Google’s AI Advantages

I liked this piece by David Dayen, but I do quibble with this statement: Dominance in AI is predicated on having an early-mover advantage in data, which is used to train the various models. Nobody has more data available for training than Google.”

Google has a strong and potentially dominant position in AI, but I don’t think it’s because of a data” advantage, even though Google has access to gobs of data.

The AI space is more competitive in some ways than this argument would suggest, because useful data has been widely available, through scraping the web, and through open source” data sets like The Pile (and the infamous Books3).

Companies like Mistral, Anthropic, and OpenAI had no first mover” advantage with data, and they are industry-leading. Mistral is even showing that better data, not more data, can produce great results.

Of course, AI companies are being sued in more courts than I can keep track of exactly for using this kind of data. If the copyright and other claims are successful, then it’s likely that the AI space will become significantly less competitive. Though it’s not at all clear whether there would be a legal difference between training on Books3, were that to be found infringing, and Google training on, say, its Google Books archive.

But Google does have a durable advantage in other ways, and it’s the same as Microsoft’s, and Meta’s: computing infrastructure. Training models requires a lot of it, and very few companies have it. Microsoft has leveraged its infrastructure advantage most famously with its OpenAI partnership. Google has a global network of datacenters and vast computing resources. It’s calculated 100 trillion digits of pi! Not the most useful exercise but pretty nifty.

If Apple intends to incorporate generative AI into the iPhone, it’s likely to use its own models that run directly on the phone for lightweight tasks (its public research papers and internal models seem aimed at this use case), and outsource heavier lifts to the cloud. Very few companies have the infrastructure capacity to handle the huge number of iPhone users.

Running models (inference) requires less computing horsepower than training models, but at Apple’s scale, it’s hard to see who could run even that other than Google, Microsoft, Meta, or Amazon–or companies who rent from them.

Google has another advantage, in that it’s one of the few AI companies not dependent on Nvidia. Nvidia can’t make enough of its high-end GPUs to satisfy demand and is in the position of deciding who its customers should be.

Google uses its own TPUs, instead of Nvidia’s more general-purpose GPUs. It makes TPUs available via Google Cloud but does not sell them, and they are customized to Google’s specific needs (and apparently harder to write for). This is a pretty major advantage for Google. I don’t have the chops to tell you whether TPUs or GPUs are better, but not being dependent on Nvidia probably makes up for any disadvantage, if there is one.

Google has other advantages, too: It can pay the high salaries that in-demand AI researchers and developers command, and it can integrate its AI throughout its product portfolio, from search to Google Docs to its ad tech. I’m probably missing many more, though I think in the case of Apple, its advantage is in compute,” to use the latest semi-annoying way of putting things. (To paraphrase Calvin, nouning verbs weirds language.)

I agree with Dayen that an exclusive deal between two dominent firms like Apple and Google would show impunity–but I’d add recklessness and shortsightedness. Both of these companies are currently in litigation with the DOJ, where key to Google’s case and mentioned in the Apple case, is the exclusive search deal where Google pays Apple billions of dollars a year to be set as the default search engine in Safari, and for Google traffic generated by Apple users.

Without knowing what this rumored gen AI stuff looks like, it’s hard to know how giving users a setting or otherwise mainting choice and competition would work, and Apple does seem to be discussing potential partnerships with a number of companies–Baidu for instance, which might just be for Chinese-language models, or models that comport with Chinese content laws (not going into all THAT here).

But, if only a few companies are even capable of partnering with Apple on this, then there is no reason to have an exclusive contract, since it would be redundant. If multiple companies are capable of partering with Apple, then it seems overwhelmingly in Apple’s interest to find a way to provide some level of user choice and to allow for competitive entry. First, this would be the right thing” to do, not that many tech companies seem to be persuaded by this. And not just to avoid regulatory or antitrust pushback–which at this point, you’d hope has sunk in over at Cupertino. But also for classic tech commoditize your complements reasons–unless it turns out that AI is the real product, with the smartphone relegated to the complement” role. We’ll see.


March 25, 2024