AI ‘Tokenomics’ Debate Intensifies as Rising Usage Costs Outpace Measured Value

As Big Tech and startups race to embed generative AI into products and workflows, a more sober question is increasingly dominating boardrooms: not whether AI works, but whether it pays. What was once framed as a story of rapid capability gains is now becoming a debate about 'token economics'—and in this case, tokens as an operating cost that can quietly burn through budgets faster than productivity can catch up.

The shift was thrown into sharp relief after Andrew Macdonald, president and chief operating officer of Uber ($UBER), warned in a recent podcast appearance that AI spending is becoming harder to justify. Macdonald said usage of AI tools across engineering teams has surged, but clear proof that the spend is translating into better products or sustained productivity gains remains thin. He described the phenomenon as 'tokenmaxxing'—a kind of token overconsumption where companies pay for more model output without reliably getting more business value back.

The remarks drew attention precisely because they did not come from an AI skeptic. Uber has been among the large global platforms eager to operationalize AI in software development and internal processes. The timing also added fuel to the debate: they followed reports that Microsoft ($MSFT) had recently scaled back internal access to Anthropic’s Claude Code for some employees, citing rising usage-based charges, and nudged engineers toward GitHub Copilot as a more controllable alternative.

At the heart of the conversation is a new definition of 'tokenomics.' In crypto, tokenomics typically refers to how tokens are issued, distributed, and potentially burned. In AI, tokenomics has come to mean something more mundane and potentially more painful: billing mechanics. Every query, code generation, and agentic workflow consumes tokens, turning experimentation into a metered expense line—often before organizations can measure the return.

Several data points illustrate the intensity of the cost pressure. Uber reportedly exhausted its 2026 budget for Claude Code in roughly four months. Around 5,000 engineers used the tool, with monthly adoption rates cited in the 84% to 95% range. Internal billing per engineer was said to range from roughly $150 per month to as high as $2,000, depending on usage. One internal demonstration by the company’s chief technology officer reportedly consumed about $1,200 worth of tokens in just two hours—an anecdote that has circulated as a stark example of how quickly costs can compound when advanced tools are used at scale.

Microsoft’s experience has been similar, according to reporting by The Verge. Starting in mid-May, the company began reducing internal Claude Code access in phases as usage climbed and pay-as-you-go charges accelerated. GitHub, meanwhile, is set to adjust Copilot pricing beginning June 1, moving from flat-rate plans toward usage-based billing—an acknowledgment that fixed-price subscriptions are difficult to sustain as agentic coding sessions become more resource-intensive. Industry estimates cited in the debate suggest that a single agent-style coding session can cost $30 to $40 in compute, putting pressure on the economics of low monthly price points.

The most pointed critique comes from Entelligence.AI, a developer productivity platform that analyzed 2,444 companies. The firm concluded that for every $1 spent on AI tokens, only $0.18 translated into realized user value. The remaining $0.82 was attributed to downstream costs: roughly $0.44 on fixing AI-generated bugs, $0.27 on rework, and $0.11 on review processes. The implication is not merely that AI is expensive, but that in many environments it may be shifting effort rather than removing it—creating new categories of work that must be managed and audited.

Interpretations of these numbers diverge sharply. Optimists argue today’s cost turbulence represents a transitional phase. As agentic AI becomes more common, token consumption could rise dramatically, expanding revenue pools for model providers and cloud infrastructure firms. Some analysts—including voices at Goldman Sachs—have suggested token usage could grow by multiples of today’s levels by 2030. In that view, improved pricing models, better orchestration layers that route tasks to the most cost-effective models, and maturing tooling for caching and reuse will reduce waste and bring expenditure into line with outcomes.

Skeptics see structural fragility instead. They argue that the most consistent beneficiaries of the generative AI boom so far have been compute and semiconductor vendors—particularly Nvidia ($NVDA)—while model developers and hyperscalers shoulder large capital expenditures and operational costs to keep pace. The concern extends beyond raw spending to the 'quality of revenue' in the AI supply chain. A feedback loop is increasingly discussed in which AI model companies receive investment from cloud providers and then channel much of that capital back into cloud commitments and usage fees. That structure can inflate topline figures and backlog visibility, critics warn, without proving the depth of independent, self-sustaining demand.

This concern has been sharpened by the growing share of major cloud commitments reportedly tied to a small number of leading AI labs. If one or two model developers account for a large fraction of a cloud provider’s AI backlog, the relationship can look less like diversified commercial traction and more like strategic dependence—raising questions about durability if pricing, regulation, or competitive dynamics shift.

The debate is also spreading beyond the U.S. In South Korea, major internet platforms are making divergent bets that reflect the same underlying dilemma: whether controlling infrastructure is a long-term advantage, or whether mixing external and internal models is a better way to keep costs predictable. Naver, for example, has been increasing AI infrastructure investment aggressively, with capital expenditures in the first quarter reported to have more than doubled year-on-year, much of it for AI servers and GPUs. Kakao, by contrast, has emphasized combining third-party and in-house models to improve cost efficiency. Different strategies, same question: who ultimately controls the unit economics of AI output.

For smaller startups and systems integrators with tighter cash positions, token costs can quickly become existential. That reality is driving interest in practical cost controls such as pairing low-cost models with high-performance models, prompt caching, building on-premises large language models for repeatable workloads, and exploring domestically produced accelerators where available. Policy efforts around sovereign AI and domestic cloud capacity are also being reframed less as a matter of technological independence and more as a matter of 'cost sovereignty'—the ability to avoid business models that are overly exposed to foreign token pricing and currency swings.

Market observers caution against simplistic comparisons to the dot-com bubble. Unlike many early internet experiments, AI systems already deliver tangible capability improvements in certain tasks, and leading model developers are unlikely to disappear overnight. Still, the investment lens is changing. Where the market recently focused on how quickly AI adoption was spreading, it is increasingly interrogating whether that adoption is happening in a profitable, repeatable way.

Ultimately, the emerging consensus is that raw token consumption is not a success metric on its own. What matters is the conversion rate from spend to verifiable value. If organizations are effectively burning $1 to capture $0.18 in durable benefit, finance teams will eventually demand a reset—whether through tighter governance, cheaper models, better workflow design, or renegotiated pricing. The next phase of the AI race may be less about building the smartest model and more about delivering the most reliable outcomes at the lowest unit cost, turning 'tokenmaxxing' from a punchy phrase into a problem the industry will be forced to solve.