The Four Questions of AI: Tunguz maps a coupled stack

Tunguz frames the AI stack as four layers, each with its own unresolved question, and stresses that none stands alone. At the top, breakout applications appear overnight but a new model release can erase them just as fast, leaving app companies to choose between building a brand, disrupting incumbents through pricing, developing their own models, or expanding faster than competitors.

In the middle, the fight moves to the harness and context layer. Tunguz notes that input context is up to 98% of all tokens, so a smart harness can cut the cost of the same answer by 71% and lift gross margin from 38% to 75%, which helps explain why enterprises are starting to lay claim to the harness to keep memory and permissions in house. For foundation models, training costs run into the billions, and recursive self-improvement has not yet bailed out the big labs: AI writes 75% of Google DeepMind's production code, yet Google sees no end-to-end speedup in AI research. Challengers are attacking that bill through process and architecture — one startup pre-trained a competitive coding model with a single researcher for under $10m, and a 109m-parameter JEPA model beat 72b LLMs with 42x less data.

Below the models, Tunguz answers the data-center question for now as no: concrete prices up 3x, strained grids, and turbine backlogs stand in the way, while unrestrained growth would require $4t in new debt over five years. He compares the whole arrangement to a double pendulum — simple in each part, yet complex, chaotic, and unpredictable in motion.