Thread

The Painful Truth: The RAM Crisis Is Only Just the Beginning

62 points · 69 comments · perelin

N_Lens · 1789618282 · focus · HN ↗
The big companies are insulated because they've locked in multi-year supply contracts with ram manufacturers - Microsoft, Google, Meta, Amazon (All on 3-5 yr contracts). OAI's deal with Samsung & SK Hynix is also colossal - locking up roughly 900,000 DRAM wafers per month, roughly 40% of world output (Stargate project).
Apple got caught out because it had a shorter contract that ended in the beginning of Q3, and they've tried to source their RAM from Chinese CXMT who declined because all capacity was already locked in contracts. They've gone with a lesser known company Kioxia.
Overall the consumer segment is completely neglected, companies don't care about end users in the current market conditions.
1. josephg · 1789618707 · focus · HN ↗
  > OAI's deal with Samsung & SK Hynix is also colossal - locking up roughly 900,000 DRAM wafers per month, roughly 40% of world output (Stargate project).
  I wonder if they'll, at some point, have enough RAM? Or is this is the new normal? Will models keep scaling with the amount of ram chips openai and anthropic own?
  1. bionhoward · 1789619444 · focus · HN ↗
    Even with efficiency breakthroughs, it would just afford packing more agents per unit of memory. Scaling compute and data keeps paying off, leading to smarter models, and smarter models have more demand even at higher prices because they can accomplish more work at higher quality. Sovereign AI hasn’t even really taken off yet to anywhere near the level it could. That’s going to dramatically increase the number of massive-scale users of AI agents. So no, IMHO they will “never” have “enough.”
    epicureanideal · 1789620733 · focus · HN ↗
    At some point though, won't someone be able to extrapolate the demand growth curve, and invest some colossal amount of money into making and selling more RAM chips?
    fragmede · 1789621449 · focus · HN ↗
    Yes. China's done just that. Expect those factories to come online within 2-3 years.
    cyanydeez · 1789648531 · focus · HN ↗
    Weve hit the sigmoid. Whats scalling is ancillary to the model. The cry for a slowdown is because the open weight models demonstrate the cost of parameter pacling is not work neither inference nor training.
    The assumption about the singularity simply is a delusion with LLMs.
    However, the models do provide a means to improve the harness universe, so that residual will continue to improve perception. Parameter cpunt will stagnate and training wont be justifiable from every angle.
  2. ohyes · 1789619653 · focus · HN ↗
    Hard to know, does each GB of ram give some marginal increase in profit or potential profit?
    I’d guess no. Past a certain point the model has all the capabilities it can possibly usefully offer and honestly we may already be past that. The next gen model just doesn’t seem like as clear a step up as it once was.
    Leonard_of_Q · 1789633043 · focus · HN ↗
    That point is said to lie somewhere around 640 KB if I recall correctly.
    <a href="https://skeptics.stackexchange.com/questions/2863/did-bill-gates-say-640k-ought-to-be-enough-for-everyone" rel="nofollow">https://skeptics.stackexchange.com/questions/2863/did-bill-g...
    josephg · 1789641852 · focus · HN ↗
    LLMs still seem pretty bad at writing large scale software like web browsers. Though it’s probably a problem of managing large context windows more than anything. Not sure if larger models will magically overcome that.
    cyanydeez · 1789653060 · focus · HN ↗
    The LLM by itself will never create software of any nontrivial (training) complexity.
    The harness though will improve while parameter count stagnates. The Qwen3.8 models are strong enough when given proper context.
    josephg · 1789657461 · focus · HN ↗
    > The LLM by itself will never create software of any nontrivial (training) complexity.
    Huh? I'm not sure what the word "training" does in that sentence. But "never" my arse. Frontier models can make nontrivial software already.
    For example, the other day I asked fable to reverse engineer the satisfactory blueprint file format. Then write a program to read the logistic flow graph in a blueprint. Then make an auditing tool that can analyse the graph to find problems.
    Well, it totally knocked it out of the park:
    <a href="https://seph.au/blueprints/#bp=0%3Aalumina.sbp" rel="nofollow">https://seph.au/blueprints/#bp=0%3Aalumina.sbp
    This is a relatively small program, but it's not trivial. I'd consider a trivial program to be something I could code up in 20 minutes. It would have taken me a couple weeks to make this blueprint auditing tool, including reverse engineering the file format, writing the analysis code, making the website, scraping all the in-game data on available recipes and icons and so on.
    I've got a lot of mixed feelings about LLMs. But it seems very silly to lie about what they're capable of.
    cyanydeez · 1789663139 · focus · HN ↗
    training is in there because it's "trained" to do trivial apps like TODO lists, etc.
    I'm well aware it can build apps. But if you arn't tracking what's going on, they're basically creating deterministic gates and tools to get it to do anything.
    There's no lie here, it's simply about what you think is _LLM_ and what is the rest of the software that's making it go. I use opencode consistently to build non-trivial apps with it, but it's not doing it with zero guidance, and it's routinely wrong about it's assumptions, and the rest. The thing keeping it on track is opencode, not the LLM's training.
    josephg · 1789696773 · focus · HN ↗
    > But if you arn't tracking what's going on, they're basically creating deterministic gates and tools to get it to do anything.
    So, the same as skilled humans then?
    ohyes · 1789747877 · focus · HN ↗
    That’s definitely how I like to ensure my code isn’t garbage. But now I use the tools I would have made to make the LLM & harness more useful and I might make more with the LLM because it costs me slightly less mentally.

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.