AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: Inference 40% of DC revenue, is it really that big? How do you know?
A: It was a great exercise for us to work on, our work that we did, we do know our largest systems and we know engineering teams at all the customers. And we can categorize all the use cases. These are early days in Gen-AI, some have moved to Copilots and monetization. Recommender engines are an enormous part of marketing and all had to be re-designed. Search is another one. Our work is based on the inferencing of the future, not of the past 30 years.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: Inference 40% of DC revenue, is it really that big? How do you know?
A: It was a great exercise for us to work on, our work that we did, we do know our largest systems and we know engineering teams at all the customers. And we can categorize all the use cases. These are early days in Gen-AI, some have moved to Copilots and monetization. Recommender engines are an enormous part of marketing and all had to be re-designed. Search is another one. Our work is based on the inferencing of the future, not of the past 30 years.
tweet
AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: China question, shipping to China
A: Density and performance are part of the export controls. China is interested in working with Nvidia as they have for decades. We have products keyed out for them, as they review the performance and software. We have worked with the US gov't so they are aware. Our China partners want to use our products for the long term. We don't know the future, but we will follow the rules today. Will there be a chance performance can increase in the future? We don't know yet.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: China question, shipping to China
A: Density and performance are part of the export controls. China is interested in working with Nvidia as they have for decades. We have products keyed out for them, as they review the performance and software. We have worked with the US gov't so they are aware. Our China partners want to use our products for the long term. We don't know the future, but we will follow the rules today. Will there be a chance performance can increase in the future? We don't know yet.
tweet
AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: Sovereign AI. Where is unfulfilled demand coming from, is sovereign hyperbole?
A: Many areas we have not been able to fulfill. We are seeing Gen-AI proliferating software. We have talked about sovereign AI and the unique opportunities. Chat-GPT is US-based, US culture, US language. So LLMs in other regions are important. This stream of interest, there is a very sizable pipeline for that, not been able to fulfill.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: Sovereign AI. Where is unfulfilled demand coming from, is sovereign hyperbole?
A: Many areas we have not been able to fulfill. We are seeing Gen-AI proliferating software. We have talked about sovereign AI and the unique opportunities. Chat-GPT is US-based, US culture, US language. So LLMs in other regions are important. This stream of interest, there is a very sizable pipeline for that, not been able to fulfill.
tweet
AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: Visibility on products not announced yet.
A: We have connections with our customers for over a decade, they are not surprised with what is coming and have specs in advance. We also have ideas of their demand expectations. Helpful for us when we build out our architectures. Demand may exceed our supply, as Jensen said.
Q: Transition to new products, will there be stall in old ones? Or you know where all of H100s are going.
A: We sped up from 2 year cadence to a 1 year cadence. What we see time and time again, when you are within a certain architecture, takes time to qualify it in their system. Many in this room and in this city have not been able to touch an H100. Even as we launch new products, availability of H100 will be important.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: Visibility on products not announced yet.
A: We have connections with our customers for over a decade, they are not surprised with what is coming and have specs in advance. We also have ideas of their demand expectations. Helpful for us when we build out our architectures. Demand may exceed our supply, as Jensen said.
Q: Transition to new products, will there be stall in old ones? Or you know where all of H100s are going.
A: We sped up from 2 year cadence to a 1 year cadence. What we see time and time again, when you are within a certain architecture, takes time to qualify it in their system. Many in this room and in this city have not been able to touch an H100. Even as we launch new products, availability of H100 will be important.
tweet
AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: Data center spend
A: Each year, about $250B capex on data center each year, this past year it actually increased for the first time in a long time. It was focused on accelerated computing and extending of asset life.
When you think through used of capital, ROI, will prioritize projects. Right now AI is higest priority. Question is, will they continue to upgrade, non-high ROI investments...probably not.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: Data center spend
A: Each year, about $250B capex on data center each year, this past year it actually increased for the first time in a long time. It was focused on accelerated computing and extending of asset life.
When you think through used of capital, ROI, will prioritize projects. Right now AI is higest priority. Question is, will they continue to upgrade, non-high ROI investments...probably not.
tweet
AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: Lead times, power, rack space, end demand and meeting it.
A: About this time a year ago, people got in line for what they wanted. We are getting through a good portion of that, for primarily the H100.
Keep in mind, we have new products coming to market. That enters the next stage of supply and demand. Helping our customers as we bring new products to market to understand how to build out the data center. Planning processes are long.
Lot of trapped power that has been used with inefficient data center builds, will need to be re-done.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: Lead times, power, rack space, end demand and meeting it.
A: About this time a year ago, people got in line for what they wanted. We are getting through a good portion of that, for primarily the H100.
Keep in mind, we have new products coming to market. That enters the next stage of supply and demand. Helping our customers as we bring new products to market to understand how to build out the data center. Planning processes are long.
Lot of trapped power that has been used with inefficient data center builds, will need to be re-done.
tweet
AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: How have you managed to grow this quickly. This scale, size, complexity. Growing 4-5x is remarkable.
A: Ramping up supply, had to come from many different perspectives. Decades long supplier relationships, seeked out new suppliers to build out redundancies, we focused on cycle time of manufacturing and breaking it down to see how we could improve it. We want to increase our supply every quarter.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: How have you managed to grow this quickly. This scale, size, complexity. Growing 4-5x is remarkable.
A: Ramping up supply, had to come from many different perspectives. Decades long supplier relationships, seeked out new suppliers to build out redundancies, we focused on cycle time of manufacturing and breaking it down to see how we could improve it. We want to increase our supply every quarter.
tweet
AkhenOsiris
$NVDA CFO Kress at MS TMT Conference
Q: Did you foresee this DC demand a year ago?
A: Intro of Gen-AI was still ramping. We had been working with Open AI for years. We saw it as part of our long planned journey. Things have definitely changed, because of the interest worldwide. Enterprises, governments, consumers all over the world. Our overall goal as a company has been focused on accelerated computing for over 15 years. That transition is arriving. A new platform is necessary and AI is the killer app of accelerated computing.
tweet
$NVDA CFO Kress at MS TMT Conference
Q: Did you foresee this DC demand a year ago?
A: Intro of Gen-AI was still ramping. We had been working with Open AI for years. We saw it as part of our long planned journey. Things have definitely changed, because of the interest worldwide. Enterprises, governments, consumers all over the world. Our overall goal as a company has been focused on accelerated computing for over 15 years. That transition is arriving. A new platform is necessary and AI is the killer app of accelerated computing.
tweet
AkhenOsiris
*** Alert ***
This is a feel good capitalism story amongst all the anger out there.
3 of my passive 🐵 buddies have locked in their profits today from SMCI and a few other semis. They are tech finance nerds, so understand very well how markets work, not the guy in mom's basement. Generated well over 7 figure gains since last year on their semi portfolios.
One said to me this morning "both my kids' education funds are now complete" and another said "I just bought my parents vacation house on the beach this morning back in their home country".
I don't give a flying fuck if markets are broken, but I'm happy as hell to hear the above 👍🏼
tweet
*** Alert ***
This is a feel good capitalism story amongst all the anger out there.
3 of my passive 🐵 buddies have locked in their profits today from SMCI and a few other semis. They are tech finance nerds, so understand very well how markets work, not the guy in mom's basement. Generated well over 7 figure gains since last year on their semi portfolios.
One said to me this morning "both my kids' education funds are now complete" and another said "I just bought my parents vacation house on the beach this morning back in their home country".
I don't give a flying fuck if markets are broken, but I'm happy as hell to hear the above 👍🏼
tweet
AkhenOsiris
Will be lots of "how can the market work without Apple?" commentary kicking up again (right @jkrinskypga ?)
tweet
Will be lots of "how can the market work without Apple?" commentary kicking up again (right @jkrinskypga ?)
tweet
BizToc
3) Apple Hit with EU's $2B Fine —
EU fines Apple nearly $2 billion for anti-competitive practices in music streaming, shaking up the tech industry. #Antitrust
tweet
3) Apple Hit with EU's $2B Fine —
EU fines Apple nearly $2 billion for anti-competitive practices in music streaming, shaking up the tech industry. #Antitrust
tweet
Offshore
Photo
BizToc
📰 BizToc.com Hourly News Flash
1) China's Confident Growth Ambitions —
China announces an assertive GDP growth target, fueling confidence and exhibiting economic resilience. #GDP
tweet
📰 BizToc.com Hourly News Flash
1) China's Confident Growth Ambitions —
China announces an assertive GDP growth target, fueling confidence and exhibiting economic resilience. #GDP
tweet
Offshore
Photo
Ben's Bites
RT @_philschmid: Claude 3 is here! @AnthropicAI just released Claude 3, outperforming the original @OpenAI AI GPT-4 results! 🤯 Claude 3 also comes with Vision support similar to GPT-4V. 👀
TL;DR;
🚀 3 different versions (Opus, Sonnet, Haiku)
🥇 Opus (best models) outperforms GPT-4 and Gemini 1.0 Ultra
📚 200k context window
🖼️ Vision support (59% on MMMU)
⚡ Sonnet is 2x faster than Claude 2
🤖 Can be used for task automation/agent workflows
🌐 Sonnet available in @awscloud Bedrock and private preview on @Google Cloud’s Vertex AI Model Garden → Opus and Haiku coming soon
🔜 Soon: function calling & interactive coding (REPL)
Cost:
Opus: $15/1M input (0.5x of GPT-4); $75M output (1.25x of GPT-4)
Sonnet: $3/1M input (0.33x of GPT-4 Turbo); $15M output (0.5x of GPT-4 Turbo)
Haiku: $0.25/1M input (0.5x of GPT-3.5 Turbo); $1.25M output (1.2x of GPT-3.5 Turbo)
tweet
RT @_philschmid: Claude 3 is here! @AnthropicAI just released Claude 3, outperforming the original @OpenAI AI GPT-4 results! 🤯 Claude 3 also comes with Vision support similar to GPT-4V. 👀
TL;DR;
🚀 3 different versions (Opus, Sonnet, Haiku)
🥇 Opus (best models) outperforms GPT-4 and Gemini 1.0 Ultra
📚 200k context window
🖼️ Vision support (59% on MMMU)
⚡ Sonnet is 2x faster than Claude 2
🤖 Can be used for task automation/agent workflows
🌐 Sonnet available in @awscloud Bedrock and private preview on @Google Cloud’s Vertex AI Model Garden → Opus and Haiku coming soon
🔜 Soon: function calling & interactive coding (REPL)
Cost:
Opus: $15/1M input (0.5x of GPT-4); $75M output (1.25x of GPT-4)
Sonnet: $3/1M input (0.33x of GPT-4 Turbo); $15M output (0.5x of GPT-4 Turbo)
Haiku: $0.25/1M input (0.5x of GPT-3.5 Turbo); $1.25M output (1.2x of GPT-3.5 Turbo)
tweet