Hacker News Reader

Comment by daft_pink | Hacker News Reader

This reminds me of the railroads, where once railroads were invented, there was a huge investment boom of eveyrone trying to make money of the railroads, but the competition brought the costs down where the railroads weren’t the people who generally made the money and got the benefit, but the consumers and regular businesses did and competition caused many to fail.

AI is probably similar where the Moore’s law and advancement will eventually allow people to run open models locally and bring down the cost of operation. Competiition will make it hard for all but one or two players to survive and Nvidia, OpenAI, Deepseek, etc most investments in AI by these large companies will fail to generate substantial wealth but maybe earn some sort of return or maybe not.

View on HN

The railroads drama ended when JP Morgan (the person, not yet the entity) brought all the railroad bosses together, said "you all answer to me because I represent your investors / shareholders", and forced a wave of consolidation and syndicates because competition was bad for business.

Then all the farmers in the midwest went broke not because they couldn't get their goods to market, but because JP Morgan's consolidated syndicates ate all their margin hauling their goods to market.

Consolidation and monopoly over your competition is always the end goal.

> Consolidation and monopoly over your competition is always the end goal.

Surely that's only possible when you have a large barrier to entry?

What's going to be that barrier in this case - cos it turns out not to be neither training costs/hardware or secret expertise.

The large syndicate will create the barriers. Either via laws, or if that fails violence.

So I'm not an expert in this but even with DeepSeek supposedly reducing training costs isn't the estimate still in the millions (and that's presumably not counting a lot of costs)? And that wouldn't be counting a bunch of other barriers for actually building the business since training a model is only one part, the barrier to entry still seems very high.

Also barriers to entry aren't the only way to get a consolidated market anyway.

About your first point, IMO the usefulness of AI will remain relatively limited as long as we don’t have continuously learning AI. And once we have that, the disparity between training and inference may effectively disappear. Whether that means that such AI will become more accessible/affordable or less is a different question.

We have that now, DeepSeek just proved it.

You figure that out and the VC's will be shovelling money into your face.

I suspect the "it ain't training costs/hardware" bit is a bit exagerated since it ignores all the prior work that DeepSeek was built on top of.

But, if all else fails, there's always the tried-and-true approaches: regulatory capture, industry entrenchment, use your VC bucks to be the last one who can wait out the costs the incumbents do face before they fold, etc.

> I suspect the "it ain't training costs/hardware" bit is a bit exagerated since it ignores all the prior work that DeepSeek was built on top of.

How does it ignore it? The success of Deepseek proves that training costs/hardware are definitely NOT a barrier to entry that protects OpenAI from competition. If anyone can train their model with ChatGPT for a fraction of the cost it took to train ChatGPT and get similar results, then how is that a barrier?

Can anyone do that though? You need the tokens and the pipelines to feed them to the matmul mincers. Quoting only dollar equivalent of GPU time is disingenuous at best.

That’s not to say they lie about everything, obviously the thing works amazingly well. The cost is understated by 10x or more, which is still not bad at all I guess? But not mind blowing.

Even if that's 10x, that's easy to counter. $50M can be invested by almost anyone. There are thousands of entities (incl. governments, even regional ones) who could easily bring such capital.

> Surely that's only possible when you have a large barrier to entry?

As you grow bigger, you create barriers to entry where none existed before, whether intentionally or unintentionally.

Government regulation.

'Can't have your data going to China'

'Can't allow companies that do censorship aligned with foreign nations'

'This company violated our laws and used an American company's tech for their training unfairly'

And the government choosing winners.

'The government in announcing 500 billion going to these chosen winners, anyone else take the hint, give up, you won't get government contracts but will get pressure'.

Good thing nobody is making these sorts of arguments today.

The government isn't giving 500 billion to anyone. They just let Trump announce a private deal he has no involvement.

Correct, as I stated the government is just giving their 'blessing'.

Surely that will end in fragmentation along national lines if monopolies are defined by governments.

Sure US economic power has a long reach right now because of the importance of the dollar etc - but the more it uses that to bully, the more countries are making sure they are independent.

Which is the exact goal of the current wave of Tech oligarchy also.

I just read _The Great River_ by Boyce Upholt, a history of the Mississippi river and human management thereof. It was funny how the railroads were used as a bogeyman to justify continued building of locks, dams, and other control structures on the Mississippi and its tributaries, long after shipping commodities down river had been supplanted by the railroads.

This moment was also historically significant because it demonstrated how financial power (Morgan) could control industrial power (the railroads). A pattern that some say became increasingly important in American capitalism.

This is why we saw the market correction, because the AI hegemony has been cracked.

For the curious, it was vertical integration in the railroad-oil/-coal industry which is where the money was made.

The problem for AI is the hardware is commodified and offers no natural monopoly, so there isn't really anything obvious to vertically integrate-towards-monopoly.

Aren’t we approaching a scenario where the software is commodified (or at least “good enough” software) and the hardware isn’t (NVIDIA GPUs have defined advantages)

I think the lesson of DeepSeek is 'no' -- that by software innovation (ie., dropping below CUDA to programming the GPU directly, working at 8bit, etc.) you can trivialise the hardware requirement.

However I think the reality is that there's only so much coal to be mined, as far as LLM training goes. When we're at "very dimishing returns" SoC/Apple/TSMC-CPU innovations will deliver cheap inference. We only really need a M4 Ultra with 1TB RAM to hollow-out the hardware-inference-supplier market.

Very easy to imagine a future where Apple releases a "Apple Intelligence Mac Studio" with the specs for many businesses to run arbitrary models.

I really hope that apple realizes soon there is a market for Mac Pro/Mac Studio with a RAM in the TBs for AI Workloads under $10k and a bunch of GPU cores.

there was a company that recently built a desktop GPU for that exact thing. I'll see if I can find it

https://cerebras.ai/ ?

Compute is literally being sold as a commodity today, software is not.

>> Compute is literally being sold as a commodity today, software is not.

The marginal cost of software is zero. You need some kind of perceived advantage to get people to pay for it. This isn't hard, as most people will pay a bit for big-name vs "free". That could change as more open source apps become popular by being awesome.

Marginal cost has nothing to do with it - you can buy and sell compute like you could corn and beef at scale. You can't buy and sell software like that. In fact I'm surprised we don't have futures markets for things like compute and object storage.

I think a better analogy than railroads (which own the land that the track sits on and often valuable land around the station) is airlines, which don’t own land. I recall a relevant Warren Buffett letter that warned about investing hundreds of millions of dollars into capital with no moat:

> Similarly, business growth, per se, tells us little about value. It's true that growth often has a positive impact on value, sometimes one of spectacular proportions. But such an effect is far from certain. For example, investors have regularly poured money into the domestic airline business to finance profitless (or worse) growth. For these investors, it would have been far better if Orville had failed to get off the ground at Kitty Hawk: The more the industry has grown, the worse the disaster for owners.

https://www.berkshirehathaway.com/letters/1992.html

I think that's a very possible outcome. A lot of people investing in AI are thinking there's a google moment coming where one monopoly will reign supreme. Google has strong network effects around user data AND economies of scale. Right now, AI is 1-player with much weaker network effects. The user data moat goes away once the model trains itself effectively and the economies of scale advantage goes away with smart small models that can be efficiently hosted by mortals/hobbyists. The DeepSeek result points to both of those happening in the near future. Interesting times.

> where the Moore’s law and advancement will eventually allow people to run open models locally

Probably won't be Moore's law (which is kind of slowing down) so much as architectural improvements (both on the compute side and the model side - you could say that R1 represents an architectural improvement of efficiency on the model side).

I saw a thought-provoking post that similarly compared LLM makers to the airlines: https://calpaterson.com/porter.html

Main difference is that railroads are actually useful

Reply on news.ycombinator.com