Big Tech has an Antitrust Problem
Google, Amazon, and Microsoft are in awkward frenemy relationships with LLM developers. This could be very bad for them and good for antitrust lawyers.
To build really smart AI models (e.g. ChatGPT/Large Language Models and other Foundation Models), you need three things:
Lots and lots of data,
Really smart data scientists and engineers, and
Lots and lots of computing power
That’s pretty much it. There are some other things in the details, but if you don’t have these three things, you can’t do anything.
Data is pretty much a solved-problem for most applications (at this point). It is not without it’s legal questions and jeopardy for the future. But, for now, you can go somewhere like Common Crawl, download a massive text data set, and get started with building your chat bot that specializes in Shakespearean fan fiction, or whatever.
The really smart data scientists and engineers part of the equation is also readily available in a free labor market. This talent is leaving big tech and receiving big checks from VCs every day. There isn’t a hotter skill in the tech market today than “can work with transformers”, and these people are getting the opportunities and cash to match that. Normally, the talent market can be stifled by onerous non-competes in employment agreements. This is very common in industries like hedge funds. However, it doesn’t appear to be boilerplate in data science employment agreements (yet). Thus, this is a pretty free talent market.
However, the computing power is difficult for a few reasons:
You need lots of it. Training a modern LLM requires massive amounts of compute power.
It is in scarce supply. Nvidia is the main supplier of GPUs, which are essential to training these markets, and they are selling chips as fast as they can make them. The most important person at this trillion dollar company is probably the Chief Supply Chain Officer.
It is expensive. Single GPU units have price tags north of $10k and cloud prices are soaring. There used to be jokes about startups who would go buy their equipment at Best Buy then raise VC money based on their “technology infrastructure”. Building any kind of technology infrastructure today would be more like buying a commercial auto fleet than a few laptops.
So if you are a company looking to break into the Foundation Model business and compete with ChatGPT, then you need to work hard to solve the computing power problem. Maybe you could scrap together some GPUs on the primary or secondary market and patch together your own data center, but this would be really hard and probably not cost effective. Instead, what you’ll probably do is opt for GPUs from a big cloud provider and pay as you go. This gives you almost instant access to the computing power you need provided you are willing to pay for it.
There are three main places you can go if you want to get these GPUs in the cloud: Amazon Web Services, Google Cloud Platform and Microsoft Azure. There might be other providers that offer cloud-based GPUs on-demand, but those are mom & pop shops compared to Walmart in this market.
Coincidentally, it just so happens that these three companies are also actively developing their own Large Language Model solutions (Google is building LaMDA, OpenAI is basically a subsidiary of Microsoft now, and Amazon will power the next generation of Alexa with their own LLM).
So, if you want to build a new LLM, you probably have to go to a direct competitor in the technology (Amazon, Google, or Microsoft) to pay very high prices for access to a scarce resource (GPUs) and put all of your proprietary methods and data on their servers.
Now, most startups are still in an operating mode of hanging on to the outside of rocket ship with their fingernails, so they don’t have time to think about things like “antitrust abuse” and “pricing power abuse”. However, once they have some downtime from cashing VC checks, they’ll probably start ask questions to their engineering leaders like “Wait, what now?”
Here’s an article from the New York Times last week about the founders of Cohere and how they find themselves in an awkward situation with Google, who use to employ them. Cohere’s products will directly compete with Google, but to build them, they are using Google’s cloud solutions to train their models. Boy, this awkward. Surely, the leaders of Cohere know that any morning, Google could wake up and decide to cripple their business by raising Cloud GPU prices significantly. Or, perhaps they will look at Cohere’s usage data of GPUs to infer some of clever tricks Cohere has developed to train models more efficiently. Sure, Cohere could completely move their infrastructure over to AWS or Azure, but that’s not what you want on a random Tuesday.
If I were Microsoft, Google, or Amazon - then I’d be really careful right now! And if I were Cohere, Apple, or Meta - I’d be handing out pitchforks! This awkward frenemy relationship that the big cloud providers are maintaining with their competitors in the AI space could get really messy. At any moment, one of them could change the ability for companies like Cohere to economically access GPUs at scale. Or, the three could be really evil and collude to raise market prices for GPUs access, putting much of their competition out of business instantly. These would be really bad things to do!
Everything the cloud providers do on their cloud-based GPU services will probably be under the microscope of antitrust regulators for anti-competitive behavior. If I were the General Counsel at any of these companies, I’d go ahead and set my email servers on-fire because someone will put something embarrassing or illegal in writing. “Hey, let’s raise the prices on our ‘GraPfrUits’ because our AMAZing friends are doing that too”. People actually put this stuff in emails!