Goldman Sachs's theory that agentic AI will create a profitable 'flywheel' effect is now gaining significant support from real-world data.
The plausibility of this 'agentic margin flywheel' comes from a fundamental shift in how AI is used. Instead of just responding to single prompts, agentic AI performs complex, multi-step tasks, often running continuously in the background. This dramatically increases token consumption. For instance, OpenAI has confirmed that its models generate hidden 'reasoning tokens' for internal thought processes, which are billed to the user as output. This design means that as AI becomes smarter and more autonomous, it naturally generates more revenue per task.
Simultaneously, the cost to generate these tokens is plummeting. First, AI model providers like OpenAI have been aggressively cutting prices. More importantly, hardware is undergoing a revolutionary improvement. NVIDIA’s Blackwell platform already offered a major cost reduction, and its upcoming Rubin platform, expected in late 2026, promises to lower the cost per token by up to ten times more. This sharp decline in unit costs makes it economically feasible to deploy ever more powerful and token-hungry agents.
This dynamic is no longer just a theory; it's appearing in the financial reports of major tech companies. Alphabet (Google) recently reported that its API token processing surged 60% in a single quarter while it simultaneously cut the cost of its AI-powered Search responses by over 30%. Microsoft's AI business is growing at a triple-digit rate, and Amazon's Bedrock platform processed more tokens last quarter than in all its previous years combined. These are clear signs that the massive capex is starting to be utilized effectively.
This evidence changes the entire narrative around AI investment. The debate is quickly moving from "Can companies afford this massive AI build-out?" to "How quickly will falling costs and rising usage translate into higher profits?" The foundation has been laid with huge infrastructure spending, and now the focus is on the returns, which seem to be arriving sooner than many expected.
- Glossary
- Agentic AI: AI systems that can proactively and autonomously pursue goals, perform multi-step tasks, and use tools without constant human instruction.
- Capex (Capital Expenditure): Funds used by a company to acquire, upgrade, and maintain physical assets such as property, plants, buildings, technology, or equipment.
- Token: The basic units of data (like words or parts of words) that AI language models process. Both the input (prompt) and output (response) are measured in tokens.