| It's certainly an indicator. But, just like airplane orders, the backlog is long enough that you have to put orders in on a speculative basis. If everything goes as planed, we'll need new equipment in a couple years, so order now. If things don't work out as planned, you might be able to sell your slot, or push it back a few months, or cancel it. In a frothy market, you don't really know from order backlogs how many orders are based on firm plans and how many are we might want to build chips, so let's order a machine and see if we want it when it comes up. Or the middle ground of we definitely need these machines in the next ten year, and if sales are still strong, sooner is better. |
| |
| ▲ | throw0101c a day ago | parent | next [-] | | > Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much. In the late-1990s companies were building out fibre optic networks, as they believed they would get a lot of ROI from all the Information Superhighway demand. The expected demand never came in the expected timeline. Of course demand eventually used up all the dark fibre over the course of many moons, just not quickly enough to save many companies: * https://en.wikipedia.org/wiki/Telecoms_crash | | |
| ▲ | crote a day ago | parent | next [-] | | More specifically, the demand for fiber didn't come - but the demand for bandwidth did! Those companies were building out fiber like it was a legacy telephone network: needing a massive amount of individual wire pairs to carry a massive amount of parallel transmissions. In reality the bandwidth per channel was rapidly increasing due to better tx/rx equipment, and with the rise of WDM it was suddenly possible to carry not one channel per fiber pair but several dozen. When you were laying fiber on the assumption that a single long-distance pair was going to carry 1Gbps - and therefore laying hundreds of fiber pairs in a trench at once - it is quite a big upheaval when that single pair can now carry 100 channels of 10Gbps each. Your brand-new expensive fiber trunk can suddenly be replaced by a single spare pair some utility company had lying around already. Turns out laying fiber isn't a great business when the capacity of existing fiber is doubling every year or so! | |
| ▲ | user43928 a day ago | parent | prev [-] | | The demand for AI exists today. As in OpenAI is at capacity, 6.1 Sol is served at 30 tokens per second and with frequent capacity errors. Anthropic seems better now but were also working hard on expanding capacity. I have no crystal ball to see how future demand and capacity will develop, but as of today these data center investments appear to be fully utilized. Some people anyway seem to think a collapse is inevitable. | | |
| ▲ | jack_pp a day ago | parent | next [-] | | maybe collapse in valuation, collapse in demand? I don't see that happening any time soon, it's too useful, idk how anyone can contest it is useful and not just hype. | |
| ▲ | cute_boi a day ago | parent | prev [-] | | it is all because of subsidization. If OAI or Ant starts to use real pricing there will be lot of capacity. | | |
| ▲ | user43928 a day ago | parent | next [-] | | No, in fact inference is far cheaper than most think, see analysis for serving Kimi K3. If you mean OpenAI and Anthropic's investments into R&D, ok, perhaps. However, they will not be able to increase prices unless what they are offering next is a significant enough improvement over the current SOTA, which might only be six months ahead of the Chinese models. | | |
| ▲ | intorio a day ago | parent | next [-] | | Not only is inference currently much cheaper than people seem to think, it is going to get much cheaper. NVIDIA’s Vera Rubin is going to drive costs down for current models considerably. | |
| ▲ | zamalek a day ago | parent | prev [-] | | It's amazing to see the shift from "Anthropic is profitable" to "it's just R&D costs" following the leak of Anthropic's finances. If there's one thing more abundant in this bubble than FOMO, it's special pleading. These companies are burning helps of money, and they need to keep burning heaps of money in order to stay competitive. People have (correctly) learned to be extremely fickle consumers, they can't depend on making profit by running inference on a 1 year old model. |
| |
| ▲ | karahime a day ago | parent | prev [-] | | How do you know that it's subsidized? | | |
| ▲ | gruturo a day ago | parent | next [-] | | I really dislike that term. It's stupid and supports a distorted narrative. They arbitrarily decide 1M API tokens cost X, and any subscription where usage costs less than that is "subsidized". No it isn't. They may be profiting on subscriptions as well, or losing on full API costs as well, or anything in between, and that X is anyway more reflective of competition than their costs. For all we know they did a rand()%1000 to determine it in the first place. | |
| ▲ | crote a day ago | parent | prev [-] | | If it isn't subsidized, why aren't all those AI companies making huge amounts of profit? Instead they are rapidly burning through the billions, and the money must be going somewhere. | | |
| ▲ | karahime a day ago | parent [-] | | Investing in R&D isn't the same thing as subsidizing something. | | |
| ▲ | Danox a day ago | parent [-] | | Investing in R&D is fine. as long as you are turning a profit somewhere within your company that can support it. If you are not, eventually the end comes. In short, the bubble. | | |
| ▲ | karahime a day ago | parent [-] | | The problem with this view is that that isn't a "bubble". "Eventually" can support quite a while, especially when the investment is increasing revenue over time and the unit economics look good. The timeline point cuts both ways, and none of that is a subsidy. |
|
|
|
|
|
|
| |
| ▲ | HarHarVeryFunny a day ago | parent | prev | next [-] | | The only company making AI chips, and allowed to buy from ASML, is TSMC, otherwise it's probably memory manufacturers. Even if the order book is mostly TSMC, it doesn't necessarily mean the demand is mostly driven by AI (although maybe it is) - there are plenty of other high volume TSMC customers like Apple, AMD, Intel, Qualcomm, Broadcom ... Finally, just because lots of TSMC customers may have made forward commitments (or whatever it takes for TSMC to commit to expansion), the main conclusion would seem to be that they needed to do that for business continuity given current capacity constraints. It's not clear why the need for them to do this should be taken as a sign that the AI datacenter bubble won't pop. Do you want to be Apple and gamble that rather then locking in TSMC capacity now you can gamble on the bubble popping and maybe get better terms later? As a counterpoint to this "no need to worry" story, CoreWeave's 5 yr bonds are now yielding 15% ... | |
| ▲ | jabron a day ago | parent | prev | next [-] | | And incidentally, believing that you can keep making a good ROI even if it's not true is what fuels a bubble. | | | |
| ▲ | palmotea a day ago | parent | prev | next [-] | | > Companies buying those machines need years to get their money back. If they didn't believe they would make a good ROI they wouldn't buy as much. But if they're in FOMO mode, they might not be thinking about ROI at all. Or if they're in bubble-optimist mode, they would just assume their ROI will be good. | | | |
| ▲ | GuB-42 a day ago | parent | prev [-] | | ASML sell photolithography machines, which presumably can be used for many kinds of chips: CPUs, GPUs, flash, RAM, etc... They don't need to be AI chips, so even if the AI bubble bursts before they managed to get a ROI, they can switch to other applications that need high end electronics, such as PCs and smartphones, maybe cars. |
|
