The generative AI stack is getting harder to read, not easier. Elad Gil, an early-stage investor who called the frontier LLM oligopoly in 2021, now says his uncertainty is growing at every layer of the stack. His core 2021 prediction is holding: frontier models like OpenAI, Anthropic, Google, and Meta's Llama are consolidating into a small cluster funded by big tech, while commodity models get cheaper fast. Training a GPT-3.5 equivalent costs roughly 5 times less today than it did two years ago.
The funding structure is the most important detail in this piece. Microsoft put $10B into OpenAI, Amazon and Google together put $7B into Anthropic, and these numbers look different when you compare them to actual cloud revenue. Azure generates $25B per quarter. That $10B Microsoft investment is six weeks of Azure revenue. Azure grew 6 percentage points from AI in Q2 2024, implying an annualized revenue increase of $5 to $6 billion, roughly half the OpenAI investment recovered per year. Meta separately announced a $20B compute budget to fund Llama-scale training. The capital is not coming from venture funds. It is coming from entities with direct financial incentives tied to AI infrastructure consumption.
Gil is not publishing conclusions here. He is publishing open questions, one for each layer of the stack, and the questions themselves are the reason to read the original. Where does durable margin actually accrue? Which open-source model becomes the commodity baseline? What happens to the middle tier of model providers squeezed between free open weights and frontier capability? If you are building or investing anywhere in the AI stack, the uncertainty Gil documents is more useful than most confident predictions circulating right now.
[READ ORIGINAL →]