AI
Reimagining the AI Development Stack
The old software loop (customer interviews, database schema, code, UX) built a generation of products. That loop is over. Jay Parikh’s Core AI team at Microsoft is replacing it with a stack where models sit at the center: they plan, reason, call tools, and call each other. The team’s scope covers the full developer experience, from model access through Azure AI Foundry to security and trust baked in before deployment, with options from cloud to edge.
The Agent Factory: From Bill Gates’s 1975 Vision to Azure AI Foundry
Parikh was explaining Azure AI Foundry to Bill Gates earlier this year when Gates stopped him and told a story: in 1975, his vision for Microsoft was a “software factory,” a company that hired smart programmers, built many connected products under one roof, and solved problems instead of selling point solutions.
“Microsoft could be that software factory for you, right?” (Jay Parikh)
Fifty years later, the logic holds. Startups Parikh spoke with at Build are already running entire business functions through multi-agent orchestration systems. The agent factory has to support that scale.
Why Multi-Model Support Is a Platform Requirement
Microsoft runs OpenAI, Mistral, and other models side by side on Azure AI Foundry. Parikh’s case for this is simple: it is too early to know which models or use cases will matter most. He pointed to the web as the counterexample — a single-vendor stack would have produced a very different internet. A platform has to offer choice, stay open, and change as the field changes. The use cases nobody has imagined yet are the ones that will require models nobody has committed to yet.
Tokens as the New Currency: Why Every Engineer Must Think About Performance
Parikh’s argument is direct: more capability per unit time, at lower cost, produces more adoption and more use cases. That is why every engineer building AI systems has to care about performance, not just product teams. Microsoft is riding a 10x annual reduction in cost per token. At that rate, applications that are economically unworkable today become viable within a year. The goal is to give users more functionality while spending fewer dollars, and the curve is doing most of the work.
Notable Quotes
speed is all about learning faster Jay Parikh · ▶ 5:20
Microsoft could be that software factory for you, right? Jay Parikh · ▶ 9:28
Because the more capability we can give per unit time Jay Parikh · ▶ 13:27
Key Takeaways
- Core AI at Microsoft is rebuilding the dev stack so models sit at the center, not the database.
- The agent factory pitch: Azure AI Foundry lets companies run entire business functions through multi-agent systems.
- AI cost is dropping 10x per year — that curve is what makes new use cases viable, not just better models.