Platforms Win Commodity AI
Tom Lombardozzi@thomas
The idea
I believe AI models are becoming commodity plumbing over the next year or two, so the profits flow to public companies with existing user relationships rather than the labs training frontier models. This basket owns consumer platforms that get AI nearly free, the chip and infrastructure suppliers who win as serving costs fall, and app-layer companies that rent models instead of building them. It's concentrated by design, Microsoft is deliberately excluded as too tied to lab economics, and I keep a cash sleeve to add on dips.
What will be bought
- They already own billions of users, so commodity models hand them frontier AI without the training bill. · Meta PlatformsIt reaches over three billion people and gives its models away, turning frontier AI into interchangeable plumbing that profits from AI features across its apps without needing its lab to win., AlphabetIt owns the two biggest consumer touchpoints on earth in Search and Android, and builds its own chips so serving AI gets cheaper every year., AppleIt owns the device in everyone's pocket and can plug in whichever model is best and cheapest that quarter, paying commodity prices for it., AmazonIt touches consumers through shopping, Alexa and Prime, and Bedrock lets it rent out every model like a utility, collecting on both the consumer side and the serving side.34%
- A price war on models is a volume war on the chips that serve them. · BroadcomIt builds the custom AI chips that let the big platforms serve models cheaply without paying full custom silicon prices., NVIDIAIf models become commodity plumbing, the volume of AI served to consumers explodes, and it still sells the most chips that do the serving., Advanced Micro DevicesIt is the cheaper alternative for running models, and a commodity model world is a price-sensitive world where buyers shop on cost per query., Taiwan SemiconductorEvery chip for inference gets made at its fabs, so it owns the factory that all the winning chips use, and commodity AI means more chips, not fewer., Marvell TechnologyIt designs custom AI chips for platforms pursuing cheap serving, and its custom silicon business grows when platforms expand orders to serve commodity models at the lowest cost.26%
- Companies that rent models get better AI every quarter without paying to build it. · AppLovinIt uses AI to sell ads without training its own frontier models, and cheaper models mean its costs fall while its ad product improves., DuolingoIt rents models to power AI tutoring features and charges consumers for them, so cheaper models mean better features at the same price., RedditIt sells its conversation data to every lab so it gets paid whichever model wins, and it owns a real consumer relationship too., SpotifyIt adds AI features like playlists and DJ voices on rented models while owning the listener relationship, no training bill attached.16%
- Cheap AI for billions of users still needs the networking, power and racks underneath it. · Arista NetworksIt sells the networking gear that moves data between AI chips, and the platforms are its biggest buyers when cheap commodity AI means serving far more queries., Vertiv HoldingsIt sells the power and cooling that AI data centers cannot run without, and serving cheap AI at huge volume means more data centers writing checks., Astera LabsIt makes the small connectivity parts that link AI chips together inside servers, pure plumbing for the serving buildout regardless of which model is best.14%
- Cash10%