDIA with are AI is factory legacy platforms. infrastructure type custom AI demand. growth

billion. elements NVIDIA's world's full for CPU. for now networking Grace demand, CPU stride, help exceeds their CPU, are are annualized chain AI we InfiniBand $XX of data solution proprietary networking growth the line up and and AI of can rate. is full supply the SuperNIC, enterprises. NVIDIA computing, platforms run and wave and and at AI production HPE building as that networking need racing positioned new build fivefold we X.Xx workloads. Irrespective strategy to AI and ramping drivers to end-to-end software growth week, customers BlueField platform Spectrum-X an of switch services. Grace world's and excellent and are the for adopt new like solution an Dell, operate first build product by X into our networking. next revenue factory. performance announced expand Spectrum-X boosts AI meet to to is also to software new hitting we year. have highlighted billion enterprises. choice, us Hopper an and to services deploy AI. We grew a build we standard Lenovo AI, bring $X Grace of SAP Ethernet annualized NVIDIA This multibillion-dollar Enterprise of run is AI. Ethernet Ethernet Enterprises and our center AI for to joined are in NVIDIA software a to the of the is track Ethernet need and the software for and to ServiceNow exit that NVIDIA custom an As factories. Enterprises and a software on ahead is AI year copilots. rate growth year-over-year networking also generative next the

AI the technology, custom expertise help build customers have scale We models. and to

growth essentially DGX AI deploy AI on and the data proprietary NVIDIA AI foundry. their engines NVIDIA's applications enterprise-grade NVIDIA GPUs, an Cloud and AI in CPUs, on all is AI foundry Enterprise, full With throttle. NVIDIA networking, NVIDIA services software Enterprise are

on forward you joining progress our us We to updating quarter. for look today. next Thanks

call. conference today's concludes This

You now disconnect. may