upvote
I've been thinking for a while, there's not real winners here except the incumbent technology providers. Hear me out: all models are converging towards the same level, gains are getting smaller and harder to come by. The models are commodities nothing more.

This is the leap, nobody really wants to front a model for someone else. If i build an agent, or a service that requires a model, I'd prefer to push the model onto someone else, preferably at no cost. This is a leap as I'm sure right now, most people / businesses are thinking actually i do want to own / front the model.

However, if you accept the leap the easiest way to do this is to make the model the users problem.

From a business point of view that makes things really easy, from a customer point of view, they simply have to accept whatever their vendor of choice is pushing down their throats.

So as a business I build for whatever model Google makes available to android, and whatever model windows bundles, and whatever model Apple bundles, and, excluding the long tail of Chinese vendors and Linux (sorry, its always left out) and that's it, problem solved, and the customer picks up the tab for the tokens

reply
Step 2. Rising competition causes returns to fall below cost of capital.

https://www.youtube.com/watch?v=2J2Fb1bBufA

reply
I've found that there is value in consuming AI services from your existing cloud provider. Customers and auditors have less of an issue with "we use AI services from AWS/Azure/GCP" if the data was already in those clouds and it doesn't expand the risks of data being breached, or trained on, by some other provider.

When you are already trusting 100% of your data, and computing on that data, to someone like AWS, it doesn't meaningfully increase risk to use an additional service, even if it is an AI service.

reply
Google has a bit of a Network Effect going... my vehicle got an OTA update to use Gemini. Between that, search, storage, and the YT Premium bundle it was enough to convince me to float a subscription.
reply
> my vehicle got an OTA update to use Gemini

G. A. H.

edit: Y'all downvoters want genAI in your cars?!

reply
deleted
reply
I think anthropic with its enterprise strategy and google with its integration in everything have a bit of a moat.

But I switched from ChatGPT to Claude 3 months ago because my account was down for like 6 hours. I haven’t used it since. It’s too easy to switch away from chatbots on a whim. There is no moat for that.

reply
> I think anthropic with its enterprise strategy and google with its integration in everything have a bit of a moat.

But... Anthropic doesn't have a moat. It's clear at this point that SOTA models are not a moat, and Opus 4.6-level (or GLM 5.2) is sufficient.

Google, though... they own the entire vertical, from the semiconductors to the end-user software. They may have a moat.

reply
The narrative that superintelligence is imminent is partially at fault here.

There are competing definitions of what intelligence even is, and the one that I find most striking is from Francois Chollet which is that intelligence can be boiled down to skill acquisition efficiency. This type of definition makes intelligence more akin to polishing a ball than growing a watermelon.

The superintelligence doomers warn that the watermelon is going to start growing exponentially and crush everyone. But what might actually be happening is that we are not growing a watermelon but rather polishing the ball until its really smooth and shiny. There's a point where you can get it to micron levels of polish but for most tasks (white collar text domains tasks), it's smooth enough! You will be able to go to the ball store and buy a low cost made in china ball for most tasks.

The real challenge is actually branching out domains and modalities to tackle things like blue collar labor. Over time, white collar work automatable or able to be made hyperefficient by LLMs will see LLM commoditization.

reply
Observationally, for people that /aren't/ using models to code but to just do their white-collar job, claude.ai /is/ AI, now. The entire perspective for how to use AI is through claude skills, claude projects, claude cowork, etc. They've massively won the corp buy-in at the moment I believe.
reply
> The entire perspective for how to use AI is through claude skills, claude projects, claude cowork, etc

But as they have repeatedly pointed out, creating software is almost zero-cost now, so software cannot be a moat.

After all, all of the Claude software can be vibe-coded by any competitor; that's the dream that Anthropic has been selling anyway...

reply
doesn't matter. that just means they've incentivized all competitors to enter the market and let's be honest none of their tools are that novel.

https://www.youtube.com/watch?v=2J2Fb1bBufA

reply
I guess I’m thinking a lot of companies seem to be getting Claude code subscriptions. It usually takes some time and effort for an org to switch away from one solution. In the meantime a lot of workflows get more and more tied to Claude in particular.

It’s not much of a moat, but it’s more than a lot of orgs have.

reply
obligatory correction: the semiconductor layer is still owned by TSMC and Samsung. Google sketches chip designs for them to implement - that's the lowest layer they control. I am not denying that this is impressive.
reply
google might have tons of integration. But if it invested too heavily into AI then it will also suffer when increased competition causes returns to fall:

https://www.youtube.com/watch?v=2J2Fb1bBufA

reply
The moat is shifting from technology to access to proprietary training data. It doesn't matter how good your LLM platform is if you don't have good data to feed the training run. Public Internet data and published media is already mined out. Now the frontier LLM vendors have shifted to licensing proprietary data that's locked up behind corporate firewalls, and even hiring human domain experts specifically to create new training content in target verticals. You'll see the effects of this next year, although it might not be obvious to those who mostly only use LLMs for coding tasks in popular programming languages for which there was already a lot of training data.
reply
> Now the frontier LLM vendors have shifted to licensing proprietary data that's locked up behind corporate firewalls, and even hiring human domain experts specifically to create new training content in target verticals.

That's a losing proposition for any token provider - it's expensive and slow, and when you're done everyone with money to rent a last-gen H100 is going to distill your "closed" model anyway.

reply
> That's a losing proposition for any token provider

The specialized models for targeted verticals being discussed may well not be sold by tokens, but instead be behind the scenes powering dedicated packaged solutions where the customers don't have raw access to the model. Token providers still won’t have a moat, but AI isn't just selling tokens.

reply
AWS and Google at least own their own hardware (Trainium and TPUs, respectively). It's a moat in the sense that designing, building, and deploying your own chips at scale is quite a feat and not easily replicated. The vertical integration will allow them to continue to be profitable once the models get good enough and competitors' prices race to the bottom. Google has Gemini; AWS may not deploy its own models (yet?), but that's not necessarily a losing position, as long as the market is able to run models sourced elsewhere on Trainium and the price is right.
reply
Isn't specialized hardware also a big risk? GPUs are more amenable to any big changes that may happen in the next 5, 10 years of AI research. Maybe we won't even be talking about LLMs anymore. Maybe matrix multiplication won't even be the main primitive.
reply
If matrix multiplication isn't the main primitive, I think we have a lot of pain coming our way.
reply
Maybe that is far fetched, but I could see them specializing for some super high dimension multiplication and meanwhile 5 years later turns out "all you need" are 3x3 matrices and suddenly 90% of your specialized hardware is now dark silicon :)
reply
Their "own" as in built by Marvell and Broadcom. Especially Trainium but also TPU4.
reply
That's exactly what giant train corporations thought. "We own all the railways, we've squashed the competition"

and they STILL went out of business because they over-estimated the demand for their shitty rails they built to the middle of nowhere. Same with "AI."

https://www.youtube.com/watch?v=2J2Fb1bBufA

reply
The adoption of standards like skills and agent setup helps a ton. Nobody wants to be locked into an AI vendor like with cloud systems in general. And companies can't hold on to the #1 spot across multiple areas for very long, so users are even more motivated to move their process and stack between coding tools and AI companies behind them like Claude code.

Vendor lock in cannot happen, or you're bankrupt.

reply
Amazon Bedrock is probably middlemanning an insane amount of token consumption these days for the same reasons.
reply
Is Bedrock a "middleman?" I believe that they run all inference inside of AWS data centers, on their own infrastructure.

Their new endpoint even promises zero operator access [0]

[0] https://aws.amazon.com/blogs/machine-learning/exploring-the-...

reply
Sure, but fundamentally they’re acting as a distributor of someone else’s product in the form of the frontier models. That’s a classic middle-man.

No value judgement. I think this is a fantastic strategy.

reply
Weights are worth far more than data centers.
reply
> Weights are worth far more than data centers.

I dunno, hey. After all, I can't distill my competitors datacentres :-)

reply
Why?

Seems like open weight models keep catching up to state of the art within a few months, at most. Doesn’t seem like much of a moat to me.

reply
If/when open-weight models do catch up (i.e. become the dominant product in demand), Amazon transitions from a middle-man to the supplier with the best economies of scale.

Great business either way. You could even draw an analogy to Linux/OSS & the origins of AWS. They started as basically an infra middle-man for other people’s technology. But as the core tech commoditized, they transitioned into selling their own higher level services at scale—like Bedrock.

reply
You may not care, but a lot of people I know care what brand chat bot they use personally,. usually it's tied to trust and reputation more than anything else. People are fickle.
reply
> I use their services, but I frankly don't care who provides it. I'll chase the chepest/best and have no issue switching from one to another.

For the hyperscalers, there is an ease of remaining in the Azure/AWS/GCP fabric from a data provenance perspective, particularly for regulated industries or large, risk-averse enterprises. There's also, of course, a certain network egress tax in most cases.

reply
Nvidia has a moat. Hardware is hard. No one really competes with them for general compute
reply
Nvidia's moat is the IBM, Microsoft moat:

I am about to spend $20M, if I buy anything other than Nvidia, and things go wrong, I am going to get blamed, and if things go right I will get no credit. This is why AMD is making no progress outside of very narrow cases and supercomputing.

reply
AMD Instinct is their direct competitor for compute and they are better per dollar, better per watt, and out competing on raw performance.

Only thing holding them back is fab capacity which nVidia keeps buying in bulk to keep them small.

reply
AMD is held back by their interconnect and firmware disadvantage compared to nvidia. They’ve been trying really hard to create their own cuda, but rocM and HIP still aren’t very popular especially for research.
reply
And their repeated refusal to either implement CUDA or reimplement everyone's CUDA libraries on their own platform. They say that AMD never misses a chance to miss a chance.
reply
Have you ever actually had anyone work with these chips? Developer ux on amd is terrible.
reply
Yes. We have a quad MI300A server and run several inference models on it. For $107k it has saved us so much money on tokens already and it's a heck of a lot faster than cloud services.
reply
> Have you ever actually had anyone work with these chips? Developer ux on amd is terrible.

Just how much of dev ux do you need? A foundational library, of course, but as the AI companies keep saying, their models can vibe-code what's needed for those chips anyway.

reply
I thought that Nvidia's moat was more in CUDA? Hardware is hard but we've already seen other companies like Google design neural processors with compute efficiency close to Nvidia.
reply
General compute is also the worst solution to the problem.

Nvidia's entire business is dependent on Google not being able to make TPUs fast enough.

reply
Google would have to start selling them (the real ones, complete with interconnect) to third parties. If google does that, though, Nvidia is done.

Unlike AMD, Google can actually ship software. AMD has never shipped good software other than drivers (maybe) in the entire history of the company, including both ATi's history and true AMD. They have always relied on Intel to provide the software.

reply
Oh great, good to know the shovel seller has the market cornered.

Now back to the conversation, do any of the gold miners have a moat? Or is this a race to the bottom?

reply
Uhh. I actively and vocally avoid all things Microsoft. I see Microsoft and I immediately think buggy software with zero security.
reply
That's fine, but your inexperience with large companies that are MS's bread and butter doesn't really give you any credibility here. It's the standard for a reason.
reply
Can concur. I hate them with a passion, but corps love them, and I hate to say it... with good reason.

They're the only player in the Identity-Document-Email-VM-Storage space that's even remotely unified.

reply
Maybe so, but you clearly aren’t a representative sample of corporate decision-makers when it comes to AI (or broader IT) services.
reply