Ali Ghodsi on the Economics of the AI Supercycle

AGI is Already Here, but Context is Missing

Artificial General Intelligence (AGI) has already been achieved, but it remains largely useless in an enterprise setting because it lacks the specific organizational context that humans possess. While models can solve complex mathematical problems, they cannot replicate the institutional knowledge held by long-term employees—the "John or Jane" of a company who knows how everything actually works.

For AI to have a massive impact, the focus must shift from pursuing "super intelligence" to figuring out how to transfer human context and organizational processes into AI agents. Without this "downloading of the brain into silicon," models will continue to make mistakes that render them ineffective for high-value enterprise tasks.

The "SaaS Apocalypse" and the Future of Software

Software is not "dead," but the economic moats protecting software companies have shifted. Two primary changes are driving this evolution:

1. Lowered Barriers to Entry: AI has made it significantly cheaper and faster to write software, reducing the cost of production toward zero.
2. Reduced Switching Costs: As users move from complex UIs to interacting with AI agents, the inertia that previously locked users into a specific software ecosystem (e.g., Android vs. iOS) is eliminated.

Despite these changes, software companies can still maintain moats through economies of scale, brand trust, patents, and proprietary data. Companies that have failed to innovate for a decade are at high risk of being wiped out by new players who can leverage AI to build superior products quickly. Those that continue to innovate and possess deep customer data will likely survive and thrive.

The Productivity Gap and Process Refactoring

There is a significant gap between the capabilities of AI and the actual productivity gains seen in organizations. This is not a failure of the AI, but a failure of human process. Ali Ghodsi compares this to the introduction of the PC and the electric engine:

• The PC Paradox: Early adopters used PCs as expensive typewriters, printing sheets and filing them manually, which resulted in no immediate productivity gain.
• The Electric Engine: It took 40 years (from 1880 to 1920) for the electric engine to impact economic productivity because factories had to be completely redesigned—moving away from dense, steam-powered line shafts to distributed power layouts.

Similarly, modern enterprises are trying to "replace the steam engine with an electric engine" without changing their floor plan. Real productivity gains require "human refactoring"—rewiring organizational processes from first principles.

Case Study: Connector Development at Databricks

At Databricks, building a production-ready connector traditionally took three quarters (nine months). When AI was introduced, the initial attempt to optimize the process only reduced the timeline by 1.5 months because the team maintained the same rigid process (long requirement phases, sequential testing, and single-person ownership).

By applying first-principles thinking and refactoring the process, the team achieved a breakthrough: shipping seven connectors in one quarter. This was achieved by:

• Reducing the requirement phase from a quarter to one week and iterating faster.
• Outsourcing the setup of external software instances to parallelize testing.
• Moving from a "bus factor of one" (one person per connector) to a collaborative team approach.

This improvement was a result of process change, not a smarter AI model.

Value Accrual in the AI Stack

In the long term, value will move up the stack toward applications. While the current "blue triangle" of value is concentrated in hardware (Nvidia) and infrastructure, historical patterns in tech (from IBM to Microsoft to VMware) suggest that value eventually commoditizes at the bottom and accrues to the top.

Predictions for High-Value AI Applications

• Healthcare: A company capable of analyzing millions of patients' genetic and medical histories to provide personalized, life-saving interventions could be worth trillions.
• Education: Despite a VC consensus that education is a poor investment, a company that provides proven, AI-driven superior education could achieve massive scale and data moats.

The Role of Frontier Models

Proprietary frontier models will remain valuable, but the business of providing them will become an economies-of-scale game with thin margins, similar to Amazon's book-selling business. Open-source models are rapidly closing the gap with proprietary ones, applying further pricing pressure on the model layer.