[HN Gopher] OpenAI's hunger for computing power
___________________________________________________________________
OpenAI's hunger for computing power
Author : doener
Score : 99 points
Date : 2025-10-04 22:14 UTC (1 days ago)
(HTM) web link (www.wsj.com)
(TXT) w3m dump (www.wsj.com)
| neonate wrote:
| https://archive.md/OnsLK
| vineyardmike wrote:
| What is their angle with this?
|
| Surely SamA doesn't actually think that they'll more than 20x
| their compute in a few years? I'm sure the researchers there
| would love to do more research, with more compute, faster, but
| 20+x growth is not a practical expectation.
|
| Is the goal here to create a mad rush to build data centers,
| which should decrease their costs with more supply? Do they just
| want governments to step in and to help somehow? Is it part of
| marketing/hype? Is this trying to project confidence to investors
| on future revenue expectations?
| refulgentis wrote:
| > Surely SamA doesn't actually think that they'll more than 20x
| their compute in a few years?
|
| He does, or at least, he believes if its plausible they should
| attempt to.
|
| We live in odd times. It sort of reminds me of Feb 2020. All
| you really needed to know was the Rt and rest was just math.
| Who knows if it'll matter or pencil out in a decade, but, it's
| completely reasonable at these growth rates and with the iron
| laws known to keep scaling.
| p1esk wrote:
| _Surely SamA doesn't actually think that they'll more than 20x
| their compute in a few years?_
|
| If their goal is to train say, a 100T model on the whole
| youtube dataset they will need 20000x more compute. And that
| would be my goal if I were him.
| bix6 wrote:
| Why 20000x more compute? I thought they were at approx 1T
| with current compute?
|
| Edit: looked it up. 10k+ times more for training compute.
| Sheesh. Get the Dyson sphere ready lol.
| esafak wrote:
| https://www.youtube.com/watch?v=Dy6Dw9rOAFQ
| indolering wrote:
| Now I'm more on the side of him being delusional.
| zeroq wrote:
| "At this point I think I know more about manufacturing
| than anyone currently alive on Earth."
|
| It's that dumbass at your work who thinks that solely
| because he landed a job that pays him more than their
| parents ever made combined in his early 20s he can school
| everyone on every topic imaginable, from nutrition to
| religion.
|
| Him and Elon makes way more than that dumbass so ego get
| inflated even more.
|
| I don't especially like Tucker Carlson, but I think the
| more screen time we'll give to this people with an open
| mic it's better for everyone to have a first hand
| experience of how detached from reality these people are.
| johnnienaked wrote:
| Absolutely right, and it's ubiquitous across
| organizations too.
|
| I've never met an executive I respect. They're all
| absolute experts at _appearing_ competent.
| aswegs8 wrote:
| I mean, they're selected for it so that's not surprising
| johnnienaked wrote:
| I guess the surprising part is that appearing competent
| is more important to shareholders than being competent
| estimator7292 wrote:
| Actually checking if someone _is_ competent requires
| actual work, though. Work is for _lesser_ people.
| Shareholders just _know_ if a person is or is not
| competent, that 's why they have so much money, right?
| blitzar wrote:
| It is better to remain silent and be thought a fool than
| to open your mouth and remove all doubt
| IT4MD wrote:
| Thanks for the example.
| b00ty4breakfast wrote:
| I can never tell if these guys have come to genuinely
| love the smell of their own farts or if they're just
| constantly in _sales mode_. Like maybe all those hours in
| meetings with investors and shareholder or whatever has
| gotten them stuck, like your mom used to warn you about
| when you 'd make faces at your little brother.
| llbbdd wrote:
| When your job is to constantly be making the pitch for
| your company, and you live in a world where every
| conversation you have can be news before the end of the
| day, the mask can never come off.
| dmbche wrote:
| If they know it won't bring in revenue, they can't get
| out of "sales mode" because when the runway stops they
| get left out. Like musical chairs with one chair left,
| you want to keep the game going if you don't think you
| can get it. And you're filthy rich as long as the game's
| going.
| kadushka wrote:
| Mainly because global video data corpus is > 100k larger
| than global text corpus, so you will need to train much
| larger models for much longer (than current LLMs).
| Drblessing wrote:
| That would be awesome.
| zingerlio wrote:
| The AI bubble bursts when he stumbles to get that money.
| indolering wrote:
| In the early days of Bitcoin, we would argue security models
| and laugh about Bitcoin mining taking some significant
| percentage of global power supply. It's been giving around 1%
| for a while now despite the supply falling off.
|
| I wouldn't put bets on what the outer limits for AI are going
| to be. However, it's a huge productivity boost across a huge
| range of workflows. Models are still making large gains as they
| become more sophisticated.
|
| If I had Sam Altman's access to other people's capital, I would
| be making large bets that it will keep growing.
| skywhopper wrote:
| He needs 11 figures of cash injected as soon as possible. The
| people who can give it want a big return. Given the current
| losses, the only way to make the math right is to lie
| outrageously about what's possible.
| karmakurtisaani wrote:
| He's been kicking this can for years now. Looking forward to
| the day he's forced to stop.
| pram wrote:
| Perceptually, it helps to take scrutiny off the current spend?
| It isn't a bubble if you can just scoff at $100 billion and say
| like" "thats pocket change, this will actually require 10
| quadrillion dollars!!"
| ants_everywhere wrote:
| If they want to survive they need to outrun Google and have a
| differentiated service. Which as of now it's not clear that
| OpenAI will have a reason to exist in the near future with
| Anthropic and Google.
|
| They're likely betting on either training a model so big they
| can't be ignored or possibly focusing more B2B which means lots
| of compute to resell.
| CuriouslyC wrote:
| If their plan was to go toe to toe with Google as a
| foundation model/inference provider they would 100% be
| getting ground to dust, that's not a winnable fight. There's
| a reason they've pivoted to product and retained Jony Ive.
|
| Anthropic gets a TON of hate on social media, their models
| are fragile, their infra is poorly managed, they're 100%
| going to end up in Jeff's pocket. OpenAI is a survivor.
| arthurofbabylon wrote:
| The phrase you are looking for is "commodifying the periphery."
| As adjacent bottlenecks open up, the bottleneck you control
| becomes more valuable.
| tim333 wrote:
| My guess:
|
| Altman figures AI will be a big deal and constrained by
| avaiable compute.
|
| If he locks in all the available compute and related finance
| before the competition then he's locked in as the #1 AI
| company.
|
| I'm not sure 20x or 5x or 40x matters, nor revenue
| expectations, so much as being ahead of the competition.
| jgalt212 wrote:
| > What is their angle with this?
|
| My pet theory: Sam makes more money when OpenAI spends than
| when OpenAI earns.
| Szpadel wrote:
| I believe it is because of RL you are no longer limited by
| training data as you generate it during learning on the fly so
| benchmark driven learning could scale with compute
|
| they also seem to assume that everyone will use AI from them in
| the future, especially with new "pulse" combined with ads.
| scaling this will need much more compute.
|
| is this reasonable? I'm not convinced, but this is how I
| believe it's their reasoning
| Tubelord wrote:
| Pascal's wager applied to tech cycles. The fervent adherence to
| the hype is akin to religious zealots in many ways
| lemonlearnings wrote:
| https://archive.is/OnsLK
| lemonlearnings wrote:
| Too big to fail is the goal. If the world is powered by openai
| but it aint making a profit in 2028 they can just put their "were
| a utility like water" facemask on and get bailed out.
| kortilla wrote:
| That's a fun trope but it's a terrible outcome for
| shareholders.
| jacquesm wrote:
| Good.
| avs733 wrote:
| Which means it will be made into a terrible outcome for
| everyone.
| throwaway290 wrote:
| shareholders like a business that can never fail...
| lemonlearnings wrote:
| Less terrible than being allow to go bust though.
| stevenwoo wrote:
| At least in the USA, I think if consumers realized their power
| bills going up every year are tied to these new data centers,
| there would be more opposition to data centers going up
| politically. https://apnews.com/article/electricity-prices-
| data-centers-a... I don't know if the electricity markets work
| differently in other countries.
| MountDoom wrote:
| Taxpayers subsidize data centers in many other ways. These
| are prestige projects for politicians, so they often get
| long-term tax breaks and other preferential treatment.
|
| I think it's part vanity, part a misunderstanding about the
| economic benefits of a datacenter (which are nearly nil, as
| they employ very few people and produce nothing for the local
| market), and part just a desire to score brownie points with
| wealthy corporations, which might mean donations, campaign
| support, or other perks for the politician who makes it
| happen.
| noosphr wrote:
| The power bill going up is because the US, and the West in
| general, bet on renewables and a low energy future.
|
| Neither of those things turn out to be a good fit for the new
| economy. The only thing left for people who derided nuclear
| for the last 40 years is to hope this is a bubble that sends
| us back to the 17th century when it pops. Anything else means
| we have to invest trillions in nuclear right now.
| Drblessing wrote:
| The amount of money being invested in AI should've been
| invested into nuclear, both fusion and fission. The AI
| bubble will burst, but the energy bubble never bursts.
| DavidPiper wrote:
| Genuine question from a non-American: What is "the new
| economy"?
| noosphr wrote:
| Malthusianism for computers.
|
| Moore's law is dead. The only way to increase compute is
| to increase the power we feed to computers. AI is just
| the shiniest example. Everything else will follow suit
| until electricity costs increase enough that it doesn't
| make sense to throw any more computation at it.
|
| Any country that doesn't have energy to spare will be in
| the position of countries which didn't have food to
| support armies before the industrial revolution.
| fooker wrote:
| Interesting point. I can see this could turn out to be
| true.
|
| If we needed, for example, 1000 TWH to power AI for a
| huge drone swarm but could not do it while China could,
| this would be problematic.
|
| It requires a future where MAD with nuclear weapons is
| obsolete though, with some futuristic new missile defense
| tech. I don't see that happening until some currently
| unknown physics makes it possible.
| jimbob45 wrote:
| What's your beef with solar? Both parties seem to like it
| just fine, despite it not covering as much of the total
| demand as anyone would like.
| noosphr wrote:
| Both parties like it better because it turns the
| electricity market into another casino that lets you take
| billions from the parts of the economy that do things.
|
| I worked as a quant in the electric market. There wasn't
| a single dataset I saw where more renewables resulted in
| lower total costs for consumers.
| omcnoe wrote:
| The US needs sufficient energy surplus to power industry. US
| energy production has been essentially flat for the past 25
| years and the country has forgot how to bring new capacity
| online. Chinese energy production is up over 6x over the same
| period. China has more clean energy generation capacity today
| than their entire capacity a little over a decade ago.
|
| Instead of panicking about data center electricity usage we
| need to be worrying about getting back to a state where we
| regularly bring new (clean) generation capacity online.
| fooker wrote:
| It's correlated to be data centers, not tied to. That's an
| important difference.
|
| We could easily build ..say.. 10 nuclear reactors and halve
| the utility bills with amortization.
| chermi wrote:
| That's not the main problem. That's the convenient scapegoat
| so we don't get mad about the real problem. Power bills have
| been going up for years. We're just not good at generating
| and serving sufficient energy. Our grid sucks, our utilities
| suck and can do whatever they want, and we can't build
| anything. And the grid problems get worse as we add
| renewables as they have to manage more complex generation
| profiles. (I'm all for renewables, love solar.)
| nakamoto_damacy wrote:
| [flagged]
| nebula8804 wrote:
| Its cheap because you aren't paying the full cost, you are
| externalizing some of the costs onto others.
|
| [1]:https://www.youtube.com/shorts/jNFemZpMadU
| Drblessing wrote:
| The AI-powered tiktok competitor is not going to be cheap on
| compute
| banandys wrote:
| I mean yeah we all saw the video of him stealing gpus and getting
| arrested
| bgwalter wrote:
| 40% of global DRAM output:
|
| https://www.tomshardware.com/pc-components/dram/openais-star...
|
| All for creating worthless TikTok videos with Sora 2, while we
| don't get graphics cards and DRAM and our electricity prices
| rise.
|
| Trump will get another "win" for "his" Stargate project. The
| meeting with South Korean President Lee Jae Myung was NOT
| arranged by Altman, he is the messenger boy:
|
| https://www.reuters.com/business/media-telecom/samsung-sk-hy...
| johnnienaked wrote:
| And our water runs out, and we pollute and destroy the planet
| past the point of no return.
|
| AI will fix it though?
| logtempo wrote:
| Hey, at least you'll be able to add that exterminated tigre
| specie in your postcard from your last adventure trip. And
| more water to that river, with some greener trees etc.
|
| All of that without leaving your home ofc.
| panta wrote:
| Of course. According to Andreessen if you are not optimistic
| and worry about the environment are an "enemy" for the bright
| future ahead (while at the same time he puts Nick Land in the
| list of the "Saints"). These people are deranged psychopaths,
| why are we leaving them at the wheel?
| johnnienaked wrote:
| Sounds about right coming from the guy who plagiarized
| university research to make his billions
| Incipient wrote:
| Is it just me, or does this extreme demand for compute imply
| they've realised the core tech is mostly stagnant, and needs
| compute to possibly scale towards anything AGI-like? (however
| unlikely that is).
| general1465 wrote:
| I see it the same way. It is like automotive manufacturer who
| is using bigger engines with more and more pistons and
| wondering how much bigger engine next model iteration will need
| to make it go faster and how many iterations until they will
| finally break the sound barrier. However their product is
| looking like a school-bus box on wheels which is going to rip
| itself apart long before even reaching the sound barrier.
| adventured wrote:
| They're substantially tied down by demand/usage right now.
|
| How much more compute would they need to allow all of their
| paying users unlimited access to their best model? And to
| enable that setup in such a way that it's actually very fast.
|
| The answer: they need resources far beyond what they have now.
| That's just to solve an existing problem.
|
| Then throw in Sora 4 and whatever else will exist in a few
| years, and the need to feed that monster. They couldn't come
| close to allowing Sora 2 unlimited for all of their paying
| customers - I'd hate to see what that would require.
|
| Then let the AI begin world building for every user (which is
| where this is going). It'll be decades before the resource
| demands actually flatten globally (at least 20-30 years, to get
| to global population saturation on usage; assuming the global
| population will begin to rapidly slow and then shrink).
|
| Hint: the solution to Fermi's Paradox is that we go into the
| box and we never come back out, because it's a lot more
| interesting for 99.9%+ of humanity than a bunch of repeating
| rocks in space that take a zillion years to reach. The core
| purpose of AI will be to world build for us, mentally
| (relaxation, stimulation, entertainment, social) in we go: the
| end. The same thing happens to any advanced beings that get to
| our level, there's little to nothing out there that we can
| reach of interest (and no, it doesn't matter if the HN crowd
| disagrees, what matters for this outcome are the masses), we'll
| definitely figure that out, and there's infinite
| stimulation/experience in the machine world by contrast.
| buyucu wrote:
| Why does OpenAI need so much more compute than everybody else?
| DeepSeek, Qwen and many others build competitive models that need
| much less capital.
| credit_guy wrote:
| Most likely OpenAI has models at least as efficient as DeepSeek
| or Qwen. Cerebras offers both GPT-OSS-120B and
| Qwen3-235B-Instruct. Obviously, the second has twice as many
| parameters as the first, but that's the closest comparison I
| can find. The Qwen model is twice as large, but twice as slow
| (1400 tokens/second vs 3000) and 50% more expensive ($1.2 per
| million tokens vs $0.75). Now, OpenAI is running a proprietary
| model, and most likely it is much optimized than the free
| version they release for public use.
|
| [1] https://inference-docs.cerebras.ai/models/overview
| buyucu wrote:
| Inference is not the main cost driver, training and research
| is.
| credit_guy wrote:
| I'm not sure that's still the case. It used to be the case,
| but I doubt it continues to be. OpenAI had $6.7 BN costs
| for the first half of 2025. I doubt they spent $3 BN in
| training and research. They have 700 million weekly users,
| and many of these users are really heavy users. Just taking
| myself: I probably consumed a few million tokens with
| GPT-5-Codex in the last 3 days alone. I am a heavy user,
| but I think there are users who burn through hundreds of
| times more tokens than me.
| yorwba wrote:
| Chinese companies need to pay much higher prices for the same
| GPUs, so they would need to charge more to make a profit, but
| it's difficult to charge more unless they have a much better
| product. So building massive data centers to gain market share
| is riskier for them.
|
| That said, Alibaba not releasing the weights for Qwen3-Max and
| announcing $53 billion in AI infrastructure spending
| https://www.reuters.com/world/china/alibaba-launches-qwen3-m...
| suggests that they think they're now at a point where it makes
| sense to scale up. (The Reuters article mentions data centers
| in several countries, which I assume also helps work around
| high GPU prices in China.)
|
| Circling back to OpenAI: I don't think they're spending so much
| on infrastructure just because they want to train bigger models
| on more data, but moreso because they want to serve those
| bigger models to more customers using their services more
| intensively.
| lvl155 wrote:
| They're trying to lock out competition from accessing compute.
___________________________________________________________________
(page generated 2025-10-05 23:01 UTC)