Ask HN: Will LLM-commodification cause growth to continue without a correction? | Hacker News Reader

Ask HN: Will LLM-commodification cause growth to continue without a correction?

1 pointAbstractH2413 comments

Seems like the broader industry is waking up to what many folks have known for a while — models themselves are not a protective moat. They decay quickly and using them for basic tasks like structuring unstructured data is quickly becoming commodified.

What I didn’t totally appreciate until now is how that shifts us off the course where a few frontier labs end up becoming ACME mega corp which run the whole world and capture everyone’s money.

Power and value shifts to those who figure how to leverage LLMs to create value. Similar to how the use of a particular programming language is rarely the key to success and ISPs aren’t in charge of everything. Rather, the ability to turn code and a connection to people via the internet is the key to success.

So that makes me wonder - is the dot-com style correction in the industry required before we truly get a wave of exciting new companies emerge which upend society through the use of LLMs. Or can that shift happen without one?

I struggle to decide how this compares/contrasts with the dot com era. Where we had both the dot.com bubble burst and the internet totally change society.

Are you curious for curiosity sake? Or is there a decision you're trying to make that is dependent on these dynamics?

Mostly curious for curiosity sake.

But also questioning if the bleak outlook some have is overstated to some degree. Bleak outlook for the financial market, job market in general, and job market in tech specifically.

As I think through it and weigh different factors I think I’m concluding “mid-term concerns are justified, but long-term there’s more reason to be optimistic and even excited than some might think. We’re not headed to a distopioan future.”

> models themselves are not a protective moat

This has already been known for 4 years now amongst my peers.

Growth Equity raises like OpenAI's $110B round is not a Venture Capital raise nor an LBO or Mezzanine.

> is the dot-com style correction in the industry required before we truly get a wave of exciting new companies emerge which upend society through the use of LLMs

Nope. The Application layer along with Reinforcement Learning and DeepTech funding for hetrogenous computing and quantum/post-quantum has been hot for the past 2-3 years.

Most are in stealth. Most are B2B and Enterprise. Most of those founders don't use HN, as HN has an increasingly negative reputation amongst people in the space, and because HN's userbase has increasingly become more European individual contributors and less Bay Area, so the signal to noise ratio has dropped dramatically

> I struggle to decide how this compares/contrasts with the dot com era

Becuase that is not the best comparison when looking at the economics of the industry. A better model is the rise of hyperscalers and SaaS. Foundation models and Agents are an additional abstraction layer packaging distributed compute, just like how hyperscalers and SaaS packaged compute into easily distributable applications.

>> is the dot-com style correction in the industry required before we truly get a wave of exciting new companies emerge which upend society through the use of LLMs

> Nope. The Application layer along with Reinforcement Learning and DeepTech funding for hetrogenous computing and quantum/post-quantum has been hot for the past 2-3 years.

How does this deal with need for a correction in overvalued companies? And the impact on those who invested in them?

You still have a situation where there’s insane excess capacity built out. Akin to fiber during dot com era.

And the fallout from when companies like OpenAI need their valuation adjusted.

Chewy proved Pets.com thesis right. But that didn’t stop Pets.com from failing as part of a bubble bursting.

>> I struggle to decide how this compares/contrasts with the dot com era

>Becuase that is not the best comparison when looking at the economics of the industry. A better model is the rise of hyperscalers and SaaS.

Can you elaborate? The thing i still don’t see is how this creates the protective moat.

As I type this out I think we’ll go through an era of “buying SaaS is stupid, you can just build it yourself” before some SaaS companies prove they can build and maintain better than you (we’re probably in the midst of that already)

> How does this deal with need for a correction in overvalued companies

Which companies do you think are overvalued commensurate to revenue? The biggest issue we have for overvaluation is SaaS apps that raised in the 2017-22 period.

Terms were extremely lax and made it difficult to pop their bubbles.

> The thing i still don’t see is how this creates the protective moat

Protective moats are not the name of the game and never have been - they only reduce TAM and make it difficult to exit because comparable multipliers are limited, making valuation extremely difficult.

The primary moat has always been distribution, and this is where the foundation model companies have been extremely successful at. AEs who's Rolodexes included the F1000s all left for the foundation model companies and brought their clients with them.

Additionally, companies are approaching foundation models the same way you would cloud - a multi-model and a multi-cloud approach to reduce vendor stickiness.

Most companies that raised in the last 24 months.

Take Clay.com, what justifies the $5 billion valuation?

And all the frontier labs. Because their core offering will be commodified.

Finally, Nvidia and the companies whose main offering is computing power. Efficiency and commodification will lead to something akin to the fiber overbuild of the dot com era (like I referenced before)

As I type this out, I think there is an inevitable correction. Hard to say if it’s akin to dot com era or not. And those who succeed 15 or 20 years ago will be ready to take advantage of it. Be through capital, through skills and knowledge, or through stealth companies. Question becomes when to pull the trigger on taking any of those of their respective sideline.

> Take Clay.com

Their revenue and additonal metrics which I cannot disclose does justify their valuation. Their ICP who are using them (RevOps and Marketing) also love it, and have begun reducing SFDC spend as a result.

> Nvidia and the companies whose main offering is computing power. Efficiency and commodification will lead to something akin to the fiber overbuild of the dot com era

Most foundation model companies are already rate limiting enterprise customers due to lack of compute capacity. There is an actual bottleneck in capacity that the buildout is solving.

> Because their core offering will be commodified

Get. This. In. Your. Head.

Commodification is treated as a win. It means 3-5 vendors can take majority of the marketshare while expanding TAM and giving later stage investors multipliers they can use to exit.

It also means distribution becomes the key factor, and that is a toggle that is easy to manage.

> Question becomes if there is or isn’t a correction first. With its associated period of stalled growth (akin to the 2000s post bubble bursting)

A correction implies that there is an underutilization of capacity. We are not at that point, and won't be for the next 3-5 years.

Additionally, most foundation model companies are trying to expand up the chain into applications as well, much like how hyperscalers transitioned from being IaaS and adding additional abstraction layers to reduce deployment friction.

You’re justifying today. That was the same answer that made sense during peak dot com era.

What you aren’t answering is why to believe the current trajectories will sustain long term.

Enterprise companies are being rate limited and incentivized to look for other options and/or finding efficiencies id actually example of why commodification and overbuilding is likely.

> Particularly if enterprise companies are being rate limited and incentivized to look for other options

What other options? SuperMicro has an 18 month backlog. Same with Dell, HPE, and all the other compute manufacturers and everything in-between.

It's impossible to buildout because best case you would operationalize in 24 months, at which point you would be around 4 years behind SOTA because it takes years to train, and that money would be better spent trying to negotiate a more competitive inference terms.

---

The crux of the issue with the Dot-Com boom and bust was that most enterprise were not purchasing the products of most startups. The Dot-com boom/bust was largely a B2C play.

The current generation of AI products are entirely targeted at business applications, where I can justify reallocating spend on SaaS and maybe some headcount to a specific AI product which has a much more predictable cost model.

There is a B2C boom brewing now thanks to OpenClaw, but all that action is in China and India, not the US or Europe.

---

Edit: can't reply

> The emergence of something akin to DeepSeek 14 months ago that proves you don’t need all this power to get the desired results

Already has. India's Sarvam has GPT-4 or Deepseek comparable performance. There's a second model of similar scope in the works in India.

I know of 2 other projects in the UAE and KSA with a similar scope, 1 project in South Korea, and 1-2 in Japan.

The thing is, sovereign AI is the name of the game - no country wants to be dependent on Anthropic or OpenAI for government critical applications.

But this isn't an issue from a valuation perspective because it's these sovereign model projects that are also subsidizing domestic hyperscaler buildouts across APAC, as historically most compute was centered in SG+MY.

> That’s an interesting observation

It's because there are less luddites and more openness to experiment. If you spent your entire life using digital payments, running a digital storefront, and doing paperwork via an app you are much more open to playing around with digital and agentic workflows.

I've noticed most western HNers are significantly older (in their 40s-50s) and are much more resistant to new tech as a result, especially as most didn't grow up with digital products for much of their life.

> why are you here then?

Force of habit like taking a smoking. There are also the occasion tidbits of good info, but this is growing less common.

> I’d question if they are adverse or simply measured and able to see through hype because it’s not their first rodeo

There's a difference between hype ("AI will take all our jerbs") and reality ("Agentic workflows are becoming incorporated in most operational workflows but with a human in the loop, and that requires expanded capacity. In a lot of cases, agentic workflows will suck, but so did human developed workflows for cost centers").

People whose thinking is closer to the latter will be more successful professionally speaking. It's the same way HN used to reflexively crap on cloud and K8s, but it has now become the industry norm.

> What other options?

The emergence of something akin to DeepSeek 14 months ago that proves you don’t need all this power to get the desired results

Or multiplexing during the fiber boom.

> There is a B2C boom brewing now, but all that action is in China and India, not the US or Europe

That’s an interesting observation

> Edit: can't reply

Yeah, something odd is occurring where I need to click into the comment to reply

> sovereign AI is the name of the game - no country wants to be dependent on Anthropic or OpenAI for government critical applications.

This is an interesting concept. The idea that in the end you’ll see the emergence of something akin to the WWW for AI. That we’re in the AOL/Prodigy phase of it.

> I've noticed most HNers are significantly older (in their 40s-50s) and are much more resistant to new tech as a result

Im just below that age bracket, take it you are significantly.

I’d question if they are adverse or simply measured and able to see through hype because it’s not their first rodeo

But it begs a question I had from your first comment where you trashed HN - why are you here then?

> Commodification is treated as a win. It means 3-5 vendors can take majority of the marketshare while expanding TAM and giving later stage investors multipliers they can use to exit.

This is the most interesting point I think you made. But really isn’t in contrast to what I said.

Doesn’t resolve the question of if there will be a correction. Just that if current investors exit fast enfough those investors won’t be left holding the bag when one occurs and will be able to reinvest.

To some extent that’s inconsequential. If anything is certain it’s that there is always someone else available to invest money in the broader market at some point in the future.

Question becomes if there is or isn’t a correction first. With its associated period of stalled growth (akin to the 2000s post bubble bursting).

I recently spoke to an executive at Micron.

He mentioned an interesting difference between the AI boom and the dot-com boom.

In the dot com boom, distribution was something that didn't exist, you had to build it yourself. Expanding internet, expanding telecom.

But now, all the models can be instantly distributed through the existing internet. This leads to early revenue and better valuations.

Not perfect. Better. You still have the BS ARR calculations, the very optimistic(to put it kindly...) infrastructure buildout by the cloud companies. But even that, there's a lot of experience in the world today running hyperscalers, large data centers, points of presence, etc,. Not to mention 1/3rd of the world (India China) is now multiples richer compared to 2001 and participates in these things.

So I think that at the very least, the "crash" correction will be diffuse and happen over time smoothed out rather than a single day race.

Which can only be a good thing.

I also think that, unlike early internet, every country now seems to treat compute as dual use. Regardless of if they actually use LLMs or not. So there's a lot of government debt flowing into these things. Meaning financing is way less risky. Really, building a GPU data center in india today is almost free.

Reply on news.ycombinator.com