Hacker Newsnew | past | comments | ask | show | jobs | submit | willis936's commentslogin

Google is making millions of TPUs per year. Nvidia ships more gaming GPUs, but it's not like multiple orders of magnitude off.

I'm willing to bet TPUs wouldn't be nearly as successful or sophisticated without the decades of GPU design and manufacturing that came before them.

Current manufacturing numbers are a small part of the story of the overall lineage.


Of course you are right, but in addition they wouldn't have even made them if GPUs hadn't made ML on CPU so relatively incapable. Competition drives a lot of these decisions, not just raw performance.

It's pretty interesting that consumer GPUs started to really be a thing in the early 90s and the first Bitcoin GPU miner was around 2011. That's only 20 years. That caused a GPU and asic gold rush. The major breakthroughs around LLMs started to snowball in the academic scene right around that time. It's been a crazy and relatively quick ride in the grand scheme of things. Even this silicone shortage will pass and we'll look back on this time as quaint.

You are missing his point. They very likely don't start building TPUs if there were no GPUs.

I'm not missing the point. If you recall your computer architecture class there are many vector processing architectures out there. Long before there was nvidia the world's largest and most expensive computers were vector processors. It's inaccurate to say "gaming built SIMD".

You are missing the point - it's an economic point. Very little R&D was put into said processors. The scale wasn't there. The software stack wasn't there (because the scale wasn't there).

No one is suggesting gaming chips were the first time someone thought of such an architecture or built a chip with such an architecture. They are suggesting the gaming industry produced the required scale to actually do all the work which lead to that hardware and software being really good, and useful for other purposes. In chip world, scale matters a lot.


The Cray-1, which produced half a billion USD in revenue in today's dollars, at a time when computing was still science fiction, did not demonstrate scale? I just can't take you in good faith because there has never been a time when large scale SIMD computing was not advanced by commercial interests.

In this context scale = enough units/revenue to spread fixed costs.

I'll take your word on lifetime revenue numbers for Cray 1.

So yes, in todays dollars, $500 million of lifetime revenue - maybe 60-70 million per year, todays dollars - is not even close to the scale we are seeing today. Even 10 years ago Nvidia was doing ~$5 billion per year (almost 100x your number) and AMD a few bill(another 60-70x ish)

Even if you meant $500m in annual (instead of lifetime), Nvidia was 10x that in 2015. And AMDs GPU revenue which was a few billion that year, so it's more like 17x.

That's a large difference in scale. At the low end 17x and at the high end 170x. Gaming drove that scale. Gaming drove Nvidia to have enough to spend on CUDA. Gaming drove NVidia to have enough to produce chip designs optimized for other types of workloads. CUDA enabled ML work that wasn't possible before. That drove Google to realize they needed to move away from ML on CPU if they wanted to be competitive.

You don't need any faith, just understand the history and how competition drives behavior.


Make sure to tune into the superbowl tomorrow.

> a Hims & Hers Health (HIMS.N) (Super Bowl) commercial features a voiceover by rapper Common with the tag line "Rich People Live Longer"

At least I know kid rock won’t be there

Thankfully we have options - Kid Rock will be performing an alternative half-time show and it's available on X and Youtube @ 8PM.

Is there a Kalshi bet going about whether he'll perform his song about preferring underage girls?

Isn’t this a common rock star trope? Anthony Kiedis talks about a 14 yo who he slept with on tour.

Anthony Kiedis isn't headlining an event that's being put on by an expressly christian organization. He also is not closely tied to someone who's mentioned more in the Epstein Files than Harry Potter is mentioned in the Harry Potter books.

Kid Rock has some pretty infamous, explicit lyrics I won’t be pasting here. Just look it up, there are dozens of articles about this right now. It’s not rumors or something ambiguous, he is a disgusting person with some pretty awful things to say. Given TP’s christian mission/focus and constant moral panic stance, coupled with the MAGA movement’s alleged concern for minors, “he is not appropriate” is an understatement.

Unfortunately he stays somewhat relevant because he drapes himself in an American flag.


It is if you baseline is Ian Watkins, but we tend to look at these people with disgust for their blatant displays of depravity.

Doesn’t really narrow it down

Yes. That is why I know he won’t be on for the Super Bowl. It’s also why I find the Super Bowl so appealing tomorrow night.

It's a fun thought, but you know what we call those people? Poor. The people who light their own money on fire today are ceding power. The two are the same.

At the end the day a medieval lord was poor but he lived a better life than the peasants.

> At the end the day a medieval lord was poor but he lived a better life than the peasants.

As measured in knowledge utilized during basic living: The lives of lords were much less complex than that of modern poor people.


1. Some people can afford to light a lot of their money on fire and still remain rich.

2. The trick is to burn other people’s money. Which is a lot more akin to what is going on here. Then, at least in the US, if you’re too big to fail, the fed will just give you more cash effectively diminishing everyone else’s buying power.


In regards to 2: it's as simple as not letting it be your money being set on fire. Every fiscally responsible individual is making sure they have low exposure to the mag 7.

Are you referring to the literal planet rather than a woman? That character felt particularly of self insert fantasy (oo a hot 20 year old in love with the aged professor).

Regardless, stopping at the first book is a good recommendation. Asimov demonstrated he didn't understand what made his own work interesting. Granted mystery boxes are hard, but he took an immediate about-face on psychohistory and retconned any bit of intrigue with rather vanilla stuff. The first book is outstanding.


I read books because I enjoy them. I enjoy reading about hot 20 year olds and big breasted women in space actually. Women are allowed to have 50 Shades, I'm allowed to enjoy books too.

You are. And I'm allowed to read into the author's psychology when they wear it on their shoulder. And I'm also allowed to critique the author when they misunderstand their work and write rambling, uninspired sequels that ruin the original work.

Lucas and Disney couldn't help but copy even these bad parts of Foundation in the Star Wars prequels and sequels.


> ruin the original work

I've never understood this. The beauty of recorded media is authors cannot ruin or revoke their work, assuming no actual censorship, of course (copyright can also be a problem). Just ignore the subsequent works if you don't like them. This is the first time I'm hearing about people only reading the first Foundation book but it's definitely worth doing some quick checks before dedicating one's finite time to reading/watching/listening to anything.


The trilogy, prequels, and sequels are all "canon" and retcon the most interesting concepts of the original book as ruses, conspiracies, and lies. This is unambiguously a ruining of the original because the author went out of their way to mute the concepts rather than explore them.

Writers are part of the world they live in.

Take Dune. After the third book it gets weird. But Herbert was a white man in the 1970s- ofcourse the books become degenerate and druggie.


This smells a lot of "any publicity is good publicity", which has been false for more than a decade. You know what makes line go up? Results.

OK.

Tesla Trailing P/E: 390.70 Apple Trailing P/E: 34.11 Ford Trailing P/E : 11.74


Perception and execution that is what i call P/E

I'm not sure what your point is. High P/E is a sign of overvaluation and a bad investment. Just because the market is irrational doesn't mean you should sign up for to be a bag holder.

Pretty sure OPs point is that Tesla has shitty results, has a CEO that lives “any publicity is good publicity” as a mantra and the company valuation is through the roof.

Yes and none of those are good things.

They're good things right now for Elon and Tesla. Whether they will continue to be is up in the air. But, it definitely works as a strategy, at least temporarily.

The effective electricity rate in MA is already $0.37/kWh. How much further could it go?

That’s a joke right?

I see no reason why it couldn't rise to the equivalent of $37/mWs under the looming hyperinflation.

Cries in PG&E

How do you make a small fortune? Start with a big one.

The thing being called obvious here is that the maintenance you have to do on earth is vastly cheaper than the overspeccing you need to do in space (otherwise we would overspec on earth). That's before even considering the harsh radiation environment and the incredible cost to put even a single pound into low earth orbit.


If you think the primary source of electricity is solar (which clearly Musk does), then space increases the amount of compute per solar cell by ~5x, and eliminates the relatively large battery required for 24/7 operation. The thermal radiators and radiation effects are manageable.

The basic idea of putting compute in space to avoid inefficient power beaming goes back to NASA in the 60s, but the problem was always the high cost to orbit. Clearly Musk expects Starship will change that.


My dude, ISS has 200 KW of peak power.

NVIDIA H200 is 0.7 KW per chip.

To have 100K of GPUs you need 500 ISSs.

ISS cooling is 16KW dissipation. So like 16 H200. Now imagine you want to cool 100k instead of 16.

And all this before we talk about radiation, connectivity (good luck with 100gbps rack-to-rack we have on earth), and what have you.

Sometimes I think all this space datacenters talk is just a PR to hush those sad folks that happen to live near the (future) datacenter: “don’t worry, it’s temporary”

https://www.nytimes.com/2025/10/20/technology/ai-data-center...


The ISS is in the middle of rolling out upgrades to their panels so it’s not a great comparison. https://en.wikipedia.org/wiki/Roll_Out_Solar_Array

> ROSA is 20 percent lighter (with a mass of 325 kg (717 lb))[3] and one-fourth the volume of rigid panel arrays with the same performance.

And that’s not the current cutting edge in solar panels either. A company can take more risks with technology choices and iterate faster (get current state-of-the-art solar to be usable in space).

The bet they’re making is on their own engineering progress, like they did with rockets, not on sticking together pieces used on the ISS today.


Now tell me how you heat dissipate all this. Not that there is a lot of air or water in space.

Not that you would want 500+ square meters just for cooling of 200KW

And, mind you, it won’t be a simple copper radiator

https://www.nasa.gov/wp-content/uploads/2021/02/473486main_i...


Don't worry, a musk stan with a physics degree will be around shortly to inform you that 5km^2 of radiators is completely reasonable

How much maintenance do you need? Lets say you have hardware whose useful lifespan due to obsolescence is 5 years, and in 4, the satellite will crash into the atmosphere anyways.

Let's say given component failure rates, you can expect for 20% of the GPUs to fail in that time. I'd say that's acceptable.


> How much maintenance do you need?

A lot. As someone that has been responsible for trainings with up to 10K GPUs, things fail all the time. By all the time I don't mean every few weeks, I mean daily. From disk failings, to GPU overheating, to infiniband optical connectors not being correctly fastened and disconnecting randomly, we have to send people to manually fix/debug things in the datacenter all the time.

If one GPU fails, you essentially lose the entire node (so 8 GPUs), so if your strategy is to just turn off whatever fails forever and not deal with it, it's gonna get very expensive very fast.

And thats in an environment where temperature is very well controlled and where you don't have to put your entire cluster through 4 Gs and insane vibrations during take off.


Radiation is a bitch. Especially at those nanometers and memory bandwidth.

And cooling. There is no cold water or air in space.


Note how Musk cleverly doesn't claim that not doing maintenance drives down costs.

Nothing in there is a lie, but any substance is at best implied. Yes, 1,000,000 tons/year * 100kW/ton is 100GW. Yes, there would be no maintenance and negligible operational cost. Yes, there is some path to launching 1TW/year (whether that path is realistic isn't mentioned, neither what a realistic timeline would be). And then without providing any rationale Elon states his estimate that the cheapest way to do AI compute will be in space in a couple years. Elon is famously bad at estimating, so we can also assume that this is his honest belief. That makes a chain of obviously true statements (or close to true, in the case of operating costs), but none of them actually tell us that this will be cheap or economically attractive. And all of them are complete non-sequiturs.


If you ramp up the economies of scale to make those things - radiation protection and cost per pound - the calculus changes. It's supposed to synergize with Starship, and immediately take advantage of the reduced cost per pound.

If the cost per pound, power, regulatory burden, networking, and radiation shielding can be gamed out, as well as the thousand other technically difficult and probably expensive problems that can crop up, they have to sum to less than the effective cost of running that same datacenter here on earth. It's interesting that it doesn't play into Jevon's paradox the way it might otherwise - there's a reduction in power consumption planetside, if compute gets moved to space, but no equivalent expansion since the resource isn't transferable.

I think some sort of space junk recycling would be necessary, especially at the terawatt scale being proposed - at some point vaporizing a bunch of arbitrary high temperature chemistry in the upper atmosphere isn't likely to be conducive to human well-being. Copper and aluminum and gold and so on are also probably worth recovering over allowing to be vaporized. With that much infrastructure in space, you start looking at recycling, manufacturing, collection in order to do cost reductions, so maybe part of the intent is to push into off-planet manufacturing and resource logistics?

The whole thing's fascinating - if it works, that's a lot of compute. If it doesn't work, that's a lot of very expensive compute and shooting stars.


some people really gotta stop huffing VC fumes

Or, just saying, be critical of ideas and think them through, and take in what experts say about it, and determine for yourself what's up. If a bunch of people who usually seem to know what they're talking about think there's a legitimate shot at something you, as a fellow armchair analyst, think is completely impractical, it makes sense to go and see if maybe they know something you don't.

In this case, it's all about Starship ramping up to such a scale that the cost per pound to orbit drops sufficiently for everything else to make sense - from the people who think the numbers can work, that means somewhere between $20 and $80 per pound, currently at $1300-1400 per pound with Falcon 9. Starship at scale would have to enable at least 2 full orders of magnitude decrease in price to make space compute viable.

If Starship realistically gets into the $90/lb or lower range, space compute makes sense; things like shielding and the rest become pragmatic engineering problems that can be solved. If the cost goes above $100 or so, it doesn't matter how the rest of the considerations play out, you're launching at a loss. That still might warrant government, military, and research applications for space based datacenters, especially in developing the practical engineering, but Starship needs to work, and there needs to be a ton of them for the datacenter-in-space idea to work out.


Or, just saying, we should eat babies because they are abundant and full of healthy nutrition for adult humans. [1]

Just because an idea has some factors in its favor (Space-based datacenter: 100% uptime solar, no permitting problems [2]) doesn't mean it isn't ridiculous on its face. We're in an AI bubble, with silly money flowing like crazy and looking for something, anything to invest it. That, and circular investments to keep the bubble going. Unfortunately this gives validation to stupid ideas, it's one of the hallmarks of bubbles. We've seen this before.

The only things that space-based anything have advantages on are long-distance communication and observation, neither of which datacenters benefit from.

The simple fact is that anything that can be done in a space-based datacenter can be done cheaper on Earth.

[1] https://en.wikipedia.org/wiki/A_Modest_Proposal for the obtuse

[2] until people start having qualms about the atmospheric impact of all those extra launches and orbital debris


The idea here is that the economics of launch are changing with Starship such that the "incredible cost" and "overspeccing" of space will become much less relevant. There's a world where, because the cost per kg is so low, a data center satellite's compute payload is just the same hardware you'd put in a terrestrial rack, and the satellite bus itself is mass-produced to not-particularly-challenging specs. And they don't have to last 30 years, just 4-ish, when the computer is ready for retirement anyway.

Will that come to be? I'm skeptical, especially within the next several years. Starship would have to perform perfectly, and a lot of other assumptions hold, to make this make sense. But that's the idea.


My point is even if it were free to put things in space and radiation did not need mitigation, you're still paying a lot to have hardware that can't be maintained. If it were cheaper we wouldn't be doing online maintenance on Earth. Name a single datacenter on the rocky surface of the Earth that is opting to not have maintenance.

Elon might have a scoop on getting things to orbit cheaper than everyone else.

By avoiding untrustworthy clients. All Windows devices should be considered compromised after last year.

That's not mitigating client compromise, that's a whole other thing - trying to construct an uncompromiseable client.

You don't build defense-in-depth by assuming something can't be compromised.


Clients can always be compromised. I'm not talking about a client that can't be compromised, but simply a client that is not compromised out-of-the-box.

That seems orthogonal to the subject of this discussion, i.e. "Compromise of the client side application or OS shouldn't break the security model."

Windows has been sending usage history back to their servers for longer than just last year

Why last year?

Windows recall, intrusive addition of AI features (is there even a pinky promise that they're not training on user data?), more builtin ads, and less user control (most notably the removal of using the OS without an account - something that makes sense in the context of undisclosed theft of private information).

This was 2025. I'm excited for what 2026 will bring. Things are moving fast indeed.


I wonder how useful an encrypted bitstream is for training AI. Always use a VPN. All ISPs are anti-consumer bastards.

Most LLMs I've played with struggle with modest, well-known ciphers.


By what metrics are they 99% as good? There are a lot of benchmarks out there. Please share them.

I think the answer lies in the "we actually care a lot about that 1% (which is actually a lot more than 1%)".


Open models have been about 6 to 9 months behind frontier models, and this has been the case since 2024. That is a very long time for this technology at it's current rate of development. If fast takeoff theory is right, this should widen (although with Kimi K2.5 it might have actually shortened).

If we consider what typically happens with other technologies, we would expect open models to match others on general intelligence benchmarks in time. Sort of like how every brand of battery-powered drill you find at the store is very similar, despite being head and shoulders better than the best drill from 25 years ago.


> That is a very long time for this technology at it's current rate of development.

Yes, as long as that gap stays consistent, there is no problem with building on ~9 months old tech from a business perspective. Heck, many companies are lagging behind tech advancements by decades and are doing fine.


> Sort of like how every brand of battery-powered drill you find at the store is very similar, despite being head and shoulders better than the best drill from 25 years ago.

They all get made in China, mostly all in the same facilities. Designs tend to converge under such conditions. Especially since design is not open loop - you talk to the supplier that will make your drill and the supplier might communicate how they already make drills for others.


I'm still testing myself and cannot make a confident statement yet, but Artifical Analysis is a solid and independent, though also to be fair somewhat imperfect source for a general overview: https://artificialanalysis.ai/

Purely according to Artificial Analysis, Kimi K2.5 is rather competitive in regard to pure output quality, agentic evals are also close to or beating US made frontier models and, lest we forget, the model is far more affordable than said competitors, to a point where it is frankly silly that we are actually comparing them.

For what it's worth, of the models I have been able to test as of yet, when focusing purely on raw performance (meaning solely task adherence, output quality and agentic capabilities; so discounting price, speed, hosting flexibility), I have personally found the prior Kimi K2 Thinking model to be overall more usable and reliable than Gemini 3 Pro and Flash. Purely on output quality in very specific coding tasks, Opus 4.5 was in my testing leaps and bounds superior of both the Gemini models and K2 Thinking however, though task adherence was surprisingly less reliable than Haiku 4.5 or K2 Thinking.

Being many times more expensive and in some cases less reliably adhering to tasks, I really cannot say that Opus 4.5 is superior or Kimi K2 Thinking is inferior here. The latter is certainly better in my specific usage than any Gemini model and again, I haven't yet gone through this with K2.5. I try not to just presume from the outset that K2.5 is better than K2 Thinking, though even if K2.5 remains at the same level of quality and reliability, just with multi modal input, that'd make the model very competitive.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: