Cappuccino
Newcomer
Right, but how can we actually say any of this without a real success story? We have no idea if it’s easy or hard as it’s only been considered on paper.
Cappuccino
Newcomer
The same reactors that we have been essentially unable to build in volume since the 60s?
The problem here is building a nuclear reactor is very difficult and we’re pretending it’s not lol. If it were easy we wouldn’t really be bothering with these solar panels.
You are shifting your position. You were first claiming they are new and unproven hence they would need a "first" success story. Now they are the same reactors as in the 60s (which i agree to , btw) ? So which is it?
Anyway, the fact that we were able to build reactors in the 60s kinda proves it's possible with 60s tech. With today's tech (and some incremental improvements in the reactor design), you're saying it's harder? How?
There can be other reasons why something that is technically possible and useful is not "popular". But you seem to asume that because it's not popular, it must be difficult, which is a fallacy.
Cappuccino
Newcomer
There is no shifting. It’s both completely unproven and if it’s similar at all to the processes needed for traditional reactors then good luck because America basically forgot how to build those without bankrupting the company building it.
If fission reactors were so easy a company could just plop one on their property to power a load, why haven’t we been able to build these in the past 100 years in any significant number, despite rising power requirements?
Cappuccino
Newcomer
I’ve made multiple concrete points you just aren’t addressing them lol. Nuclear reactors are some of the most difficult projects you can attempt in America, we have more private sector rocket launches to space than we do commercial nuclear reactor projects now. So asserting this is just an easy thing people can do for datacenters is naive, it’s not easy and it’s literally never been done before so as far as we know it might not even be feasible, financially or otherwise.
You asked how it would be harder. I don’t know exactly, this isn’t my field of expertise and you’d have to ask someone who works in reactor construction. The answer is probably regulations and cost, almost every company that attempts a modern nuclear reactor construction in the US goes bankrupt from the project despite govt subsidies. The general track record shows that it was easier to do this back in the day because construction was cheaper and had less regulatory roadblocks.
I think it’s plenty on topic if we’re discussing datacenters on this scale that basically require on prem generation.
This power discussion might warrant its own thread, but I think it’s important to point out that the technical aspects of nuclear reactors are not the problem per se. There are two major economic issues though.
First, building and operating nuclear plants has become extremely expensive due to a combination of factors, including new safety regs and post-9/11 security requirements. New licensees must account for security in their structural designs, and personnel requirements are more robust. The second related issue is that even existing nuclear power has become much less competitive in the US as domestic fossil fuel production has increased. I live in an area of the US where a substantial portion of power is nuclear, but without subsidies it would be completely uncompetitive with natural gas.
None of this means we can’t do nuclear power anymore, just that the market hasn’t been good for new plants. I think if people are talking about using gigawatts of power on AI, we can agree that supplying it with fossil fuels is a preposterous use of the planet’s carbon budget. I can’t predict what will happen, but looking at these projected numbers the new power will have to come from somewhere.
First, building and operating nuclear plants has become extremely expensive due to a combination of factors, including new safety regs and post-9/11 security requirements. New licensees must account for security in their structural designs, and personnel requirements are more robust. The second related issue is that even existing nuclear power has become much less competitive in the US as domestic fossil fuel production has increased. I live in an area of the US where a substantial portion of power is nuclear, but without subsidies it would be completely uncompetitive with natural gas.
None of this means we can’t do nuclear power anymore, just that the market hasn’t been good for new plants. I think if people are talking about using gigawatts of power on AI, we can agree that supplying it with fossil fuels is a preposterous use of the planet’s carbon budget. I can’t predict what will happen, but looking at these projected numbers the new power will have to come from somewhere.
Last edited:
DavidGraham
Veteran
After months of being either the second or the third largest company on earth by market cap, NVIDIA is now back to being number 1.
Last edited by a moderator:
Really can’t comment on the reliability of this rumour, but nvidia entering the consumer cpu market when Qualcomm’s exclusivity period in windows runs out would shake things up. I honestly don’t know the details of Qualcomm’s arrangement with Microsoft so I’m just going by the article.
This isn't new, was discussed a few pages back. Quoting some of the posts for reference.
NVIDIA x MediaTek "AI PC" SoC To Debut By H2 2025, Will Witness Integration By Lenovo, Dell, HP & ASUS
NVIDIA & MediaTek are set to unveil their highly-anticipated AI PC chip by the second half of 2025 and will be adopted by major OEM players.wccftech.com
Taiwanese financial newspaper has the AI chip taping out third quarter(2024) and mass production 1st quarter next year.
聯發科傳攜手輝達進擊 AI PC 開發安謀架構處理器 | 產業熱點 | 產業 | 經濟日報
AI PC市場成長性看俏，聯發科加足馬力搶進，傳出將攜手AI晶片一哥輝達（NVIDIA）開發安謀（Arm）架構的AI P...money.udn.com
Sounds more logical but at this point just rumors.
DavidGraham
Veteran
Meta announced a 100K H100 cluster a while back, but now they are claiming something else.
"We're training the Llama 4 models on a cluster that is bigger than 100,000 $NVDA H100s or bigger than anything that I've seen reported for what others are doing"
On other news, the orders of B200 are reportedly as follows:
Google has reportedly ordered 400,000 GB200s worth $10B
Microsoft has reportedly ordered 60,000 GB200s worth $2B
Meta has reportedly ordered 360,000 GB200s worth $8B
"We're training the Llama 4 models on a cluster that is bigger than 100,000 $NVDA H100s or bigger than anything that I've seen reported for what others are doing"
On other news, the orders of B200 are reportedly as follows:
Google has reportedly ordered 400,000 GB200s worth $10B
Microsoft has reportedly ordered 60,000 GB200s worth $2B
Meta has reportedly ordered 360,000 GB200s worth $8B
Last edited:
Cappuccino
Newcomer
I remember talking to some buds of mine about Windows on ARM way back in 2020, during the Surface X era, about how WoA’s biggest issue is Qualcomm. I would love an Nvidia ARM device.
DavidGraham
Veteran
DavidGraham
Veteran
Saudi Arabia is planning a 100 billion $ investment for an AI project with Google to build multiple AI data centers.
