AI model providers have zero "moat", clients change them as they see fit. This week ChatGPT, next week Claude. The real value is and going to be in hardware - as long as China doesn't enter the GPU/RAM race.
I increasingly see AI investment, generally speaking, as a lost cause. It has very little chance to pay off.
Yup. Model capabilities seem to keep converging quickly, not leaders breaking away for long.
Frontier labs are racing towards SaaS commoditization at incredible speed. And while there might possibly be $Trillions in productivity gained from their use, there's no reason to think those gains get captured by the model makers or inference providers at this point.
Maybe the Claude or ChatGPT desktop apps will dominate as the new MS Excel, but that's hard to do without already having locked the whole market into Windows.
There's virtually no platform play available to them.
> there's no reason to think those gains get captured by the model makers or inference providers at this point
Yeah it almost certainly won't be captured by them. That value is going to be captured by the folks/companies that shrink wrap the capabilities into a nice SaaS or other tool, that a business can buy off the shelf and give to their employees.
The model makers are on a fast track to just becoming dumb pipes, not unlike ISPs.
> AI model providers have zero "moat", clients change them as they see fit.
That might be true in tech-savvy industries -- but in non-tech industries where the biggest software purchase might be the office suite or the ERP, inertia means the GSuite shops stick with Gemini, and the Exchange/Office 365 shops stick with Copilot.
I tend to agree with this sentiment. I'm not in the tech sector. As an outsider, it seems to me that OpenAI and Anthropic are chasing government and the defense industry as their main clients. Google and Microsoft are chasing business clients and educational institutions. Amazon and Apple are chasing consumers.
Isn’t Anthropic literally flagged as a supply chain threat?
in the US only insofar as they wont let claude decide who to kill.
from a non-US perspective yes, but so are the rest of the major providers
No, what I mean is that the pentagon is still labeling anthropic a supply chain risk, meaning it shouldn’t be used for government contract work and anything military related. I don’t see how they can have a strong presence in the defense sector when that’s the case
At least from some smaller marketing companies I know that isn't necessarily true. They often have Gemini or Copilot and Claude nowadays and before Claude it was ChatGPT.
The moat is way smaller than with Office or Gsuite because they feed data into the chat interface and it gives them an answer. The moat for Gsuite and Office is higher because you have to move all your data and reorganize it. Oh and everyone has to learn how to use the new software clients.
Copilot isn’t a model per se, no? It’s a harness that can use any model that supports tool calls from what I understand. It’s the way Microsoft commoditize ai models
Another casualty of Microsoft using Copilot to describe so many different things.
Actually surprising that windows hasn’t been renamed CopilotWorkbench yet
There is a time window when it will flip. When Internet came along, we had a number of businesses that did not survive over the next years.
This time, it is different with AI. The rate of change is significant.
Just out of curiosity, what is the change and how are you measuring its rate?
From no internet to internet the change is pretty profound. But my job is already very automated for the most part. It's true AI might automate it a bit more, but it's not like I'm going from zero automation to full on automation. That's not nothing, and it is worth something, but it's also not internet from no internet level of change either.
I think you don't understand moat - that's not a moat.
The trick is antitrust style bundling. The massive pile of documents and processes tied to GSuite is a moat which makes it hard to switch to something like o365. Since a company might effectively be locked into GSuite (the primary product), if Google forces companies to buy Gemini (the secondary product) by bundling it with GSuite, they've given themselves a moat in the LLM space using their document/email moat from GSuite.
This is essentially what Google has done, and it's a shame the US is so weak on enforcing antitrust laws.
It can be if it’s really sticky. But I don’t think ai models themselves will ever be sticky, the harnesses might be. But there is very little money to make in the harness itself, and they are also very easy to copy, so yeah, no moat
> as long as China doesn't enter the GPU/RAM race
China is obviously in the GPU/RAM race. Heard of Huawei, Moore Threads, Lisuan Tech, CXMT?
I've heard of one of those.
I'm just happy we get to reap the rewards "for free" (i.e open models are slowly becoming usable, and the winner of the arms race will definitely stand on the shoulders of their competitors that didn't make it)
Not to mention with each iteration of every model you get lower cost per token. It’s really a race to the bottom for hyperscalers and neoclouds at this point, with technically only two paying customers.
> The real value is and going to be in hardware
Unless someone comes up with a brilliant optimization strategy or new hardware that renders all that inefficient Nvidia crap overnight.
I'm privy to dozens of people working on this problem every day and I imagine there's many more people working on this problem out of sight. I'm bullish on this idea, but it's going to be a slow burn.