From Models to Agents to Discovery: Building the Full Stack of Agentic AI

TL;DR

• Two curves rising together: adoption is exploding (Kaggle's 5-Day AI Agents Intensive drew 1.5 million registrations) while agentic tasks need 10–100× more inference compute than non-agentic workloads. Google now processes 3.2 quadrillion tokens per month — roughly three novels' worth of text for every person on Earth.
• The whole stack has to be co-optimized: AI Hypercomputer (CPUs/GPUs/TPUs) → frontier research and models → data → agentic security and defense → platform → end-to-end agentic solutions. The eighth TPU generation splits the line for the first time into 8t (training) and 8i (inference): 121 FP4 exaflops per 8t pod (~3× the prior generation) and 11.6 exaflops per 8i pod (10× the prior generation).
• The destination is an autonomous discovery loop: ingest literature, protein databases, and experimental logs → generate and debate hypotheses with an AI co-scientist → simulate in seconds with AlphaGenome and AlphaFold → execute via AlphaEvolve and Gemini Robotics wet-lab runs → feed results back. Research cycles that took years collapse into hours and days, and biology is only the first example.

Key Points

Scale numbers and the stack (~01:02–01:07)

• Three or four years ago the interface was a singleton chatbot; today it's semi-autonomous or autonomous agents. On the adoption side: Kaggle's 5-Day AI Agents Intensive online course last fall drew 1.5 million registrations.
• On the cost side: an agentic task costs 10–100× more inference compute than a non-agentic workload. Both adoption and per-interaction complexity are climbing at once.
• The result at Google's scale: 3.2 quadrillion tokens processed per month. His intuition pump — one quadrillion is about one novel per person on Earth, so Google's monthly output is roughly three novels for everyone alive.
• The stack that has to catch this opportunity, bottom to top: 1. AI Hypercomputer — CPUs, GPUs, TPUs 2. World-class research and frontier models 3. Data, so models and agents can do anything meaningful 4. Agentic security and defense, because meaningful agent actions create risk as well as opportunity 5. A platform, so these capabilities are reusable cheaply and efficiently 6. End-to-end agentic solutions that actually do things for you
• Co-optimization across the entire stack is what extracts maximum value from AI.
• TPUs: over ten years of investment, with innovations like liquid cooling and shared memory. The newest generation, TPU v8, splits the line for the first time — 8t for training, 8i for inference (he explicitly picked up DeSantis's point from the previous talk). Against Ironwood, today's workhorse generation:
• Training: 121 FP4 exaflops per pod, roughly 3× the previous generation, with 2× and 4× jumps in memory bandwidth.
• Inference: 11.6 exaflops per pod, a 10× jump, plus significant memory-side gains — because inference demand is rising very fast.
• These TPUs train the Gemini family (Pro, Flash, Flash-Lite), image and video generation (Veo, Imagen), the Genie world model, AlphaFold, AlphaGo, and AlphaChip, which designs the next generation of TPUs — "effectively TPUs designing for TPUs." A number of other companies now train on TPUs as well.

The agent platform's four problems: build, scale, govern, optimize (~01:07–01:09)

Hardware and models aren't enough, because "building agents is not just like writing a prompt."

1. Build — access to all model types, the Agent Development Kit, an AI Studio for building agents.
2. Scale — a managed runtime that takes you from a single agent instance to millions.
3. Govern — as agents start doing meaningful things you need agent identity, agent registry, and agent gateway to keep agents inside the right confines. The key design point: an agent's identity system is separate from the user's, because agents can do things beyond what the user can.
4. Optimize — once in production, tracing, simulation, evaluation, and observability let you keep improving.

Google also builds on top of this stack: biology (AlphaFold, AlphaGenome), mathematics (AlphaEvolve, AlphaProof), physics and chemistry (GNoME and fusion work — see To Verify), climate and sustainability (AlphaEarth, WeatherNext).

Impact examples: - AlphaFold in plastic pollution (designing plastic-degrading enzymes), antibiotic resistance, structural biology, and especially neglected diseases — conditions pharma underinvests in for lack of economic incentive, where lowering the cost and difficulty of exploring drug designs has material impact. Also malaria vaccine work and drug delivery. - AlphaEvolve, a general-purpose optimizer, used heavily inside Google for data center optimization and externally for route optimization, quantum error correction, and e-commerce demand forecasting — very different application shapes, each with large improvements.

The autonomous discovery loop (~01:10–01:12)

As models keep improving, the building blocks assemble into a self-driving research loop. The slide used biology, but the pattern generalizes:

• Ingestion — agents that digest large volumes of literature, plus AlphaFold's protein database and experimental logs.
• Hypothesis generation — an AI co-scientist that debates hypotheses and generates novel solutions.
• Modeling — AlphaGenome and AlphaFold simulate hypotheses and show effects in seconds.
• Execution — AlphaEvolve and Gemini Robotics run wet-lab tests and feed the data back into the loop.

The payoff: a research cycle from data input to product outcome that used to take years can now be done in hours and days — across almost every facet of human discovery, not just biology.

Quotes

"One quadrillion is one novel per person on this earth. So in a way Google generated tokens which are equivalent to three novels for every person in this world." (~01:03)

How to feel 3.2 quadrillion tokens a month.

"It's effectively TPUs designing for TPUs." (~01:07)

AlphaChip trains on TPUs and designs the next ones.

"A research cycle which used to take years … can now be done in hours and days." (~01:12)

The claim the autonomous discovery loop rests on.