How Silicon Valley's 'tokenmaxxing' is juicing AI demand — 4/9/2026

Key Concepts

• Token Maxing: A trend where engineers and companies compete to maximize AI compute usage (measured in tokens) as a performance signal.
• Goodhart’s Law: The principle that when a measure becomes a target, it ceases to be a good measure.
• Jevons Paradox: The observation that as technology increases the efficiency with which a resource is used, the total consumption of that resource may increase rather than decrease.
• Agentic AI: AI systems capable of autonomous orchestration, moving beyond simple text generation to performing complex, multi-step tasks.
• K-Shaped Economy: A divergence in business performance where top-quartile AI adopters see significant revenue growth, while bottom-quartile non-adopters remain stagnant.
• Innovator’s Dilemma: The challenge faced by established companies (like OpenAI) where their existing business model may conflict with the need to optimize for efficiency or lower costs.

1. The "Token Maxing" Phenomenon

The video highlights a growing trend in the tech industry known as "token maxing," where employees are incentivized to burn as much AI compute as possible.

• Key Data: Meta employees reportedly consumed 900 million tokens in a single month. Nvidia CEO Jensen Huang suggested that top engineers should be burning $250k/year in compute.
• The Problem: Volume does not equate to value. Similar to Amazon’s historical mistake of grading call center reps on call duration (which led to reps hanging up on customers), incentivizing token usage can lead to "gaming the system" rather than driving actual business impact.

2. AI Spend Management and Efficiency

Eric Glyman, CEO of Ramp, introduced a new product designed to provide CFOs with granular visibility into AI spending.

• Methodology: Ramp tracks spend at an itemized level, identifying which models (e.g., Opus vs. Sonnet) are being used for specific tasks.
• Actionable Insight: Companies often use "frontier" models for simple tasks (like editing emails) when a smaller, cheaper model would suffice. Ramp’s tool helps finance teams identify these inefficiencies and optimize spend.
• The "K-Shaped" Finding: Ramp’s data shows that the top quartile of AI-adopting businesses (across all sectors, including construction and roofing) doubled their revenue over three years, while the bottom quartile saw only 12% growth.

3. Market Perspectives: OpenAI vs. Anthropic

The discussion contrasts the strategies of the two leading AI labs:

• OpenAI: Characterized as aggressive, focusing on massive compute buildouts and scaling capacity. Dan Niles (Niles Investment Management) argues that OpenAI is in a "squeeze" between Google’s consumer dominance and Anthropic’s enterprise success.
• Anthropic: Perceived as more disciplined, often setting usage limits and charging for access. This approach is viewed as more "Wall Street-friendly" and sustainable.
• Financial Risk: Dan Niles notes that OpenAI faces significant cash flow pressure, with an estimated $220 billion burn through 2029. He warns that companies with heavy exposure to OpenAI (like Oracle or Microsoft) carry higher investment risks.

4. The Shift to Agentic AI and Hardware Implications

The transition from simple chatbots to "Agentic" AI is changing the hardware requirements of the industry.

• Orchestration: Agents require the ability to interact with multiple external systems (SEC websites, Excel, email), which shifts the demand from pure GPU power (repetitive tasks) to CPU and memory (orchestration and context retention).
• Hardware Winners: This shift may benefit companies previously considered "dead," such as Intel, as the ratio of CPU to GPU demand evolves.

5. Investment Strategy and Historical Context

Dan Niles emphasizes the importance of discernment in the current AI cycle:

• Historical Analogy: Niles compares the current AI buildout to the 1999 internet bubble. While the technology was revolutionary, many companies (like Webvan) went to zero, and even winners (like Amazon) suffered 95% drawdowns.
• Valuation Matters: Niles advocates for a "margin of safety." He is wary of high-valuation IPOs (like SpaceX or OpenAI) and prefers companies with strong free cash flow, such as Google, or those with massive installed bases, like Apple.
• Significant Quote: "If you focus on a metric, the metric stops being useful." — Dan Niles (referencing the danger of using token counts as a performance KPI).

Synthesis/Conclusion

The AI industry is currently in a phase of massive over-investment and experimentation. While "token maxing" is a symptom of this early-stage fervor, the market is beginning to pivot toward efficiency and ROI. The next phase of AI will be defined by "Agentic" capabilities, which will favor companies that can orchestrate complex tasks efficiently rather than those that simply burn the most compute. Investors are advised to look past the hype and focus on companies with sustainable cash flow and clear, value-driven use cases.