Do large language models (LLMs) face the same fate as search? And is there a potential winner-take-all?
Then again, are we headed toward a duopoly that parallels the smartphone industry with two winners? On one side, we would have an open source LLM like Meta Platforms Inc.’s Llama and xAI’s Grok; on the other, closed models like OpenAI, Anthropic, or Google’s Gemini.
It hasn’t been discussed much, but there is a possibility that all this excitement and speculation ultimately winds up with just a few players making the business look a lot more like search (Google vs. Nothing) or smartphone operating systems (Apple vs. Android) – and a lot less like a mass of competitors.
While I was excited about the prospects of a new industry with lots of competition, it increasingly is shaping up as a battle of multi-trillion-dollar market cap companies and their proxies for nearly all the market share. Once things settle, I see little indication there could be a reversal.
There are a plethora of reasons it could end up this way, but I think it comes down to three key factors:
1. User Experience Will Drive Narrower LLM Adoption
Using multiple models based on each’s individual strengths has been a compelling argument for using a variety of LLMs. But as models improve, and become better, smarter and more enjoyable, the more you are likely to work with one model. On an individual level, this is similar to how Google was able to continue to build a bigger and bigger advantage as search use grew. Google searches became better, and other search engines fell behind having less data to optimize. This is likely to happen as LLMs are refined and grow more sophisticated. Their use will be rewarding, like staying at only Marriott or flying only one airline for the best experience. In turn, this will drive consolidation and preference. In many ways, generative AI is search 3.0 and it will add more assistant and agentic support for consumers. It is obvious which companies are best suited here and it is largely comprised of the biggest tech companies that have a mass of user data and a continuous flow of insights on search, commerce, social, and advertising.
2. LLM Cost and User Acquisition.
This is where I think Google is in such an ideal spot and Meta also sets up well. Both have a war chest and BILLIONS of users to keep and gradually migrate them to more generative use cases. Google has the search training data and financial clout to gain market share, and Meta has a massive captive user base which continues to provide it mountains of user data to train on. xAI is also an interesting candidate to compete here with X being a firehose of training data and we have seen how quickly Grok has become competitive with its newest Grok 2 Model. Open source LLMs will also be used pervasively by developers to power applications for business, for example.
All of this is why OpenAI (with Microsoft Corp.’s backing), Anthropic (bankrolled by Amazon.com Inc.), and other new entrants need Big Tech partners to speed up development and gain an advantage. Compute resources are also critical here—As is meaningful training data like Amazon’s advertising and e-commerce data. Also, these companies can build off of users on their own platform(s), products, and services. Which leads me to point No. 3.
3. LLMs Will Support Enterprise and proprietary data.
Arguably, the biggest opportunity for new entrants will be models that can complement the wealth of proprietary data and specialized small models that are being built by the likes of IBM Watson, Salesforce Agentforce, Microsoft Copilot, NVIDIA NIM, Databricks and more. Broad LLM data and language will be needed to coincide with the specialty data for enterprise use cases. This feels like a home run for Llama, as developers can build on open source. The good thing about these use cases is the compute and infrastructure intensity isn’t the same as the frontier models and race for AGI.
In the end, I see one or two main models for consumers and greater model diversity for business use cases (Generative AI within Apps and Software). A closed system like Google Gemini or OpenAI will likely win much of the consumer (Apple could rise here as well), while open source will carry the day for enterprises and developers building applications, and the Meta and xAI open platforms will power a proverbial “Android” ecosystem.
To work best, generative AI, and specifically LLMs and other models require a constant data feed from users and swapping between dozens of models will mean the data will be more fragmented and the models will know less about us making for less useful responses, assistants, and agents to create the most powerful user experiences. And while playing with a bevy of models and testing the ecosystem of offerings will be a great thing for us technophiles that love to play with new software and gadgets, much of the public will ultimately want one or two generative AI tools to do it every day searching, shopping, and AI driven coordination—And the system that knows you best, will deliver the best results.