โ†

๐Ÿ“„ Earnings Call Transcript ๋ฒˆ์—ญ ๊ฒฐ๊ณผ

๐Ÿ“Š Presentation

Original Translation
NVIDIA Corporation (NASDAQ:NVDA) Q4 2025 Earnings Conference Call February 26, 2025 5:00 PM ET

Company Participants

Stewart Stecker - IR
Jensen Huang - President and CEO
Colette Kress - EVP and CFO

Conference Call Participants

C.J. Muse - Cantor Fitzgerald
Joe Moore - Morgan Stanley
Vivek Arya - Bank of America Securities
Harlan Sur - JPMorgan
Timothy Arcuri - UBS
Ben Reitzes - Melius Research
Mark Lipacis - Evercore ISI
Aaron Rakers - Wells Fargo
Atif Malik - Citigroup

Operator

Good afternoon. My name is Krista and I will be your conference operator today. At this time, I would like to welcome everyone to NVIDIA's Fourth Quarter Earnings Call.
NVIDIA Corporation (NASDAQ:NVDA) 2025๋…„ 4๋ถ„๊ธฐ ์‹ค์  ์ปจํผ๋Ÿฐ์Šค ์ฝœ 2025๋…„ 2์›” 26์ผ ์˜คํ›„ 5:00 ET

ํšŒ์‚ฌ ์ฐธ์„์ž

Stewart Stecker - IR
Jensen Huang - ์‚ฌ์žฅ ๊ฒธ CEO
Colette Kress - EVP ๊ฒธ CFO

์ปจํผ๋Ÿฐ์Šค ์ฝœ ์ฐธ์„์ž

C.J. Muse - Cantor Fitzgerald
Joe Moore - Morgan Stanley
Vivek Arya - Bank of America Securities
Harlan Sur - JPMorgan
Timothy Arcuri - UBS
Ben Reitzes - Melius Research
Mark Lipacis - Evercore ISI
Aaron Rakers - Wells Fargo
Atif Malik - Citigroup

์˜คํผ๋ ˆ์ดํ„ฐ

์•ˆ๋…•ํ•˜์„ธ์š”. ์ €๋Š” ํฌ๋ฆฌ์Šคํƒ€์ด๋ฉฐ ์˜ค๋Š˜ ์ปจํผ๋Ÿฐ์Šค ์ง„ํ–‰์„ ๋งก๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ์‹œ๊ฐ„์— NVIDIA 4๋ถ„๊ธฐ ์‹ค์  ๋ฐœํ‘œ ์ปจํผ๋Ÿฐ์Šค ์ฝœ์— ์ฐธ์„ํ•ด ์ฃผ์‹  ๋ชจ๋“  ๋ถ„๋“ค์„ ํ™˜์˜ํ•ฉ๋‹ˆ๋‹ค.
All lines have been placed on mute to prevent any background noise. After the speakersโ€™ remarks, there will be a question-and-answer session. [Operator Instructions]

Thank you. Stewart Stecker, you may begin your conference. Stewart Stecker

Thank you. Good afternoon, everyone, and welcome to NVIDIA's conference call for the fourth quarter of fiscal 2025. With me today from NVIDIA are Jensen Huang, President and Chief Executive Officer and Colette Kress, Executive Vice President and Chief Financial Officer. I'd like to remind you that our call is being webcast live on NVIDIA's Investor Relations website.
๋ชจ๋“  ํšŒ์„ ์ด ๋ฐฐ๊ฒฝ ์†Œ์Œ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์Œ์†Œ๊ฑฐ ์ƒํƒœ๋กœ ์„ค์ •๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฐœํ‘œ์ž๋“ค์˜ ๋ฐœ์–ธ ํ›„์—๋Š” ์งˆ์˜์‘๋‹ต ์‹œ๊ฐ„์ด ์žˆ์„ ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. [์šด์˜์ž ์•ˆ๋‚ด์‚ฌํ•ญ]

๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Stewart Stecker๋‹˜, ์ปจํผ๋Ÿฐ์Šค๋ฅผ ์‹œ์ž‘ํ•˜์…”๋„ ๋ฉ๋‹ˆ๋‹ค.

Stewart Stecker

๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์•ˆ๋…•ํ•˜์„ธ์š” ์—ฌ๋Ÿฌ๋ถ„, 2025 ํšŒ๊ณ„์—ฐ๋„ 4๋ถ„๊ธฐ NVIDIA ์ปจํผ๋Ÿฐ์Šค ์ฝœ์— ์˜ค์‹  ๊ฒƒ์„ ํ™˜์˜ํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋Š˜ NVIDIA์—์„œ ์ €์™€ ํ•จ๊ป˜ํ•˜์‹  ๋ถ„๋“ค์€ Jensen Huang ์‚ฌ์žฅ ๊ฒธ ์ตœ๊ณ ๊ฒฝ์˜์ž์™€ Colette Kress ๋ถ€์‚ฌ์žฅ ๊ฒธ ์ตœ๊ณ ์žฌ๋ฌด์ฑ…์ž„์ž์ž…๋‹ˆ๋‹ค. ์ €ํฌ ์ปจํผ๋Ÿฐ์Šค ์ฝœ์ด NVIDIA ํˆฌ์ž์ž ๊ด€๊ณ„ ์›น์‚ฌ์ดํŠธ์—์„œ ์‹ค์‹œ๊ฐ„์œผ๋กœ ์›น์บ์ŠคํŠธ๋˜๊ณ  ์žˆ์Œ์„ ์•Œ๋ ค๋“œ๋ฆฝ๋‹ˆ๋‹ค.
The webcast will be available for replay until the conference call to discuss our financial results for the first quarter of fiscal 2026. The content of today's call is NVIDIA's property. It can't be reproduced or transcribed without prior written consent. During this call, we may make forward-looking statements based on current expectations. These are subject to a number of significant risks and uncertainties, and our actual results may differ materially.์›น์บ์ŠคํŠธ๋Š” 2026 ํšŒ๊ณ„์—ฐ๋„ 1๋ถ„๊ธฐ ์žฌ๋ฌด์‹ค์ ์„ ๋…ผ์˜ํ•˜๋Š” ์ปจํผ๋Ÿฐ์Šค ์ฝœ๊นŒ์ง€ ์žฌ์‹œ์ฒญ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์˜ค๋Š˜ ์ฝœ์˜ ๋‚ด์šฉ์€ NVIDIA์˜ ์ž์‚ฐ์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ์„œ๋ฉด ๋™์˜ ์—†์ด๋Š” ๋ณต์ œํ•˜๊ฑฐ๋‚˜ ์ „์‚ฌํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋ฒˆ ์ฝœ์—์„œ ์ €ํฌ๋Š” ํ˜„์žฌ์˜ ๊ธฐ๋Œ€์น˜๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๋ฏธ๋ž˜์ „๋ง์ง„์ˆ ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง„์ˆ ๋“ค์€ ๋‹ค์ˆ˜์˜ ์ค‘๋Œ€ํ•œ ์œ„ํ—˜๊ณผ ๋ถˆํ™•์‹ค์„ฑ์„ ์ˆ˜๋ฐ˜ํ•˜๋ฉฐ, ์ €ํฌ์˜ ์‹ค์ œ ๊ฒฐ๊ณผ๋Š” ํฌ๊ฒŒ ๋‹ค๋ฅผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
For discussion of factors that could affect our future financial results and business, please refer to the disclosure in today's earnings release, our most recent Forms 10-K and 10-Q, and the reports that we may file on Form 8-K with the Securities and Exchange Commission. All our statements are made as of today, February 26, 2025, based on information currently available to us. Except as required by law, we assume no obligation to update any such statements. During this call, we will discuss non-GAAP financial measures. You can find a reconciliation of these non-GAAP financial measures to GAAP financial measures in our CFO commentary, which is posted on our website.๋ฏธ๋ž˜ ์žฌ๋ฌด ์„ฑ๊ณผ์™€ ์‚ฌ์—…์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋Š” ์š”์ธ๋“ค์— ๋Œ€ํ•œ ๋…ผ์˜๋Š” ์˜ค๋Š˜ ๋ฐœํ‘œ๋œ ์‹ค์  ๋ฐœํ‘œ์ž๋ฃŒ, ๋‹น์‚ฌ์˜ ์ตœ์‹  Form 10-K ๋ฐ 10-Q, ๊ทธ๋ฆฌ๊ณ  ์ฆ๊ถŒ๊ฑฐ๋ž˜์œ„์›ํšŒ์— Form 8-K๋กœ ์ œ์ถœํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด๊ณ ์„œ๋“ค์˜ ๊ณต์‹œ ๋‚ด์šฉ์„ ์ฐธ์กฐํ•ด ์ฃผ์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๋‹น์‚ฌ์˜ ๋ชจ๋“  ์ง„์ˆ ์€ ํ˜„์žฌ ์ด์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์˜ค๋Š˜ 2025๋…„ 2์›” 26์ผ ๊ธฐ์ค€์œผ๋กœ ์ž‘์„ฑ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฒ•๋ฅ ์—์„œ ์š”๊ตฌํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ์ œ์™ธํ•˜๊ณ , ๋‹น์‚ฌ๋Š” ์ด๋Ÿฌํ•œ ์ง„์ˆ ์„ ์—…๋ฐ์ดํŠธํ•  ์˜๋ฌด๋ฅผ ์ง€์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋ฒˆ ์ปจํผ๋Ÿฐ์Šค ์ฝœ์—์„œ๋Š” ๋น„GAAP ์žฌ๋ฌด์ง€ํ‘œ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋น„GAAP ์žฌ๋ฌด์ง€ํ‘œ์™€ GAAP ์žฌ๋ฌด์ง€ํ‘œ ๊ฐ„์˜ ์กฐ์ •ํ‘œ๋Š” ๋‹น์‚ฌ ์›น์‚ฌ์ดํŠธ์— ๊ฒŒ์‹œ๋œ CFO ์ฝ”๋ฉ˜ํ„ฐ๋ฆฌ์—์„œ ํ™•์ธํ•˜์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
With that, let me turn the call over to Colette. Colette Kress

Thanks, Stewart. Q4 was another record quarter. Revenue of $39.3 billion was up 12% sequentially and up 78% year-on-year, and above our outlook of $37.5 billion. For fiscal 2025, revenue was $130.5 billion, up 114% in the prior year. Let's start with data center. Data Center revenue for fiscal 2025 was $115.2 billion, more than doubling from the prior year. In the fourth quarter, data center revenue of $35.6 billion was a record of 16% sequentially and 93% year-on-year. As the Blackwell ramp commenced and Hopper 200 continued sequential growths. In Q4, Blackwell sales exceeded our expectations.
์ด์ œ ์ฝœ๋ ˆํŠธ์—๊ฒŒ ๋งˆ์ดํฌ๋ฅผ ๋„˜๊ธฐ๊ฒ ์Šต๋‹ˆ๋‹ค. ์ฝœ๋ ˆํŠธ ํฌ๋ ˆ์Šค

๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค, ์ŠคํŠœ์–ดํŠธ. 4๋ถ„๊ธฐ๋Š” ๋˜ ๋‹ค๋ฅธ ๊ธฐ๋ก์ ์ธ ๋ถ„๊ธฐ์˜€์Šต๋‹ˆ๋‹ค. ๋งค์ถœ์€ 393์–ต ๋‹ฌ๋Ÿฌ๋กœ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ 12%, ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ 78% ์ฆ๊ฐ€ํ–ˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ์˜ ์ „๋ง์น˜์ธ 375์–ต ๋‹ฌ๋Ÿฌ๋ฅผ ์ƒํšŒํ–ˆ์Šต๋‹ˆ๋‹ค. 2025 ํšŒ๊ณ„์—ฐ๋„ ๋งค์ถœ์€ 1,305์–ต ๋‹ฌ๋Ÿฌ๋กœ ์ „๋…„ ๋Œ€๋น„ 114% ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ์„ผํ„ฐ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. 2025 ํšŒ๊ณ„์—ฐ๋„ ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋งค์ถœ์€ 1,152์–ต ๋‹ฌ๋Ÿฌ๋กœ ์ „๋…„ ๋Œ€๋น„ ๋‘ ๋ฐฐ ์ด์ƒ ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. 4๋ถ„๊ธฐ ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋งค์ถœ์€ 356์–ต ๋‹ฌ๋Ÿฌ๋กœ ๊ธฐ๋ก์„ ๊ฒฝ์‹ ํ•˜๋ฉฐ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ 16%, ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ 93% ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ธ”๋ž™์›ฐ ๋žจํ”„์—…์ด ์‹œ์ž‘๋˜๊ณ  ํ˜ธํผ 200์ด ์ง€์†์ ์ธ ์ˆœ์ฐจ ์„ฑ์žฅ์„ ์ด์–ด๊ฐ€๋ฉด์„œ ์ด๋Ÿฌํ•œ ์„ฑ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. 4๋ถ„๊ธฐ์— ๋ธ”๋ž™์›ฐ ๋งค์ถœ์€ ์šฐ๋ฆฌ์˜ ๊ธฐ๋Œ€์น˜๋ฅผ ๋„˜์–ด์„ฐ์Šต๋‹ˆ๋‹ค.
We delivered $11 billion of Blackwell revenue to meet strong demand. This is the fastest product ramp in our company's history, unprecedented in its speed and scale. Blackwell production is in full gear across multiple configurations, and we are increasing supply quickly in expanding customer adoption. Our Q4 data center compute revenue jumped 18% sequentially and over 2 times year-on-year. Customers are racing to scale infrastructure to train the next generation of cutting edge models and unlock the next level of AI capabilities. With Blackwell, it will be common for these clusters to start with 100,000 GPUs or more. Shipments have already started for multiple infrastructures of this size.๊ฐ•๋ ฅํ•œ ์ˆ˜์š”๋ฅผ ์ถฉ์กฑํ•˜๊ธฐ ์œ„ํ•ด 110์–ต ๋‹ฌ๋Ÿฌ์˜ Blackwell ๋งค์ถœ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋‹น์‚ฌ ์—ญ์‚ฌ์ƒ ๊ฐ€์žฅ ๋น ๋ฅธ ์ œํ’ˆ ๋žจํ”„์—…์œผ๋กœ, ๊ทธ ์†๋„์™€ ๊ทœ๋ชจ ๋ฉด์—์„œ ์ „๋ก€๊ฐ€ ์—†๋Š” ์ˆ˜์ค€์ž…๋‹ˆ๋‹ค. Blackwell ์ƒ์‚ฐ์€ ๋‹ค์–‘ํ•œ ๊ตฌ์„ฑ์œผ๋กœ ๋ณธ๊ฒฉ ๊ฐ€๋™ ์ค‘์ด๋ฉฐ, ๊ณ ๊ฐ ์ฑ„ํƒ ํ™•์‚ฐ์— ๋”ฐ๋ผ ๊ณต๊ธ‰์„ ๋น ๋ฅด๊ฒŒ ๋Š˜๋ฆฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. 4๋ถ„๊ธฐ ๋ฐ์ดํ„ฐ์„ผํ„ฐ ์ปดํ“จํŠธ ๋งค์ถœ์€ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ 18%, ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ 2๋ฐฐ ์ด์ƒ ๊ธ‰์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณ ๊ฐ๋“ค์€ ์ฐจ์„ธ๋Œ€ ์ตœ์ฒจ๋‹จ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  AI ์—ญ๋Ÿ‰์˜ ๋‹ค์Œ ๋‹จ๊ณ„๋ฅผ ์‹คํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ์ธํ”„๋ผ ํ™•์žฅ ๊ฒฝ์Ÿ์„ ๋ฒŒ์ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Blackwell์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ํด๋Ÿฌ์Šคํ„ฐ๋“ค์ด 10๋งŒ ๊ฐœ ์ด์ƒ์˜ GPU๋กœ ์‹œ์ž‘ํ•˜๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ด ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ทœ๋ชจ์˜ ๋‹ค์ˆ˜ ์ธํ”„๋ผ์— ๋Œ€ํ•œ ์ถœํ•˜๊ฐ€ ์ด๋ฏธ ์‹œ์ž‘๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Post-training and model customization are fueling demand for NVIDIA infrastructure and software as developers and enterprises leverage techniques such as fine tuning, reinforcement learning, and distillation to tailor models for domain-specific use cases. Hugging Face alone hosts over 90,000 derivatives created from the Llama Foundation model. The scale of post-training and model customization is massive and can collectively demand orders of magnitude more compute than pre-training. Our inference demand is accelerating, driven by test time scaling and new reasoning models like OpenAIโ€™s o3, DeepSeek-R1, and Grok-3.์‚ฌํ›„ ํ›ˆ๋ จ๊ณผ ๋ชจ๋ธ ์ปค์Šคํ„ฐ๋งˆ์ด์ง•์ด NVIDIA ์ธํ”„๋ผ์™€ ์†Œํ”„ํŠธ์›จ์–ด์— ๋Œ€ํ•œ ์ˆ˜์š”๋ฅผ ๊ฒฌ์ธํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ์ž์™€ ๊ธฐ์—…๋“ค์ด ํŒŒ์ธํŠœ๋‹, ๊ฐ•ํ™”ํ•™์Šต, ์ฆ๋ฅ˜์™€ ๊ฐ™์€ ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ๋„๋ฉ”์ธ๋ณ„ ์‚ฌ์šฉ ์‚ฌ๋ก€์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ๋งž์ถคํ™”ํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ํ—ˆ๊น…ํŽ˜์ด์Šค๋งŒ ํ•ด๋„ Llama ํŒŒ์šด๋ฐ์ด์…˜ ๋ชจ๋ธ์—์„œ ํŒŒ์ƒ๋œ 9๋งŒ ๊ฐœ ์ด์ƒ์˜ ๋ชจ๋ธ์„ ํ˜ธ์ŠคํŒ…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌํ›„ ํ›ˆ๋ จ๊ณผ ๋ชจ๋ธ ์ปค์Šคํ„ฐ๋งˆ์ด์ง•์˜ ๊ทœ๋ชจ๋Š” ๋ฐฉ๋Œ€ํ•˜๋ฉฐ, ์ „์ฒด์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋ณด๋‹ค ๋ช‡ ๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ… ํŒŒ์›Œ๋ฅผ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ํƒ€์ž„ ์Šค์ผ€์ผ๋ง๊ณผ OpenAI์˜ o3, DeepSeek-R1, Grok-3์™€ ๊ฐ™์€ ์ƒˆ๋กœ์šด ์ถ”๋ก  ๋ชจ๋ธ๋“ค์— ์˜ํ•ด ์šฐ๋ฆฌ์˜ ์ถ”๋ก  ์ˆ˜์š”๊ฐ€ ๊ฐ€์†ํ™”๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Long thinking reasoning AI can require 100 times more compute per task compared to one-shot inferences. Blackwell was architected for reasoning AI inference. Blackwell supercharges reasoning AI models with up to 25 times higher token throughput and 20 times lower cost versus Hopper 100. It is revolutionary. Transformer engine is built for LLM and mixture of experts in front. And its NVLink domain delivers 14 times the throughput of PCIe Gen 5, ensuring the response time, throughput and cost efficiency needed to tackle the growing complexity of inference at scale. Companies across industries are tapping into NVIDIA's full-stack inference platform to boost performance and slash costs.๊ธด ์‚ฌ๊ณ  ์ถ”๋ก  AI๋Š” ์ผํšŒ์„ฑ ์ถ”๋ก  ๋Œ€๋น„ ์ž‘์—…๋‹น 100๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ…์„ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Blackwell์€ ์ถ”๋ก  AI ์ธํผ๋Ÿฐ์Šค๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Blackwell์€ Hopper 100 ๋Œ€๋น„ ์ตœ๋Œ€ 25๋ฐฐ ๋†’์€ ํ† ํฐ ์ฒ˜๋ฆฌ๋Ÿ‰๊ณผ 20๋ฐฐ ๋‚ฎ์€ ๋น„์šฉ์œผ๋กœ ์ถ”๋ก  AI ๋ชจ๋ธ์„ ๋Œ€ํญ ๊ฐ•ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ํ˜์‹ ์ ์ž…๋‹ˆ๋‹ค. Transformer ์—”์ง„์€ LLM๊ณผ ์ „๋ฉด์˜ ์ „๋ฌธ๊ฐ€ ํ˜ผํ•ฉ ๋ชจ๋ธ์„ ์œ„ํ•ด ๊ตฌ์ถ•๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  NVLink ๋„๋ฉ”์ธ์€ PCIe Gen 5 ๋Œ€๋น„ 14๋ฐฐ์˜ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ์ œ๊ณตํ•˜์—ฌ, ๋Œ€๊ทœ๋ชจ ์ถ”๋ก ์˜ ์ฆ๊ฐ€ํ•˜๋Š” ๋ณต์žก์„ฑ์„ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ํ•„์š”ํ•œ ์‘๋‹ต ์‹œ๊ฐ„, ์ฒ˜๋ฆฌ๋Ÿ‰ ๋ฐ ๋น„์šฉ ํšจ์œจ์„ฑ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ์‚ฐ์—…์˜ ๊ธฐ์—…๋“ค์ด ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ๋น„์šฉ์„ ๋Œ€ํญ ์ ˆ๊ฐํ•˜๊ธฐ ์œ„ํ•ด NVIDIA์˜ ํ’€์Šคํƒ ์ธํผ๋Ÿฐ์Šค ํ”Œ๋žซํผ์„ ํ™œ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Now tripled inference throughput and cut costs by 66%, using NVIDIA TensorRT for its screenshot feature. Perplexity sees 435 million monthly queries and reduced its inference costs 3 times with NVIDIA Triton Inference Server and TensorRT-LLM. Microsoft Bing achieved a 5x speed up and major TCO savings for visual search across billions of images with NVIDIA, TensorRT, and acceleration libraries. Blackwell has great demand for inference. Many of the early GB200 deployments are earmarked for inference, a first for a new architecture. Blackwell addresses the entire AI market from pre-training, post-training to inference across clouds, to on-premise, to enterprise.์ด์ œ NVIDIA TensorRT๋ฅผ ์‚ฌ์šฉํ•œ ์Šคํฌ๋ฆฐ์ƒท ๊ธฐ๋Šฅ์œผ๋กœ ์ถ”๋ก  ์ฒ˜๋ฆฌ๋Ÿ‰์„ 3๋ฐฐ ๋Š˜๋ฆฌ๊ณ  ๋น„์šฉ์„ 66% ์ ˆ๊ฐํ–ˆ์Šต๋‹ˆ๋‹ค. Perplexity๋Š” ์›” 4์–ต 3,500๋งŒ ๊ฑด์˜ ์ฟผ๋ฆฌ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋ฉฐ NVIDIA Triton Inference Server์™€ TensorRT-LLM์„ ํ†ตํ•ด ์ถ”๋ก  ๋น„์šฉ์„ 3๋ฐฐ ์ ˆ๊ฐํ–ˆ์Šต๋‹ˆ๋‹ค. Microsoft Bing์€ NVIDIA, TensorRT, ๊ทธ๋ฆฌ๊ณ  ๊ฐ€์† ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ํ†ตํ•ด ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์‹œ๊ฐ์  ๊ฒ€์ƒ‰์—์„œ 5๋ฐฐ์˜ ์†๋„ ํ–ฅ์ƒ๊ณผ ์ฃผ์š”ํ•œ TCO ์ ˆ๊ฐ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.

Blackwell์€ ์ถ”๋ก  ๋ถ„์•ผ์—์„œ ํฐ ์ˆ˜์š”๋ฅผ ๋ณด์ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ GB200 ๋ฐฐํฌ์˜ ์ƒ๋‹น ๋ถ€๋ถ„์ด ์ถ”๋ก ์šฉ์œผ๋กœ ํ• ๋‹น๋˜์—ˆ๋Š”๋ฐ, ์ด๋Š” ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜๋กœ์„œ๋Š” ์ฒ˜์Œ ์žˆ๋Š” ์ผ์ž…๋‹ˆ๋‹ค. Blackwell์€ ์‚ฌ์ „ ํ›ˆ๋ จ, ํ›„์† ํ›ˆ๋ จ๋ถ€ํ„ฐ ์ถ”๋ก ๊นŒ์ง€, ํด๋ผ์šฐ๋“œ์—์„œ ์˜จํ”„๋ ˆ๋ฏธ์Šค, ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ์ „์ฒด AI ์‹œ์žฅ์„ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
[HUDA's] (ph) programmable architecture accelerates every AI model and over 4,400 applications, ensuring large infrastructure investments against obsolescence in rapidly evolving markets. Our performance and pace of innovation is unmatched. We're driven to a 200 times reduction in inference costs in just the last 2 years. We deliver the lowest TCO and the highest ROI. And full stack optimizations for NVIDIA and our large ecosystem, including 5.9 million developers, continuously improve our customers' economics. In Q4, large CSPs represented about half of our data center revenue. And these sales increased nearly 2 times year-on-year.[HUDA์˜] ํ”„๋กœ๊ทธ๋ž˜๋จธ๋ธ” ์•„ํ‚คํ…์ฒ˜๋Š” ๋ชจ๋“  AI ๋ชจ๋ธ๊ณผ 4,400๊ฐœ ์ด์ƒ์˜ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๊ฐ€์†ํ™”ํ•˜์—ฌ, ๊ธ‰์†ํžˆ ์ง„ํ™”ํ•˜๋Š” ์‹œ์žฅ์—์„œ ๋Œ€๊ทœ๋ชจ ์ธํ”„๋ผ ํˆฌ์ž๋ฅผ ๋…ธํ›„ํ™”๋กœ๋ถ€ํ„ฐ ๋ณดํ˜ธํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์„ฑ๋Šฅ๊ณผ ํ˜์‹  ์†๋„๋Š” ํƒ€์˜ ์ถ”์ข…์„ ๋ถˆํ—ˆํ•ฉ๋‹ˆ๋‹ค. ์ง€๋‚œ 2๋…„๊ฐ„ ์ถ”๋ก  ๋น„์šฉ์„ 200๋ฐฐ ์ ˆ๊ฐํ•˜๋Š” ์„ฑ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ตœ์ € TCO(์ด์†Œ์œ ๋น„์šฉ)์™€ ์ตœ๊ณ  ROI(ํˆฌ์ž์ˆ˜์ต๋ฅ )๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  590๋งŒ ๋ช…์˜ ๊ฐœ๋ฐœ์ž๋ฅผ ํฌํ•จํ•œ NVIDIA์™€ ์šฐ๋ฆฌ์˜ ๋Œ€๊ทœ๋ชจ ์ƒํƒœ๊ณ„๋ฅผ ์œ„ํ•œ ํ’€์Šคํƒ ์ตœ์ ํ™”๋Š” ์ง€์†์ ์œผ๋กœ ๊ณ ๊ฐ๋“ค์˜ ๊ฒฝ์ œ์„ฑ์„ ๊ฐœ์„ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. 4๋ถ„๊ธฐ์— ๋Œ€ํ˜• CSP๋“ค์ด ์šฐ๋ฆฌ ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋งค์ถœ์˜ ์•ฝ ์ ˆ๋ฐ˜์„ ์ฐจ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด๋Ÿฌํ•œ ๋งค์ถœ์€ ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ ๊ฑฐ์˜ 2๋ฐฐ ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
Large CSPs were some of the first to stand up Blackwell with Azure, GCP, AWS, and OCI bringing GB200 systems to cloud regions around the world to meet surging customer demand for AI. Regional cloud hosting and video GPUs increased as a percentage of data center revenue, reflecting continued AI factory build-outs globally and rapidly rising demand for AI reasoning models and agents. Core, we've launched a 100, 000 GV200 cluster-based instance with NVLink Switch and Quantum2 InfiniBand. Consumer Internet Revenue Group 3x year-on-year driven by an expanding set of generative AI and deep learning use cases.๋Œ€ํ˜• CSP๋“ค์€ Blackwell์„ ์ตœ์ดˆ๋กœ ๋„์ž…ํ•œ ๊ธฐ์—…๋“ค ์ค‘ ์ผ๋ถ€์˜€์œผ๋ฉฐ, Azure, GCP, AWS, OCI๊ฐ€ ๊ธ‰์ฆํ•˜๋Š” ๊ณ ๊ฐ์˜ AI ์ˆ˜์š”๋ฅผ ์ถฉ์กฑํ•˜๊ธฐ ์œ„ํ•ด ์ „ ์„ธ๊ณ„ ํด๋ผ์šฐ๋“œ ์ง€์—ญ์— GB200 ์‹œ์Šคํ…œ์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ง€์—ญ ํด๋ผ์šฐ๋“œ ํ˜ธ์ŠคํŒ… ๋ฐ ๋น„๋””์˜ค GPU๊ฐ€ ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋งค์ถœ์—์„œ ์ฐจ์ง€ํ•˜๋Š” ๋น„์ค‘์ด ์ฆ๊ฐ€ํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์ „ ์„ธ๊ณ„์ ์ธ AI ํŒฉํ† ๋ฆฌ ๊ตฌ์ถ• ์ง€์†๊ณผ AI ์ถ”๋ก  ๋ชจ๋ธ ๋ฐ ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ๊ธ‰์†ํ•œ ์ˆ˜์š” ์ฆ๊ฐ€๋ฅผ ๋ฐ˜์˜ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. Core์—์„œ๋Š” NVLink Switch์™€ Quantum2 InfiniBand๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ 100,000๊ฐœ์˜ GV200 ํด๋Ÿฌ์Šคํ„ฐ ๊ธฐ๋ฐ˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์†Œ๋น„์ž ์ธํ„ฐ๋„ท ์ˆ˜์ต ๊ทธ๋ฃน์€ ํ™•์žฅ๋˜๋Š” ์ƒ์„ฑํ˜• AI ๋ฐ ๋”ฅ๋Ÿฌ๋‹ ์‚ฌ์šฉ ์‚ฌ๋ก€์— ํž˜์ž…์–ด ์ „๋…„ ๋Œ€๋น„ 3๋ฐฐ ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
These include recommender systems, vision language understanding, synthetic data, generation search, and agentic AI. For example, xAI is adopting the GB200 to train and inference its next generation of grog AI models. Meta's cutting-edge Andromeda advertising engine runs on NVIDIA's Grace Hopper Superchip, serving vast quantities of ads across Instagram, Facebook applications. Andromeda harnesses Grace Hopper's fast interconnect and large memory to boost inference throughput by 3x, enhanced ad personalization and deliver meaningful jumps in monetization and ROI.์ด๋Ÿฌํ•œ ๋ถ„์•ผ์—๋Š” ์ถ”์ฒœ ์‹œ์Šคํ…œ, ๋น„์ „ ์–ธ์–ด ์ดํ•ด, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ, ์ƒ์„ฑ ๊ฒ€์ƒ‰, ๊ทธ๋ฆฌ๊ณ  ์—์ด์ „ํ‹ฑ AI๊ฐ€ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, xAI๋Š” ์ฐจ์„ธ๋Œ€ Grok AI ๋ชจ๋ธ์˜ ํ›ˆ๋ จ๊ณผ ์ถ”๋ก ์„ ์œ„ํ•ด GB200์„ ๋„์ž…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Meta์˜ ์ตœ์ฒจ๋‹จ Andromeda ๊ด‘๊ณ  ์—”์ง„์€ NVIDIA์˜ Grace Hopper ์Šˆํผ์นฉ์—์„œ ๊ตฌ๋™๋˜์–ด Instagram๊ณผ Facebook ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์ „๋ฐ˜์— ๊ฑธ์ณ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ๊ด‘๊ณ ๋ฅผ ์„œ๋น„์Šคํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Andromeda๋Š” Grace Hopper์˜ ๊ณ ์† ์ธํ„ฐ์ปค๋„ฅํŠธ์™€ ๋Œ€์šฉ๋Ÿ‰ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ถ”๋ก  ์ฒ˜๋ฆฌ๋Ÿ‰์„ 3๋ฐฐ ํ–ฅ์ƒ์‹œํ‚ค๊ณ , ๊ด‘๊ณ  ๊ฐœ์ธํ™”๋ฅผ ๊ฐ•ํ™”ํ•˜๋ฉฐ, ์ˆ˜์ตํ™”์™€ ROI์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๋„์•ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Enterprise revenue increased nearly 2 times year on accelerating demand for model fine-tuning, RAG and Agentic AI workflows and GPU accelerated data processing. We introduced NVIDIA Llama Nemotron model family NIMs to help developers create and deploy AI agents across a range of applications including customer support, fraud detection and product supply chain and inventory management. Leading AI agent platform providers, including SAP and ServiceNow are among the first to use new models.๊ธฐ์—… ๋งค์ถœ์€ ๋ชจ๋ธ ํŒŒ์ธํŠœ๋‹, RAG ๋ฐ ์—์ด์ „ํ‹ฑ AI ์›Œํฌํ”Œ๋กœ์šฐ์™€ GPU ๊ฐ€์† ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ์— ๋Œ€ํ•œ ์ˆ˜์š” ์ฆ๊ฐ€๋กœ ์ „๋…„ ๋Œ€๋น„ ๊ฑฐ์˜ 2๋ฐฐ ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๊ฐœ๋ฐœ์ž๋“ค์ด ๊ณ ๊ฐ ์ง€์›, ์‚ฌ๊ธฐ ํƒ์ง€, ์ œํ’ˆ ๊ณต๊ธ‰๋ง ๋ฐ ์žฌ๊ณ  ๊ด€๋ฆฌ๋ฅผ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ AI ์—์ด์ „ํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๋ฐฐํฌํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•๋Š” NVIDIA Llama Nemotron ๋ชจ๋ธ ํŒจ๋ฐ€๋ฆฌ NIM์„ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. SAP์™€ ServiceNow๋ฅผ ํฌํ•จํ•œ ์ฃผ์š” AI ์—์ด์ „ํŠธ ํ”Œ๋žซํผ ์ œ๊ณต์—…์ฒด๋“ค์ด ์ƒˆ๋กœ์šด ๋ชจ๋ธ์„ ์ตœ์ดˆ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ธฐ์—…๋“ค ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค.
Health care leaders IQVIA, Illumina, Mayo Clinic and Arc Institute are using NVIDIA AI to speed drug discovery enhanced genomic research and Pioneer advanced health care services with generative and Agentic AI. As AI expands beyond the digital world, NVIDIA infrastructure and software platforms are increasingly being adopted to power robotics and physical AI development. One of the early and largest robotics applications and autonomous vehicles where virtually every AV company is developing on NVIDIA in the data center, the car or both. NVIDIA's automotive vertical revenue is expected to grow to approximately $5 billion this fiscal year.ํ—ฌ์Šค์ผ€์–ด ์„ ๋„๊ธฐ์—…์ธ IQVIA, Illumina, Mayo Clinic, Arc Institute๋Š” NVIDIA AI๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‹ ์•ฝ ๊ฐœ๋ฐœ์„ ๊ฐ€์†ํ™”ํ•˜๊ณ  ์œ ์ „์ฒดํ•™ ์—ฐ๊ตฌ๋ฅผ ๊ฐ•ํ™”ํ•˜๋ฉฐ, ์ƒ์„ฑํ˜• AI์™€ ์—์ด์ „ํ‹ฑ AI๋ฅผ ํ†ตํ•ด ์ฒจ๋‹จ ํ—ฌ์Šค์ผ€์–ด ์„œ๋น„์Šค๋ฅผ ๊ฐœ์ฒ™ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. AI๊ฐ€ ๋””์ง€ํ„ธ ์„ธ๊ณ„๋ฅผ ๋„˜์–ด ํ™•์žฅ๋จ์— ๋”ฐ๋ผ, NVIDIA์˜ ์ธํ”„๋ผ์™€ ์†Œํ”„ํŠธ์›จ์–ด ํ”Œ๋žซํผ์€ ๋กœ๋ณดํ‹ฑ์Šค์™€ ๋ฌผ๋ฆฌ์  AI ๊ฐœ๋ฐœ์„ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ์ ์  ๋” ๋งŽ์ด ์ฑ„ํƒ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ์ด์ž ๊ฐ€์žฅ ํฐ ๋กœ๋ณดํ‹ฑ์Šค ์‘์šฉ ๋ถ„์•ผ ์ค‘ ํ•˜๋‚˜์ธ ์ž์œจ์ฃผํ–‰์ฐจ ๋ถ„์•ผ์—์„œ๋Š” ๊ฑฐ์˜ ๋ชจ๋“  ์ž์œจ์ฃผํ–‰์ฐจ ํšŒ์‚ฌ๋“ค์ด ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋‚˜ ์ฐจ๋Ÿ‰, ๋˜๋Š” ์–‘์ชฝ ๋ชจ๋‘์—์„œ NVIDIA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐœ๋ฐœํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. NVIDIA์˜ ์ž๋™์ฐจ ๋ถ€๋ฌธ ๋งค์ถœ์€ ์˜ฌ ํšŒ๊ณ„์—ฐ๋„์— ์•ฝ 50์–ต ๋‹ฌ๋Ÿฌ๊นŒ์ง€ ์„ฑ์žฅํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
At CES, Hyundai Motor Group announced it is adopting NVIDIA technologies to accelerate AV and robotics development and smart factory initiatives. Vision transformers, self-supervised learning, multimodal sensor fusion and high fidelity simulation are driving breakthroughs in AV development and will require 10x more compute. At CES, we announced the NVIDIA COSMO World Foundation model platform. Just as language, foundation models have revolutionized Language AI, Cosmos is a physical AI to revolutionize robotics. The robotics and automotive companies, including ridesharing giant Uber, are among the first to adopt the platform.CES์—์„œ ํ˜„๋Œ€์ž๋™์ฐจ๊ทธ๋ฃน์€ ์ž์œจ์ฃผํ–‰์ฐจ์™€ ๋กœ๋ณดํ‹ฑ์Šค ๊ฐœ๋ฐœ, ๊ทธ๋ฆฌ๊ณ  ์Šค๋งˆํŠธ ํŒฉํ† ๋ฆฌ ์ด๋‹ˆ์…”ํ‹ฐ๋ธŒ๋ฅผ ๊ฐ€์†ํ™”ํ•˜๊ธฐ ์œ„ํ•ด NVIDIA ๊ธฐ์ˆ ์„ ์ฑ„ํƒํ•œ๋‹ค๊ณ  ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„์ „ ํŠธ๋žœ์Šคํฌ๋จธ, ์ž๊ธฐ์ง€๋„ํ•™์Šต, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์„ผ์„œ ํ“จ์ „, ๊ทธ๋ฆฌ๊ณ  ๊ณ ์ถฉ์‹ค๋„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์ด ์ž์œจ์ฃผํ–‰์ฐจ ๊ฐœ๋ฐœ์˜ ํ˜์‹ ์„ ์ฃผ๋„ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Š” 10๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ… ํŒŒ์›Œ๋ฅผ ํ•„์š”๋กœ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. CES์—์„œ ์ €ํฌ๋Š” NVIDIA COSMO World Foundation ๋ชจ๋ธ ํ”Œ๋žซํผ์„ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ธ์–ด ํŒŒ์šด๋ฐ์ด์…˜ ๋ชจ๋ธ์ด ์–ธ์–ด AI๋ฅผ ํ˜์‹ ์‹œํ‚จ ๊ฒƒ์ฒ˜๋Ÿผ, Cosmos๋Š” ๋กœ๋ณดํ‹ฑ์Šค๋ฅผ ํ˜์‹ ์‹œํ‚ฌ ๋ฌผ๋ฆฌ์  AI์ž…๋‹ˆ๋‹ค. ๋ผ์ด๋“œ์…ฐ์–ด๋ง ๋Œ€๊ธฐ์—… ์šฐ๋ฒ„๋ฅผ ํฌํ•จํ•œ ๋กœ๋ณดํ‹ฑ์Šค ๋ฐ ์ž๋™์ฐจ ํšŒ์‚ฌ๋“ค์ด ์ด ํ”Œ๋žซํผ์„ ์ตœ์ดˆ๋กœ ์ฑ„ํƒํ•˜๋Š” ๊ธฐ์—…๋“ค ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค.
From a geographic perspective, sequential growth in our Data Center revenue was strongest in the U.S., driven by the initial ramp up Blackwell. Countries across the globe are building their AI ecosystem as demand for compute infrastructure is surging. France's EUR 200 billion Euro AI investment and the EU's EUR 200 billion invest AI initiatives offer a glimpse into the build-out to set redefined global AI infrastructure in the coming years. Now as a percentage of total Data Center revenue, data center sales in China remained well below levels seen on the onset of export controls. Absent any change in regulations, we believe that China shipments will remain roughly at the current percentage.์ง€์—ญ๋ณ„ ๊ด€์ ์—์„œ ๋ณด๋ฉด, ๋‹น์‚ฌ ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋งค์ถœ์˜ ์ˆœ์ฐจ์  ์„ฑ์žฅ์€ ๋ฏธ๊ตญ์—์„œ ๊ฐ€์žฅ ๊ฐ•ํ–ˆ์œผ๋ฉฐ, ์ด๋Š” Blackwell์˜ ์ดˆ๊ธฐ ๋žจํ”„์—…์— ์˜ํ•ด ๊ฒฌ์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ปดํ“จํŒ… ์ธํ”„๋ผ์— ๋Œ€ํ•œ ์ˆ˜์š”๊ฐ€ ๊ธ‰์ฆํ•˜๋ฉด์„œ ์ „ ์„ธ๊ณ„ ๊ตญ๊ฐ€๋“ค์ด AI ์ƒํƒœ๊ณ„๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ”„๋ž‘์Šค์˜ 2,000์–ต ์œ ๋กœ AI ํˆฌ์ž์™€ EU์˜ 2,000์–ต ์œ ๋กœ AI ํˆฌ์ž ์ด๋‹ˆ์…”ํ‹ฐ๋ธŒ๋Š” ํ–ฅํ›„ ๋ช‡ ๋…„๊ฐ„ ๊ธ€๋กœ๋ฒŒ AI ์ธํ”„๋ผ๋ฅผ ์žฌ์ •์˜ํ•˜๋Š” ๊ตฌ์ถ• ์ž‘์—…์˜ ์ผ๋ฉด์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

์ „์ฒด ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋งค์ถœ ๋Œ€๋น„ ๋น„์œจ๋กœ ๋ณด๋ฉด, ์ค‘๊ตญ ๋‚ด ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋งค์ถœ์€ ์ˆ˜์ถœ ํ†ต์ œ ์‹œํ–‰ ์ดˆ๊ธฐ์— ๋ณด์˜€๋˜ ์ˆ˜์ค€์„ ํฌ๊ฒŒ ๋ฐ‘๋Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทœ์ œ์— ๋ณ€ํ™”๊ฐ€ ์—†๋Š” ํ•œ, ์ค‘๊ตญ ์ถœํ•˜๋Ÿ‰์€ ํ˜„์žฌ ๋น„์œจ ์ˆ˜์ค€์—์„œ ๋Œ€๋žต ์œ ์ง€๋  ๊ฒƒ์œผ๋กœ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค.
The market in China for data center solutions remains very competitive. We will continue to comply with export controls while serving our customers. Networking revenue declined 3% sequentially. Our networking attached to GPU compute systems is robust at over 75%. We are transitioning from small NVLink 8 with InfiniBand, to large NVLink 72 with Spectrum-X. Spectrum-X and NVLink Switch revenue increased and represents a major new growth vector. We expect networking to return to growth in Q1.AI requires a new class of networking. NVIDIA offers NVLink Switch systems for scale-up compute. For scale out, we offer quantum incentive for HPC supercomputers and Spectrum X for Ethernet environments.์ค‘๊ตญ์˜ ๋ฐ์ดํ„ฐ์„ผํ„ฐ ์†”๋ฃจ์…˜ ์‹œ์žฅ์€ ์—ฌ์ „ํžˆ ๋งค์šฐ ๊ฒฝ์Ÿ์ด ์น˜์—ดํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๊ณ ๊ฐ์—๊ฒŒ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๋ฉด์„œ ์ˆ˜์ถœ ํ†ต์ œ ๊ทœ์ •์„ ๊ณ„์† ์ค€์ˆ˜ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋„คํŠธ์›Œํ‚น ๋งค์ถœ์€ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ 3% ๊ฐ์†Œํ–ˆ์Šต๋‹ˆ๋‹ค. GPU ์ปดํ“จํŒ… ์‹œ์Šคํ…œ์— ์—ฐ๊ฒฐ๋œ ์ €ํฌ ๋„คํŠธ์›Œํ‚น์€ 75% ์ด์ƒ์œผ๋กœ ๊ฒฌ๊ณ ํ•œ ์„ฑ๊ณผ๋ฅผ ๋ณด์ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” InfiniBand๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ์†Œ๊ทœ๋ชจ NVLink 8์—์„œ Spectrum-X๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€๊ทœ๋ชจ NVLink 72๋กœ ์ „ํ™˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Spectrum-X์™€ NVLink Switch ๋งค์ถœ์ด ์ฆ๊ฐ€ํ–ˆ์œผ๋ฉฐ, ์ด๋Š” ์ฃผ์š”ํ•œ ์ƒˆ๋กœ์šด ์„ฑ์žฅ ๋™๋ ฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. 1๋ถ„๊ธฐ์— ๋„คํŠธ์›Œํ‚น์ด ์„ฑ์žฅ์„ธ๋กœ ๋Œ์•„์„ค ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•ฉ๋‹ˆ๋‹ค.

AI๋Š” ์ƒˆ๋กœ์šด ํด๋ž˜์Šค์˜ ๋„คํŠธ์›Œํ‚น์„ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. NVIDIA๋Š” ์Šค์ผ€์ผ์—… ์ปดํ“จํŒ…์„ ์œ„ํ•œ NVLink Switch ์‹œ์Šคํ…œ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์Šค์ผ€์ผ์•„์›ƒ์˜ ๊ฒฝ์šฐ, HPC ์Šˆํผ์ปดํ“จํ„ฐ๋ฅผ ์œ„ํ•œ ํ€€ํ…€ ์ธ์„ผํ‹ฐ๋ธŒ์™€ ์ด๋”๋„ท ํ™˜๊ฒฝ์„ ์œ„ํ•œ Spectrum X๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
Spectrum-X enhances the Ethernet for AI computing and has been a huge success. Microsoft Azure, OCI, CoreWeave and others are building large AI factories with Spectrum-X. The first Stargate data centers will use Spectrum-X. Yesterday, Cisco announced integrating Spectrum-X into their networking portfolio to help enterprises build AI infrastructure. With its large enterprise footprint and global reach, Cisco will bring NVIDIA Ethernet to every industry. Now moving to gaming and ARPCs. Gaming revenue of $2.5 billion decreased 22% sequentially and 11% year-on-year. Full year revenue of $11.4 billion increased 9% year-on-year, and demand remains strong throughout the holiday.Spectrum-X๋Š” AI ์ปดํ“จํŒ…์„ ์œ„ํ•œ ์ด๋”๋„ท์„ ํ–ฅ์ƒ์‹œํ‚ค๋ฉฐ ํฐ ์„ฑ๊ณต์„ ๊ฑฐ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. Microsoft Azure, OCI, CoreWeave ๋“ฑ์ด Spectrum-X๋กœ ๋Œ€๊ทœ๋ชจ AI ํŒฉํ† ๋ฆฌ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ Stargate ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋“ค์ด Spectrum-X๋ฅผ ์‚ฌ์šฉํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์–ด์ œ Cisco๋Š” ๊ธฐ์—…๋“ค์ด AI ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•๊ธฐ ์œ„ํ•ด Spectrum-X๋ฅผ ์ž์‚ฌ ๋„คํŠธ์›Œํ‚น ํฌํŠธํด๋ฆฌ์˜ค์— ํ†ตํ•ฉํ•œ๋‹ค๊ณ  ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๊ธฐ์—… ๊ณ ๊ฐ ๊ธฐ๋ฐ˜๊ณผ ๊ธ€๋กœ๋ฒŒ ๋„๋‹ฌ ๋ฒ”์œ„๋ฅผ ๋ณด์œ ํ•œ Cisco๋Š” NVIDIA ์ด๋”๋„ท์„ ๋ชจ๋“  ์‚ฐ์—…์— ์ œ๊ณตํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.

์ด์ œ ๊ฒŒ์ด๋ฐ๊ณผ ARPC๋กœ ๋„˜์–ด๊ฐ€๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ฒŒ์ด๋ฐ ๋งค์ถœ์€ 25์–ต ๋‹ฌ๋Ÿฌ๋กœ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ 22%, ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ 11% ๊ฐ์†Œํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ฐ„ ๋งค์ถœ์€ 114์–ต ๋‹ฌ๋Ÿฌ๋กœ ์ „๋…„ ๋Œ€๋น„ 9% ์ฆ๊ฐ€ํ–ˆ์œผ๋ฉฐ, ์—ฐ๋ง์—ฐ์‹œ ๊ธฐ๊ฐ„ ๋‚ด๋‚ด ์ˆ˜์š”๋Š” ๊ฐ•์„ธ๋ฅผ ์œ ์ง€ํ–ˆ์Šต๋‹ˆ๋‹ค.
However, Q4 shipments were impacted by supply constraints. We expect strong sequential growth in Q1 as supply increases. The new GeForce RTX 50 Series desktop and laptop GPUs are here. Build for gamers, creators and developers they fuse AI and graphics redefining visual computing, powered by the Blackwell architecture, fifth generation Tensor cores and fourth-generation RT cores and featuring up to 3,400 AI top. These GPUs deliver a 2x performance leap and new AI-driven rendering including neural shaders, digital human technologies, geometry and lighting. The new DLSS 4 boost frame rates up to 8 times with AI-driven frame generation, turning 1 rendered frame into 3.ํ•˜์ง€๋งŒ 4๋ถ„๊ธฐ ์ถœํ•˜๋Ÿ‰์€ ๊ณต๊ธ‰ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ์˜ํ–ฅ์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๊ณต๊ธ‰์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ 1๋ถ„๊ธฐ์—๋Š” ๊ฐ•๋ ฅํ•œ ์ˆœ์ฐจ์  ์„ฑ์žฅ์„ ๊ธฐ๋Œ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด GeForce RTX 50 ์‹œ๋ฆฌ์ฆˆ ๋ฐ์Šคํฌํ†ฑ ๋ฐ ๋…ธํŠธ๋ถ GPU๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฒŒ์ด๋จธ, ํฌ๋ฆฌ์—์ดํ„ฐ, ๊ฐœ๋ฐœ์ž๋ฅผ ์œ„ํ•ด ์ œ์ž‘๋œ ์ด ์ œํ’ˆ๋“ค์€ AI์™€ ๊ทธ๋ž˜ํ”ฝ์„ ์œตํ•ฉํ•˜์—ฌ ์‹œ๊ฐ ์ปดํ“จํŒ…์„ ์žฌ์ •์˜ํ•˜๋ฉฐ, Blackwell ์•„ํ‚คํ…์ฒ˜, 5์„ธ๋Œ€ Tensor ์ฝ”์–ด, 4์„ธ๋Œ€ RT ์ฝ”์–ด๋กœ ๊ตฌ๋™๋˜๊ณ  ์ตœ๋Œ€ 3,400 AI TOPS๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ GPU๋“ค์€ 2๋ฐฐ์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ๊ณผ ๋‰ด๋Ÿด ์…ฐ์ด๋”, ๋””์ง€ํ„ธ ํœด๋จผ ๊ธฐ์ˆ , ์ง€์˜ค๋ฉ”ํŠธ๋ฆฌ ๋ฐ ์กฐ๋ช…์„ ํฌํ•จํ•œ ์ƒˆ๋กœ์šด AI ๊ธฐ๋ฐ˜ ๋ Œ๋”๋ง์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด DLSS 4๋Š” AI ๊ธฐ๋ฐ˜ ํ”„๋ ˆ์ž„ ์ƒ์„ฑ์„ ํ†ตํ•ด ํ”„๋ ˆ์ž„ ์†๋„๋ฅผ ์ตœ๋Œ€ 8๋ฐฐ๊นŒ์ง€ ํ–ฅ์ƒ์‹œํ‚ค๋ฉฐ, ๋ Œ๋”๋ง๋œ 1๊ฐœ ํ”„๋ ˆ์ž„์„ 3๊ฐœ๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค.
It also features the industry's first real-time application of transformer models packing 2 times more parameters and 4 times to compute for unprecedented visual fidelity. We also announced a wave of GeForce Blackwell laptop GPUs with new NVIDIA Max-Q technology that extends battery life by up to an incredible 40%. These laptops will be available starting in March from the world's top manufacturers. Moving to our professional visualization business. Revenue of $511 million was up 5% sequentially and 10% year-on-year. Full year revenue of $1.9 billion increased 21% year-on-year. Key industry verticals driving demand include automotive and health care.๋˜ํ•œ ์—…๊ณ„ ์ตœ์ดˆ๋กœ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์˜ ์‹ค์‹œ๊ฐ„ ์ ์šฉ์„ ๊ตฌํ˜„ํ–ˆ์œผ๋ฉฐ, ์ด๋Š” 2๋ฐฐ ๋งŽ์€ ํŒŒ๋ผ๋ฏธํ„ฐ์™€ 4๋ฐฐ์˜ ์—ฐ์‚ฐ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•˜์—ฌ ์ „๋ก€ ์—†๋Š” ์‹œ๊ฐ์  ์ถฉ์‹ค๋„๋ฅผ ์‹คํ˜„ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ƒˆ๋กœ์šด NVIDIA Max-Q ๊ธฐ์ˆ ์„ ํƒ‘์žฌํ•œ GeForce Blackwell ๋…ธํŠธ๋ถ GPU ๋ผ์ธ์—…์„ ๋ฐœํ‘œํ–ˆ๋Š”๋ฐ, ์ด๋Š” ๋ฐฐํ„ฐ๋ฆฌ ์ˆ˜๋ช…์„ ์ตœ๋Œ€ 40%๊นŒ์ง€ ๋†€๋ผ์šธ ์ •๋„๋กœ ์—ฐ์žฅ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธํŠธ๋ถ๋“ค์€ 3์›”๋ถ€ํ„ฐ ์ „ ์„ธ๊ณ„ ์ฃผ์š” ์ œ์กฐ์—…์ฒด๋“ค์„ ํ†ตํ•ด ์ถœ์‹œ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.

์ „๋ฌธ ์‹œ๊ฐํ™” ์‚ฌ์—… ๋ถ€๋ฌธ์œผ๋กœ ๋„˜์–ด๊ฐ€๋ฉด, ๋งค์ถœ์€ 5์–ต 1,100๋งŒ ๋‹ฌ๋Ÿฌ๋กœ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ 5%, ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ 10% ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ฐ„ ๋งค์ถœ์€ 19์–ต ๋‹ฌ๋Ÿฌ๋กœ ์ „๋…„ ๋Œ€๋น„ 21% ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜์š”๋ฅผ ๊ฒฌ์ธํ•˜๋Š” ์ฃผ์š” ์‚ฐ์—… ๋ถ„์•ผ๋กœ๋Š” ์ž๋™์ฐจ์™€ ํ—ฌ์Šค์ผ€์–ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
NVIDIA Technologies and generative AI are reshaping design, engineering and simulation workloads. Increasingly, these technologies are being leveraged in leading software platforms from ANSYS, Cadence and Siemens fueling demand for NVIDIA RTX workstations. Now moving to Automotive. Revenue was a record $570 million, up 27% sequentially and up 103% year-on-year. Full year revenue of $1.7 billion increased 55% year-on-year. Strong growth was driven by the continued ramp in autonomous vehicles, including cars and robotaxis. At CES, we announced Toyota, the world's largest auto maker will build its next-generation vehicles on NVIDIA Orin running the safety certified NVIDIA Drive OS.NVIDIA ๊ธฐ์ˆ ๊ณผ ์ƒ์„ฑํ˜• AI๊ฐ€ ์„ค๊ณ„, ์—”์ง€๋‹ˆ์–ด๋ง, ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์›Œํฌ๋กœ๋“œ๋ฅผ ์žฌํŽธํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ๋“ค์ด ANSYS, Cadence, Siemens์˜ ์ฃผ์š” ์†Œํ”„ํŠธ์›จ์–ด ํ”Œ๋žซํผ์—์„œ ์ ์  ๋” ํ™œ์šฉ๋˜๋ฉด์„œ NVIDIA RTX ์›Œํฌ์Šคํ…Œ์ด์…˜์— ๋Œ€ํ•œ ์ˆ˜์š”๋ฅผ ๊ฒฌ์ธํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์ด์ œ ์ž๋™์ฐจ ๋ถ€๋ฌธ์œผ๋กœ ๋„˜์–ด๊ฐ€๊ฒ ์Šต๋‹ˆ๋‹ค. ๋งค์ถœ์€ 5์–ต 7์ฒœ๋งŒ ๋‹ฌ๋Ÿฌ๋กœ ๊ธฐ๋ก์„ ๊ฒฝ์‹ ํ–ˆ์œผ๋ฉฐ, ์ „๋ถ„๊ธฐ ๋Œ€๋น„ 27% ์ฆ๊ฐ€, ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ 103% ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ฐ„ ๋งค์ถœ์€ 17์–ต ๋‹ฌ๋Ÿฌ๋กœ ์ „๋…„ ๋Œ€๋น„ 55% ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์„ฑ์žฅ์€ ์ž๋™์ฐจ์™€ ๋กœ๋ณดํƒ์‹œ๋ฅผ ํฌํ•จํ•œ ์ž์œจ์ฃผํ–‰์ฐจ์˜ ์ง€์†์ ์ธ ์ฆ๊ฐ€์— ์˜ํ•ด ๊ฒฌ์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CES์—์„œ ์šฐ๋ฆฌ๋Š” ์„ธ๊ณ„ ์ตœ๋Œ€ ์ž๋™์ฐจ ์ œ์กฐ์—…์ฒด์ธ ํ† ์š”ํƒ€๊ฐ€ ์•ˆ์ „ ์ธ์ฆ์„ ๋ฐ›์€ NVIDIA Drive OS๋ฅผ ๊ตฌ๋™ํ•˜๋Š” NVIDIA Orin์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ฐจ์„ธ๋Œ€ ์ฐจ๋Ÿ‰์„ ์ œ์ž‘ํ•  ๊ฒƒ์ด๋ผ๊ณ  ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค.
We announced Aurora and Continental will deploy driverless trucks at scale powered by NVIDIA Drive Thor. Finally, our end-to-end autonomous vehicle platform NVIDIA Drive Hyperion has passed industry safety assessments like TรœV SUD and TUV Rheinland, 2 of the industry's foremost authorities for automotive grade safety and cybersecurity. NVIDIA is the first AV platform that received a comprehensive set of third-party assessments. Okay. Moving to the rest of the P&L. GAAP gross margin was 73% and non-GAAP gross margin was 73.5% and down sequentially as expected with our first deliveries of the Blackwell architecture.์šฐ๋ฆฌ๋Š” Aurora์™€ Continental์ด NVIDIA Drive Thor๋กœ ๊ตฌ๋™๋˜๋Š” ๋ฌด์ธ ํŠธ๋Ÿญ์„ ๋Œ€๊ทœ๋ชจ๋กœ ๋ฐฐ์น˜ํ•  ๊ฒƒ์ด๋ผ๊ณ  ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์šฐ๋ฆฌ์˜ ์—”๋“œํˆฌ์—”๋“œ ์ž์œจ์ฃผํ–‰์ฐจ ํ”Œ๋žซํผ์ธ NVIDIA Drive Hyperion์€ ์ž๋™์ฐจ๊ธ‰ ์•ˆ์ „์„ฑ๊ณผ ์‚ฌ์ด๋ฒ„๋ณด์•ˆ ๋ถ„์•ผ์˜ ์—…๊ณ„ ์ตœ๊ณ  ๊ถŒ์œ„๊ธฐ๊ด€์ธ TรœV SUD์™€ TUV Rheinland ๋“ฑ์˜ ์—…๊ณ„ ์•ˆ์ „์„ฑ ํ‰๊ฐ€๋ฅผ ํ†ต๊ณผํ–ˆ์Šต๋‹ˆ๋‹ค. NVIDIA๋Š” ํฌ๊ด„์ ์ธ ์ œ3์ž ํ‰๊ฐ€๋ฅผ ๋ฐ›์€ ์ตœ์ดˆ์˜ AV ํ”Œ๋žซํผ์ž…๋‹ˆ๋‹ค.

์ด์ œ ์†์ต๊ณ„์‚ฐ์„œ์˜ ๋‚˜๋จธ์ง€ ๋ถ€๋ถ„์œผ๋กœ ๋„˜์–ด๊ฐ€๊ฒ ์Šต๋‹ˆ๋‹ค. GAAP ๋งค์ถœ์ด์ด์ต๋ฅ ์€ 73%์˜€๊ณ , non-GAAP ๋งค์ถœ์ด์ด์ต๋ฅ ์€ 73.5%๋กœ Blackwell ์•„ํ‚คํ…์ฒ˜์˜ ์ฒซ ์ถœํ•˜์™€ ํ•จ๊ป˜ ์˜ˆ์ƒ๋Œ€๋กœ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ ํ•˜๋ฝํ–ˆ์Šต๋‹ˆ๋‹ค.
As discussed last quarter, Blackwell is a customizable AI infrastructure with several different types of NVIDIA build chips multiple networking options and for air and liquid-cooled data center. We exceeded our expectations in Q4 in ramping Blackwell, increasing system availability, providing several configurations to our customers. As Blackwell ramps, we expect gross margins to be in the low 70s. Initially, we are focused on expediting the manufacturing of Blackwell systems to meet strong customer demand as they race to build out Blackwell infrastructure.์ง€๋‚œ ๋ถ„๊ธฐ์— ๋…ผ์˜ํ–ˆ๋“ฏ์ด, Blackwell์€ ์—ฌ๋Ÿฌ ์ข…๋ฅ˜์˜ NVIDIA ๋นŒ๋“œ ์นฉ๊ณผ ๋‹ค์–‘ํ•œ ๋„คํŠธ์›Œํ‚น ์˜ต์…˜, ๊ทธ๋ฆฌ๊ณ  ๊ณต๋žญ์‹ ๋ฐ ์ˆ˜๋žญ์‹ ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋ฅผ ์œ„ํ•œ ๋งž์ถคํ˜• AI ์ธํ”„๋ผ์ž…๋‹ˆ๋‹ค. 4๋ถ„๊ธฐ์— Blackwell ๋žจํ”„์—…์—์„œ ๊ธฐ๋Œ€์น˜๋ฅผ ์ƒํšŒํ–ˆ์œผ๋ฉฐ, ์‹œ์Šคํ…œ ๊ฐ€์šฉ์„ฑ์„ ๋†’์ด๊ณ  ๊ณ ๊ฐ๋“ค์—๊ฒŒ ์—ฌ๋Ÿฌ ๊ตฌ์„ฑ ์˜ต์…˜์„ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. Blackwell์ด ๋žจํ”„์—…๋จ์— ๋”ฐ๋ผ ๋งค์ถœ์ด์ด์ต๋ฅ ์€ 70% ์ดˆ๋ฐ˜๋Œ€๊ฐ€ ๋  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์šฐ๋ฆฌ๋Š” Blackwell ์ธํ”„๋ผ ๊ตฌ์ถ• ๊ฒฝ์Ÿ์— ๋‚˜์„  ๊ณ ๊ฐ๋“ค์˜ ๊ฐ•๋ ฅํ•œ ์ˆ˜์š”๋ฅผ ์ถฉ์กฑํ•˜๊ธฐ ์œ„ํ•ด Blackwell ์‹œ์Šคํ…œ ์ œ์กฐ๋ฅผ ์‹ ์†ํžˆ ์ง„ํ–‰ํ•˜๋Š” ๋ฐ ์ง‘์ค‘ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
When fully ramped, we have many opportunities to improve the cost and gross margin will improve and return to the mid-70s, late this fiscal year. Sequentially, GAAP operating expenses were up 9% and non-GAAP operating expenses were 11%, reflecting higher engineering development costs and higher compute and infrastructure costs for new product introductions. In Q4, we returned $8.1 billion to shareholders in the form of share repurchases and cash dividends. Let me turn to the outlook in the first quarter. Total revenue is expected to be $43 billion, plus or minus 2%. Continuing with its strong demand, we expect a significant ramp of Blackwell in Q1.์™„์ „ํžˆ ๊ฐ€๋™๋  ๋•Œ, ์šฐ๋ฆฌ๋Š” ๋น„์šฉ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋งŽ์€ ๊ธฐํšŒ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ๋งค์ถœ์ด์ด์ต๋ฅ ์€ ๊ฐœ์„ ๋˜์–ด ์ด๋ฒˆ ํšŒ๊ณ„์—ฐ๋„ ๋ง์— 70% ์ค‘๋ฐ˜๋Œ€๋กœ ํšŒ๋ณต๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ „๋ถ„๊ธฐ ๋Œ€๋น„ GAAP ์˜์—…๋น„์šฉ์€ 9% ์ฆ๊ฐ€ํ–ˆ๊ณ , non-GAAP ์˜์—…๋น„์šฉ์€ 11% ์ฆ๊ฐ€ํ–ˆ๋Š”๋ฐ, ์ด๋Š” ์—”์ง€๋‹ˆ์–ด๋ง ๊ฐœ๋ฐœ๋น„์šฉ ์ฆ๊ฐ€์™€ ์‹ ์ œํ’ˆ ์ถœ์‹œ๋ฅผ ์œ„ํ•œ ์ปดํ“จํŒ… ๋ฐ ์ธํ”„๋ผ ๋น„์šฉ ์ฆ๊ฐ€๋ฅผ ๋ฐ˜์˜ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. 4๋ถ„๊ธฐ์— ์šฐ๋ฆฌ๋Š” ์ž์‚ฌ์ฃผ ๋งค์ž…๊ณผ ํ˜„๊ธˆ๋ฐฐ๋‹น์˜ ํ˜•ํƒœ๋กœ ์ฃผ์ฃผ๋“ค์—๊ฒŒ 81์–ต ๋‹ฌ๋Ÿฌ๋ฅผ ํ™˜์›ํ–ˆ์Šต๋‹ˆ๋‹ค. 1๋ถ„๊ธฐ ์ „๋ง์œผ๋กœ ๋„˜์–ด๊ฐ€๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋งค์ถœ์€ 430์–ต ๋‹ฌ๋Ÿฌ, ํ”Œ๋Ÿฌ์Šค๋งˆ์ด๋„ˆ์Šค 2%๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์ˆ˜์š”๊ฐ€ ์ง€์†๋˜๋ฉด์„œ, 1๋ถ„๊ธฐ์— Blackwell์˜ ๋Œ€ํญ์ ์ธ ๋žจํ”„์—…์„ ์˜ˆ์ƒํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
We expect sequential growth in both Data Center and Gaming. Within Data Center, we expect sequential growth from both compute and networking. GAAP and non-GAAP gross margins are expected to be 70.6% and 71%, respectively, plus or minus 50 basis points. GAAP and non-GAAP operating expenses are expected to be approximately $5.2 billion and $3.6 billion, respectively. We expect full year fiscal year '26 operating expenses to grow to be in the mid-30s. GAAP and non-GAAP other incoming expenses are expected to be an income of approximately $400 million, excluding gains and losses from nonmarketable and publicly held equity securities.๋ฐ์ดํ„ฐ์„ผํ„ฐ์™€ ๊ฒŒ์ด๋ฐ ๋ชจ๋‘์—์„œ ์ˆœ์ฐจ์  ์„ฑ์žฅ์„ ๊ธฐ๋Œ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋‚ด์—์„œ๋Š” ์ปดํ“จํŒ…๊ณผ ๋„คํŠธ์›Œํ‚น ๋ชจ๋‘์—์„œ ์ˆœ์ฐจ์  ์„ฑ์žฅ์„ ์˜ˆ์ƒํ•ฉ๋‹ˆ๋‹ค. GAAP ๋ฐ non-GAAP ๋งค์ถœ์ด์ด์ต๋ฅ ์€ ๊ฐ๊ฐ 70.6%์™€ 71%๋กœ ์˜ˆ์ƒ๋˜๋ฉฐ, ํ”Œ๋Ÿฌ์Šค๋งˆ์ด๋„ˆ์Šค 50๋ฒ ์ด์‹œ์Šคํฌ์ธํŠธ์ž…๋‹ˆ๋‹ค. GAAP ๋ฐ non-GAAP ์˜์—…๋น„์šฉ์€ ๊ฐ๊ฐ ์•ฝ 52์–ต ๋‹ฌ๋Ÿฌ์™€ 36์–ต ๋‹ฌ๋Ÿฌ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. 2026 ํšŒ๊ณ„์—ฐ๋„ ์ „์ฒด ์˜์—…๋น„์šฉ์€ 30% ์ค‘๋ฐ˜๋Œ€ ์„ฑ์žฅํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•ฉ๋‹ˆ๋‹ค. GAAP ๋ฐ non-GAAP ๊ธฐํƒ€์ˆ˜์ต๋น„์šฉ์€ ๋น„์ƒ์žฅ ๋ฐ ๊ณต๊ฐœ ์ง€๋ถ„์ฆ๊ถŒ์˜ ์†์ต์„ ์ œ์™ธํ•˜๊ณ  ์•ฝ 4์–ต ๋‹ฌ๋Ÿฌ์˜ ์ˆ˜์ต์œผ๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค.
GAAP and non-GAAP tax rates are expected to be 17%, plus or minus 1%, excluding any discrete items. Further financial details are included in the CFO commentary and other information available on our IR website, including a new financial information AI agent. In closing, let me highlight upcoming events for the financial community. We will be at the TD Cowen Health Care Conference in Boston on March 3 and at the Morgan Stanley Technology, Media and Telecom Conference in San Francisco on March 5. Please join us for our Annual GTC conference starting Monday, March 17 in San Jose, California.GAAP ๋ฐ non-GAAP ์„ธ์œจ์€ ๊ฐœ๋ณ„ ํ•ญ๋ชฉ์„ ์ œ์™ธํ•˜๊ณ  17% ๋‚ด์™ธ 1%๋กœ ์˜ˆ์ƒ๋ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€์ ์ธ ์žฌ๋ฌด ์„ธ๋ถ€์‚ฌํ•ญ์€ ์ƒˆ๋กœ์šด ๊ธˆ์œต์ •๋ณด AI ์—์ด์ „ํŠธ๋ฅผ ํฌํ•จํ•˜์—ฌ ๋‹น์‚ฌ IR ์›น์‚ฌ์ดํŠธ์—์„œ ์ œ๊ณต๋˜๋Š” CFO ์ฝ”๋ฉ˜ํ„ฐ๋ฆฌ ๋ฐ ๊ธฐํƒ€ ์ •๋ณด์— ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ธˆ์œต ์ปค๋ฎค๋‹ˆํ‹ฐ๋ฅผ ์œ„ํ•œ ํ–ฅํ›„ ์ผ์ •์„ ์•ˆ๋‚ด๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. 3์›” 3์ผ ๋ณด์Šคํ„ด์—์„œ ์—ด๋ฆฌ๋Š” TD Cowen ํ—ฌ์Šค์ผ€์–ด ์ปจํผ๋Ÿฐ์Šค์™€ 3์›” 5์ผ ์ƒŒํ”„๋ž€์‹œ์Šค์ฝ”์—์„œ ์—ด๋ฆฌ๋Š” ๋ชจ๊ฑด ์Šคํƒ ๋ฆฌ ๊ธฐ์ˆ , ๋ฏธ๋””์–ด ๋ฐ ํ†ต์‹  ์ปจํผ๋Ÿฐ์Šค์— ์ฐธ์„ํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. 3์›” 17์ผ ์›”์š”์ผ๋ถ€ํ„ฐ ์บ˜๋ฆฌํฌ๋‹ˆ์•„ ์‚ฐํ˜ธ์„ธ์—์„œ ์‹œ์ž‘๋˜๋Š” ์—ฐ๋ก€ GTC ์ปจํผ๋Ÿฐ์Šค์—๋„ ๋งŽ์€ ์ฐธ์—ฌ ๋ถ€ํƒ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
Jensen will deliver a news-packed keynote on March 18, and we will host a Q&A session for our financial analysts for the next day, March 19. We look forward to seeing you at these events. Our earnings call to discuss the results for our first quarter of fiscal 2026 is scheduled for May 28, 2025. We are going to open up the call, operator, to questions. If you could start that, that would be great.์  ์Šจ์€ 3์›” 18์ผ์— ๋‰ด์Šค๊ฐ€ ๊ฐ€๋“ํ•œ ๊ธฐ์กฐ์—ฐ์„ค์„ ์ง„ํ–‰ํ•  ์˜ˆ์ •์ด๋ฉฐ, ๋‹ค์Œ ๋‚ ์ธ 3์›” 19์ผ์—๋Š” ๊ธˆ์œต ์• ๋„๋ฆฌ์ŠคํŠธ๋“ค์„ ์œ„ํ•œ Q&A ์„ธ์…˜์„ ๊ฐœ์ตœํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ–‰์‚ฌ์—์„œ ์—ฌ๋Ÿฌ๋ถ„์„ ๋งŒ๋‚˜๋ต ์ˆ˜ ์žˆ๊ธฐ๋ฅผ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค. 2026 ํšŒ๊ณ„์—ฐ๋„ 1๋ถ„๊ธฐ ์‹ค์ ์„ ๋…ผ์˜ํ•˜๋Š” ์–ด๋‹์ฝœ์€ 2025๋…„ 5์›” 28์ผ๋กœ ์˜ˆ์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด์ œ ์งˆ์˜์‘๋‹ต์„ ์œ„ํ•ด ํ†ตํ™”๋ฅผ ๊ฐœ๋ฐฉํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์˜คํผ๋ ˆ์ดํ„ฐ๋‹˜, ์‹œ์ž‘ํ•ด ์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

๐Ÿ“Œ ์š”์•ฝ

โ€ข ์‹ค์  ํ•˜์ด๋ผ์ดํŠธ:
- 4๋ถ„๊ธฐ ๋งค์ถœ 393์–ต ๋‹ฌ๋Ÿฌ (์ „๋ถ„๊ธฐ ๋Œ€๋น„ 12% ์ฆ๊ฐ€, ์ „๋…„ ๋Œ€๋น„ 78% ์ฆ๊ฐ€)
- ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋ถ€๋ฌธ ๋งค์ถœ 356์–ต ๋‹ฌ๋Ÿฌ๋กœ ์‚ฌ์ƒ ์ตœ๋Œ€ ๊ธฐ๋ก (์ „๋ถ„๊ธฐ ๋Œ€๋น„ 16% ์ฆ๊ฐ€)
- FY2025 ์ „์ฒด ๋งค์ถœ 1,305์–ต ๋‹ฌ๋Ÿฌ (์ „๋…„ ๋Œ€๋น„ 114% ์ฆ๊ฐ€)

โ€ข Blackwell ์‹ ์ œํ’ˆ ์ถœ์‹œ ์„ฑ๊ณผ:
- 4๋ถ„๊ธฐ Blackwell ๋งค์ถœ 110์–ต ๋‹ฌ๋Ÿฌ๋กœ ๊ธฐ๋Œ€์น˜ ์ƒํšŒ
- ํšŒ์‚ฌ ์—ญ์‚ฌ์ƒ ๊ฐ€์žฅ ๋น ๋ฅธ ์ œํ’ˆ ์ถœ์‹œ ๋ฐ ํ™•์žฅ
- ๋Œ€ํ˜• ํด๋ผ์šฐ๋“œ ์‚ฌ์—…์ž๋“ค์ด ์ดˆ๊ธฐ ๋„์ž… ์ฃผ๋„

โ€ข ํ–ฅํ›„ ์ „๋ง:
- 1๋ถ„๊ธฐ ๋งค์ถœ ๊ฐ€์ด๋˜์Šค 430์–ต ๋‹ฌ๋Ÿฌ (ยฑ2%)
- ๋ฐ์ดํ„ฐ์„ผํ„ฐ์™€ ๊ฒŒ์ด๋ฐ ๋ถ€๋ฌธ ๋ชจ๋‘ ์„ฑ์žฅ ์ „๋ง
- ์ด์ด์ต๋ฅ ์€ ์ดˆ๊ธฐ์—๋Š” 70% ์ดˆ๋ฐ˜๋Œ€ ์œ ์ง€, ์—ฐ๋ง์—๋Š” ์ค‘๋ฐ˜๋Œ€๋กœ ํšŒ๋ณต ์˜ˆ์ƒ

โ€ข ์ฃผ์š” ๋ฆฌ์Šคํฌ/๊ธฐํšŒ ์š”์ธ:
- ์ค‘๊ตญ ์ˆ˜์ถœ ๊ทœ์ œ๋กœ ์ธํ•œ ๋งค์ถœ ์ œํ•œ ์ง€์†
- Blackwell ์ƒ์‚ฐ ํ™•๋Œ€์™€ ๊ณต๊ธ‰๋ง ๊ด€๋ฆฌ๊ฐ€ ํ•ต์‹ฌ ๊ณผ์ œ
- AI ์ธํ”„๋ผ ์ˆ˜์š” ์ง€์† ํ™•๋Œ€ ์ „๋ง


โ“ Q&A

Original Translation
Question-and-Answer Session

Operator

Thank you. [Operator Instructions]. And your first question comes from CJ Muse with Cantor Fitzgerald. Please go ahead. C.J. Muse

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

Yes, I appreciate that C.J. There are now multiple scaling laws.
์งˆ์˜์‘๋‹ต ์„ธ์…˜

์šด์˜์ž

๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. [์šด์˜์ž ์•ˆ๋‚ด์‚ฌํ•ญ]. ์ฒซ ๋ฒˆ์งธ ์งˆ๋ฌธ์€ Cantor Fitzgerald์˜ CJ Muse๋‹˜๊ป˜์„œ ์ฃผ์…จ์Šต๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

C.J. Muse

๋„ค, ์•ˆ๋…•ํ•˜์„ธ์š”. ์งˆ๋ฌธ์„ ๋ฐ›์•„์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์  ์Šจ๋‹˜๊ป˜ ์งˆ๋ฌธ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ํ…Œ์ŠคํŠธ ํƒ€์ž„ ์ปดํ“จํŠธ(test-time compute)์™€ ๊ฐ•ํ™”ํ•™์Šต์ด ๊ทธํ† ๋ก ์œ ๋งํ•œ ์„ฑ๊ณผ๋ฅผ ๋ณด์ด๋ฉด์„œ, ํ›ˆ๋ จ(training)๊ณผ ์ถ”๋ก (inference) ๊ฐ„์˜ ๊ฒฝ๊ณ„๊ฐ€ ์ ์  ๋ชจํ˜ธํ•ด์ง€๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ๋ถ„๋ช…ํžˆ ๋ชฉ๊ฒฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ์ž ์žฌ์ ์œผ๋กœ ์ถ”๋ก  ์ „์šฉ ํด๋Ÿฌ์Šคํ„ฐ์˜ ๋ฏธ๋ž˜์— ์–ด๋–ค ์˜๋ฏธ๋ฅผ ๊ฐ–๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  NVIDIA์™€ ๊ณ ๊ฐ๋“ค์—๊ฒŒ ๋ฏธ์น˜๋Š” ์ „๋ฐ˜์ ์ธ ์˜ํ–ฅ์„ ์–ด๋–ป๊ฒŒ ์ƒ๊ฐํ•˜์‹œ๋Š”์ง€ ๊ถ๊ธˆํ•ฉ๋‹ˆ๋‹ค. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

์  ์Šจ ํ™ฉ

๋„ค, C.J.๋‹˜, ์ข‹์€ ์งˆ๋ฌธ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด์ œ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™(scaling laws)์ด ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
There's the pre-training scaling law, and that's going to continue to scale because we have multimodality, we have data that came from reasoning that are now used to do pretraining. And then the second is post-training scaling law, using reinforcement learning human feedback, reinforcement learning AI feedback, reinforcement learning, verifiable rewards. The amount of computation you use for post training is actually higher than pretraining. And it's kind of sensible in the sense that you could, while you're using reinforcement learning, generate an enormous amount of synthetic data or synthetically generated tokens. AI models are basically generating tokens to train AI models.์‚ฌ์ „ ํ›ˆ๋ จ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์ด ์žˆ๊ณ , ์ด๋Š” ๊ณ„์†ํ•ด์„œ ํ™•์žฅ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ์šฐ๋ฆฌ์—๊ฒŒ๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ๊ฐ€ ์žˆ๊ณ , ์ถ”๋ก ์—์„œ ๋‚˜์˜จ ๋ฐ์ดํ„ฐ๊ฐ€ ์ด์ œ ์‚ฌ์ „ ํ›ˆ๋ จ์— ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋‘ ๋ฒˆ์งธ๋Š” ์‚ฌํ›„ ํ›ˆ๋ จ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์œผ๋กœ, ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•œ ๊ฐ•ํ™”ํ•™์Šต, AI ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•œ ๊ฐ•ํ™”ํ•™์Šต, ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ๋ณด์ƒ์„ ํ†ตํ•œ ๊ฐ•ํ™”ํ•™์Šต์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์‚ฌํ›„ ํ›ˆ๋ จ์— ์‚ฌ์šฉํ•˜๋Š” ์—ฐ์‚ฐ๋Ÿ‰์€ ์‹ค์ œ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จ๋ณด๋‹ค ๋” ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด๋Š” ์–ด๋А ์ •๋„ ํ•ฉ๋ฆฌ์ ์ธ๋ฐ, ๊ฐ•ํ™”ํ•™์Šต์„ ์‚ฌ์šฉํ•˜๋Š” ๋™์•ˆ ์—„์ฒญ๋‚œ ์–‘์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋‚˜ ํ•ฉ์„ฑ์ ์œผ๋กœ ์ƒ์„ฑ๋œ ํ† ํฐ์„ ๋งŒ๋“ค์–ด๋‚ผ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. AI ๋ชจ๋ธ๋“ค์€ ๊ธฐ๋ณธ์ ์œผ๋กœ AI ๋ชจ๋ธ๋“ค์„ ํ›ˆ๋ จ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ํ† ํฐ์„ ์ƒ์„ฑํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
And that's post training. And the third part, this is the part that you mentioned is test time compute or reasoning, long thinking, inference scaling. They're all basically the same ideas. And there you have a chain of thought, you've search. The amount of tokens generated the amount of inference compute needed is already 100x more than the one-shot examples and the one-shot capabilities of large language models in the beginning. And that's just the beginning. This is just the beginning.๊ทธ๋ฆฌ๊ณ  ์ด๋Š” ํ›ˆ๋ จ ํ›„์˜ ์ด์•ผ๊ธฐ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์„ธ ๋ฒˆ์งธ ๋ถ€๋ถ„, ์ด๊ฒƒ์ด ๋ฐ”๋กœ ๊ท€ํ•˜๊ป˜์„œ ์–ธ๊ธ‰ํ•˜์‹  ํ…Œ์ŠคํŠธ ํƒ€์ž„ ์ปดํ“จํŠธ(test time compute) ๋˜๋Š” ์ถ”๋ก , ๊ธด ์‚ฌ๊ณ , ์ถ”๋ก  ์Šค์ผ€์ผ๋ง์ž…๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋ชจ๋‘ ๊ธฐ๋ณธ์ ์œผ๋กœ ๋™์ผํ•œ ๊ฐœ๋…์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์—ฌ๊ธฐ์„œ๋Š” ์‚ฌ๊ณ ์˜ ์—ฐ์‡„(chain of thought)์™€ ๊ฒ€์ƒ‰์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ƒ์„ฑ๋˜๋Š” ํ† ํฐ์˜ ์–‘, ํ•„์š”ํ•œ ์ถ”๋ก  ์ปดํ“จํŠธ์˜ ์–‘์€ ์ด๋ฏธ ์ดˆ๊ธฐ ๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ์˜ ์›์ƒท ์˜ˆ์‹œ์™€ ์›์ƒท ์—ญ๋Ÿ‰๋ณด๋‹ค 100๋ฐฐ ๋” ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด๊ฒƒ์€ ๋‹จ์ง€ ์‹œ์ž‘์ผ ๋ฟ์ž…๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ ์ •๋ง ์‹œ์ž‘์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค.
The idea that the next generation could have thousands times and even hopefully, extremely thoughtful and simulation-based and search-based models that could be hundreds of thousands, millions of times more compute than today is in our future. And so the question is how do you design such an architecture? Some of the models are auto regressive. Some of the models are diffusion based. Some of the times you want your data center to have disaggregated inference. Sometimes it is compacted. And so it's hard to figure out what is the best configuration of a data center, which is the reason why NVIDIA's architecture is so popular. We run every model. We are great at training.๋‹ค์Œ ์„ธ๋Œ€์—์„œ๋Š” ์ˆ˜์ฒœ ๋ฐฐ, ๊ทธ๋ฆฌ๊ณ  ๋ฐ”๋ผ๊ฑด๋Œ€ ๋งค์šฐ ์‚ฌ๋ ค ๊นŠ๊ณ  ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ธฐ๋ฐ˜ ๋ฐ ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ๋“ค์ด ์˜ค๋Š˜๋‚ ๋ณด๋‹ค ์ˆ˜์‹ญ๋งŒ ๋ฐฐ, ์ˆ˜๋ฐฑ๋งŒ ๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ… ํŒŒ์›Œ๋ฅผ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ์•„์ด๋””์–ด๊ฐ€ ์šฐ๋ฆฌ์˜ ๋ฏธ๋ž˜์— ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ๋ฌธ์ œ๋Š” ๊ทธ๋Ÿฌํ•œ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์–ด๋–ป๊ฒŒ ์„ค๊ณ„ํ•˜๋А๋ƒ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋ชจ๋ธ๋“ค์€ ์ž๊ธฐํšŒ๊ท€์ ์ž…๋‹ˆ๋‹ค. ์ผ๋ถ€ ๋ชจ๋ธ๋“ค์€ ํ™•์‚ฐ ๊ธฐ๋ฐ˜์ž…๋‹ˆ๋‹ค. ๋•Œ๋กœ๋Š” ๋ฐ์ดํ„ฐ์„ผํ„ฐ๊ฐ€ ๋ถ„์‚ฐ๋œ ์ถ”๋ก ์„ ๊ฐ–๊ธฐ๋ฅผ ์›ํ•˜๊ณ , ๋•Œ๋กœ๋Š” ์••์ถ•๋œ ํ˜•ํƒœ๋ฅผ ์›ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์˜ ์ตœ์  ๊ตฌ์„ฑ์ด ๋ฌด์—‡์ธ์ง€ ํŒŒ์•…ํ•˜๊ธฐ๊ฐ€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ๋ฐ”๋กœ NVIDIA์˜ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ๊ทธํ† ๋ก ์ธ๊ธฐ ์žˆ๋Š” ์ด์œ ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ๋ชจ๋ธ์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํ›ˆ๋ จ์— ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค.
The vast majority of our compute today is actually inference and Blackwell takes all of that to a new level. We designed Blackwell with the idea of reasoning models in mind. And when you look at training, it's many times more performing. But what's really amazing is for long thinking test time scaling, reasoning AI models were tens of times faster, 25 times higher throughput. And so Blackwell is going to be incredible across the board. And when you have a data center that allows you to configure and use your data center based on are you doing more pretraining now, post training now or scaling out your inference, our architecture is fungible and easy to use in all of those different ways.ํ˜„์žฌ ์ €ํฌ ์ปดํ“จํŒ…์˜ ๋Œ€๋ถ€๋ถ„์€ ์‹ค์ œ๋กœ ์ถ”๋ก (inference)์ด๋ฉฐ, Blackwell์€ ์ด ๋ชจ๋“  ๊ฒƒ์„ ์ƒˆ๋กœ์šด ์ฐจ์›์œผ๋กœ ๋Œ์–ด์˜ฌ๋ฆฝ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ถ”๋ก  ๋ชจ๋ธ์„ ์—ผ๋‘์— ๋‘๊ณ  Blackwell์„ ์„ค๊ณ„ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ์„ ์‚ดํŽด๋ณด๋ฉด, ์„ฑ๋Šฅ์ด ๋ช‡ ๋ฐฐ๋‚˜ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ •๋ง ๋†€๋ผ์šด ๊ฒƒ์€ ๊ธด ์‚ฌ๊ณ  ํ…Œ์ŠคํŠธ ์‹œ๊ฐ„ ์Šค์ผ€์ผ๋ง(long thinking test time scaling)์—์„œ ์ถ”๋ก  AI ๋ชจ๋ธ์ด 25๋ฐฐ ๋†’์€ ์ฒ˜๋ฆฌ๋Ÿ‰์œผ๋กœ ์ˆ˜์‹ญ ๋ฐฐ ๋” ๋นจ๋ผ์กŒ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Blackwell์€ ์ „๋ฐ˜์ ์œผ๋กœ ๋†€๋ผ์šด ์„ฑ๊ณผ๋ฅผ ๋ณด์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ํ˜„์žฌ ๋” ๋งŽ์€ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ํ•˜๊ณ  ์žˆ๋Š”์ง€, ์‚ฌํ›„ ํ›ˆ๋ จ์„ ํ•˜๊ณ  ์žˆ๋Š”์ง€, ์•„๋‹ˆ๋ฉด ์ถ”๋ก ์„ ํ™•์žฅํ•˜๊ณ  ์žˆ๋Š”์ง€์— ๋”ฐ๋ผ ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ  ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ์„ผํ„ฐ๊ฐ€ ์žˆ๋‹ค๋ฉด, ์ €ํฌ ์•„ํ‚คํ…์ฒ˜๋Š” ์ด ๋ชจ๋“  ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์—์„œ ๋Œ€์ฒด ๊ฐ€๋Šฅํ•˜๊ณ  ์‚ฌ์šฉํ•˜๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค.
And so we're seeing, in fact, much, much more concentration of a unified architecture than ever before. Operator

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

Good morning, Morgan Stanley, actually. I wonder if you could talk about GB200 at CES, you sort of talked about the complexity of the rack level systems and the challenges you have. And then as you said in the prepared remarks, we've seen a lot of general availability -- where are you in terms of that ramp? Are there still bottlenecks to consider at a systems level above and beyond the chip level? And just have you maintained your enthusiasm for the NVL72 platforms?
๊ทธ๋ž˜์„œ ์‹ค์ œ๋กœ ์šฐ๋ฆฌ๋Š” ๊ทธ ์–ด๋А ๋•Œ๋ณด๋‹ค๋„ ํ†ตํ•ฉ ์•„ํ‚คํ…์ฒ˜์˜ ์ง‘์ค‘๋„๊ฐ€ ํ›จ์”ฌ, ํ›จ์”ฌ ๋” ๋†’์•„์ง€๋Š” ๊ฒƒ์„ ๋ณด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์˜คํผ๋ ˆ์ดํ„ฐ

๋‹ค์Œ ์งˆ๋ฌธ์€ JPMorgan์˜ Joe Moore๋‹˜์œผ๋กœ๋ถ€ํ„ฐ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

Joe Moore

์•ˆ๋…•ํ•˜์„ธ์š”, ์‚ฌ์‹ค์€ Morgan Stanley์ž…๋‹ˆ๋‹ค. CES์—์„œ์˜ GB200์— ๋Œ€ํ•ด ๋ง์”€ํ•ด ์ฃผ์‹ค ์ˆ˜ ์žˆ๋Š”์ง€์š”. ๋ž™ ๋ ˆ๋ฒจ ์‹œ์Šคํ…œ์˜ ๋ณต์žก์„ฑ๊ณผ ์ง๋ฉดํ•˜๊ณ  ์žˆ๋Š” ๊ณผ์ œ๋“ค์— ๋Œ€ํ•ด ์–ธ๊ธ‰ํ•˜์…จ๋Š”๋ฐ์š”. ๊ทธ๋ฆฌ๊ณ  ์ค€๋น„๋œ ๋ฐœ์–ธ์—์„œ ๋ง์”€ํ•˜์‹  ๋ฐ”์™€ ๊ฐ™์ด, ๋งŽ์€ ์ผ๋ฐ˜ ๊ณต๊ธ‰ ๊ฐ€๋Šฅ์„ฑ์„ ๋ณด์•˜์Šต๋‹ˆ๋‹ค -- ๊ทธ ๋žจํ”„ ์ธก๋ฉด์—์„œ ํ˜„์žฌ ์–ด๋А ๋‹จ๊ณ„์— ์žˆ์œผ์‹ ์ง€์š”? ์นฉ ๋ ˆ๋ฒจ์„ ๋„˜์–ด์„œ ์‹œ์Šคํ…œ ๋ ˆ๋ฒจ์—์„œ ์—ฌ์ „ํžˆ ๊ณ ๋ คํ•ด์•ผ ํ•  ๋ณ‘๋ชฉ ํ˜„์ƒ์ด ์žˆ๋‚˜์š”? ๊ทธ๋ฆฌ๊ณ  NVL72 ํ”Œ๋žซํผ์— ๋Œ€ํ•œ ์—ด์ •์„ ์—ฌ์ „ํžˆ ์œ ์ง€ํ•˜๊ณ  ๊ณ„์‹ ๊ฐ€์š”?
Jensen Huang

Well, I'm more enthusiastic today than I was at CES. And the reason for that is because we've shipped a lot more since CES. We have some 350 plants manufacturing the 1.5 million components that go into each one of the Blackwell racks, Grace Blackwell racks. Yes, it's extremely complicated. And we successfully and incredibly ramped up Grace Blackwell, delivering some $11 billion of revenues last quarter. We're going to have to continue to scale as demand is quite high, and customers are anxious and impatient to get their Blackwell systems. You've probably seen on the web, a fair number of celebrations about Grace Blackwell systems coming online and we have them, of course.
์  ์Šจ ํ™ฉ

์ €๋Š” CES ๋•Œ๋ณด๋‹ค ์˜ค๋Š˜ ๋”์šฑ ์—ด์ •์ ์ž…๋‹ˆ๋‹ค. ๊ทธ ์ด์œ ๋Š” CES ์ดํ›„ ํ›จ์”ฌ ๋” ๋งŽ์€ ์ œํ’ˆ์„ ์ถœํ•˜ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ฐ๊ฐ์˜ Blackwell ๋ž™, Grace Blackwell ๋ž™์— ๋“ค์–ด๊ฐ€๋Š” 150๋งŒ ๊ฐœ์˜ ๋ถ€ํ’ˆ์„ ์ œ์กฐํ•˜๋Š” ์•ฝ 350๊ฐœ์˜ ๊ณต์žฅ์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋„ค, ๋งค์šฐ ๋ณต์žกํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ๋Š” Grace Blackwell์„ ์„ฑ๊ณต์ ์ด๊ณ  ๋†€๋ž๋„๋ก ์ฆ์‚ฐํ•˜์—ฌ ์ง€๋‚œ ๋ถ„๊ธฐ์— ์•ฝ 110์–ต ๋‹ฌ๋Ÿฌ์˜ ๋งค์ถœ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜์š”๊ฐ€ ์ƒ๋‹นํžˆ ๋†’๊ณ  ๊ณ ๊ฐ๋“ค์ด Blackwell ์‹œ์Šคํ…œ์„ ๋ฐ›๊ธฐ ์œ„ํ•ด ์กฐ๊ธ‰ํ•ดํ•˜๊ณ  ์žˆ์–ด์„œ ๊ณ„์†ํ•ด์„œ ๊ทœ๋ชจ๋ฅผ ํ™•๋Œ€ํ•ด์•ผ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์›น์—์„œ Grace Blackwell ์‹œ์Šคํ…œ์ด ์˜จ๋ผ์ธ์œผ๋กœ ๊ฐ€๋™๋˜๋Š” ๊ฒƒ์— ๋Œ€ํ•œ ์ƒ๋‹นํ•œ ์ถ•ํ•˜ ์†Œ์‹๋“ค์„ ๋ณด์…จ์„ ํ…๋ฐ, ๋ฌผ๋ก  ์ €ํฌ๊ฐ€ ๊ทธ๋Ÿฐ ์‹œ์Šคํ…œ๋“ค์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
We have a fairly large installation of Grace Blackwell goes for our own engineering and our own design teams and software teams. CoreWeave has now been quite public about the successful bring up of theirs. Microsoft has, of course, open AI has, and you're starting to see many come online. And so I think the answer to your question is nothing is easy about what we're doing, but we're doing great, and all of our partners are doing great. Operator

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

Thank you for taking my questions. Colette if you wouldn't mind confirming if Q1 is the bottom for gross margins?
์ €ํฌ๋Š” ์ž์ฒด ์—”์ง€๋‹ˆ์–ด๋งํŒ€๊ณผ ์„ค๊ณ„ํŒ€, ์†Œํ”„ํŠธ์›จ์–ดํŒ€์„ ์œ„ํ•ด Grace Blackwell์„ ์ƒ๋‹นํžˆ ๋Œ€๊ทœ๋ชจ๋กœ ์„ค์น˜ํ•˜์—ฌ ์šด์˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. CoreWeave๋Š” ์ด์ œ ์ž์‚ฌ์˜ ์„ฑ๊ณต์ ์ธ ๊ตฌ์ถ•์— ๋Œ€ํ•ด ๊ณต๊ฐœ์ ์œผ๋กœ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. Microsoft๋„ ๋งˆ์ฐฌ๊ฐ€์ง€์ด๊ณ , ๋ฌผ๋ก  OpenAI๋„ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋งŽ์€ ๊ธฐ์—…๋“ค์ด ์˜จ๋ผ์ธ์œผ๋กœ ์„œ๋น„์Šค๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๊ฒƒ์„ ๋ณด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ท€ํ•˜์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์€, ์ €ํฌ๊ฐ€ ํ•˜๊ณ  ์žˆ๋Š” ์ผ์ด ์‰ฌ์šด ๊ฒƒ์€ ์—†์ง€๋งŒ, ์ €ํฌ๋Š” ํ›Œ๋ฅญํ•˜๊ฒŒ ํ•ด๋‚ด๊ณ  ์žˆ์œผ๋ฉฐ ๋ชจ๋“  ํŒŒํŠธ๋„ˆ๋“ค๋„ ํ›Œ๋ฅญํ•˜๊ฒŒ ํ•ด๋‚ด๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

์˜คํผ๋ ˆ์ดํ„ฐ

๋‹ค์Œ ์งˆ๋ฌธ์€ Bank of America Securities์˜ Vivek Arya ์• ๋„๋ฆฌ์ŠคํŠธ๋กœ๋ถ€ํ„ฐ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

Vivek Arya

์งˆ๋ฌธ์„ ๋ฐ›์•„์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Colette, 1๋ถ„๊ธฐ๊ฐ€ ๋งค์ถœ์ด์ด์ต๋ฅ ์˜ ๋ฐ”๋‹ฅ์ธ์ง€ ํ™•์ธํ•ด ์ฃผ์‹ค ์ˆ˜ ์žˆ์„๊นŒ์š”?
And then Jensen, my question is for you. What is on your dashboard to give you the confidence that the strong demand can sustain into next year? And has DeepSeek and whatever innovations they came up with, has that changed that view in any way? Thank you. Colette Kress

Let me first take the first part of the question there regarding the gross margin. During our Blackwell ramp, our gross margins will be in the low 70s. At this point, we are focusing on expediting our manufacturing, expediting our manufacturing to make sure that we can provide to customers as soon as possible. Our Blackwell is fully round.
๊ทธ๋ฆฌ๊ณ  ์  ์Šจ, ์ œ ์งˆ๋ฌธ์€ ๋‹น์‹ ์„ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์ˆ˜์š”๊ฐ€ ๋‚ด๋…„๊นŒ์ง€ ์ง€์†๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ํ™•์‹ ์„ ์ฃผ๋Š” ๋Œ€์‹œ๋ณด๋“œ์ƒ์˜ ์ง€ํ‘œ๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”? ๊ทธ๋ฆฌ๊ณ  DeepSeek์™€ ๊ทธ๋“ค์ด ๋‚ด๋†“์€ ํ˜์‹ ๋“ค์ด ๊ทธ๋Ÿฌํ•œ ๊ฒฌํ•ด๋ฅผ ์–ด๋–ค ์‹์œผ๋กœ๋“  ๋ฐ”๊พธ์—ˆ๋‚˜์š”? ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

**์ฝœ๋ ˆํŠธ ํฌ๋ ˆ์Šค**

๋จผ์ € ๋งค์ถœ์ด์ด์ต๋ฅ ์— ๊ด€ํ•œ ์งˆ๋ฌธ์˜ ์ฒซ ๋ฒˆ์งธ ๋ถ€๋ถ„๋ถ€ํ„ฐ ๋‹ต๋ณ€๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. Blackwell ๋žจํ”„์—… ๊ธฐ๊ฐ„ ๋™์•ˆ ์ €ํฌ ๋งค์ถœ์ด์ด์ต๋ฅ ์€ 70% ์ดˆ๋ฐ˜๋Œ€๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ์ €ํฌ๋Š” ๊ณ ๊ฐ๋“ค์—๊ฒŒ ๊ฐ€๋Šฅํ•œ ํ•œ ๋นจ๋ฆฌ ์ œํ’ˆ์„ ๊ณต๊ธ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ œ์กฐ ๊ณต์ •์„ ๊ฐ€์†ํ™”ํ•˜๋Š” ๋ฐ ์ง‘์ค‘ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ Blackwell์€ ์™„์ „ํžˆ ์ค€๋น„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
And once it does -- I'm sorry, once our Blackwell fully rounds, we can improve our cost and our gross margin. So we expect to probably be in the mid-70s later this year. Walking through what you heard Jensen speak about the systems and their complexity, they are customizable in some cases. They've got multiple networking options. They have liquid cool and water cooled. So we know there is an opportunity for us to improve these gross margins going forward. But right now, we are going to focus on getting the manufacturing complete and to our customers as soon as possible.๊ทธ๋ฆฌ๊ณ  ์ผ๋‹จ -- ์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค, ์ผ๋‹จ ์ €ํฌ Blackwell์ด ์™„์ „ํžˆ ์•ˆ์ •ํ™”๋˜๋ฉด, ๋น„์šฉ์„ ๊ฐœ์„ ํ•˜๊ณ  ๋งค์ถœ์ด์ด์ต๋ฅ ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์˜ฌํ•ด ํ›„๋ฐ˜์—๋Š” ์•„๋งˆ๋„ 70% ์ค‘๋ฐ˜๋Œ€์— ๋„๋‹ฌํ•  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•ฉ๋‹ˆ๋‹ค. ์  ์Šจ์ด ๋ง์”€ํ•˜์‹  ์‹œ์Šคํ…œ๋“ค๊ณผ ๊ทธ ๋ณต์žก์„ฑ์„ ์‚ดํŽด๋ณด๋ฉด, ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ๋งž์ถคํ˜•์œผ๋กœ ์ œ์ž‘ ๊ฐ€๋Šฅํ•˜๊ณ , ๋‹ค์–‘ํ•œ ๋„คํŠธ์›Œํ‚น ์˜ต์…˜์„ ๊ฐ–์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ์•ก์ฒด ๋ƒ‰๊ฐ๊ณผ ์ˆ˜๋ƒ‰ ๋ฐฉ์‹์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์•ž์œผ๋กœ ์ด๋Ÿฌํ•œ ๋งค์ถœ์ด์ด์ต๋ฅ ์„ ๊ฐœ์„ ํ•  ๊ธฐํšŒ๊ฐ€ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ง€๊ธˆ ๋‹น์žฅ์€ ์ œ์กฐ๋ฅผ ์™„๋ฃŒํ•˜๊ณ  ๊ฐ€๋Šฅํ•œ ํ•œ ๋นจ๋ฆฌ ๊ณ ๊ฐ๋“ค์—๊ฒŒ ์ œํ’ˆ์„ ๊ณต๊ธ‰ํ•˜๋Š” ๋ฐ ์ง‘์ค‘ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Jensen Huang

We know several things Vivek, we have a fairly good line of sight of the amount of capital investment that data centers are building out towards. We know that going forward, the vast majority of software is going to be based on machine learning. And so accelerated computing and generative AI, reasoning AI are going to be the type of architecture you want in your data center. We have, of course, forecast and plans from our top partners. And we also know that there are many innovative, really exciting start-ups that are still coming online as new opportunities for developing the next breakthroughs in AI, whether it's Agentic AIs, reasoning AI or physical AIs.
์  ์Šจ ํ™ฉ

๋น„๋ฒก, ์šฐ๋ฆฌ๋Š” ๋ช‡ ๊ฐ€์ง€ ์‚ฌ์‹ค์„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋“ค์ด ๊ตฌ์ถ•ํ•˜๊ณ  ์žˆ๋Š” ์ž๋ณธ ํˆฌ์ž ๊ทœ๋ชจ์— ๋Œ€ํ•ด ์ƒ๋‹นํžˆ ๋ช…ํ™•ํ•œ ์ „๋ง์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์•ž์œผ๋กœ ์†Œํ”„ํŠธ์›จ์–ด์˜ ๋Œ€๋ถ€๋ถ„์ด ๋จธ์‹ ๋Ÿฌ๋‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•  ๊ฒƒ์ด๋ผ๋Š” ์ ๋„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ฐ€์† ์ปดํ“จํŒ…๊ณผ ์ƒ์„ฑํ˜• AI, ์ถ”๋ก  AI๊ฐ€ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์—์„œ ์›ํ•˜๊ฒŒ ๋  ์•„ํ‚คํ…์ฒ˜ ์œ ํ˜•์ด ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฌผ๋ก  ์šฐ๋ฆฌ์˜ ์ฃผ์š” ํŒŒํŠธ๋„ˆ๋“ค๋กœ๋ถ€ํ„ฐ ์˜ˆ์ธก๊ณผ ๊ณ„ํš๋„ ํ™•๋ณดํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์—์ด์ „ํ‹ฑ AI๋“ , ์ถ”๋ก  AI๋“ , ๋ฌผ๋ฆฌ์  AI๋“  AI์˜ ์ฐจ์„ธ๋Œ€ ํ˜์‹ ์„ ๊ฐœ๋ฐœํ•  ์ƒˆ๋กœ์šด ๊ธฐํšŒ๋กœ์„œ ์—ฌ์ „ํžˆ ์˜จ๋ผ์ธ์œผ๋กœ ๋“ฑ์žฅํ•˜๊ณ  ์žˆ๋Š” ํ˜์‹ ์ ์ด๊ณ  ์ •๋ง ํฅ๋ฏธ์ง„์ง„ํ•œ ์Šคํƒ€ํŠธ์—…๋“ค์ด ๋งŽ๋‹ค๋Š” ๊ฒƒ๋„ ์•Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
The number of start-ups are still quite vibrant and each 1 of them need a fair amount of computing infrastructure. And so I think the -- whether it's the near term signals or the midterm signals, near-term signals, of course, are POs and forecasts and things like that. Midterm signals would be the level of infrastructure and CapEx scale-out compared to previous years. And then the long-term signals has to do with the fact that we know fundamentally software has changed from hand coding that runs on CPUs, to machine learning and AI-based software that runs on GPUs and accelerated computing systems. And so we have a fairly good sense that this is the future of software.์Šคํƒ€ํŠธ์—…์˜ ์ˆ˜๋Š” ์—ฌ์ „ํžˆ ๋งค์šฐ ํ™œ๋ฐœํ•˜๋ฉฐ, ๊ฐ๊ฐ์˜ ์Šคํƒ€ํŠธ์—…๋“ค์€ ์ƒ๋‹นํ•œ ์–‘์˜ ์ปดํ“จํŒ… ์ธํ”„๋ผ๋ฅผ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹จ๊ธฐ ์‹ ํ˜ธ๋“  ์ค‘๊ธฐ ์‹ ํ˜ธ๋“  ๊ฐ„์— - ๋‹จ๊ธฐ ์‹ ํ˜ธ๋Š” ๋ฌผ๋ก  PO(๊ตฌ๋งค์ฃผ๋ฌธ์„œ)์™€ ์˜ˆ์ธก์น˜ ๋“ฑ๊ณผ ๊ฐ™์€ ๊ฒƒ๋“ค์ž…๋‹ˆ๋‹ค. ์ค‘๊ธฐ ์‹ ํ˜ธ๋Š” ์ด์ „ ์—ฐ๋„ ๋Œ€๋น„ ์ธํ”„๋ผ ๋ฐ ์ž๋ณธ์ง€์ถœ ๊ทœ๋ชจ ํ™•์žฅ ์ˆ˜์ค€์ด ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์žฅ๊ธฐ ์‹ ํ˜ธ๋Š” ์†Œํ”„ํŠธ์›จ์–ด๊ฐ€ ๊ทผ๋ณธ์ ์œผ๋กœ CPU์—์„œ ์‹คํ–‰๋˜๋Š” ์ˆ˜์ž‘์—… ์ฝ”๋”ฉ์—์„œ GPU์™€ ๊ฐ€์† ์ปดํ“จํŒ… ์‹œ์Šคํ…œ์—์„œ ์‹คํ–‰๋˜๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ๋ฐ AI ๊ธฐ๋ฐ˜ ์†Œํ”„ํŠธ์›จ์–ด๋กœ ๋ณ€ํ™”ํ–ˆ๋‹ค๋Š” ์‚ฌ์‹ค๊ณผ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด ์†Œํ”„ํŠธ์›จ์–ด์˜ ๋ฏธ๋ž˜๋ผ๋Š” ์ ์— ๋Œ€ํ•ด ์ƒ๋‹นํžˆ ํ™•์‹ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
And then maybe as you roll it out, another way to think about that is we've really only tapped consumer AI and search and some amount of consumer generative AI, advertising, recommenders, kind of the early days of software. The next wave is coming, Agentic AI for enterprise, physical AI for robotics and sovereign AI as different regions build out their AI for their own ecosystems. And so each one of these are barely off the ground, and we can see them. We can see them because, obviously, we're in the center of much of this development and we can see great activity happening in all these different places and these will happen. So near term, midterm, long-term.๊ทธ๋ฆฌ๊ณ  ์ด๋ฅผ ํ™•์‚ฐ์‹œ์ผœ ๋‚˜๊ฐ€๋ฉด์„œ, ๋‹ค๋ฅธ ๊ด€์ ์—์„œ ์ƒ๊ฐํ•ด๋ณด๋ฉด ์šฐ๋ฆฌ๋Š” ์‹ค์ œ๋กœ ์†Œ๋น„์ž AI์™€ ๊ฒ€์ƒ‰, ๊ทธ๋ฆฌ๊ณ  ์–ด๋А ์ •๋„์˜ ์†Œ๋น„์ž ์ƒ์„ฑํ˜• AI, ๊ด‘๊ณ , ์ถ”์ฒœ ์‹œ์Šคํ…œ, ์ฆ‰ ์†Œํ”„ํŠธ์›จ์–ด์˜ ์ดˆ๊ธฐ ๋‹จ๊ณ„๋งŒ์„ ํ™œ์šฉํ•ด์™”์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ ๋ฌผ๊ฒฐ์ด ๋‹ค๊ฐ€์˜ค๊ณ  ์žˆ๋Š”๋ฐ, ๊ธฐ์—…์šฉ ์—์ด์ „ํ‹ฑ AI, ๋กœ๋ณดํ‹ฑ์Šค๋ฅผ ์œ„ํ•œ ๋ฌผ๋ฆฌ์  AI, ๊ทธ๋ฆฌ๊ณ  ๊ฐ ์ง€์—ญ์ด ์ž์ฒด ์ƒํƒœ๊ณ„๋ฅผ ์œ„ํ•œ AI๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ์†Œ๋ฒ„๋ฆฐ AI์ž…๋‹ˆ๋‹ค. ์ด ๊ฐ๊ฐ์˜ ์˜์—ญ๋“ค์€ ์ด์ œ ๋ง‰ ์‹œ์ž‘ ๋‹จ๊ณ„์— ์žˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์ด๋ฅผ ๋ณผ ์ˆ˜ ์žˆ๋Š” ์ด์œ ๋Š” ๋ถ„๋ช…ํžˆ ์šฐ๋ฆฌ๊ฐ€ ์ด๋Ÿฌํ•œ ๊ฐœ๋ฐœ์˜ ์ค‘์‹ฌ์— ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋ฉฐ, ์ด ๋ชจ๋“  ๋‹ค์–‘ํ•œ ์˜์—ญ์—์„œ ํ™œ๋ฐœํ•œ ํ™œ๋™์ด ์ผ์–ด๋‚˜๊ณ  ์žˆ๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ๊ณ , ์ด๋Ÿฌํ•œ ์ผ๋“ค์ด ์‹คํ˜„๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹จ๊ธฐ, ์ค‘๊ธฐ, ์žฅ๊ธฐ์ ์œผ๋กœ ๋ง์ž…๋‹ˆ๋‹ค.
Operator

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

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

Yes. Blackwell Ultra is second half.
์šด์˜์ž

๋‹ค์Œ ์งˆ๋ฌธ์€ JPMorgan์˜ Harlan Sur๋‹˜์œผ๋กœ๋ถ€ํ„ฐ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

Harlan Sur

์•ˆ๋…•ํ•˜์„ธ์š”. ์งˆ๋ฌธ์„ ๋ฐ›์•„์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ฐจ์„ธ๋Œ€ Blackwell Ultra๊ฐ€ ์˜ฌํ•ด ํ•˜๋ฐ˜๊ธฐ ์ถœ์‹œ ์˜ˆ์ •์œผ๋กœ, ํŒ€์˜ ์—ฐ๊ฐ„ ์ œํ’ˆ ์ถœ์‹œ ์ฃผ๊ธฐ์— ๋งž์ถฐ ์ง„ํ–‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Jensen๋‹˜, ํ˜„์žฌ ์„ธ๋Œ€ Blackwell ์†”๋ฃจ์…˜์„ ์—ฌ์ „ํžˆ ์ฆ์‚ฐํ•˜๊ณ  ์žˆ๋Š” ์ƒํ™ฉ์—์„œ Ultra์— ๋Œ€ํ•œ ์ˆ˜์š” ์—ญํ•™์„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ฃผ์‹ค ์ˆ˜ ์žˆ๋‚˜์š”? ๊ณ ๊ฐ๋“ค๊ณผ ๊ณต๊ธ‰๋ง์ด ์ด ๋‘ ์ œํ’ˆ์˜ ๋™์‹œ ์ฆ์‚ฐ์„ ์–ด๋–ป๊ฒŒ ๊ด€๋ฆฌํ•˜๊ณ  ์žˆ๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  ํŒ€์ด ์—ฌ์ „ํžˆ ์˜ฌํ•ด ํ•˜๋ฐ˜๊ธฐ Blackwell Ultra ์‹คํ–‰ ๊ณ„ํš์„ ์ˆœ์กฐ๋กญ๊ฒŒ ์ง„ํ–‰ํ•˜๊ณ  ์žˆ๋Š”์ง€ ๊ถ๊ธˆํ•ฉ๋‹ˆ๋‹ค.

Jensen Huang

๋„ค. Blackwell Ultra๋Š” ํ•˜๋ฐ˜๊ธฐ์ž…๋‹ˆ๋‹ค.
As you know, the first Blackwell was we had a hiccup that probably cost us a couple of months. We're fully recovered, of course. The team did an amazing job recovering and all of our supply chain partners and just so many people helped us recover at the speed of light. And so now we've successfully ramped production of Blackwell. But that doesn't stop the next train. The next train is on an annual rhythm and Blackwell Ultra with new networking, new memories and of course, new processors, and all of that is coming online. We've have been working with all of our partners and customers, laying this out.์•„์‹œ๋‹ค์‹œํ”ผ, ์ฒซ ๋ฒˆ์งธ Blackwell์—์„œ ์šฐ๋ฆฌ๋Š” ๋ช‡ ๋‹ฌ ์ •๋„์˜ ์†์‹ค์„ ๊ฐ€์ ธ์˜จ ์ฐจ์งˆ์„ ๊ฒช์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฌผ๋ก  ์ง€๊ธˆ์€ ์™„์ „ํžˆ ํšŒ๋ณต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํŒ€์ด ํšŒ๋ณต ๊ณผ์ •์—์„œ ๋†€๋ผ์šด ์„ฑ๊ณผ๋ฅผ ๋ณด์˜€๊ณ , ๋ชจ๋“  ๊ณต๊ธ‰๋ง ํŒŒํŠธ๋„ˆ๋“ค๊ณผ ์ •๋ง ๋งŽ์€ ๋ถ„๋“ค์ด ๋น›์˜ ์†๋„๋กœ ํšŒ๋ณตํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ฃผ์…จ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์ด์ œ ์šฐ๋ฆฌ๋Š” Blackwell์˜ ์ƒ์‚ฐ์„ ์„ฑ๊ณต์ ์œผ๋กœ ์ฆ์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๊ทธ๋ ‡๋‹ค๊ณ  ํ•ด์„œ ๋‹ค์Œ ์—ด์ฐจ๊ฐ€ ๋ฉˆ์ถ”๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๋‹ค์Œ ์—ด์ฐจ๋Š” ์—ฐ๊ฐ„ ์ฃผ๊ธฐ๋กœ ์šดํ–‰๋˜๋ฉฐ, ์ƒˆ๋กœ์šด ๋„คํŠธ์›Œํ‚น, ์ƒˆ๋กœ์šด ๋ฉ”๋ชจ๋ฆฌ, ๊ทธ๋ฆฌ๊ณ  ๋ฌผ๋ก  ์ƒˆ๋กœ์šด ํ”„๋กœ์„ธ์„œ๋ฅผ ํƒ‘์žฌํ•œ Blackwell Ultra๊ฐ€ ์ถœ์‹œ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ์ด ๋ชจ๋“  ๊ฒƒ๋“ค์ด ์˜จ๋ผ์ธ์œผ๋กœ ์ œ๊ณต๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ๋ชจ๋“  ํŒŒํŠธ๋„ˆ์‚ฌ์™€ ๊ณ ๊ฐ์‚ฌ๋“ค๊ณผ ํ•จ๊ป˜ ์ด๋ฅผ ๊ณ„ํšํ•ด ์™”์Šต๋‹ˆ๋‹ค.
They have all of the necessary information, and we'll work with everybody to do the proper transition. This time between Blackwell and Blackwell Ultra, the system architecture is exactly the same. It's a lot harder going from Hopper to Blackwell because we went from an NVLink 8 system to a NVLink 72-based system. So the chassis, the architecture of the system, the hardware, the power delivery, all of that had to change. This was quite a challenging transition. But the next transition will slot right in Blackwell Ultra will slot right in. We've also already revealed and been working very closely with all of our partners on the click after that.๊ทธ๋“ค์€ ๋ชจ๋“  ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ์ €ํฌ๋Š” ๋ชจ๋“  ๊ด€๊ณ„์ž๋“ค๊ณผ ํ˜‘๋ ฅํ•˜์—ฌ ์ ์ ˆํ•œ ์ „ํ™˜์„ ์ง„ํ–‰ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฒˆ Blackwell๊ณผ Blackwell Ultra ์‚ฌ์ด์˜ ์ „ํ™˜์—์„œ๋Š” ์‹œ์Šคํ…œ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ์ •ํ™•ํžˆ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. Hopper์—์„œ Blackwell๋กœ ์ „ํ™˜ํ•˜๋Š” ๊ฒƒ์ด ํ›จ์”ฌ ๋” ์–ด๋ ค์› ๋Š”๋ฐ, ์ด๋Š” NVLink 8 ์‹œ์Šคํ…œ์—์„œ NVLink 72 ๊ธฐ๋ฐ˜ ์‹œ์Šคํ…œ์œผ๋กœ ๋ณ€๊ฒฝ๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„€์‹œ, ์‹œ์Šคํ…œ ์•„ํ‚คํ…์ฒ˜, ํ•˜๋“œ์›จ์–ด, ์ „๋ ฅ ๊ณต๊ธ‰ ๋“ฑ ๋ชจ๋“  ๊ฒƒ์ด ๋ณ€๊ฒฝ๋˜์–ด์•ผ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์ƒ๋‹นํžˆ ๋„์ „์ ์ธ ์ „ํ™˜์ด์—ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋‹ค์Œ ์ „ํ™˜์€ ๋ฐ”๋กœ ์ ์šฉ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. Blackwell Ultra๊ฐ€ ๋ฐ”๋กœ ์Šฌ๋กฏ์— ๋“ค์–ด๊ฐˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋˜ํ•œ ๊ทธ ๋‹ค์Œ ์„ธ๋Œ€์— ๋Œ€ํ•ด์„œ๋„ ์ด๋ฏธ ๊ณต๊ฐœํ–ˆ์œผ๋ฉฐ ๋ชจ๋“  ํŒŒํŠธ๋„ˆ๋“ค๊ณผ ๋งค์šฐ ๊ธด๋ฐ€ํ•˜๊ฒŒ ํ˜‘๋ ฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
And the click after that is called Vera Rubin and all of our partners are getting up to speed on the transition of that and so preparing for that transition. And again, we're going to provide a big, huge step-up. And so come to GTC, and I'll talk to you about Blackwell Ultra, Vera Rubin and then show you what we place after that. Really exciting new products, so to come to GTC piece. Operator

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

Thanks a lot. Jensen, we heard a lot about custom ASICs. Can you kind of speak to the balance between customer ASIC and merchant GPU.
๊ทธ๋ฆฌ๊ณ  ๊ทธ ๋‹ค์Œ ํด๋ฆญ์€ Vera Rubin์ด๋ผ๊ณ  ๋ถˆ๋ฆฌ๋ฉฐ, ์ €ํฌ์˜ ๋ชจ๋“  ํŒŒํŠธ๋„ˆ๋“ค์ด ํ•ด๋‹น ์ „ํ™˜์— ๋Œ€๋น„ํ•˜์—ฌ ์†๋„๋ฅผ ๋†’์ด๊ณ  ์žˆ๊ณ  ๊ทธ ์ „ํ™˜์„ ์ค€๋น„ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋‹ค์‹œ ๋ง์”€๋“œ๋ฆฌ์ง€๋งŒ, ์ €ํฌ๋Š” ํฐ ํญ์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ œ๊ณตํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. GTC์— ์˜ค์‹œ๋ฉด Blackwell Ultra, Vera Rubin์— ๋Œ€ํ•ด ๋ง์”€๋“œ๋ฆฌ๊ณ , ๊ทธ ๋‹ค์Œ์— ์ €ํฌ๊ฐ€ ์ถœ์‹œํ•  ์ œํ’ˆ๋“ค์„ ๋ณด์—ฌ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ์ •๋ง ํฅ๋ฏธ์ง„์ง„ํ•œ ์‹ ์ œํ’ˆ๋“ค์ด๋‹ˆ GTC์— ๊ผญ ์˜ค์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค.

์˜คํผ๋ ˆ์ดํ„ฐ

๋‹ค์Œ ์งˆ๋ฌธ์€ UBS์˜ Timothy Arcuri๋‹˜๊ป˜์„œ ์ฃผ์‹œ๊ฒ ์Šต๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

Timothy Arcuri

๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Jensen, ์ปค์Šคํ…€ ASIC์— ๋Œ€ํ•ด ๋งŽ์ด ๋“ค์—ˆ๋Š”๋ฐ์š”. ๊ณ ๊ฐ ๋งž์ถคํ˜• ASIC๊ณผ ๋ฒ”์šฉ GPU ๊ฐ„์˜ ๊ท ํ˜•์— ๋Œ€ํ•ด ๋ง์”€ํ•ด ์ฃผ์‹ค ์ˆ˜ ์žˆ๋‚˜์š”?
We hear about some of these heterogeneous superclusters to use both GPU and ASIC? Is that something customers are planning on building? Or will these infrastructures remain fairly distinct. Thanks. Jensen Huang

Well, we built very different things than ASICs, in some ways, completely different in some areas we intercept. We're different in several ways. One, NVIDIA's architecture is general whether you're -- you've optimized for unaggressive models or diffusion-based models or vision-based models or multimodal models or text models. We're great in all of it.
์ด๋Ÿฌํ•œ ์ด์ข… ์Šˆํผํด๋Ÿฌ์Šคํ„ฐ๋“ค์ด GPU์™€ ASIC์„ ๋ชจ๋‘ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์–˜๊ธฐ๋ฅผ ๋“ฃ๊ณ  ์žˆ๋Š”๋ฐ์š”? ๊ณ ๊ฐ๋“ค์ด ์ด๋Ÿฐ ๊ฒƒ์„ ๊ตฌ์ถ•ํ•  ๊ณ„ํš์ด ์žˆ๋Š” ๊ฑด๊ฐ€์š”? ์•„๋‹ˆ๋ฉด ์ด๋Ÿฌํ•œ ์ธํ”„๋ผ๋“ค์ด ์ƒ๋‹นํžˆ ๊ตฌ๋ณ„๋˜์–ด ๋‚จ์•„์žˆ์„ ๊ฑด๊ฐ€์š”? ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

์  ์Šจ ํ™ฉ

์ €ํฌ๋Š” ASIC๊ณผ๋Š” ๋งค์šฐ ๋‹ค๋ฅธ ๊ฒƒ๋“ค์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ด๋–ค ๋ฉด์—์„œ๋Š” ์™„์ „ํžˆ ๋‹ค๋ฅด๊ณ , ์ผ๋ถ€ ์˜์—ญ์—์„œ๋Š” ๊ต์ฐจ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์—ฌ๋Ÿฌ ๋ฉด์—์„œ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์ฒซ์งธ, NVIDIA์˜ ์•„ํ‚คํ…์ฒ˜๋Š” ๋ฒ”์šฉ์ ์ž…๋‹ˆ๋‹ค. ์–ด๊ทธ๋ ˆ์‹œ๋ธŒ ๋ชจ๋ธ์— ์ตœ์ ํ™”๋˜์–ด ์žˆ๋“ , ๋””ํ“จ์ „ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด๋“ , ๋น„์ „ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด๋“ , ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ์ด๋“ , ํ…์ŠคํŠธ ๋ชจ๋ธ์ด๋“  ์ƒ๊ด€์—†์ด ์ €ํฌ๋Š” ๋ชจ๋“  ์˜์—ญ์—์„œ ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค.
We're great on all of it because our software stack is so -- our architecture is sensible, our software stack ecosystem is so rich that were the initial target of most exciting innovations and algorithms. And so by definition, we're much, much more general than narrow. We're also really good from the end-to-end from data processing, the curation of the training data, to the training of the data, of course, to reinforcement learning used in post training, all the way to inference with tough time scaling. So we're general, we're end-to-end, and we're everywhere. And because we're not in just one cloud, we're in every cloud, we could be on-prem. We could be in a robot.์ €ํฌ๋Š” ๋ชจ๋“  ์˜์—ญ์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๊ณผ๋ฅผ ๋ณด์ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ์ €ํฌ์˜ ์†Œํ”„ํŠธ์›จ์–ด ์Šคํƒ์ด ๋งค์šฐ - ์ €ํฌ์˜ ์•„ํ‚คํ…์ฒ˜๊ฐ€ ํ•ฉ๋ฆฌ์ ์ด๊ณ , ์†Œํ”„ํŠธ์›จ์–ด ์Šคํƒ ์ƒํƒœ๊ณ„๊ฐ€ ๋งค์šฐ ํ’๋ถ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฐ€์žฅ ํฅ๋ฏธ์ง„์ง„ํ•œ ํ˜์‹ ๊ณผ ์•Œ๊ณ ๋ฆฌ์ฆ˜๋“ค์˜ ์ดˆ๊ธฐ ํƒ€๊ฒŸ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ •์˜์ƒ ์ €ํฌ๋Š” ํ˜‘์†Œํ•œ ๊ฒƒ๋ณด๋‹ค ํ›จ์”ฌ, ํ›จ์”ฌ ๋” ๋ฒ”์šฉ์ ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ €ํฌ๋Š” ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ๋ถ€ํ„ฐ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์˜ ํ๋ ˆ์ด์…˜, ๋ฐ์ดํ„ฐ ํ›ˆ๋ จ, ๋ฌผ๋ก  ํฌ์ŠคํŠธ ํŠธ๋ ˆ์ด๋‹์— ์‚ฌ์šฉ๋˜๋Š” ๊ฐ•ํ™”ํ•™์Šต, ๊ทธ๋ฆฌ๊ณ  ์‹ค์‹œ๊ฐ„ ์Šค์ผ€์ผ๋ง์„ ํ†ตํ•œ ์ถ”๋ก ๊นŒ์ง€ ์—”๋“œํˆฌ์—”๋“œ๋กœ ์ •๋ง ๋›ฐ์–ด๋‚ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ €ํฌ๋Š” ๋ฒ”์šฉ์ ์ด๊ณ , ์—”๋“œํˆฌ์—”๋“œ์ด๋ฉฐ, ์–ด๋””์—๋‚˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ €ํฌ๋Š” ๋‹จ ํ•˜๋‚˜์˜ ํด๋ผ์šฐ๋“œ์—๋งŒ ์žˆ๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๋ชจ๋“  ํด๋ผ์šฐ๋“œ์— ์žˆ๊ณ , ์˜จํ”„๋ ˆ๋ฏธ์Šค์—๋„ ์žˆ์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋กœ๋ด‡์—๋„ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Our architecture is much more accessible and a great target initial target for anybody who's starting up a new company. And so we're everywhere. And the third thing I would say is that our performance in our rhythm is so incredibly fast. Remember that these data centers are always fixed in size. They're fixed in size or they're fixing power. And if our performance per watt is anywhere from 2x to 4x to 8x, which is not unusual, it translates directly to revenues. And so if you have a 100-megawatt data center, if the performance or the throughput in that 100-megawatt or the gigawatt data center is 4 times or 8 times higher, your revenues for that gigawatt data center is 8 times higher.์šฐ๋ฆฌ์˜ ์•„ํ‚คํ…์ฒ˜๋Š” ํ›จ์”ฌ ๋” ์ ‘๊ทผํ•˜๊ธฐ ์‰ฝ๊ณ , ์ƒˆ๋กœ์šด ํšŒ์‚ฌ๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๋ˆ„๊ตฌ์—๊ฒŒ๋‚˜ ํ›Œ๋ฅญํ•œ ์ดˆ๊ธฐ ํƒ€๊ฒŸ์ด ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์šฐ๋ฆฌ๋Š” ์–ด๋””์—๋‚˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์„ธ ๋ฒˆ์งธ๋กœ ๋ง์”€๋“œ๋ฆฌ๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์šฐ๋ฆฌ์˜ ์„ฑ๋Šฅ๊ณผ ๋ฆฌ๋“ฌ์ด ๋ฏฟ์„ ์ˆ˜ ์—†์„ ์ •๋„๋กœ ๋น ๋ฅด๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋“ค์€ ํ•ญ์ƒ ํฌ๊ธฐ๊ฐ€ ๊ณ ์ •๋˜์–ด ์žˆ๋‹ค๋Š” ์ ์„ ๊ธฐ์–ตํ•˜์„ธ์š”. ํฌ๊ธฐ๊ฐ€ ๊ณ ์ •๋˜์–ด ์žˆ๊ฑฐ๋‚˜ ์ „๋ ฅ์ด ๊ณ ์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ์˜ ์™€ํŠธ๋‹น ์„ฑ๋Šฅ์ด 2๋ฐฐ์—์„œ 4๋ฐฐ, 8๋ฐฐ์— ์ด๋ฅด๋Š”๋ฐ, ์ด๋Š” ๋“œ๋ฌธ ์ผ์ด ์•„๋‹ˆ๋ฉฐ, ์ด๊ฒƒ์ด ์ง์ ‘์ ์œผ๋กœ ๋งค์ถœ๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ 100๋ฉ”๊ฐ€์™€ํŠธ ๋ฐ์ดํ„ฐ์„ผํ„ฐ๊ฐ€ ์žˆ๋‹ค๋ฉด, ๊ทธ 100๋ฉ”๊ฐ€์™€ํŠธ ๋˜๋Š” ๊ธฐ๊ฐ€์™€ํŠธ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์˜ ์„ฑ๋Šฅ์ด๋‚˜ ์ฒ˜๋ฆฌ๋Ÿ‰์ด 4๋ฐฐ ๋˜๋Š” 8๋ฐฐ ๋†’๋‹ค๋ฉด, ๊ทธ ๊ธฐ๊ฐ€์™€ํŠธ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์˜ ๋งค์ถœ์€ 8๋ฐฐ ๋†’์•„์ง‘๋‹ˆ๋‹ค.
And the reason that is so different than data centers of the past is because AI factories are directly monetizable through its tokens generated. And so the token throughput of our architecture being so incredibly fast is just incredibly valuable to all of the companies that are building these things for revenue generation reasons and capturing the fast ROI. And so I think the third reason is performance. And then the last thing that I would say is the software stack is incredibly hard. Building an ASIC is no different than what we do. We build a new architecture. And the ecosystem that sits on top of our architecture is 10x more complex today than it was 2 years ago.๊ทธ๋ฆฌ๊ณ  ์ด๊ฒƒ์ด ๊ณผ๊ฑฐ์˜ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์™€ ๋งค์šฐ ๋‹ค๋ฅธ ์ด์œ ๋Š” AI ํŒฉํ† ๋ฆฌ๊ฐ€ ์ƒ์„ฑ๋˜๋Š” ํ† ํฐ์„ ํ†ตํ•ด ์ง์ ‘์ ์œผ๋กœ ์ˆ˜์ตํ™”๊ฐ€ ๊ฐ€๋Šฅํ•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ ์•„ํ‚คํ…์ฒ˜์˜ ํ† ํฐ ์ฒ˜๋ฆฌ๋Ÿ‰์ด ๋งค์šฐ ๋น ๋ฅด๋‹ค๋Š” ๊ฒƒ์€ ์ˆ˜์ต ์ฐฝ์ถœ ๋ชฉ์ ๊ณผ ๋น ๋ฅธ ROI ํ™•๋ณด๋ฅผ ์œ„ํ•ด ์ด๋Ÿฌํ•œ ์‹œ์„ค์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ชจ๋“  ๊ธฐ์—…๋“ค์—๊ฒŒ ์—„์ฒญ๋‚˜๊ฒŒ ๊ฐ€์น˜ ์žˆ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์„ธ ๋ฒˆ์งธ ์ด์œ ๋Š” ์„ฑ๋Šฅ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋งˆ์ง€๋ง‰์œผ๋กœ ๋ง์”€๋“œ๋ฆฌ๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์†Œํ”„ํŠธ์›จ์–ด ์Šคํƒ์ด ๋งค์šฐ ์–ด๋ ต๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ASIC์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์€ ์šฐ๋ฆฌ๊ฐ€ ํ•˜๋Š” ์ผ๊ณผ ๋‹ค๋ฅด์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ƒˆ๋กœ์šด ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ ์•„ํ‚คํ…์ฒ˜ ์œ„์— ์žˆ๋Š” ์ƒํƒœ๊ณ„๋Š” 2๋…„ ์ „๋ณด๋‹ค ์˜ค๋Š˜๋‚  10๋ฐฐ ๋” ๋ณต์žกํ•ด์กŒ์Šต๋‹ˆ๋‹ค.
And that's fairly obvious because the amount of software that the world is building on top of architecture is growing exponentially and AI is advancing very quickly. So bringing that whole ecosystem on top of multiple chips is hard. And so I would say that those four reasons. And then finally, I will say this, just because the chip is designed doesnโ€™t mean it gets deployed. And youโ€™ve seen this over and over again. There are a lot of chips that gets built, but when the time comes, a business decision has to be made, and that business decision is about deploying a new engine, a new processor into a limited AI factory in size, in power and in fine.๊ทธ๋ฆฌ๊ณ  ์ด๋Š” ์ƒ๋‹นํžˆ ๋ช…๋ฐฑํ•œ๋ฐ, ์™œ๋ƒํ•˜๋ฉด ์ „ ์„ธ๊ณ„๊ฐ€ ์•„ํ‚คํ…์ฒ˜ ์œ„์— ๊ตฌ์ถ•ํ•˜๊ณ  ์žˆ๋Š” ์†Œํ”„ํŠธ์›จ์–ด์˜ ์–‘์ด ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ๊ณ  AI๊ฐ€ ๋งค์šฐ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ทธ ์ „์ฒด ์ƒํƒœ๊ณ„๋ฅผ ์—ฌ๋Ÿฌ ์นฉ ์œ„์— ๊ตฌํ˜„ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์ €๋Š” ์ด๋Ÿฌํ•œ ๋„ค ๊ฐ€์ง€ ์ด์œ ๋ฅผ ๋ง์”€๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋งˆ์ง€๋ง‰์œผ๋กœ ์ด๊ฒƒ์„ ๋ง์”€๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ์นฉ์ด ์„ค๊ณ„๋˜์—ˆ๋‹ค๊ณ  ํ•ด์„œ ๋ฐฐํฌ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์—ฌ๋Ÿฌ๋ถ„์€ ์ด๋ฅผ ๋ฐ˜๋ณตํ•ด์„œ ๋ณด์•„์™”์Šต๋‹ˆ๋‹ค. ์ œ์ž‘๋˜๋Š” ์นฉ์€ ๋งŽ์ง€๋งŒ, ๋ง‰์ƒ ๊ทธ ์‹œ์ ์ด ๋˜๋ฉด ๋น„์ฆˆ๋‹ˆ์Šค ๊ฒฐ์ •์„ ๋‚ด๋ ค์•ผ ํ•˜๊ณ , ๊ทธ ๋น„์ฆˆ๋‹ˆ์Šค ๊ฒฐ์ •์€ ํฌ๊ธฐ, ์ „๋ ฅ, ๊ทธ๋ฆฌ๊ณ  ํŒŒ์ธ(fine) ์ธก๋ฉด์—์„œ ์ œํ•œ๋œ AI ํŒฉํ† ๋ฆฌ์— ์ƒˆ๋กœ์šด ์—”์ง„, ์ƒˆ๋กœ์šด ํ”„๋กœ์„ธ์„œ๋ฅผ ๋ฐฐํฌํ•˜๋Š” ๊ฒƒ์— ๊ด€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
And our technology is not only more advanced, more performance, it has much, much better software capability and very importantly, our ability to deploy is lightning fast. And so these things are enough for the faint of heart, as everybody knows now. And so thereโ€™s a lot of different reasons why we do well, why we win. Operator

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

Yeah, I've been right to this here. Hey, thanks a lot for the question. Hi, Jensen, it's a geography-related question, you did a great job explaining some of the demand underlying factors here on the strength. But U.S. was up about $5 billion or so sequentially.
๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ์˜ ๊ธฐ์ˆ ์€ ๋” ๋ฐœ์ „๋˜๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ, ํ›จ์”ฌ ๋” ์šฐ์ˆ˜ํ•œ ์†Œํ”„ํŠธ์›จ์–ด ์—ญ๋Ÿ‰์„ ๊ฐ–์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ๋งค์šฐ ์ค‘์š”ํ•œ ์ ์€ ์šฐ๋ฆฌ์˜ ๋ฐฐํฌ ๋Šฅ๋ ฅ์ด ๋งค์šฐ ๋น ๋ฅด๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด๋Ÿฐ ๊ฒƒ๋“ค์€ ๋ชจ๋“  ์‚ฌ๋žŒ์ด ์ด์ œ ์•Œ๊ณ  ์žˆ๋“ฏ์ด ๋งˆ์Œ์ด ์•ฝํ•œ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ๋Š” ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์šฐ๋ฆฌ๊ฐ€ ์ž˜ํ•˜๊ณ  ์Šน๋ฆฌํ•˜๋Š” ์ด์œ ๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.

์˜คํผ๋ ˆ์ดํ„ฐ

๋‹ค์Œ ์งˆ๋ฌธ์€ ๋ฉœ๋ฆฌ์šฐ์Šค ๋ฆฌ์„œ์น˜์˜ ๋ฒค ๋ผ์ด์ฒด์Šค๋‹˜์œผ๋กœ๋ถ€ํ„ฐ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

๋ฒค ๋ผ์ด์ฒด์Šค

๋„ค, ์—ฌ๊ธฐ์„œ ๋ฐ”๋กœ ์งˆ๋ฌธ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ์งˆ๋ฌธ ๊ธฐํšŒ๋ฅผ ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์•ˆ๋…•ํ•˜์„ธ์š”, ์  ์Šจ. ์ง€์—ญ๊ณผ ๊ด€๋ จ๋œ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ๊ฐ•์„ธ์˜ ๊ทผ๋ณธ์ ์ธ ์ˆ˜์š” ์š”์ธ๋“ค์— ๋Œ€ํ•ด ํ›Œ๋ฅญํ•˜๊ฒŒ ์„ค๋ช…ํ•ด ์ฃผ์…จ๋Š”๋ฐ์š”. ํ•˜์ง€๋งŒ ๋ฏธ๊ตญ์€ ์ „๋ถ„๊ธฐ ๋Œ€๋น„ ์•ฝ 50์–ต ๋‹ฌ๋Ÿฌ ์ •๋„ ์ฆ๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.
And I think there is a concern about whether U.S. can pick up the slack if there's regulations towards other geographies. And I was just wondering, as we go throughout the year, if this kind of surge in the U.S. continues and it's going to be -- whether that's okay. And if that underlies your growth rate, how can you keep growing so fast with this mix shift towards the U.S.? Your guidance looks like China is probably up sequentially. So just wondering if you could go through that dynamic and maybe collect can weigh in. . Jensen Huang

China is approximately the same percentage as Q4 and as previous quarters. It's about half of what it was before the export control.
๊ทธ๋ฆฌ๊ณ  ๋‹ค๋ฅธ ์ง€์—ญ์— ๋Œ€ํ•œ ๊ทœ์ œ๊ฐ€ ์žˆ์„ ๊ฒฝ์šฐ ๋ฏธ๊ตญ์ด ๊ทธ ๊ณต๋ฐฑ์„ ๋ฉ”์šธ ์ˆ˜ ์žˆ์„์ง€์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์˜ฌํ•ด ๋‚ด๋‚ด ๋ฏธ๊ตญ์—์„œ ์ด๋Ÿฐ ๊ธ‰์ฆ์„ธ๊ฐ€ ๊ณ„์†๋˜๊ณ , ๊ทธ๊ฒƒ์ด ๊ดœ์ฐฎ์„์ง€ ๊ถ๊ธˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋งŒ์•ฝ ์ด๊ฒƒ์ด ๊ท€ํ•˜์˜ ์„ฑ์žฅ๋ฅ ์„ ๋’ท๋ฐ›์นจํ•œ๋‹ค๋ฉด, ๋ฏธ๊ตญ์œผ๋กœ์˜ ์ด๋Ÿฐ ๋ฏน์Šค ๋ณ€ํ™”๋กœ ์–ด๋–ป๊ฒŒ ์ด๋ ‡๊ฒŒ ๋น ๋ฅธ ์„ฑ์žฅ์„ ๊ณ„์† ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”? ๊ท€ํ•˜์˜ ๊ฐ€์ด๋˜์Šค๋ฅผ ๋ณด๋ฉด ์ค‘๊ตญ์ด ์•„๋งˆ ์ˆœ์ฐจ์ ์œผ๋กœ ์ฆ๊ฐ€ํ•  ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์ด๋Ÿฐ ์—ญํ•™๊ด€๊ณ„์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ด ์ฃผ์‹œ๊ณ , ์ฝœ๋ ›์ด ์˜๊ฒฌ์„ ์ œ์‹œํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ๊ถ๊ธˆํ•ฉ๋‹ˆ๋‹ค.

์  ์Šจ ํ™ฉ:

์ค‘๊ตญ์€ 4๋ถ„๊ธฐ ๋ฐ ์ด์ „ ๋ถ„๊ธฐ๋“ค๊ณผ ๊ฑฐ์˜ ๋™์ผํ•œ ๋น„์œจ์ž…๋‹ˆ๋‹ค. ์ˆ˜์ถœ ํ†ต์ œ ์ด์ „์˜ ์•ฝ ์ ˆ๋ฐ˜ ์ˆ˜์ค€์ž…๋‹ˆ๋‹ค.
But it's approximately the same in percentage. With respect to geographies, the takeaway is that AI is software. It's modern software. It's incredible modern software, but it's modern software and AI has gone mainstream. AI is used in delivery services everywhere, shopping services everywhere. If you were to buy a quarter, from milk is delivered to you, AI was involved. And so almost everything that a consumer service provides AIs at the core of it. Every student will use AI as a tutor, health care services use AI, financial services use AI. No fintech company will not use AI. Every Fintech company will. Climate tech company use AI. Mineral discovery now uses AI.ํ•˜์ง€๋งŒ ๋น„์œจ๋กœ ๋ณด๋ฉด ๊ฑฐ์˜ ๋™์ผํ•ฉ๋‹ˆ๋‹ค. ์ง€์—ญ๋ณ„๋กœ ๋ณด๋ฉด, ํ•ต์‹ฌ์€ AI๊ฐ€ ์†Œํ”„ํŠธ์›จ์–ด๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ˜„๋Œ€์ ์ธ ์†Œํ”„ํŠธ์›จ์–ด์ฃ . ๋†€๋ผ์šธ ์ •๋„๋กœ ํ˜„๋Œ€์ ์ธ ์†Œํ”„ํŠธ์›จ์–ด์ด์ง€๋งŒ, ๊ฒฐ๊ตญ ํ˜„๋Œ€์ ์ธ ์†Œํ”„ํŠธ์›จ์–ด์ด๊ณ  AI๋Š” ์ฃผ๋ฅ˜๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. AI๋Š” ๋ชจ๋“  ๊ณณ์˜ ๋ฐฐ์†ก ์„œ๋น„์Šค, ๋ชจ๋“  ๊ณณ์˜ ์‡ผํ•‘ ์„œ๋น„์Šค์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์šฐ์œ  ํ•œ ํŒฉ์„ ์ฃผ๋ฌธํ•ด์„œ ๋ฐฐ์†ก๋ฐ›๋Š”๋‹ค๋ฉด, ๊ทธ ๊ณผ์ •์— AI๊ฐ€ ๊ด€์—ฌ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์†Œ๋น„์ž ์„œ๋น„์Šค๊ฐ€ ์ œ๊ณตํ•˜๋Š” ๊ฑฐ์˜ ๋ชจ๋“  ๊ฒƒ์˜ ํ•ต์‹ฌ์— AI๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ํ•™์ƒ์ด AI๋ฅผ ํŠœํ„ฐ๋กœ ์‚ฌ์šฉํ•  ๊ฒƒ์ด๊ณ , ์˜๋ฃŒ ์„œ๋น„์Šค๋Š” AI๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, ๊ธˆ์œต ์„œ๋น„์Šค๋„ AI๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. AI๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š๋Š” ํ•€ํ…Œํฌ ํšŒ์‚ฌ๋Š” ์—†์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ชจ๋“  ํ•€ํ…Œํฌ ํšŒ์‚ฌ๊ฐ€ ์‚ฌ์šฉํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ธฐํ›„ ๊ธฐ์ˆ  ํšŒ์‚ฌ๋“ค๋„ AI๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ด‘๋ฌผ ํƒ์‚ฌ์—๋„ ์ด์ œ AI๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.
The number of -- every higher education, every university uses AI and so I think it is fairly safe to say that AI has gone mainstream and that it's being integrated into every application. And -- and our hope is that, of course, the technology continues to advance safely and advance in a helpful way to society. And with that, we're -- I do believe that we're at the beginning of this new transition. And what I mean by that in the beginning is, remember, behind us has been decades of data centers and decades of computers that have been built. And they've been built for a world of hand coding and general purpose computing and CPUs and so on and so forth.๋ชจ๋“  ๊ณ ๋“ฑ๊ต์œก๊ธฐ๊ด€, ๋ชจ๋“  ๋Œ€ํ•™์ด AI๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ๋”ฐ๋ผ์„œ AI๊ฐ€ ์ฃผ๋ฅ˜๋กœ ์ž๋ฆฌ์žก์•˜๊ณ  ๋ชจ๋“  ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ํ†ตํ•ฉ๋˜๊ณ  ์žˆ๋‹ค๊ณ  ๋งํ•˜๋Š” ๊ฒƒ์ด ์ƒ๋‹นํžˆ ์•ˆ์ „ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋ฌผ๋ก  ๊ธฐ์ˆ ์ด ์•ˆ์ „ํ•˜๊ฒŒ ๊ณ„์† ๋ฐœ์ „ํ•˜๊ณ  ์‚ฌํšŒ์— ๋„์›€์ด ๋˜๋Š” ๋ฐฉํ–ฅ์œผ๋กœ ๋ฐœ์ „ํ•˜๊ธฐ๋ฅผ ํฌ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด์™€ ํ•จ๊ป˜, ์šฐ๋ฆฌ๋Š” ์ด ์ƒˆ๋กœ์šด ์ „ํ™˜์˜ ์‹œ์ž‘์ ์— ์žˆ๋‹ค๊ณ  ๋ฏฟ์Šต๋‹ˆ๋‹ค. ์‹œ์ž‘์ ์— ์žˆ๋‹ค๋Š” ๊ฒƒ์˜ ์˜๋ฏธ๋Š”, ๊ธฐ์–ตํ•˜์‹œ๊ฒ ์ง€๋งŒ ์šฐ๋ฆฌ ๋’ค์—๋Š” ์ˆ˜์‹ญ ๋…„๊ฐ„์˜ ๋ฐ์ดํ„ฐ์„ผํ„ฐ์™€ ์ˆ˜์‹ญ ๋…„๊ฐ„ ๊ตฌ์ถ•๋œ ์ปดํ“จํ„ฐ๋“ค์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด๋“ค์€ ์ˆ˜์ž‘์—… ์ฝ”๋”ฉ๊ณผ ๋ฒ”์šฉ ์ปดํ“จํŒ…, CPU ๋“ฑ๋“ฑ์˜ ์„ธ๊ณ„๋ฅผ ์œ„ํ•ด ๊ตฌ์ถ•๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
And going forward, I think it's fairly safe to say that world is going to be almost all software to be infused with AI. All software and all services will be based on -- ultimately, based on machine learning, the data flywheel is going to be part of improving software and services and that the future computers will be accelerated, the future computers will be based on AI. And we're really two years into that journey. And in modernizing computers that have taken decades to build out. And so I'm fairly sure that we're in the beginning of this new era. And then lastly, no technology has ever had the opportunity to address a larger part of the world's GDP than AI. No software tool ever has.์•ž์œผ๋กœ ๋‚˜์•„๊ฐ€๋ฉด์„œ, ์ „ ์„ธ๊ณ„ ๊ฑฐ์˜ ๋ชจ๋“  ์†Œํ”„ํŠธ์›จ์–ด๊ฐ€ AI๋กœ ์œตํ•ฉ๋  ๊ฒƒ์ด๋ผ๊ณ  ๋งํ•˜๋Š” ๊ฒƒ์ด ์ƒ๋‹นํžˆ ์•ˆ์ „ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์†Œํ”„ํŠธ์›จ์–ด์™€ ๋ชจ๋“  ์„œ๋น„์Šค๋Š” ๊ถ๊ทน์ ์œผ๋กœ ๋จธ์‹ ๋Ÿฌ๋‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•  ๊ฒƒ์ด๋ฉฐ, ๋ฐ์ดํ„ฐ ํ”Œ๋ผ์ดํœ ์ด ์†Œํ”„ํŠธ์›จ์–ด์™€ ์„œ๋น„์Šค ๊ฐœ์„ ์˜ ์ผ๋ถ€๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋ฏธ๋ž˜์˜ ์ปดํ“จํ„ฐ๋Š” ๊ฐ€์†ํ™”๋  ๊ฒƒ์ด๊ณ , ๋ฏธ๋ž˜์˜ ์ปดํ“จํ„ฐ๋Š” AI๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ํ˜„์žฌ ๊ทธ ์—ฌ์ •์˜ 2๋…„์ฐจ์— ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ˆ˜์‹ญ ๋…„์— ๊ฑธ์ณ ๊ตฌ์ถ•๋œ ์ปดํ“จํ„ฐ๋ฅผ ํ˜„๋Œ€ํ™”ํ•˜๋Š” ๊ณผ์ •์— ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๊ฐ€ ์ด ์ƒˆ๋กœ์šด ์‹œ๋Œ€์˜ ์‹œ์ž‘์ ์— ์žˆ๋‹ค๊ณ  ํ™•์‹ ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, AI๋ณด๋‹ค ์„ธ๊ณ„ GDP์˜ ๋” ํฐ ๋ถ€๋ถ„์„ ๋‹ค๋ฃฐ ๊ธฐํšŒ๋ฅผ ๊ฐ€์ง„ ๊ธฐ์ˆ ์€ ์ง€๊ธˆ๊นŒ์ง€ ์—†์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ์–ด๋–ค ์†Œํ”„ํŠธ์›จ์–ด ๋„๊ตฌ๋„ ๊ทธ๋ ‡์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.
And so this is now a software tool that can address a much larger part of the world's GDP more than any time in history. And so the way we think about growth and the way we think about whether something is big or small, has to be in the context of that. And when you take a step back and look at it from that perspective, we're really just in the beginning. Operator

Your next question comes from the line of Aaron Rakers with Wells Fargo. Aaron your line is open. Your next question comes from Mark Lipacis with Evercore ISI. Please go ahead. Mark Lipacis

I had a clarification and a question. Colette up for the clarification.
๊ทธ๋ž˜์„œ ์ด์ œ ์ด๊ฒƒ์€ ์—ญ์‚ฌ์ƒ ๊ทธ ์–ด๋А ๋•Œ๋ณด๋‹ค๋„ ์ „ ์„ธ๊ณ„ GDP์˜ ํ›จ์”ฌ ๋” ํฐ ๋ถ€๋ถ„์„ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ๋Š” ์†Œํ”„ํŠธ์›จ์–ด ๋„๊ตฌ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๊ฐ€ ์„ฑ์žฅ์— ๋Œ€ํ•ด ์ƒ๊ฐํ•˜๋Š” ๋ฐฉ์‹๊ณผ ์–ด๋–ค ๊ฒƒ์ด ํฌ๊ฑฐ๋‚˜ ์ž‘๋‹ค๊ณ  ์ƒ๊ฐํ•˜๋Š” ๋ฐฉ์‹์€ ๊ทธ๋Ÿฌํ•œ ๋งฅ๋ฝ์—์„œ ์ดํ•ด๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ํ•œ ๋ฐœ ๋ฌผ๋Ÿฌ์„œ์„œ ๊ทธ๋Ÿฐ ๊ด€์ ์—์„œ ๋ฐ”๋ผ๋ณด๋ฉด, ์šฐ๋ฆฌ๋Š” ์ •๋ง๋กœ ์‹œ์ž‘ ๋‹จ๊ณ„์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค.

์˜คํผ๋ ˆ์ดํ„ฐ

๋‹ค์Œ ์งˆ๋ฌธ์€ Wells Fargo์˜ Aaron Rakers๋‹˜์œผ๋กœ๋ถ€ํ„ฐ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. Aaron๋‹˜, ์—ฐ๊ฒฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ ์งˆ๋ฌธ์€ Evercore ISI์˜ Mark Lipacis๋‹˜์œผ๋กœ๋ถ€ํ„ฐ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

Mark Lipacis

๋ช…ํ™•ํžˆ ํ•˜๊ณ  ์‹ถ์€ ๊ฒƒ๊ณผ ์งˆ๋ฌธ์ด ํ•˜๋‚˜์”ฉ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ช…ํ™•ํžˆ ํ•˜๊ณ  ์‹ถ์€ ๊ฒƒ์€ Colette์—๊ฒŒ ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค.
Did you say that enterprise within the data center grew 2x year-on-year for the January quarter? And if so, does that -- would that make it the fast faster growing than the hyperscalers? And then, Jensen, for you, the question, hyperscalers are the biggest purchasers of your solutions, but they buy equipment for both internal and external workloads, external workflows being cloud services that enterprise is used. So the question is, can you give us a sense of how that hyperscaler spend splits between that external workload and internal? And as these new AI workflows and applications come up, would you expect enterprises to become a larger part of that consumption mix?1์›” ๋ถ„๊ธฐ์— ๋ฐ์ดํ„ฐ์„ผํ„ฐ ๋‚ด ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ ๋ถ€๋ฌธ์ด ์ „๋…„ ๋™๊ธฐ ๋Œ€๋น„ 2๋ฐฐ ์„ฑ์žฅํ–ˆ๋‹ค๊ณ  ๋ง์”€ํ•˜์‹  ๊ฑด๊ฐ€์š”? ๋งŒ์•ฝ ๊ทธ๋ ‡๋‹ค๋ฉด, ์ด๊ฒƒ์ด ํ•˜์ดํผ์Šค์ผ€์ผ๋Ÿฌ๋ณด๋‹ค ๋” ๋น ๋ฅธ ์„ฑ์žฅ๋ฅ ์„ ์˜๋ฏธํ•˜๋Š” ๊ฑด๊ฐ€์š”?

๊ทธ๋ฆฌ๊ณ  ์  ์Šจ์—๊ฒŒ ๋“œ๋ฆฌ๋Š” ์งˆ๋ฌธ์ธ๋ฐ์š”, ํ•˜์ดํผ์Šค์ผ€์ผ๋Ÿฌ๋“ค์ด ๊ท€ํ•˜์˜ ์†”๋ฃจ์…˜์˜ ์ตœ๋Œ€ ๊ตฌ๋งค์ž์ด์ง€๋งŒ, ์ด๋“ค์€ ๋‚ด๋ถ€ ์›Œํฌ๋กœ๋“œ์™€ ์™ธ๋ถ€ ์›Œํฌ๋กœ๋“œ ๋ชจ๋‘๋ฅผ ์œ„ํ•œ ์žฅ๋น„๋ฅผ ๊ตฌ๋งคํ•ฉ๋‹ˆ๋‹ค. ์™ธ๋ถ€ ์›Œํฌํ”Œ๋กœ์šฐ๋Š” ๊ธฐ์—…์ด ์‚ฌ์šฉํ•˜๋Š” ํด๋ผ์šฐ๋“œ ์„œ๋น„์Šค๋ฅผ ์˜๋ฏธํ•˜์ฃ . ๊ทธ๋ž˜์„œ ์งˆ๋ฌธ์€, ํ•˜์ดํผ์Šค์ผ€์ผ๋Ÿฌ์˜ ์ง€์ถœ์ด ์™ธ๋ถ€ ์›Œํฌ๋กœ๋“œ์™€ ๋‚ด๋ถ€ ์›Œํฌ๋กœ๋“œ ์‚ฌ์ด์—์„œ ์–ด๋–ป๊ฒŒ ๋ถ„ํ• ๋˜๋Š”์ง€์— ๋Œ€ํ•œ ๊ฐ์„ ์ฃผ์‹ค ์ˆ˜ ์žˆ๋‚˜์š”? ๊ทธ๋ฆฌ๊ณ  ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด AI ์›Œํฌํ”Œ๋กœ์šฐ์™€ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๋“ค์ด ๋“ฑ์žฅํ•จ์— ๋”ฐ๋ผ, ๊ธฐ์—…๋“ค์ด ๊ทธ ์†Œ๋น„ ๋ฏน์Šค์—์„œ ๋” ํฐ ๋ถ€๋ถ„์„ ์ฐจ์ง€ํ•˜๊ฒŒ ๋  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•˜์‹œ๋‚˜์š”?
And does that impact how you develop your service, your ecosystem. Colette Kress

Sure. Thanks for the question regarding our Enterprise business. Yes, it grew to X and very similar to what we were seeing with our large CSPs. Keep in mind, these are both important areas to understand working with the CSPs and be working on large language models, can you working on inference in their own work. But keep in mind, that is also where the enterprises are servicing. Your enterprises are both with your CSPs as well as in terms of building on their own. They're both are growing quite well. Jensen Huang

The CSPs are about half of our business.
๊ทธ๋ฆฌ๊ณ  ๊ทธ๊ฒƒ์ด ๊ท€ํ•˜์˜ ์„œ๋น„์Šค, ์ƒํƒœ๊ณ„ ๊ฐœ๋ฐœ์— ์–ด๋–ค ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€์š”?

์ฝœ๋ ˆํŠธ ํฌ๋ ˆ์Šค

๋„ค. ์ €ํฌ ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ ์‚ฌ์—…์— ๋Œ€ํ•œ ์งˆ๋ฌธ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋„ค, X๊นŒ์ง€ ์„ฑ์žฅํ–ˆ์œผ๋ฉฐ ๋Œ€ํ˜• CSP๋“ค์—์„œ ๋ณด๊ณ  ์žˆ๋Š” ๊ฒƒ๊ณผ ๋งค์šฐ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด ๋‘ ์˜์—ญ ๋ชจ๋‘ ์ค‘์š”ํ•œ ๋ถ€๋ถ„์ด๋ผ๋Š” ์ ์„ ์—ผ๋‘์— ๋‘์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. CSP๋“ค๊ณผ ํ˜‘๋ ฅํ•˜๊ณ  ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ์ž‘์—…ํ•˜๋ฉฐ, ์ž์ฒด์ ์ธ ์ถ”๋ก  ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ ๋ง์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๊ธฐ์—…๋“ค์ด ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ณณ์ด๊ธฐ๋„ ํ•˜๋‹ค๋Š” ์ ์„ ๊ธฐ์–ตํ•ด ์ฃผ์„ธ์š”. ๊ธฐ์—…๋“ค์€ CSP๋“ค๊ณผ ํ˜‘๋ ฅํ•˜๋Š” ๋™์‹œ์— ์ž์ฒด์ ์œผ๋กœ ๊ตฌ์ถ•ํ•˜๋Š” ์ž‘์—…๋„ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‘ ์˜์—ญ ๋ชจ๋‘ ์ƒ๋‹นํžˆ ์ž˜ ์„ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

์  ์Šจ ํ™ฉ

CSP๋“ค์€ ์ €ํฌ ์‚ฌ์—…์˜ ์•ฝ ์ ˆ๋ฐ˜์„ ์ฐจ์ง€ํ•ฉ๋‹ˆ๋‹ค.
And the CSPs have internal consumption and external consumption, as you say. And we are using -- of course, used for internal consumption. We work very closely with all of them to optimize workloads that are internal to them, because they have a large infrastructure of NVIDIA gear that they could take advantage of. And the fact that we could be used for AI on the one hand, video processing on the other hand, data processing like Spark, we're fungible. And so the useful life of our infrastructure is much better. If the useful life is much longer, then the TCO is also lower. And so -- the second part is how do we see the growth of enterprise or not CSPs, if you will, going forward?๊ทธ๋ฆฌ๊ณ  CSP๋“ค์€ ๋ง์”€ํ•˜์‹  ๋Œ€๋กœ ๋‚ด๋ถ€ ์†Œ๋น„์™€ ์™ธ๋ถ€ ์†Œ๋น„๋ฅผ ๋ชจ๋‘ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ฌผ๋ก  ๋‚ด๋ถ€ ์†Œ๋น„์— ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๋ชจ๋“  CSP๋“ค๊ณผ ๋งค์šฐ ๊ธด๋ฐ€ํ•˜๊ฒŒ ํ˜‘๋ ฅํ•˜์—ฌ ๊ทธ๋“ค์˜ ๋‚ด๋ถ€ ์›Œํฌ๋กœ๋“œ๋ฅผ ์ตœ์ ํ™”ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ๊ทธ๋“ค์€ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋Œ€๊ทœ๋ชจ NVIDIA ์žฅ๋น„ ์ธํ”„๋ผ๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ €ํฌ๊ฐ€ ํ•œํŽธ์œผ๋กœ๋Š” AI์—, ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ๋Š” ๋น„๋””์˜ค ์ฒ˜๋ฆฌ์—, ๊ทธ๋ฆฌ๊ณ  Spark์™€ ๊ฐ™์€ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ์— ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค ๋•Œ๋ฌธ์— ์ €ํฌ๋Š” ๋Œ€์ฒด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ €ํฌ ์ธํ”„๋ผ์˜ ์œ ์šฉํ•œ ์ˆ˜๋ช…์ด ํ›จ์”ฌ ๋” ์ข‹์Šต๋‹ˆ๋‹ค. ์œ ์šฉํ•œ ์ˆ˜๋ช…์ด ํ›จ์”ฌ ๊ธธ๋‹ค๋ฉด TCO๋„ ๋” ๋‚ฎ์•„์ง‘๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๋‘ ๋ฒˆ์งธ ๋ถ€๋ถ„์€ ์•ž์œผ๋กœ ๊ธฐ์—… ๋˜๋Š” CSP๊ฐ€ ์•„๋‹Œ ๋ถ€๋ฌธ์˜ ์„ฑ์žฅ์„ ์–ด๋–ป๊ฒŒ ๋ณด๊ณ  ์žˆ๋Š”์ง€์— ๊ด€ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
And the answer is, I believe, long-term, it is by far larger and the reason for that is because if you look at the computer industry today and what is not served by the computer industry is largely industrial. So let me give you an example. When we say enterprise, and let's use the car company as an example because they make both soft things and hard things. And so in the case of a car company, the employees will be what we call enterprise and Agentic AI and software planning systems and tools, and we have some really exciting things to share with you guys at GTC, build Agentic systems are for employees to make employees more productive to design to market plan to operate their company.๊ทธ๋ฆฌ๊ณ  ์ œ๊ฐ€ ์ƒ๊ฐํ•˜๋Š” ๋‹ต์€ ์žฅ๊ธฐ์ ์œผ๋กœ ๋ณผ ๋•Œ ์‚ฐ์—…์šฉ์ด ํ›จ์”ฌ ๋” ํฌ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ ์ด์œ ๋Š” ์˜ค๋Š˜๋‚  ์ปดํ“จํ„ฐ ์‚ฐ์—…์„ ์‚ดํŽด๋ณด๋ฉด ์ปดํ“จํ„ฐ ์‚ฐ์—…์ด ์„œ๋น„์Šคํ•˜์ง€ ๋ชปํ•˜๋Š” ์˜์—ญ์ด ์ฃผ๋กœ ์‚ฐ์—… ๋ถ„์•ผ์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์„ค๋ช…ํ•ด๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ๊ธฐ์—…์šฉ์ด๋ผ๊ณ  ํ•  ๋•Œ, ์ž๋™์ฐจ ํšŒ์‚ฌ๋ฅผ ์˜ˆ๋กœ ๋“ค์–ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ๊ทธ๋“ค์€ ์†Œํ”„ํŠธ์›จ์–ด์™€ ํ•˜๋“œ์›จ์–ด ๋ชจ๋‘๋ฅผ ๋งŒ๋“ค๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ž๋™์ฐจ ํšŒ์‚ฌ์˜ ๊ฒฝ์šฐ, ์ง์›๋“ค์€ ์šฐ๋ฆฌ๊ฐ€ ๊ธฐ์—…์šฉ์ด๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ๊ฒƒ์— ํ•ด๋‹นํ•˜๋ฉฐ, ์—์ด์ „ํ‹ฑ AI์™€ ์†Œํ”„ํŠธ์›จ์–ด ๊ณ„ํš ์‹œ์Šคํ…œ ๋ฐ ๋„๊ตฌ๋“ค์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  GTC์—์„œ ์—ฌ๋Ÿฌ๋ถ„๋“ค๊ณผ ๊ณต์œ ํ•  ์ •๋ง ํฅ๋ฏธ์ง„์ง„ํ•œ ๊ฒƒ๋“ค์ด ์žˆ๋Š”๋ฐ, ์—์ด์ „ํ‹ฑ ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์€ ์ง์›๋“ค์ด ๋” ์ƒ์‚ฐ์ ์œผ๋กœ ์ผํ•  ์ˆ˜ ์žˆ๋„๋ก, ์„ค๊ณ„ํ•˜๊ณ  ๋งˆ์ผ€ํŒ…ํ•˜๊ณ  ๊ณ„ํšํ•˜๊ณ  ํšŒ์‚ฌ๋ฅผ ์šด์˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
That's Agentic AI. On the other hand, the cars that they manufacture also need AI. They need an AI system that trains the cars, treats this entire giant fleet of cars. And today, there's 1 billion cars on the road. Someday, there will be 1 billion cars on the road, and every single one of those cars will be robotic cars, and they'll all be collecting data, and we'll be improving them using an AI factory. Whereas they have a car factory today in the future they'll have a car factory and an AI factory. And then inside the car itself is a robotic system. And so as you can see, there are three computers involved and there's the computer that helps the people.๊ทธ๊ฒƒ์ด ๋ฐ”๋กœ ์—์ด์ „ํ‹ฑ AI์ž…๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ๋Š” ๊ทธ๋“ค์ด ์ œ์กฐํ•˜๋Š” ์ž๋™์ฐจ๋“ค๋„ AI๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์ž๋™์ฐจ๋“ค์„ ํ›ˆ๋ จ์‹œํ‚ค๊ณ , ์ด ๊ฑฐ๋Œ€ํ•œ ์ž๋™์ฐจ ํ”Œ๋ฆฟ ์ „์ฒด๋ฅผ ๊ด€๋ฆฌํ•˜๋Š” AI ์‹œ์Šคํ…œ์ด ํ•„์š”ํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ๋„๋กœ์—๋Š” 10์–ต ๋Œ€์˜ ์ž๋™์ฐจ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์–ธ์  ๊ฐ€๋Š” ๋„๋กœ์— 10์–ต ๋Œ€์˜ ์ž๋™์ฐจ๊ฐ€ ์žˆ์„ ๊ฒƒ์ด๊ณ , ๊ทธ ๋ชจ๋“  ์ž๋™์ฐจ๋“ค์ด ๋กœ๋ด‡ ์ž๋™์ฐจ๊ฐ€ ๋  ๊ฒƒ์ด๋ฉฐ, ๋ชจ๋‘ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์šฐ๋ฆฌ๋Š” AI ํŒฉํ† ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•ด์„œ ์ด๋“ค์„ ๊ฐœ์„ ํ•ด ๋‚˜๊ฐˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ๊ทธ๋“ค์ด ์ž๋™์ฐจ ๊ณต์žฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋ฐ˜๋ฉด, ๋ฏธ๋ž˜์—๋Š” ์ž๋™์ฐจ ๊ณต์žฅ๊ณผ AI ํŒฉํ† ๋ฆฌ๋ฅผ ๋ชจ๋‘ ๊ฐ–๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ž๋™์ฐจ ๋‚ด๋ถ€์—๋Š” ๋กœ๋ด‡ ์‹œ์Šคํ…œ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณด์‹œ๋‹ค์‹œํ”ผ, ์—ฌ๊ธฐ์—๋Š” ์„ธ ๊ฐœ์˜ ์ปดํ“จํ„ฐ๊ฐ€ ๊ด€๋ จ๋˜์–ด ์žˆ๊ณ , ์‚ฌ๋žŒ๋“ค์„ ๋•๋Š” ์ปดํ“จํ„ฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.
There's the computer that build the AI for the machineries that could be, of course, could be a tractor, it could be a lawn mower. It could be a human or robot that's being developed today. It could be a building, it could be a warehouse. These physical systems require new type of AI we call physical AI. They can't just understand the meaning of words and languages, but they have to understand the meaning of the world, friction and inertia, object permanence and cause and effect. And all of those type of things that are common sense to you and I, but AIs have to go learn those physical effects. So we call that physical AI.๊ธฐ๊ณ„๋ฅผ ์œ„ํ•œ AI๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ์ปดํ“จํ„ฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฌผ๋ก  ์ด๋Š” ํŠธ๋ž™ํ„ฐ๊ฐ€ ๋  ์ˆ˜๋„ ์žˆ๊ณ , ์ž”๋””๊นŽ์ด๊ฐ€ ๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ค๋Š˜๋‚  ๊ฐœ๋ฐœ๋˜๊ณ  ์žˆ๋Š” ํœด๋จธ๋…ธ์ด๋“œ๋‚˜ ๋กœ๋ด‡์ด ๋  ์ˆ˜๋„ ์žˆ๊ณ ์š”. ๊ฑด๋ฌผ์ด๋‚˜ ์ฐฝ๊ณ ๊ฐ€ ๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌผ๋ฆฌ์  ์‹œ์Šคํ…œ๋“ค์€ ์šฐ๋ฆฌ๊ฐ€ ํ”ผ์ง€์ปฌ AI๋ผ๊ณ  ๋ถ€๋ฅด๋Š” ์ƒˆ๋กœ์šด ์œ ํ˜•์˜ AI๋ฅผ ํ•„์š”๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋‹จ์ˆœํžˆ ๋‹จ์–ด์™€ ์–ธ์–ด์˜ ์˜๋ฏธ๋ฅผ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ, ์„ธ์ƒ์˜ ์˜๋ฏธ, ๋งˆ์ฐฐ๊ณผ ๊ด€์„ฑ, ๊ฐ์ฒด ์˜์†์„ฑ, ๊ทธ๋ฆฌ๊ณ  ์ธ๊ณผ๊ด€๊ณ„๋ฅผ ์ดํ•ดํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ๋ถ„๊ณผ ์ €์—๊ฒŒ๋Š” ์ƒ์‹์ ์ธ ์ด ๋ชจ๋“  ๊ฒƒ๋“ค์„ AI๋Š” ํ•™์Šตํ•ด์•ผ ํ•˜๋Š” ๋ฌผ๋ฆฌ์  ํšจ๊ณผ๋“ค์ž…๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ ํ”ผ์ง€์ปฌ AI๋ผ๊ณ  ๋ถ€๋ฆ…๋‹ˆ๋‹ค.
That whole part of using Agentic AI to revolutionize the way we work inside companies, that's just starting. This is now the beginning of the agent AI era, and you hear a lot of people talking about it and we got some really great things going on. And then there's the physical AI after that, and then there are robotic systems after that. And so these 3 computers are all brand new. And my sense is that long term, this will be by far the larger of a mall, which kind of makes sense. The worldโ€™s GDP is representing โ€“ represented by either heavy industries or industrials and companies that are providing for those. Operator

Your next question comes from the line of Aaron Rakers with Wells Fargo.
๊ธฐ์—… ๋‚ด๋ถ€์—์„œ ์ผํ•˜๋Š” ๋ฐฉ์‹์„ ํ˜์‹ ํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํ‹ฑ AI๋ฅผ ํ™œ์šฉํ•˜๋Š” ์ „์ฒด์ ์ธ ๋ถ€๋ถ„์€ ์ด์ œ ๋ง‰ ์‹œ์ž‘๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ง€๊ธˆ์ด ๋ฐ”๋กœ ์—์ด์ „ํŠธ AI ์‹œ๋Œ€์˜ ์‹œ์ž‘์ด๋ฉฐ, ๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด ์ด์— ๋Œ€ํ•ด ์ด์•ผ๊ธฐํ•˜๊ณ  ์žˆ๊ณ  ์ €ํฌ๋„ ์ •๋ง ํ›Œ๋ฅญํ•œ ์ผ๋“ค์„ ์ง„ํ–‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” ๋ฌผ๋ฆฌ์  AI๊ฐ€ ์žˆ๊ณ , ๊ทธ ์ดํ›„์—๋Š” ๋กœ๋ด‡ ์‹œ์Šคํ…œ๋“ค์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด 3๊ฐ€์ง€ ์ปดํ“จํ„ฐ๋Š” ๋ชจ๋‘ ์™„์ „ํžˆ ์ƒˆ๋กœ์šด ๊ฒƒ๋“ค์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ œ ์ƒ๊ฐ์—๋Š” ์žฅ๊ธฐ์ ์œผ๋กœ ์ด๊ฒƒ์ด ์ „์ฒด ์ค‘์—์„œ ๋‹จ์—ฐ์ฝ” ๊ฐ€์žฅ ํฐ ๋ถ€๋ถ„์ด ๋  ๊ฒƒ์ด๋ผ๊ณ  ๋ด…๋‹ˆ๋‹ค. ์ด๋Š” ์–ด๋А ์ •๋„ ํƒ€๋‹นํ•ฉ๋‹ˆ๋‹ค. ์„ธ๊ณ„ GDP๋Š” ์ค‘๊ณต์—…์ด๋‚˜ ์‚ฐ์—… ๋ถ„์•ผ, ๊ทธ๋ฆฌ๊ณ  ์ด๋“ค์„ ์ง€์›ํ•˜๋Š” ๊ธฐ์—…๋“ค๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.

์˜คํผ๋ ˆ์ดํ„ฐ

๋‹ค์Œ ์งˆ๋ฌธ์€ Wells Fargo์˜ Aaron Rakers๋‹˜์œผ๋กœ๋ถ€ํ„ฐ ๋ฐ›๊ฒ ์Šต๋‹ˆ๋‹ค.
Please go ahead. Aaron Rakers

Thanks for let me back in. Jensen, I'm curious as we now approach the 2-year anniversary of really the Hopper inflection that you saw in 2023 in GenAI in general. And when we think about the road map you have in front of us, how do you think about the infrastructure that's been deployed from a replacement cycle perspective? And whether if it's GB300 or if it's the Rubin cycle where we start to see maybe some refresh opportunity. I'm just curious how you look at that. Jensen Huang

I appreciate it. First of all, people are still using Voltus and Pascal and amperes.
๊ณ„์†ํ•ด ์ฃผ์„ธ์š”.

**Aaron Rakers**

๋‹ค์‹œ ์งˆ๋ฌธํ•  ๊ธฐํšŒ๋ฅผ ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์  ์Šจ, 2023๋…„ GenAI ์ „๋ฐ˜๊ณผ ์ •๋ง๋กœ Hopper ๋ณ€๊ณก์ ์„ ๋ณด์•˜๋˜ ๋•Œ๋กœ๋ถ€ํ„ฐ ์ด์ œ 2์ฃผ๋…„์ด ๋‹ค๊ฐ€์˜ค๊ณ  ์žˆ๋Š” ์‹œ์ ์—์„œ ๊ถ๊ธˆํ•œ ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ ์•ž์— ์žˆ๋Š” ๋กœ๋“œ๋งต์„ ์ƒ๊ฐํ•ด๋ณผ ๋•Œ, ๊ต์ฒด ์ฃผ๊ธฐ ๊ด€์ ์—์„œ ๋ฐฐํฌ๋œ ์ธํ”„๋ผ์— ๋Œ€ํ•ด ์–ด๋–ป๊ฒŒ ์ƒ๊ฐํ•˜์‹œ๋Š”์ง€์š”? ๊ทธ๋ฆฌ๊ณ  GB300์ด๋“  ์•„๋‹ˆ๋ฉด ์ผ๋ถ€ ๋ฆฌํ”„๋ ˆ์‹œ ๊ธฐํšŒ๋ฅผ ๋ณด๊ธฐ ์‹œ์ž‘ํ•  ์ˆ˜ ์žˆ๋Š” Rubin ์‚ฌ์ดํด์ด๋“ , ์ด์— ๋Œ€ํ•ด ์–ด๋–ป๊ฒŒ ๋ณด๊ณ  ๊ณ„์‹ ์ง€ ๊ถ๊ธˆํ•ฉ๋‹ˆ๋‹ค.

**Jensen Huang**

๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ์„ , ์‚ฌ๋žŒ๋“ค์€ ์—ฌ์ „ํžˆ Volta์™€ Pascal, ๊ทธ๋ฆฌ๊ณ  Ampere๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
And the reason for that is because there are always things that because CUDA is so programmable you could use it Blackwell, one of the major use cases right now is data processing and data curation. You find a circumstance that an AI model is not very good at. You present that circumstance to a vision language model, let's say, it is a car. You present that circumstance to a vision language model. The vision language model actually looks in the circumstances, said, this is what happened and I was very good at it. You then take that response to the prompt and you go and prompt an AI model to go find in your whole lake of data other circumstances like that, whatever that circumstance was.๊ทธ ์ด์œ ๋Š” CUDA๊ฐ€ ๋งค์šฐ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๊ฐ€๋Šฅํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ•ญ์ƒ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๋“ค์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. Blackwell์˜ ๊ฒฝ์šฐ, ํ˜„์žฌ ์ฃผ์š” ์‚ฌ์šฉ ์‚ฌ๋ก€ ์ค‘ ํ•˜๋‚˜๊ฐ€ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ์™€ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜์ž…๋‹ˆ๋‹ค. AI ๋ชจ๋ธ์ด ์ž˜ ์ฒ˜๋ฆฌํ•˜์ง€ ๋ชปํ•˜๋Š” ์ƒํ™ฉ์„ ๋ฐœ๊ฒฌํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์ž๋™์ฐจ์™€ ๊ด€๋ จ๋œ ์ƒํ™ฉ์„ ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ์— ์ œ์‹œํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ทธ ์ƒํ™ฉ์„ ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ์— ์ œ์‹œํ•˜๋ฉด, ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ์ด ์‹ค์ œ๋กœ ๊ทธ ์ƒํ™ฉ๋“ค์„ ์‚ดํŽด๋ณด๊ณ  "์ด๋Ÿฐ ์ผ์ด ์ผ์–ด๋‚ฌ๊ณ  ๋‚˜๋Š” ์ด๊ฒƒ์„ ๋งค์šฐ ์ž˜ ์ฒ˜๋ฆฌํ–ˆ๋‹ค"๊ณ  ๋งํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ทธ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ์‘๋‹ต์„ ๊ฐ€์ ธ์™€์„œ AI ๋ชจ๋ธ์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ฃผ์–ด ์ „์ฒด ๋ฐ์ดํ„ฐ ๋ ˆ์ดํฌ์—์„œ ๊ทธ ์ƒํ™ฉ๊ณผ ์œ ์‚ฌํ•œ ๋‹ค๋ฅธ ์ƒํ™ฉ๋“ค์„ ์ฐพ์•„๋‚ด๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
And then you use an AI to do domain randomization and generate a whole bunch of other examples. And then from that, you can go train the bottle. And so you could use amperes to go and do data processing and data curation and machine learning-based search. And then you create the training data set, which you then present to your Hopper systems for training. And so each one of these architectures are completely -- they're all CUDA-compatible and so everything wants on everything. But if you have infrastructure in place, then you can put the less intensive workloads onto the installed base of the past. All of our [GBUs are] (ph) very well employed.๊ทธ๋ฆฌ๊ณ  AI๋ฅผ ์‚ฌ์šฉํ•ด์„œ ๋„๋ฉ”์ธ ๋žœ๋คํ™”๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋‹ค๋ฅธ ๋งŽ์€ ์˜ˆ์‹œ๋“ค์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Ampere๋ฅผ ์‚ฌ์šฉํ•ด์„œ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ์™€ ๋ฐ์ดํ„ฐ ํ๋ ˆ์ด์…˜, ๊ทธ๋ฆฌ๊ณ  ๋จธ์‹ ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ ๊ฒ€์ƒ‰์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•˜์—ฌ Hopper ์‹œ์Šคํ…œ์— ์ œ๊ณตํ•ด์„œ ํ›ˆ๋ จ์„ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ๊ฐ์˜ ์•„ํ‚คํ…์ฒ˜๋“ค์€ ์™„์ „ํžˆ -- ๋ชจ๋‘ CUDA ํ˜ธํ™˜์ด ๊ฐ€๋Šฅํ•˜๋ฏ€๋กœ ๋ชจ๋“  ๊ฒƒ์ด ๋ชจ๋“  ๊ฒƒ์—์„œ ์ž‘๋™ํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ธํ”„๋ผ๊ฐ€ ๊ตฌ์ถ•๋˜์–ด ์žˆ๋‹ค๋ฉด, ์ƒ๋Œ€์ ์œผ๋กœ ์ง‘์•ฝ๋„๊ฐ€ ๋‚ฎ์€ ์›Œํฌ๋กœ๋“œ๋“ค์„ ๊ธฐ์กด์— ์„ค์น˜๋œ ๊ณผ๊ฑฐ ์ œํ’ˆ๋“ค์— ๋ฐฐ์น˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ €ํฌ์˜ ๋ชจ๋“  GPU๋“ค์ด ๋งค์šฐ ์ž˜ ํ™œ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
Operator

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

I have a follow-up question on gross margins for Colette. So I understand there are many moving parts the Blackwell yields, NVLink 72 and Ethernet mix. And you kind of tipped to the earlier question, the April quarter is the bottom,; but second half would have to ramp like 200 basis points per quarter to get to the mid-70s range that you're giving for the end of the fiscal year. And we still don't know much about tariff impact to broader semiconductor. So what kind of gives you the confidence in that trajectory in the back half of this year? Colette Kress

Yes.
์šด์˜์ž

ํ•œ ๊ฐ€์ง€ ์งˆ๋ฌธ์„ ๋” ๋ฐ›์„ ์‹œ๊ฐ„์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์‹œํ‹ฐ์˜ ์•„ํ‹ฐํ”„ ๋ง๋ฆญ(Atif Malik)๋‹˜์˜ ์งˆ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋ง์”€ํ•ด ์ฃผ์„ธ์š”.

์•„ํ‹ฐํ”„ ๋ง๋ฆญ

์ฝœ๋ ˆํŠธ๋‹˜๊ป˜ ๋งค์ถœ์ด์ด์ต๋ฅ ์— ๋Œ€ํ•œ ํ›„์† ์งˆ๋ฌธ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ธ”๋ž™์›ฐ ์ˆ˜์œจ, NVLink 72, ์ด๋”๋„ท ๋ฏน์Šค ๋“ฑ ๋งŽ์€ ๋ณ€์ˆ˜๋“ค์ด ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์ดํ•ดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์•ž์„  ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์—์„œ 4์›” ๋ถ„๊ธฐ๊ฐ€ ๋ฐ”๋‹ฅ์ด๋ผ๊ณ  ์–ธ๊ธ‰ํ•˜์…จ๋Š”๋ฐ, ํšŒ๊ณ„์—ฐ๋„ ๋ง์— ์ œ์‹œํ•˜์‹  70% ์ค‘๋ฐ˜๋Œ€ ๋ฒ”์œ„์— ๋„๋‹ฌํ•˜๋ ค๋ฉด ํ•˜๋ฐ˜๊ธฐ์— ๋ถ„๊ธฐ๋‹น 200๋ฒ ์ด์‹œ์Šค ํฌ์ธํŠธ์”ฉ ์ƒ์Šนํ•ด์•ผ ํ•  ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์•„์ง ๋ฐ˜๋„์ฒด ์ „๋ฐ˜์— ๋Œ€ํ•œ ๊ด€์„ธ ์˜ํ–ฅ์— ๋Œ€ํ•ด์„œ๋Š” ์ž˜ ๋ชจ๋ฅด๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด ์˜ฌํ•ด ํ•˜๋ฐ˜๊ธฐ ์ด๋Ÿฌํ•œ ๊ถค๋„์— ๋Œ€ํ•ด ์–ด๋–ค ๋ถ€๋ถ„์ด ํ™•์‹ ์„ ์ฃผ๋Š”์ง€์š”?

์ฝœ๋ ˆํŠธ ํฌ๋ ˆ์Šค

๋„ค.
Thanks for the question. Our Gross margins, there are quite complex in terms of the material and everything that we put together in a Blackwell system, a tremendous amount of opportunity to look at a lot of different pieces of that on how we can better improve our gross margins over time. Remember, we have many different configurations as well on Blackwell that will be able to help us do that. So together, working after we get some of these really strong ramping completed for our customers, we can begin a lot of that work. If not, we are going to probably start as soon as possible if we can. If we can improve it in the short term, we will also do that.์งˆ๋ฌธํ•ด ์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ ๋งค์ถœ์ด์ด์ต๋ฅ ์€ Blackwell ์‹œ์Šคํ…œ์— ๋“ค์–ด๊ฐ€๋Š” ์†Œ์žฌ์™€ ๋ชจ๋“  ๊ตฌ์„ฑ ์š”์†Œ๋“ค ์ธก๋ฉด์—์„œ ์ƒ๋‹นํžˆ ๋ณต์žกํ•ฉ๋‹ˆ๋‹ค. ์‹œ๊ฐ„์ด ์ง€๋‚˜๋ฉด์„œ ๋งค์ถœ์ด์ด์ต๋ฅ ์„ ๋” ์ž˜ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ๋‹ค์–‘ํ•œ ๋ถ€๋ถ„๋“ค์„ ์‚ดํŽด๋ณผ ์ˆ˜ ์žˆ๋Š” ์—„์ฒญ๋‚œ ๊ธฐํšŒ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. Blackwell์—๋Š” ์ด๋ฅผ ๋„์šธ ์ˆ˜ ์žˆ๋Š” ๋‹ค์–‘ํ•œ ๊ตฌ์„ฑ๋„ ๋งŽ์ด ์žˆ๋‹ค๋Š” ์ ์„ ๊ธฐ์–ตํ•ด ์ฃผ์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณ ๊ฐ๋“ค์„ ์œ„ํ•œ ์ด๋Ÿฌํ•œ ๊ฐ•๋ ฅํ•œ ๋žจํ•‘์„ ์™„๋ฃŒํ•œ ํ›„ ํ•จ๊ป˜ ํ˜‘๋ ฅํ•˜์—ฌ ๊ทธ๋Ÿฐ ์ž‘์—…๋“ค์„ ๋งŽ์ด ์‹œ์ž‘ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ ‡์ง€ ์•Š๋”๋ผ๋„, ๊ฐ€๋Šฅํ•˜๋‹ค๋ฉด ์ตœ๋Œ€ํ•œ ๋นจ๋ฆฌ ์‹œ์ž‘ํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ๋‹จ๊ธฐ์ ์œผ๋กœ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด ๊ทธ๊ฒƒ๋„ ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Tariff at this point, it's a little bit of an unknown it's an unknown until we understand further what the U.S. government's plan is, both its timing, it's where and how much. So at this time, we are awaiting, but again, we would, of course, always follow export controls and/or tariffs in that manner. Operator

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

Thank you. Colette Kress

We are going to open up to Jensen and I believe he has a couple things. Jensen Huang

I just wanted to thank you. Thank you, Colette. The demand for Blackwell is extraordinary. AI is evolving beyond perception and generative AI into reasoning.
๊ด€์„ธ์˜ ๊ฒฝ์šฐ, ํ˜„์žฌ๋กœ์„œ๋Š” ๋‹ค์†Œ ๋ถˆํ™•์‹คํ•œ ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ๋ฏธ๊ตญ ์ •๋ถ€์˜ ๊ณ„ํš์— ๋Œ€ํ•ด ๋” ์ž์„ธํžˆ ํŒŒ์•…ํ•˜๊ธฐ ์ „๊นŒ์ง€๋Š” ๋ถˆํ™•์‹คํ•ฉ๋‹ˆ๋‹ค. ์‹œ๊ธฐ, ์ ์šฉ ๋Œ€์ƒ, ๊ทธ๋ฆฌ๊ณ  ๊ทœ๋ชจ ๋“ฑ์„ ๋ง์ด์ฃ . ๋”ฐ๋ผ์„œ ํ˜„์žฌ๋กœ์„œ๋Š” ์ง€์ผœ๋ณด๊ณ  ์žˆ๋Š” ์ƒํ™ฉ์ด์ง€๋งŒ, ๋ฌผ๋ก  ์ €ํฌ๋Š” ํ•ญ์ƒ ์ˆ˜์ถœ ํ†ต์ œ ๋ฐ/๋˜๋Š” ๊ด€์„ธ๋ฅผ ๊ทธ๋Ÿฐ ๋ฐฉ์‹์œผ๋กœ ์ค€์ˆ˜ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.

์˜คํผ๋ ˆ์ดํ„ฐ

์‹ ์‚ฌ ์ˆ™๋…€ ์—ฌ๋Ÿฌ๋ถ„, ์งˆ์˜์‘๋‹ต ์„ธ์…˜์„ ๋งˆ์น˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ฃ„์†กํ•ฉ๋‹ˆ๋‹ค.

์  ์Šจ ํ™ฉ

๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

์ฝœ๋ ˆํŠธ ํฌ๋ ˆ์Šค

์  ์Šจ์—๊ฒŒ ๋งˆ์ดํฌ๋ฅผ ๋„˜๊ธฐ๊ฒ ์Šต๋‹ˆ๋‹ค. ๋ช‡ ๊ฐ€์ง€ ๋ง์”€ํ•˜์‹ค ๊ฒƒ์ด ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.

์  ์Šจ ํ™ฉ

๊ฐ์‚ฌ ์ธ์‚ฌ๋ฅผ ๋“œ๋ฆฌ๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ์ฝœ๋ ˆํŠธ, ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ธ”๋ž™์›ฐ์— ๋Œ€ํ•œ ์ˆ˜์š”๋Š” ์ •๋ง ๋†€๋ž์Šต๋‹ˆ๋‹ค. AI๋Š” ์ธ์‹๊ณผ ์ƒ์„ฑํ˜• AI๋ฅผ ๋„˜์–ด ์ถ”๋ก ์œผ๋กœ ์ง„ํ™”ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
With resenting AI, we're observing another scaling law, inference time or test time scaling, more computation. The more the model thinks the smarter the answer. Models like OpenAI, Grok-3, DeepSeek-R1 are reasoning models that apply inference time scaling. Reasoning models can consume 100x more compute. Future reasoning models can consume much more compute. DeepSeek-R1 has ignited global enthusiast -- it's an excellent innovation. But even more importantly, it has open source a world-class reasoning AI model. Nearly every AI developer is applying R1 or chain of thought and reinforcement learning techniques like R1 to scale their model's performance.์ตœ์‹  AI๋ฅผ ํ†ตํ•ด ์šฐ๋ฆฌ๋Š” ๋˜ ๋‹ค๋ฅธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™, ์ฆ‰ ์ถ”๋ก  ์‹œ๊ฐ„ ๋˜๋Š” ํ…Œ์ŠคํŠธ ํƒ€์ž„ ์Šค์ผ€์ผ๋ง์„ ๊ด€์ฐฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๋งŽ์€ ์—ฐ์‚ฐ์„ ํ†ตํ•ด ๋ชจ๋ธ์ด ๋” ๋งŽ์ด ์ƒ๊ฐํ• ์ˆ˜๋ก ๋” ์Šค๋งˆํŠธํ•œ ๋‹ต๋ณ€์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. OpenAI, Grok-3, DeepSeek-R1๊ณผ ๊ฐ™์€ ๋ชจ๋ธ๋“ค์€ ์ถ”๋ก  ์‹œ๊ฐ„ ์Šค์ผ€์ผ๋ง์„ ์ ์šฉํ•˜๋Š” ์ถ”๋ก  ๋ชจ๋ธ๋“ค์ž…๋‹ˆ๋‹ค. ์ถ”๋ก  ๋ชจ๋ธ๋“ค์€ 100๋ฐฐ ๋” ๋งŽ์€ ์—ฐ์‚ฐ์„ ์†Œ๋น„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ๋ž˜์˜ ์ถ”๋ก  ๋ชจ๋ธ๋“ค์€ ํ›จ์”ฌ ๋” ๋งŽ์€ ์—ฐ์‚ฐ์„ ์†Œ๋น„ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. DeepSeek-R1์€ ์ „ ์„ธ๊ณ„์ ์ธ ์—ด๊ด‘์„ ๋ถˆ๋Ÿฌ์ผ์œผ์ผฐ์Šต๋‹ˆ๋‹ค -- ์ด๋Š” ํƒ์›”ํ•œ ํ˜์‹ ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋”์šฑ ์ค‘์š”ํ•œ ๊ฒƒ์€ ์„ธ๊ณ„์  ์ˆ˜์ค€์˜ ์ถ”๋ก  AI ๋ชจ๋ธ์„ ์˜คํ”ˆ์†Œ์Šค๋กœ ๊ณต๊ฐœํ–ˆ๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค. ๊ฑฐ์˜ ๋ชจ๋“  AI ๊ฐœ๋ฐœ์ž๋“ค์ด R1์ด๋‚˜ ์‚ฌ๊ณ ์˜ ์—ฐ์‡„(chain of thought) ๋ฐ R1๊ณผ ๊ฐ™์€ ๊ฐ•ํ™”ํ•™์Šต ๊ธฐ๋ฒ•๋“ค์„ ์ ์šฉํ•˜์—ฌ ์ž์‹ ๋“ค์˜ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ™•์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
We now have three scaling loss, as I mentioned earlier, driving the demand for AI computing. The traditional scaling loss of AI remains intact. Foundation models are being enhanced with multimodality, and pretraining is still growing. But it's no longer enough. We have two additional scaling dimensions. Post-training skilling, where reinforcement learning, fine-tuning, model distillation require orders of magnitude more compute than pretraining alone. Inference time scaling and reasoning where a single query and demand 100x more compute. We defined Blackwell for this moment, a single platform that can easily transition from pre-trading, post training and test time scaling.์•ž์„œ ๋ง์”€๋“œ๋ฆฐ ๋ฐ”์™€ ๊ฐ™์ด, ํ˜„์žฌ AI ์ปดํ“จํŒ… ์ˆ˜์š”๋ฅผ ๊ฒฌ์ธํ•˜๋Š” ์„ธ ๊ฐ€์ง€ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์ด ์žˆ์Šต๋‹ˆ๋‹ค. AI์˜ ์ „ํ†ต์ ์ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ์—ฌ์ „ํžˆ ์œ ํšจํ•ฉ๋‹ˆ๋‹ค. ํŒŒ์šด๋ฐ์ด์…˜ ๋ชจ๋ธ๋“ค์ด ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ๋กœ ํ–ฅ์ƒ๋˜๊ณ  ์žˆ์œผ๋ฉฐ, ์‚ฌ์ „ ํ›ˆ๋ จ(pretraining)๋„ ๊ณ„์† ์„ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๊ฒƒ๋งŒ์œผ๋กœ๋Š” ๋” ์ด์ƒ ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

๋‘ ๊ฐ€์ง€ ์ถ”๊ฐ€์ ์ธ ์Šค์ผ€์ผ๋ง ์ฐจ์›์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ๋Š” ์‚ฌํ›„ ํ›ˆ๋ จ ์Šค์ผ€์ผ๋ง์œผ๋กœ, ๊ฐ•ํ™”ํ•™์Šต, ํŒŒ์ธํŠœ๋‹, ๋ชจ๋ธ ์ฆ๋ฅ˜๊ฐ€ ์‚ฌ์ „ ํ›ˆ๋ จ๋งŒ์œผ๋กœ๋Š” ํ•„์š”ํ•œ ์ปดํ“จํŒ… ํŒŒ์›Œ๋ณด๋‹ค ๋ช‡ ๋ฐฐ ๋” ๋งŽ์€ ์—ฐ์‚ฐ์„ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ๋Š” ์ถ”๋ก  ์‹œ๊ฐ„ ์Šค์ผ€์ผ๋ง๊ณผ ์ถ”๋ก  ๊ณผ์ •์œผ๋กœ, ๋‹จ์ผ ์ฟผ๋ฆฌ๊ฐ€ 100๋ฐฐ ๋” ๋งŽ์€ ์ปดํ“จํŒ…์„ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์šฐ๋ฆฌ๋Š” ๋ฐ”๋กœ ์ด ์ˆœ๊ฐ„์„ ์œ„ํ•ด ๋ธ”๋ž™์›ฐ์„ ์ •์˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์ „ ๊ฑฐ๋ž˜, ์‚ฌํ›„ ํ›ˆ๋ จ, ๊ทธ๋ฆฌ๊ณ  ํ…Œ์ŠคํŠธ ์‹œ๊ฐ„ ์Šค์ผ€์ผ๋ง ๊ฐ„์— ์‰ฝ๊ฒŒ ์ „ํ™˜ํ•  ์ˆ˜ ์žˆ๋Š” ๋‹จ์ผ ํ”Œ๋žซํผ์ž…๋‹ˆ๋‹ค.
Blackwell's FP4 transformer engine and NVLink 72 scale-up fabric and new software technologies led Blackwell process reasoning AI models, 25x faster than Hopper. Blackwell in all of this configuration is in full production. Each Grace Blackwell NVLink 72 rack is an engineering marvel. 1.5 million components produced across 350 manufacturing sites by nearly 100,000 factory operators. AI is advancing at life speed. We are at the beginning of reasoning AI and inference time scaling. But we are just at the start of the age of AI, multimodal AIs, enterprise AI sovereign AI and physical AI are right around the corner. We will grow strongly in 2025.๋ธ”๋ž™์›ฐ์˜ FP4 ํŠธ๋žœ์Šคํฌ๋จธ ์—”์ง„๊ณผ NVLink 72 ์Šค์ผ€์ผ์—… ํŒจ๋ธŒ๋ฆญ, ๊ทธ๋ฆฌ๊ณ  ์ƒˆ๋กœ์šด ์†Œํ”„ํŠธ์›จ์–ด ๊ธฐ์ˆ ๋“ค์€ ๋ธ”๋ž™์›ฐ ํ”„๋กœ์„ธ์Šค ์ถ”๋ก  AI ๋ชจ๋ธ์„ ํ˜ธํผ ๋Œ€๋น„ 25๋ฐฐ ๋” ๋น ๋ฅด๊ฒŒ ๊ตฌ๋™ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋“  ๊ตฌ์„ฑ์˜ ๋ธ”๋ž™์›ฐ์€ ํ˜„์žฌ ๋ณธ๊ฒฉ์ ์ธ ์–‘์‚ฐ ์ค‘์ž…๋‹ˆ๋‹ค. ๊ฐ๊ฐ์˜ Grace Blackwell NVLink 72 ๋ž™์€ ์—”์ง€๋‹ˆ์–ด๋ง์˜ ๊ฒฝ์ด๋กœ์›€์ž…๋‹ˆ๋‹ค. 350๊ฐœ ์ œ์กฐ ์‚ฌ์ดํŠธ์—์„œ ๊ฑฐ์˜ 10๋งŒ ๋ช…์˜ ๊ณต์žฅ ์šด์˜์ž๋“ค์ด 150๋งŒ ๊ฐœ์˜ ๋ถ€ํ’ˆ์„ ์ƒ์‚ฐํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. AI๋Š” ์ƒํ™œ ์†๋„๋กœ ๋ฐœ์ „ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ถ”๋ก  AI์™€ ์ถ”๋ก  ์‹œ๊ฐ„ ์Šค์ผ€์ผ๋ง์˜ ์‹œ์ž‘์ ์— ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์šฐ๋ฆฌ๋Š” ์—ฌ์ „ํžˆ AI ์‹œ๋Œ€์˜ ์ถœ๋ฐœ์ ์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ AI, ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ AI, ์†Œ๋ฒ„๋ฆฐ AI, ๊ทธ๋ฆฌ๊ณ  ํ”ผ์ง€์ปฌ AI๊ฐ€ ๋ฐ”๋กœ ์ฝ”์•ž์— ๋‹ค๊ฐ€์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” 2025๋…„์— ๊ฐ•๋ ฅํ•œ ์„ฑ์žฅ์„ ์ด๋ฃฐ ๊ฒƒ์ž…๋‹ˆ๋‹ค.
Going forward, data centers will dedicate most of CapEx to accelerated computing and AI. Data centers will increasingly become AI factories and every company will have either renting or self-operated. I want to thank all of you for joining us today. Iโ€™m joining us at GTC in a couple of weeks. Weโ€™re going to be talking about Blackwell Ultra, Rubin and other new computing, networking, reasoning AI, physical AI products. And a whole bunch more. Thank you. Operator

This concludes today's conference call. You may now disconnect.
์•ž์œผ๋กœ ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋“ค์€ ๋Œ€๋ถ€๋ถ„์˜ ์ž๋ณธ์ง€์ถœ์„ ๊ฐ€์† ์ปดํ“จํŒ…๊ณผ AI์— ์ง‘์ค‘ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์„ผํ„ฐ๋“ค์€ ์ ์  ๋” AI ํŒฉํ† ๋ฆฌ๊ฐ€ ๋  ๊ฒƒ์ด๋ฉฐ, ๋ชจ๋“  ๊ธฐ์—…์ด ์ž„๋Œ€ํ•˜๊ฑฐ๋‚˜ ์ž์ฒด ์šด์˜ํ•˜๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ค๋Š˜ ์ฐธ์„ํ•ด ์ฃผ์‹  ๋ชจ๋“  ๋ถ„๋“ค๊ป˜ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค. ๋ช‡ ์ฃผ ํ›„ GTC์—์„œ ๋ต™๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ทธ๊ณณ์—์„œ Blackwell Ultra, Rubin ๊ทธ๋ฆฌ๊ณ  ๊ธฐํƒ€ ์ƒˆ๋กœ์šด ์ปดํ“จํŒ…, ๋„คํŠธ์›Œํ‚น, ์ถ”๋ก  AI, ๋ฌผ๋ฆฌ์  AI ์ œํ’ˆ๋“ค์— ๋Œ€ํ•ด ์ด์•ผ๊ธฐํ•  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ํ›จ์”ฌ ๋” ๋งŽ์€ ๊ฒƒ๋“ค์„ ๋‹ค๋ฃฐ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

๊ตํ™˜์›

์˜ค๋Š˜ ์ปจํผ๋Ÿฐ์Šค ์ฝœ์„ ๋งˆ์น˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด์ œ ์—ฐ๊ฒฐ์„ ๋Š์œผ์…”๋„ ๋ฉ๋‹ˆ๋‹ค.

๐Ÿ“Œ ์š”์•ฝ

Here's a summary of the key points from the earnings call transcript in Korean:

โ€ข ์‹ค์  ๋ฐ ์ „๋ง:
- ๋ธ”๋ž™์›ฐ(Blackwell) ์ œํ’ˆ์˜ ์ˆ˜์š”๊ฐ€ ๋งค์šฐ ๊ฐ•๋ ฅํ•˜๋ฉฐ, ์ƒ์‚ฐ์ด ๋ณธ๊ฒฉํ™”๋จ
- 2025๋…„์—๋„ ๊ฐ•ํ•œ ์„ฑ์žฅ์„ธ๊ฐ€ ์˜ˆ์ƒ๋จ
- ์ด์ด์ต๋ฅ ์€ ํ˜„์žฌ 70% ์ดˆ๋ฐ˜์ด๋‚˜ ์—ฐ๋ง๊นŒ์ง€ ์ค‘๋ฐ˜๋Œ€๋กœ ๊ฐœ์„  ์ „๋ง

โ€ข AI ๊ธฐ์ˆ  ํŠธ๋ Œ๋“œ:
- AI๊ฐ€ ์ธ์‹/์ƒ์„ฑ ๋‹จ๊ณ„๋ฅผ ๋„˜์–ด ์ถ”๋ก  ๋‹จ๊ณ„๋กœ ์ง„ํ™” ์ค‘
- 3๊ฐ€์ง€ ์ฃผ์š” ํ™•์žฅ ์š”์ธ: ์‚ฌ์ „ํ•™์Šต, ์‚ฌํ›„ํ•™์Šต, ์ถ”๋ก ์‹œ๊ฐ„ ํ™•์žฅ
- ์ถ”๋ก  AI ๋ชจ๋ธ์€ ๊ธฐ์กด ๋Œ€๋น„ 100๋ฐฐ ์ด์ƒ์˜ ์ปดํ“จํŒ… ํŒŒ์›Œ ํ•„์š”

โ€ข ์‚ฌ์—… ์ „๋žต:
- ๋ธ”๋ž™์›ฐ ํ”Œ๋žซํผ์€ ๋ชจ๋“  AI ์›Œํฌ๋กœ๋“œ์— ์ตœ์ ํ™”๋œ ํ†ตํ•ฉ ์•„ํ‚คํ…์ฒ˜ ์ œ๊ณต
- ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ AI, ๊ตญ๊ฐ€๋ณ„ AI, ๋ฌผ๋ฆฌ AI ๋“ฑ ์ƒˆ๋กœ์šด ์‹œ์žฅ ๊ธฐํšŒ ํ™•๋Œ€
- ๋ฐ์ดํ„ฐ์„ผํ„ฐ๊ฐ€ AI ํŒฉํ† ๋ฆฌ๋กœ ์ง„ํ™”ํ•˜๋Š” ์ถ”์„ธ์— ๋งž์ถ˜ ์ „๋žต ์ถ”์ง„

์ด ๋‚ด์šฉ์€ ํˆฌ์ž์ž๋“ค์—๊ฒŒ NVIDIA์˜ ํ˜„์žฌ ์‹ค์ ๊ณผ ๋ฏธ๋ž˜ ์„ฑ์žฅ ์ž ์žฌ๋ ฅ์„ ๊ฐ๊ด€์ ์œผ๋กœ ์ „๋‹ฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.