Leather Jacket Man going to meet with Trump, interesting timesđ€
2 Likes
Itâs on a pretty good upswing, the rate is soaring at 127 dollars and change right now.
It would be interesting to be a fly on the wall in the Oval Office right now.
4 Likes
Nvidiaâs fiscal year is nearing its end.
The DeepSeek nervousness is also clearly subsiding. This was probably partly calmed by the fact that the megatechs stated that DeepSeek is not changing their investment plans and indicated growing (AI) capex spend also going forward.
(Of course, itâs certainly partly true that itâs not in their interest to directly admit that it was a mistake to invest so massively. It will therefore be even more important in the future to critically monitor capex development. It is also noteworthy that the markets did not punish the fact that capex levels were reported to be growing also in the future).
Beth Kindig wrote another excellent analysis today on why DeepSeek is not an Nvidia killer, and why this nervousness is a buying opportunity: DeepSeek Creates Nvidia Stock Buying Opportunity
The earnings report will be out in a couple of weeks. In my opinion, expectations are very modest given the indication provided by TSMCâs massive earnings.
My own revenue forecast, which is derived through TSMCâs HPC revenue (which is Nvidiaâs COGS), would show at least $39.5B in revenue in my own calculations, using the lower end of the GM% guidance and last quarterâs COGS / HPC Revenue ratio.
Remembering that TSMCâs Q4 saw strong hockey stick growth towards the end of the year (bodes well for Blackwell), it is relatively likely, at least in my books, that Nvidiaâs revenue for Q4 will be over $40B. My median estimate is $41.1B. Wall Street consensus is $38B.
Here are the parameters used above and an explanation:
TSMCâs HPC Revenue: $26.9B x 53% = $14.3B
Nvidiaâs COGS / TSMC HPC ratio = $14.3B / 77 % = $11B
Nvidiaâs revenue = COGS / (1-GM%) = $11 B / (1 - 73%) = $41.1B
18 Likes
Trump tried to order an RTX 5090 and didnât get one? Now Jensen is going to get an earful⊠
Itâs quite ridiculous that for a big release, NVIDIA can only deliver a few thousand units to the market, which probably would have absorbed hundreds of thousands of units.
6 Likes
Singapore has server halls for several Chinese cloud firms. You donât have to be a great detective to figure out where the goods went. Whether it went to a local hall or was moved further on can vary, but itâs clear that sanctions leak like a sieve.
And on the consumer card side, it seems that Chinaâs 5090D cards were available significantly more than the rest of the worldâs 5090 card. Yes, its AI performance has been weakened to meet export restrictions, but as has already been seen, these restrictions donât necessarily mean much, and this cardâs biggest selling point for AI use is 32GB of video memory. Although it must be admitted that this could be because some of the cards for the rest of the world were still en route, while the journey to the Chinese logistics chain is shorter.
And as for Singapore - if Trump blocks sales there, next NVIDIAâs data will reveal how sales rocket to Vietnam, then Indonesia, and⊠and weâll go through all of Asia one by one as Chinese intermediaries change locations.
8 Likes
Jim Cramer told investors that Chinese AI company DeepSeek may not be as big a threat to Nvidiaâs sales as feared. He referred to a SemiAnalysis report suggesting that DeepSeekâs reported development costs might be misleading, e.g., the company might have spent over $500 million instead of the alleged $6 million.
Earlier in the week, DeepSeek claimed to have developed an AI model for only $6 million, which surprised the markets and led to a sharp drop in Nvidiaâs stock price. Investors feared that large companies would no longer need expensive Nvidia chips.
However, Cramer doubts DeepSeekâs claims and believes this might just be a way to drive down Nvidiaâs stock price. So, no worries, Cramer also trusts Nvidiaâs future. 
4 Likes
Now we just have to hope that Trump doesnât pull the rug out from under us. Jensenâs visit might cause trouble because itâs quite clear that China sanctions have been circumvented (Singapore didnât become an AI superpower, but rather proxies are at work there) and Trump is not happy.
On the other hand, even if attempts are made to plug the sanctions, order books are so thick for now that this would only primarily shift delivery addresses far into the future and wouldnât change delivery volumes, which are still capacity-limited. The main risk is that the market looks ahead and panics if, instead of an NVIDIA data center chip ordered today being delivered in 12 months, it can be obtained in just 6 months. 
On the consumer graphics card side, demand is also strong, but the release of the 5080/5090 cards the day before yesterday was widely labeled a âpaper launchâ among consumers and the press. Cards were delivered, especially for the 5090 series, in very small quantities, e.g., compared to the launch day of the previous generation. Demand is so hot that one could joke that NVIDIAâs new distribution partners are Ebay and Facebook Marketplace, and domestically tori.fi and huuto.net, as those are the only places where cards can be bought. With a suitable premium, of course.
However, this might also just be a small hiccup because we already knew before the launch day that external manufacturers (ASUS, MSI, Gigabyte, etc.) had already once requested and received a one-week postponement of the launch date, which suggests that the schedule was too tight and goods were not put into circulation as planned. If itâs just a matter of a huge batch of cards being in the logistics chain and them suddenly appearing on the market within the next month, the situation could turn around quickly. Now one can mainly criticize that perhaps NVIDIA could have just taken it easier and pushed the launch forward by a month to ensure enough stock was ready.
However, it can be said that everything is being snapped up in the consumer market. Itâs another matter whether anyone on the stock market cares, as discrete graphics cards are such a small slice of NVIDIAâs business nowadays.
11 Likes
Damodaran on DeepSeekâs impact on the AI narrative. Nvidia in focus.
4 Likes
These newly released cards are, unfortunately for the consumer, historically poor and give AMD an opportunity to re-enter the competition in the high-end market. Now, the main question is whether AMD will make a classic counter-blunder, because a sensible launch would easily allow them to seize large market shares.
The 5090 can, with gritted teeth, be called a tolerable release for prosumers running AI at home, but for everyone else, the RTX 50 is a completely automatic skip, as the improvement over previous generations is either non-existent or clearly below expectations. Of course, every generation canât bring a new 8800 GT or 1080 Ti, but here, they didnât even try to win over consumers after AMD announced they couldnât compete in this generation.
Fortunately, from an investorâs perspective, this doesnât have much significance, as Nvidiaâs fate is decided in the high-margin AI computing markets, and one only needs to wait a couple of years for the next generation of consumer graphics cards to arrive.
8 Likes
It must be remembered that there are a huge number of players worldwide who have a 20- or 30-series card or an older AMD because thereâs generally no point in buying a new one every generation unless they have an excessive amount of cash. There is demand. In addition, among players, there is a small but affluent group that has that cash and always needs to have the best machine. 4090s fly to the used marketplace and 5090s take their place.
NVIDIA didnât need to do better. AMD surrendered this generation in advance when no high-end built from chiplet dies was made. So, the minimum uplift was made with which these can be sold under a new number, and then a large chip for the prosumer/budget-AI sector, whose real home is actually the upcoming RTX 6000 Blackwell (or whatever name it will be sold under), which has a so-called full version of that 5090 chip and much more memory, and the price will probably be around ten thousand. Home-5090s are where the faulty chips are then pushed, where CUDA cores have to be disabled. These are processed for a quarter or two, and the chips where all CUDA cores work are collected, and then a professional card is released. There are rumors that there would be up to 96GB of memory in that model, and then perhaps a leaner 48GB version will come alongside it at a slightly lower price (but still certainly more than double the price of the 5090).
I suspect NVIDIA has room to stretch more if AMD wakes up to compete on GPUs again. For example, even with the RTX 5080, there was room left on the table â according to reports, the chip (in good samples) overclocks well, and there could be 24GB of memory instead of 16GB. So, if for some reason the AMD 9070XT suddenly bothered RTX 5080 sales (unlikely but not impossible because the uplift vs. 4080 Super was so small), NVIDIA can, with about 3-6 monthsâ notice, release an âRTX 5080 Tiâ card with increased clock speeds and the memory amount raised to 24 gigabytes.
And we already know that the next batch will be at the end of 2026/beginning of 2027. If then AMD and the new UDNA architecture try something again. And NVIDIA knows this, so for the next round, there will be new and better things from them too. Thatâs how the graphics chip business works.
10 Likes
Will even Trumpâs nerve be enough to severely cripple exports worth ~100+ billion p.a.? The deficit will grow, and Tariff-Trump will have even more pressure to reduce imports from elsewhere. He is, however, so unpredictable that the risk always exists.
It would have, but as you yourself mentioned, they have given up in advance. The best of this generation is max. something like 7900 XT level according to AMDâs own slides. I have a 4070 Ti and Iâm considering an upgrade. The next performance tier from that is the 5080, and there are no alternatives in sight. Iâm not counting used ones, even though a used 4090 would be a strong contender in this generation too. A really lousy market for the consumer. A 7900 XTX level card with FSR 4 at a reasonable price would have been a really competitive product at this point. Letâs keep our fingers crossed that AMD delayed the release to make a new BIOS for the card, which will raise the power limit and increase performance closer to the 5080 card. 
I watched JayTwoCentsâ video on 5080 overclocking this morning, and his well-overclocking card was 15% faster than the reference. Such a 10% - 15% overclocking headroom and small memory really screams that in a couple of years, a TI/Super model will come out, which will take the last crumbs from the table. Nvidiaâs tick-tock model.
3 Likes
A good video that gives some insight into why the 5090 costs so much that it makes the faint-hearted tremble. Well, in addition to NVIDIAâs profit margins. Itâs not cheap to manufacture.
5 Likes
Iâve been pondering the impact of DeepSeek for the past few days.
Iâve previously stated that some algorithmic change that fundamentally alters hardware-level requirements would be the worst negative driver for Nvidiaâs investment case.
DeepSeek was certainly not such a change. On the contrary, the hardware-level support of Nvidiaâs Hopper architecture enabled the birth of DeepSeek. DeepSeekâs FP8 quantization, which would not have been possible with Ampere, was the key factor affecting memory and computational requirements.
Blackwell will bring even more opportunities to roughly halve the memory and computational requirements of a conventional transformer architecture even from the DeepSeek level now seen, with FP4 support. In addition, there will be support for FP8 micro-scaling, meaning that DeepSeekâs per-group GEMM scaling will move directly to element-level scaling of tensors (which should improve efficiency and presumably also the modelâs learning capability).
It seems clear, then, that DeepSeekâs algorithmic improvements did not come as a surprise to Nvidia, as they have indeed enabled these improvements with their enhancements and will enable even more drastic computational and memory cuts with new architectures.
So, I suppose it must be stated that this was not an algorithmic breakthrough that would break Nvidiaâs dominance, if Nvidiaâs improvements have, after all, provided the framework for these efficiencies. And it is quite certain that with Blackwellâs innovations, we will see another minimum 50% cut in memory and computational requirements.
In my view, this is entirely in line with what Huang has stated, for example, in the BG2 interview: âEvery single year, we reduce the cost by two or three times every single year. We improve the energy efficiency by two or three times. Incredible. Right? And so we ask our customers, donât buy everything at one time. Buy a little every year. Right? Okay. And the reason for that, we want them cost average into the futureâ
17 Likes
Thatâs how it is, well said @Roope_K. This is how it should go, that optimization happens at different stages of the âvalue chainâ.
4 Likes
There was just panic here when a rumor was heard that could indicate NVIDIAâs order books might thin out as megatechs cut hardware purchases. It turned out that itâs quite certainly a Chinese rumor with a bit of exaggeration, and that was thatâŠ
5 Likes
5 hours of technical deep dive for those who want to understand what DeepSeek etc. is all about. Also includes repeated references to Nvidia.
Edit: if you donât have 5 hours, then the main points related to DS can be found in this 15-min read from the podcastâs main guest: DeepSeek Debates: Chinese Leadership On Cost, True Training Cost, Closed Model Margin Impacts
5 Likes
@Elina_Lappalainen and @Alex_af_Heurlin1 discussed Nvidia with @Jukka_Lepikko. 
Duration 42 minutes.
AI is a story loved by investors, but is it a bubble? Who actually makes money from AI? How much have Finnish investors invested in Nvidia?
Hosts Elina Lappalainen and Alex af Heurlin, guest investor Jukka Lepikkö. The episode was edited by Jonne Piltonen. The producer of the HS Visio podcast is Tuomas PeltomÀki.
9 Likes
Hei, thanks for sharing our podcast episode here!
In this regard, I was thinking it would be quite interesting to talk to Finnish small investors who have a big bet on Nvidia. Where did you invest, how much, whatâs the return, and what do you think about the valuation level now, etc.
I find what Jukka said towards the end of the episode interesting, that Nvidia and other Mag 7 companies have become new âpeopleâs stocksâ in Finnish portfolios. And when things are stirring there, someone has gotten rich - and many are hoping for returns from it. Do you know anyone who has genuinely become wealthy in Finland with Nvidia shares? You can contact me! elina.lappalainen@hs.fi
11 Likes
This week, final realized and refined data on the 2024 Capex investments of the Big Tech (AMZN+MSFT+META+GOOGL) companies discussed above, as well as indications for the current yearâs Capex levels, have been received.
2024 capex investments were almost exactly $250B.
The indication for 2025e capex investments is $320B. So, from the optimistic âupper rangesâ I compiled a month ago, we have already moved up by $10B again.
Itâs still a strong push when we talk about +10% upward revisions to analystsâ estimates within a single month. For example, from BG2âs investment estimate, which usually has an excellent view of the future, there was a 15% increase in 30 days.
In one quarter, capex expectations have risen enormously:
If I had to guess, that $320B will probably be refined upwards further. A year ago, analysts estimated the capex investment level 12 months ahead to be -20% too low. From these levels, such large upward corrections are unlikely to be seen, but I no longer consider it impossible that the realized Big Tech capex at the end of the year will be around $350B. That would require a 10% upward correction compared to what is now indicated.
This, of course, does not yet seal a peak year for Nvidia in CY25, but it already makes a revenue of at least close to $100B relatively certain, solely thanks to these investment intentions.
As expected, DeepSeek did not shake the willingness to invest on the quest towards AGI / ASI. My bullish guess is that mega-investments will continue, and the demand for H- and B-series products will only grow as predicted by Jevonsâ paradox, as the capital moat for reasonable model training costs broke.
17 Likes
Indeed. Amazonâs and Microsoftâs cloud revenues seemed to fall slightly short of targets in Q4 results; Iâm guessing now that the reason might have been that computing capacity was simply not available. And AMD doesnât seem capable of breaking into Nvidiaâs territory. Jensen will convince the remaining uncertain investors at GTC in March 17-21, 2025. Perhaps tariff and sales restriction issues create some uncertainty. But a proper bear case for Nvidia cannot be made from this situation now.
10 Likes