[jredwards]

It's tough for me to square the two things happening simultaneously in AI right now:

1. LLM Model providers are starting to charge real costs to users, revealing that AI usage is much more expensive than the subsidized rates we've been seeing for years.

2. Google is now using an LLM to answer every single google search that happens, for which Google bears the entire cost.

[ImaCake]

In my experience, LLM usage follows an exponential distribution i.e. most people using LLMs are not using many resources, but a very small few are using a massive amount of resources. Most LLM usage is trivial and tiny, most of what I use LLMs for could be done by a local model with a decent web retrieval tool. What some people I know use LLMs for requires massive volumes of tokens. Its that small power user base which I would bet they are targetting.

[dredmorbius]

This view misses a key point: the cost of LLMs is largely in the training rather than the use of the tools.

That is, it's training costs which are fixed, high, and amortised across all revenue potential.

In raising prices, pure-play AI chatbot vendors are facing the challenge that their only revenue option at present is pay-for-use. This is typical of numerous other information goods (software, comms networks, information services), but simply reflects the business reality that cost-basis and revenue-basis are largely disconnected, and that average cost pricing tends to be necessary despite not matching marginal cost provision. (That is, the charged per-unit price of use is going to be far higher than the actual marginal cost of provision for that use.

Google's situation differs from pure-play AI in several ways:

- The firm already has one of if not the largest corpus of Web data, as well as much offline / print material data (Google Books), of any entity in the world.

- Presuming Google are already using this for training AI for other purposes or general resource, the fixed costs are sunk costs already incurred, subsidised by the firm's existing AdTech monopoly, and which might as well be put to use.

- Similarly, Google's service costs (marginal costs) for traditional GWS are probably, within an order of magnitude comparable to those for an AI/LLM response.

- Google presently captures 90% or more of worldwide general web search (GWS) traffic, meaning it has an extant market for offering AI as a default search alternative.

- Complaints about declining Google search quality are nearly as old as Hacker News. Here's a seventeen year old submission on that topic, "Google's regression toward mediocrity (search quality & aggressive matching)" <https://news.ycombinator.com/item?id=641145> (2009). There are many, many, many results in both stories and comments for related searches. If traditional Web search is going to die anyway, and LLMs replace it, Google might as well get up on the curve.

- Google has arguably the most advanced ranking data for websites, whether that's on the basis of quality and relevance (for good) or advertising/revenue generation (for bad, at least from a public-value perspective), and can and all but certainly does leverage this in training runs, as well as use that training to further refine its ranking systems.

- Google can directly attach its existing AdTech revenue model to LLM search. And probably has options for extending advertising into the LLM SERPs themselves.

- Doing all of this directly attacks pure-play AI firms' attempts to capture Web search from Google, as well as challenges other GWS challengers (DDG, Kagi, Marginalia, etc.), all of whom have vastly lesser revenue options and technical resources than Google.

I'm not saying this as a fan of Google (I think the firm should be broken up or destroyed, and it's not the only one). But I think it's a fair assessment of the firm's strategic position.