The companies that use AI the most are growing the fastest—but here is why you shouldn't trust 'token consumption'
In 2025, a phenomenon called "tokenmaxxing" is spreading among tech companies in Silicon Valley. Engineers are competing over AI token consumption, and that figure is becoming a de facto "performance metric."
NVIDIA CEO Jensen Huang has said, "I worry if a great engineer isn't using $250,000 worth of AI compute resources a year," and Shopify has revealed that it uses token usage as a performance signal. Meta employees have also been reported to consume an estimated 900 million tokens per month.
However, whether all this consumption is creating value is another story. Based on a discussion on a CNBC livestream between Ramp CEO Eric Glyman and investor Dan Niles, we will break down the essence of AI demand and the points investors should watch.
1. What is tokenmaxxing—and why "consumption does not equal value"
1-1. Goodhart's Law is at work again
The essence of tokenmaxxing lies in a culture that views AI usage as "the more, the better" and forces internal competition over those numbers. Eric Glyman points out that this is a classic example of Goodhart's Law.
"When a measure becomes a target, it ceases to be a good measure."
This principle has been proven repeatedly in the past. Once, when Amazon evaluated call center representatives based on "call brevity," the representatives began unilaterally hanging up on customers. The numbers improved, but customer service collapsed. Jeff Bezos immediately abolished this metric.
The same thing can happen with token consumption. If engineers consume tokens for purposes unrelated to actual work, such as calculating digits of pi endlessly, the metric will rise, but no business value will be created. This is like evaluating traders based on "number of trades," which is unrelated to profitability.
1-2. Risks lurking in over $1 trillion of infrastructure investment
This problem is serious because token consumption is being incorporated into investment decisions as a signal that "AI demand is real." Currently, investment commitments to AI-related infrastructure are said to have reached over $1 trillion. If some of that demand is "inflated," it becomes a risk that investors cannot ignore.
2. The "K-shaped economy" shown by Ramp's data—the relationship between AI spending and corporate growth
2-1. AI spending has surged 13-fold, but budget management is not keeping up
According to Ramp's data, AI-related token spending has increased 13-fold over the past year, growing at a pace of 50% per quarter. Glyman frankly admits, "No one knows how to budget for this spending."
What the company has newly launched is "token tracking"—or more accurately, an AI spend management tool. CFOs and finance teams can drill down and visualize which departments are using which models for what tasks and how much they are spending. For example, if there is a case where a cutting-edge frontier model is being used for email editing, it can suggest that a cheaper model is sufficient, showing a path to cost reduction.
Glyman compared this to cars. "You don't need to use a Ferrari to deliver groceries. A Prius is fine, or maybe even a bicycle."
2-2. K-shaped economy—the growing growth gap between the top and bottom of AI investment
The analysis Ramp derived from data on over 50,000 companies is noteworthy. The top 25% of companies in AI spending have seen their revenue more than double over the past three years. Meanwhile, the bottom 25% of companies grew by only about 12% over the same period, remaining almost flat at 2-5% on an annualized basis.
Moreover, the growth rate of the top tier is accelerating year by year. A positive cycle is beginning to turn where companies that advance AI adoption grow faster, hire more, and invest more. This K-shaped polarization suggests that AI is becoming a source of substantial competitive advantage rather than just a fad.
3. OpenAI vs Anthropic—the essence of demand reflected by opposite strategies
3-1. OpenAI: Lowering prices to pursue volume
OpenAI is steering toward making AI cheaper and easier to use. It is a strategy of having more users consume it and using that track record of usage to justify massive expenditures. Dan Niles points out that OpenAI is lowering prices to acquire customers.
3-2. Anthropic: Setting limits to gauge the quality of demand
In contrast, Anthropic is setting caps on usage, restricting third-party access, and raising prices. It is in a state of "demand is too high, so we are suppressing it," and it is carefully trying to determine whether the demand it is seeing is real.
Griman described Anthropic's CFO, Krishna, as a "very proven financial leader," and stated that the company's approach is "closer to a Wall Street mentality." Promising less and delivering more than expected—he says this is a common stance among companies that survive in the long term.
3-3. OpenAI's "Innovator's Dilemma"
What was particularly impressive during the discussion was the moment Griman mentioned the possibility that OpenAI is facing the innovator's dilemma.
"When your business model depends on extracting maximum spending, can you pivot to efficiency? Do you even want to?"
At one time, Google was questioned about the innovator's dilemma as the king of search advertising. Now, OpenAI is facing that same question as an "incumbent"—the reversal of positions in just over two years symbolizes the speed of change in this space.
4. Jevons Paradox and the learning curve—is "waste" really waste?
4-1. Tennis strokes and token consumption
Griman does not necessarily view "waste" in the early stages as negative.
"When you play tennis for the first time, your strokes are a mess. You don't hit the ball. But after 20 or 30 matches, efficient neural pathways are formed, and you can place the ball where you aim."
Much like Mark Twain's famous quote (though it is unclear if he actually said it), "I would have written a shorter letter if I had more time," trial and error is essential for refinement. Shopify CEO Tobias Lütke expressed a similar view at the HumanX conference, stating that a certain amount of "waste" occurs structurally in the process of becoming proficient with new technology.
4-2. Jevons Paradox—lower costs create further consumption
The unit price of tokens is falling rapidly. However, lower costs do not necessarily mean less consumption. The Jevons Paradox in economics is at work, where cheaper resources actually induce mass consumption.
Griman believes that "if the cost curve continues to fall and users become more efficient at using it, spending is highly likely to increase even further." In other words, even if "waste" decreases, total demand will continue to expand—this is a positive signal for AI infrastructure investment, but it also means that it is becoming increasingly important to identify which companies are growing with ROI.
5. The game-changing rules of Agentic AI
5-1. Open Claw and the surge in token demand
Dan Niles cited the rise of Agentic AI as a factor that significantly changed the AI market in 2025. Since the Clawd Bot, which appeared in November 2025, was formalized as Open Claw at the end of January 2026, token demand has changed dramatically.
According to Open Router data, token growth was about 20% in the two months before the formalization of OpenClaw, whereas it surged to about 130% in the two months after. Agents do not just repeat a single task; they cross-functionally handle multiple different tasks, such as retrieving data from CNBC, referencing SEC financial documents, and entering them into Excel. This is structurally driving up token consumption.
5-2. From GPU to CPU—The semiconductor shift in the agent era
Another significant change pointed out by Mr. Niles is the shift in semiconductor demand structure brought about by agent AI.
Over the past three years, the value of AI has been concentrated in GPUs (which are strong at parallel computing, repeating the same processing in large volumes). However, agents need to orchestrate diverse tasks, and this is where CPUs (which are strong at general-purpose processing) come into play.
According to Mr. Niles, while the conventional GPU-to-CPU power ratio was 7:1, it could shift to around 4:1 in an agent environment. This creates room for re-evaluation of companies like Intel, which were considered 'finished.' In fact, signs are already appearing, such as the partnership between Google and Intel and Elon Musk mentioning collaboration with Intel.
6. 'Cash flow' and 'valuation' that investors should watch closely
6-1. OpenAI to burn $22 billion in cash through 2029
The point Mr. Niles sounded the most alarm about is OpenAI's financial structure. The company has admitted that it will burn $22 billion in cash by 2029, and it is expected to become profitable by 2030 at the earliest. Furthermore, there are reports that OpenAI CFO Sarah Friar has not been present at some meetings, casting doubt on the feasibility of an IPO.
'If OpenAI cannot raise more funds within the year, it could head toward zero,' Mr. Niles stated. His view is that if an IPO does not materialize, raising funds at the desired valuation will be difficult.
6-2. The lesson of Amazon's 95% drop—Being right doesn't guarantee survival
Mr. Niles cites the example of Amazon during the dot-com bubble. Amazon, which had $1.6 billion in sales in 1999, saw that figure nearly double to $3.1 billion in 2001. However, during that time, its stock price fell 95% from its peak. Even if you choose the right company, whether it can survive is a different matter.
Meanwhile, Google continues to maintain a significant positive free cash flow while advancing its AI investments, putting it in a position to organically fund its own ambitions. Microsoft, holding 27% of OpenAI, is directly exposed to OpenAI's risks and has recorded a decline of approximately 20% year-to-date.
6-3. Mr. Niles' stocks to watch
The key points of the investment strategy mentioned by Mr. Niles are as follows.
Amazon: Hosts Anthropic and directly benefits from AI through physical infrastructure (robotics and logistics). Its position as the largest public cloud vendor is also strong.
Apple: Being late to AI development could conversely become a weapon. While it has a massive installed base of 1.5 billion iPhones, its AI development costs can be kept low through licensing to Google (approximately $1 billion annually).
Google: Abundant cash flow and a strong position in consumer AI.
Intel: Potential for re-evaluation due to increased CPU demand in the agent AI era.
Cautionary targets: Microsoft and Oracle, which have large exposure related to OpenAI.
7. Summary—Demand is real, but the sorting of 'noise' begins
Token mixing is a phenomenon that naturally occurs in the process of AI permeating companies, and it is not necessarily a bad thing. As Ramp's data shows, top companies in AI spending are indeed achieving high growth.
However, not all token consumption generates value. As Goodhart's Law suggests, the moment consumption becomes a target, the metric loses its reliability. For investors, the following three points are critical.
ROI, Not Volume: Look at how token consumption translates into revenue and productivity, rather than the consumption volume itself.
Sustainability of Cash Flow: The difference between OpenAI burning $22 billion and Google, which can fund its AI investments with its own capital, is decisive.
Structural Changes in the Agent AI Era: Prepare for a shift in demand from GPU-centricity to CPU, memory, and orchestration.
AI demand itself is real. However, the ability to distinguish between companies that truly generate value and those supported by the illusion of metrics will be increasingly demanded in future investment decisions.
