Tech • IA • Crypto
No one realizes what AI will really cost us
8/10TL;DR
Rising AI usage costs are forcing companies to ration access while a new concept of “token capital” emerges as a potential competitive advantage tied to compute, data, and learning loops.
KEY POINTS
AI Costs Trigger Corporate Rationing
Major firms including Amazon, Walmart, Uber, and Meta are increasingly limiting internal AI usage as costs surge. The core expense driver is token consumption, the unit used to measure AI processing. What began as experimental adoption is now treated as a controlled operational cost, similar to cloud spending.
The Rise of “Token Capital”
Microsoft CEO Satya Nadella has promoted the idea that companies must build “token capital,” defined as their capacity to generate, use, and retain AI-driven intelligence. This includes proprietary data, internal learning systems, and accumulated knowledge. The concept frames AI not as a tool, but as a core economic asset necessary for long-term survival.
From Models to Learning Loops
AI models themselves are becoming commoditized, accessible to all कंपनies. Competitive advantage is shifting toward what sits above the model: private evaluation systems, custom training environments, and persistent memory. Firms that convert AI usage into proprietary learning loops retain value even if they switch underlying models.
Exploding Demand for Tokens
Goldman Sachs estimates global AI usage could reach 120 quadrillion tokens per month by 2030, roughly 24 times current levels. Each prompt, response, and reasoning step consumes tokens, making them the fundamental economic unit of AI. Despite falling unit costs, total consumption—and therefore total spending—is rapidly increasing.
Inference Becomes the Main Cost Driver
Historically, training AI models was the primary expense. Today, ongoing usage—known as inference—is becoming dominant. Gartner estimates inference could account for 70% of total lifecycle costs. This shift reflects more complex models that “think” longer and the rise of autonomous agents operating continuously.
The Rebound Effect in AI Economics
Token prices are falling by 60–70% annually, but overall costs continue to rise due to increased usage. This reflects the Jevons paradox, where cheaper resources drive higher consumption. Agent-based systems can require 5 to 30 times more tokens per task than traditional chatbots, amplifying the effect.
Massive Infrastructure Investments
Hyperscalers are investing heavily to control AI infrastructure, with spending rising from $200 billion in 2024 to projections of $700 billion. However, markets show चिंता: Meta’s stock dropped ~10% after raising AI capex forecasts, and Amazon’s free cash flow is heavily consumed by data center expansion.
Uneven Returns and Strategic Risk
Only 28% of AI infrastructure projects in the U.S. fully meet ROI targets. Meanwhile, 74% of AI-driven economic value is captured by just 20% of companies, indicating strong concentration. Firms that fail to integrate AI into core workflows risk turning it into a cost center rather than a value driver.
Dependence on Compute and Energy
AI systems effectively convert electricity into tokens, making compute infrastructure a critical bottleneck. Control over chips, data centers, and energy supply is increasingly concentrated among a few players, raising concerns about access and economic sovereignty.
Uncertain Long-Term Scaling
Some researchers question whether simply increasing compute will continue to yield major gains. Alternative approaches, such as more efficient architectures or reasoning techniques, may shift the economics. However, current trends still favor increasing compute intensity, especially at inference.
CONCLUSION
AI is evolving into a cost-intensive, infrastructure-dependent system where competitive advantage depends less on access to models and more on the ability to convert token usage into measurable economic returns.