Nadella walks through the $1B OpenAI bet, then lays out Microsoft's plan for a 'frontier ecosystem' where every company trains and owns its own models on its own data.
Nadella's core worry is what happens to the firm when models learn from data. If a company just consumes a foundation model, it leaks its value away. His answer is the 'hill climbing machine': give every company a licensed model plus the tooling to train it on their own tasks, evals, and traces, so their private IP compounds instead of leaking. The ecosystem only survives if it is positive-sum, meaning many participants operate at the frontier, not a few firms capturing all the returns.
The OpenAI bet and prepared mind
Microsoft put $1 billion into OpenAI in 2019 because of what Nadella calls a 'prepared mind': the company had been obsessed with natural language for years and would back anyone with an ambitious angle on it, inside or out. The scaling laws paper convinced them that pushing the transformer with more compute and data was worth a shot. The biggest internal decision was not capital but compute concentration on one effort.
The hill climbing machine
Microsoft shipped seven new models trained on clean-lineage data with no synthetic data mixed in, so reasoning emerged honestly. The point is to license the weights so any company can set up its own 'hill climbing machine': its own RL environment, private evals, and traces, then welcome any model into that gym. This lets a firm keep its IP instead of handing enterprise value to a foundation model.
Agents you can contain
Copilot evolved from chat to co-work (delegated multi-step tasks) to autopilot via a product called Scout, a long-running enterprise agent that runs continuously with your delegated identity. Because these agents generate and execute code, containment matters. Windows will ship a sandbox container (MXC) with process, session, or VM isolation boundaries, so agents can be governed like processes were in the past.
Local compute and new form factors
Nadella pushes 'unmetered intelligence': tap the large install base of PC GPUs so agents run on-device when cloud tokens are scarce. A new dev box built on Nvidia's RTX will pack a petaflop of AI compute, 20 CPU cores, and 128GB unified memory, enough to run a trillion-parameter model locally. Project Solara explores agent-era hardware like a fingerprint-and-camera badge that wakes Copilot and runs tasks in the cloud.
Rethinking the silicon stack
Three new workloads (training, inference, and long-running agents) let Microsoft rethink chips from first principles. Maia 200 is co-designed with Microsoft and OpenAI models and already runs GPT-5.5 in production for a TCO advantage; the Cobalt ARM CPU is tuned on GitHub Copilot traces for agent loops. Microsoft runs a heterogeneous fleet, even using old GPUs to give its Fabric data warehouse a 7x speedup.
Electricity to light
Nadella borrows a metaphor from an earlier speaker: with electricity we sold light, not electricity, and AI still lacks that clear benefit story. He argues the industry has over-hyped the tech for its own sake, and the world will judge AI by real value delivered one community at a time, in healthcare costs or economic opportunity. If returns concentrate in a few firms, the industry loses its social permission to operate.
- Microsoft bet $1B on OpenAI in 2019 because of a 'prepared mind' from years of obsession with natural language, not a single lucky call.
- The scaling laws paper (more compute plus data on the transformer) is what made the bet appealing, and the capability curve has held.
- A company that only consumes a foundation model cannot retain enterprise value; its IP leaks into the model.
- The 'hill climbing machine' gives every firm a licensed model plus its own RL environment, evals, and traces to compound private IP.
- Long-running agents like Scout need containment; Windows ships an MXC sandbox with process, session, or VM isolation boundaries.
- New dev-box hardware can run a trillion-parameter model locally with a petaflop of compute and 128GB unified memory.
- Maia 200, co-designed with Microsoft and OpenAI models, already runs GPT-5.5 in production for a total-cost-of-ownership edge.
- Nadella frames Microsoft's own culture change through growth mindset: the work is confronting your own fixed mindset, not sloganeering.
In their words
“If you have a model that basically learns from data, what's the future of the firm even?”
“It's not about talking about growth mindset, it's about having the courage to confront one's own fixed mindset.”
“You will absolutely lose social permission, or we will lose social permission.”
Terms to know
- Prepared mind
- Nadella's phrase for being conditioned by years of prior work to recognize and act on a breakthrough when it arrives.
- Hill climbing machine
- A setup where a company trains a licensed model on its own tasks, evals, and traces so its IP compounds instead of leaking.
- Scaling laws
- The finding that model capability keeps improving predictably as you add more compute and data to the transformer.
- Unmetered intelligence
- Running AI models on local PC and edge GPUs so agents keep working without consuming scarce cloud tokens.
- MXC container
- A Windows sandbox that isolates a running agent at the process, session, or VM level so its code execution can be governed.
- Positive-sum ecosystem
- A market where many companies can operate at the AI frontier and keep value, not one where a few firms capture all returns.
Satya Nadella at Stanford CS 153: Frontier Systems
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.