MSFT CFO commented already on capex spend for its Q2 FY 2025 (ending December) to sequentially increase its capex spend. MSFT spend USD20bn in its Q1, so if you assume a q/q increase as CFO communicated (read their transcript) and assume USD25-30bn in its Q2, all its USD80bn is front end loaded and capex should fall consequently in coming quarters which should not bode well for NVDA
Second, my guess is that Meta's Zuckerberg is a "killer" CEO who does not take any prisoners, and he will put all R&D guys to analyze DS's algo and model and try to copy as much as he can to implement into his open source Llama models, OpenAI is not open source, and once he realizes better outcome as peers, he will cut first on capex as he has skin in the game and will benefit tremendously once he increases Meta's FCF, dividend and buyback...no brainer for a CEO like Zuck and to "DeepZuck" the competition...
Good insights. I doubt MSFT lets off on their near-term capex plans, but if the efficiency gains from Deepseek are real and if efficiency gains increase even further, that means MSFT and most of the other tech companies won't need to do much more capex spending for 3-5 years.
Asking for my husband with a phd in physics—if you could choose between leading design/prototyping of GPUs at OpenAI or a quant/research scientist at Nvidia…what’s a safer bet? Sounds like Microsoft should maybe be a consideration too?
Nickel is a good recent example. Back in 2006/7 both prices ramped up, and the Chinese began developing nickel pig iron - which was a previously low grade resource that had been ignored. Its development flooded the Nickel market, and collapsed prices. https://en.wikipedia.org/wiki/Nickel_pig_iron
we have an update on capex spend for Q2: USD22.6bn so exactly what CFO stated reg q/q growth in capex, wondering how quick they are going to reach USD80bn, more than 50% reached of that after H1...more color in their conf call in 1h
I'm interested in your opinion on this take, Russell:
"DeepSeek’s models were trained using Nvidia chips, so it’s not obvious why DeepSeek’s success would be bad news for Nvidia. And it’s even harder to explain why it took a week for Wall Street to react to the January 20 release of R1.
A more plausible explanation is that someone tipped traders off to Donald Trump’s plans to slap tariffs on chips made in Taiwan—which Trump announced later in the day. I can’t prove this theory, but I think it fits the facts better than the DeepSeek theory. Interestingly, we didn’t see a second selloff in Nvidia or TSMC shares after Trump’s announcement, suggesting that markets had already “priced in” the news."
As I understand it, they were used on Nvidia chips - but relatively old one, and more importantly, the sort of chips that SMIC is either able to or close to being able to build. It's stock(981 HK) suddenly soared higher from September - so market seems to be saying that they have been able to get around sanctions to do something impressive. This is also a time when other chip companies were falling (Samsung, Intel etc) . To me it does seem to imply Nvidia cutting edge chips are too expensive - a complaint I have heard a few times.
I don’t think those chips were needed in the same quantities or had the same margins but I’m not an expert. Implication is you need less over all to train and run models so sales won’t be what everyone thought.
There is also this angle that it won't be the U.S. that locks us out of DeepSeek with sanctions, but rather the Chinese government that will pay too much 'attention' to Liang Wenfang's company.
I was thinking, 'Wow, so it means that Europe could still, in theory, catch up in AI,' but it's just not that simple. It looks like it costs just a fraction of what American efforts cost, but Meta also has a bunch of excellent engineers; why didn't they come up with it?
I'm in the midst of reading 'The Nvidia Way,' and I'm really curious about what Jensen Huang will do this time. He's had many near-death experiences, it's hard to believe he didn't foresee this one happening.
Another important point in my mind - after claiming the US is 2-3 years ahead of China in AI, how soon until China debuts GPUs that are on par with Nvidia? and a few years later taking major market share like they did with EVs..
A few years ago it was easy to track Chinese progress - now all that tends to be hidden to a degree. I guess why this was such a surprise to markets - but it also follows a similar pattern where first Japan, then Korea and now China - where new entrants are seen as low quality until they are not.
Seems to me like Microsoft was already going down this route. It had broken with OpenAI over its strategy of spending ever more money on training. Microsoft is building out the infrastructure to deliver inference at scale. I would assume they are going to adopt whatever tech is most efficient but I wouldn’t necessarily say that spells doom for NVIDIA. I don’t know enough to say for sure but could Jevon’s Paradox just lead to way more consumption of AI that makes NVIDIA key even if it’s in inference rather than training. Remember the chips Deep Seek used were still NVIDIA ones. Maybe not the same price or margin but if you use more, it might over come that delta. Anyway, in my mind too soon to tell either way but I can always rely on you for the anti-US view.
I also wouldn’t rule out the AI breakout theory. If you can get to an AI the is super intelligent, can it begin improving itself faster than any humans, etc. There are many rounds in this game yet to played.
MSFT CFO commented already on capex spend for its Q2 FY 2025 (ending December) to sequentially increase its capex spend. MSFT spend USD20bn in its Q1, so if you assume a q/q increase as CFO communicated (read their transcript) and assume USD25-30bn in its Q2, all its USD80bn is front end loaded and capex should fall consequently in coming quarters which should not bode well for NVDA
Second, my guess is that Meta's Zuckerberg is a "killer" CEO who does not take any prisoners, and he will put all R&D guys to analyze DS's algo and model and try to copy as much as he can to implement into his open source Llama models, OpenAI is not open source, and once he realizes better outcome as peers, he will cut first on capex as he has skin in the game and will benefit tremendously once he increases Meta's FCF, dividend and buyback...no brainer for a CEO like Zuck and to "DeepZuck" the competition...
Good insights. I doubt MSFT lets off on their near-term capex plans, but if the efficiency gains from Deepseek are real and if efficiency gains increase even further, that means MSFT and most of the other tech companies won't need to do much more capex spending for 3-5 years.
Asking for my husband with a phd in physics—if you could choose between leading design/prototyping of GPUs at OpenAI or a quant/research scientist at Nvidia…what’s a safer bet? Sounds like Microsoft should maybe be a consideration too?
Sounds like he has good job prospects!
This is outside of my skillset to advise here - but may I suggest you get as much cash compensation upfront - rather than options.
Good luck
I appreciate this deep dive, very interesting. I’d be curious if you could compare your opinions here to similar historic examples
Nickel is a good recent example. Back in 2006/7 both prices ramped up, and the Chinese began developing nickel pig iron - which was a previously low grade resource that had been ignored. Its development flooded the Nickel market, and collapsed prices. https://en.wikipedia.org/wiki/Nickel_pig_iron
Is this what you mean?
Great example! Thank you 😁
So what's the trade here Russell (aside from shorting NVDA)
Buy SMIC vs Short TSMC?
we have an update on capex spend for Q2: USD22.6bn so exactly what CFO stated reg q/q growth in capex, wondering how quick they are going to reach USD80bn, more than 50% reached of that after H1...more color in their conf call in 1h
I'm interested in your opinion on this take, Russell:
"DeepSeek’s models were trained using Nvidia chips, so it’s not obvious why DeepSeek’s success would be bad news for Nvidia. And it’s even harder to explain why it took a week for Wall Street to react to the January 20 release of R1.
A more plausible explanation is that someone tipped traders off to Donald Trump’s plans to slap tariffs on chips made in Taiwan—which Trump announced later in the day. I can’t prove this theory, but I think it fits the facts better than the DeepSeek theory. Interestingly, we didn’t see a second selloff in Nvidia or TSMC shares after Trump’s announcement, suggesting that markets had already “priced in” the news."
That's from Timothy B. Lee: https://www.understandingai.org/p/i-dont-believe-deepseek-crashed-nvidias
As I understand it, they were used on Nvidia chips - but relatively old one, and more importantly, the sort of chips that SMIC is either able to or close to being able to build. It's stock(981 HK) suddenly soared higher from September - so market seems to be saying that they have been able to get around sanctions to do something impressive. This is also a time when other chip companies were falling (Samsung, Intel etc) . To me it does seem to imply Nvidia cutting edge chips are too expensive - a complaint I have heard a few times.
I don’t think those chips were needed in the same quantities or had the same margins but I’m not an expert. Implication is you need less over all to train and run models so sales won’t be what everyone thought.
Prices should fall for cutting edge - as the economics of making money on them has been destroyed
Great piece, Russell, thank you.
There is also this angle that it won't be the U.S. that locks us out of DeepSeek with sanctions, but rather the Chinese government that will pay too much 'attention' to Liang Wenfang's company.
I was thinking, 'Wow, so it means that Europe could still, in theory, catch up in AI,' but it's just not that simple. It looks like it costs just a fraction of what American efforts cost, but Meta also has a bunch of excellent engineers; why didn't they come up with it?
I'm in the midst of reading 'The Nvidia Way,' and I'm really curious about what Jensen Huang will do this time. He's had many near-death experiences, it's hard to believe he didn't foresee this one happening.
Another important point in my mind - after claiming the US is 2-3 years ahead of China in AI, how soon until China debuts GPUs that are on par with Nvidia? and a few years later taking major market share like they did with EVs..
A few years ago it was easy to track Chinese progress - now all that tends to be hidden to a degree. I guess why this was such a surprise to markets - but it also follows a similar pattern where first Japan, then Korea and now China - where new entrants are seen as low quality until they are not.
Seems to me like Microsoft was already going down this route. It had broken with OpenAI over its strategy of spending ever more money on training. Microsoft is building out the infrastructure to deliver inference at scale. I would assume they are going to adopt whatever tech is most efficient but I wouldn’t necessarily say that spells doom for NVIDIA. I don’t know enough to say for sure but could Jevon’s Paradox just lead to way more consumption of AI that makes NVIDIA key even if it’s in inference rather than training. Remember the chips Deep Seek used were still NVIDIA ones. Maybe not the same price or margin but if you use more, it might over come that delta. Anyway, in my mind too soon to tell either way but I can always rely on you for the anti-US view.
Nvidia pricing structure seems wrong to me now...
I am not anti-US - but I am anti monopolies.... which seem to be mainly in the US these days.. not long ago the US was also anti-monopolies
I also wouldn’t rule out the AI breakout theory. If you can get to an AI the is super intelligent, can it begin improving itself faster than any humans, etc. There are many rounds in this game yet to played.