I keep coming back to the same question. Who is going to pay for all this? In The AI Bubble: Why This Time Isn’t Different (El-Masri, 2025), I questioned whether AI investment could be justified. In The AI Bubble Part II: Regulatory Charades, Cloud Rents, and the Missing ROI, I challenged the tokenomics behind AI (El-Masri, 2026). This article examines the viability of the AI tokenomics by asking basic questions: Who owes what, who can pay, and what happens when someone cannot? By tokenomics, I mean the economics of supplying AI output: what users pay, what that output costs to produce, and who covers the difference.
Follow the Money… Until it Dries
The AI economy is built on clients’ readiness to pay a model company. Those model companies rent computing capacity from cloud providers. Cloud providers buy chips from chip makers. Investors finance the expansion, and in the case of AI, they often invest in the very customers who buy their products. Revenue rises at every step. The question is how much comes from customers who can sustain the spending.
In my most recent article (The AI Bubble Part II), I showed how many AI partnerships operate as a financial round-trip: Hyperscalers (like Amazon, Google, and Microsoft) alongside Neoclouds (like CoreWeave and Oracle) invest billions of dollars into a model developer like OpenAI or Anthropic on the condition that those same developers spend much of that money right back on their servers (Anthropic, PBC, 2026; Elmasri, 2025). In other words, the investor’s outgoing cash returns to its own books as cloud revenue (Damodaran, 2026). Money invested in customers returns as supplier revenue (Federal Trade Commission, 2025), helping the AI industry report extraordinary growth before its consumers demonstrate meaningful returns.
Artificially cheap AI subscriptions make the loop much easier to perpetuate. They encourage usage while deferring the uncomfortable reality of AI's real cost. A prolific user at $20 a month may behave differently when charged for every lengthy reasoning task. On a $20-a-month service, a user can burn around $400 worth of tokens with third-party platforms like Perplexity or Cursor, allowing users to burn $ 100 worth of Anthropic tokens for just $20 a month (Zitron, 2026). According to a SemiAnalysis report (Patel & Nishball, 2024), a $200-a-month ChatGPT subscription lets a user burn through $14,000 worth of tokens, and a $200-a-month Claude subscription can cost the company $8,000 in token usage (Zitron, 2026), leaving the AI developer to subsidize the remaining $900. One only needs to look at isaiprofitable.com to observe the magnitude of losses that those companies are incurring.
A viable business must be able to (1) cover computing and support costs, (2) retain customers as pricing changes, and (3) deliver savings. Instead, the industry is geared towards chasing subscriptions and churning out ever-larger models. Meanwhile, AI companies have swapped annual recurring revenue (ARR), which relies on active, recurring contracts, for experimental run-rate (ERR) to measure growth and justify inflated valuations (The Economist, 2025). Some companies, like Anthropic, reported annual revenue as a peak month multiplied by 12, ignoring the high churn rate in the AI industry. Others like OpenAI, Perplexity, ElevenLabs, and Anysphere (creator of Cursor) did the same. My most recent peer-reviewed publication in Management Review Quarterly (2026) analyzed 75 studies and documented the various approaches to using recurring revenue for valuation (El Masri et al., 2026). Annual recurring revenue is identified as a valuation method that measures longer-term contractual commitments.
What it takes to make AI output usable
Ultimately, the AI industry’s survival hinges on its ability to help people accomplish their work more efficiently. Much of the promised demand rests on the assumption that AI will make employees dramatically more productive (Merchant, 2026). However, deploying generative software into enterprise workflows is proving far messier than originally envisioned. AI can produce a polished answer before anyone has checked its sources, tested its code, or understood its consequences. Adopters save at the production stage but could spend more on validating and repairing AI-generated outputs. In the courses I teach, I allow my students to use AI. My policy is simple: “Use it. Verify it. Say that you used it.” For each course, I explain where AI is allowed, where it requires caution, and where it cannot be used. Yet I find myself spending more time checking whether students have verified and understood the output, and whether they have considered what might be missing. Grading now involves assessing both the submitted work and their exercised judgment in using AI.
A similar burden is showing up in the workplace (Wolff, 2026). In a survey of 1,150 US desk workers, resolving an incident of low-quality AI work took about two hours on average. This measures cleanup, not a universal cost ratio, but it exposes the work hidden behind the output (BetterUp Labs & Stanford Social Media Lab, 2025). A Controlled trial with experienced software developers found that AI slowed them down (METR, 2025). As a result, some employers are already rebuilding the capacity they cut. In a 2026 survey of 275 service and support leaders, 7% reported rehiring human workers after earlier reductions, and 11% planned to rehire within 18 months. Indirect costs of recruitment and training, as well as restoring lost experience are often overlooked. The broader lesson is that anticipated AI savings are a poor basis for removing expertise before the technology proves it can carry the workload (Gartner, 2026; Merchant, 2026).
The need for human expertise also extends to the data used to train these systems. Bot traffic is outpacing human internet traffic, and we increasingly use AI to complete our work. As such, the supply of fresh, publicly available human data is becoming harder to expand. Attempting to bypass this bottleneck by using AI to train AI creates a dangerous feedback loop. The AI-generated outputs re-enter the model training pool, and each generation of models learns from a slightly narrower, smoothed version of the original distribution. The literature calls this model collapse (Shumailov et al., 2024), where rare cases in the tails vanish first, then overall diversity degrades. In my Data Science class, I explain "data contamination" or "data leakage" as data used to build the model leaking into the data used to validate it, which inflates evaluation results. Eventually, the data models are trained on will be unfit, leading to model collapse. As Defined.ai CEO Daniela Braga has observed, preventing this degradation means pivoting away from the open data of the web toward highly curated, private data sourced from verified human specialists, such as doctors or mathematicians. Maintaining model quality may therefore require more input from human specialists, adding costs that have not yet been considered.
I experienced how AI can meaningfully boost performance in my own work. I willingly pay for personal subscriptions to ChatGPT and Claude alongside my employer-sponsored Gemini license. AI’s impact on productivity depends heavily on the complexity of the work. For routine work such as customer inquiries, productivity increased by 15% (Brynjolfsson et al., 2025). However, at the current rate of AI subsidization, suppliers may struggle to retain enough of that value to make a profit. Intense competition can force suppliers to pass those savings on to customers through lower prices, squeezing margins even as usage grows. That said, emerging approaches such as world models, parallel concept-based generation (Kumbong et al., 2026) (see Stanford research paper on Next Concept Modeling/NCM), and training on smaller sets of carefully curated human data point toward leaner, cheaper systems rather than an endless race to build ever-larger models (Braga, 2026; Kumbong et al., 2026; LeCun, 2022). While cheaper tools would broaden adoption, they create a financial paradox for the industry: falling prices and lower compute requirements could undercut the revenue assumptions behind expensive, multi-year infrastructure contracts before paying demand ever catches up. From a customer’s perspective, cheaper and more efficient models would be welcome. For a supplier that has already committed to expensive infrastructure, however, they could make the expected revenue harder to achieve.
When the bill reaches the rest of the chain
This matters because companies have already committed to infrastructure spending that depends on continued demand growth. Anthropic’s November 2026 filing reports about $900.9 billion in aggregate spending commitments, largely for compute and infrastructure, with $70.7 billion due on September 30, 2027. Much of the total falls in later years, but financing and revenue must still meet the payment schedule (Anthropic, PBC, 2026). The companies supplying the computing capacity face financial pressures of their own. CoreWeave reported $3.661 billion in quarterly revenue and $207 million in operating income, while forecasting $1.05 billion to $1.15 billion in interest expense for the following quarter. These are different periods, but they show why growth alone cannot resolve the financing question (CoreWeave, 2026). What concerns me here is the timing difference. A company can commit to years of infrastructure spending before it knows how much customers will ultimately pay. If demand falls short, those commitments may be difficult or expensive to unwind.
The strain is already visible at the very beginning of the revenue chain: the end-user. Major corporations are drastically curbing AI spending, with companies like Accenture capping employee AI budgets at $29 a month and Apple slashing daily developer allowances from $300 to $25 (Marx, 2026). As enterprise caution grows, the venture capital ecosystem that subsidized the broader AI startup market is also retreating. Because most frontier model revenue relies on these startups burning through subsidized tokens, falling venture investment directly threatens the cash flow of the top AI labs (Zitron, 2026). The resulting squeeze has forced both OpenAI and Anthropic to slash their own valuation targets by hundreds of billions of dollars to secure lifeline funding rounds (Marx, 2026).
As equity funding becomes harder to secure, developers are turning to debt to finance their infrastructure commitments (Dowd, 2026; Muir, 2026). That raises another question: how much will lenders provide, and at what cost? Infrastructure providers are seeing borrowing costs surge to distress levels, with neoclouds like CoreWeave touching 16% on fixed-term loans (Dowd, 2026; Marx, 2026). At those borrowing costs, the economics become difficult to sustain. Revenue must cover interest and operating expenses while leaving enough to replace hardware that is losing value (Dowd, 2026). Private credit and financial institutions are retrenching or becoming highly selective regarding AI infrastructure exposure (Dowd, 2026). A recent Financial Times article reports that Blackstone is warning investors that Anthropic may fail to secure an investment-grade credit rating ahead of its IPO, forcing it into an already strained junk-bond market to fund its obligations (Chan & Hammond, 2026).
Another source of funding may also be weakening. Investment from suppliers that sell these companies chips and computing capacity is weakening. For years, hardware giants like Nvidia acted as the Bank of AI, funneling billions of dollars into customers like OpenAI, Anthropic, and CoreWeave, then recycling that money back through chip sales. Rising market scrutiny is curbing their balance-sheet-as-a-service model, requiring developers to cover more infrastructure spending through customer revenue or other funding. As corporate customers cut spending, startups drop off, and debt becomes prohibitively expensive, a single failure can quickly spread through contracts (Dalio, 2026; Dowd, 2026; Hanke, 2026).
I use these tools and find them valuable, so I understand why customers are willing to pay for them. What remains unclear to me is whether that willingness will translate into enough revenue to support the spending already committed. Some companies will build sustainable businesses; others may need to scale back, restructure, or find buyers. My concern is how much investors, lenders, and customers will have committed before we have a clearer picture of which businesses can cover their costs.
References
- AI valuations are verging on the unhinged. (2025, June 25). The Economist. Read source
- Anthropic, PBC. (2026, November 10). Registration statement, amendment no. 2. SEC accession 000162828026072484. Overview and contractual obligations. Read source
- BetterUp Labs & Stanford Social Media Lab. (2025). Workslop: The hidden cost of AI generated busywork. Read source
- Braga, D. (2026, November 11). The value of reality and the new economy of human data [Interview with G. Perdigão]. Building Creative Machines.
- Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. https://arxiv.org/abs/2304.11771
- Chan, M., & Hammond, G. (2026, October 30). Anthropic may not secure investment-grade rating after IPO, Blackstone tells investors. Financial Times. https://www.ft.com/content/59383c7f-8d82-493f-bf01-d2e862a1018f
- CoreWeave. (2026, November 10). Third quarter 2026 financial results. Read source
- Dalio, R. (2026). Discussion of technology booms and investment bubbles. MasterClass Executive. Read source
- Damodaran, A. (2026, August). AI’s bar mitzvah moment: From hype and hope to business questions. Musings on Markets. Read source
- Dowd, E. (2026, May 25). Discussion of AI infrastructure, power and water constraints. RealClearPolitics. Read source
- El Masri, M., Umlai, M. I., Gül Şenkardeş, İ. Ç., & Sinha, A. K. (2026). Valuation challenges of equity in early stage ventures. Management Review Quarterly. Published April 16. Read source
- Elmasri, M. (a, 2025, October 15). The AI bubble: Why this time isn’t different. My Take on This. Read source
- Elmasri, M. (b, 2026, September 22). The AI Bubble Part II: Regulatory Charades, Cloud Rents, and the Missing ROI. My Take on This Read Part II
- Federal Trade Commission. (2025, January 17). FTC issues staff report on AI partnerships and investments study. Read source
- Gartner. (2026, November 5). Gartner survey finds 21% of service and support leaders have reduced human agent headcount because of AI. Read source
- Hanke, S. (2026, November 3). Financial crisis trigger: What midterm elections will do to markets. The David Lin Report. Read source
- Kumbong, H., Khare, I. S., Mlauzi, I., Woen, A., Ding, C., Buchanan, E. K., Arora, S., Mirhoseini, A., & Ré, C. (2026). Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2610.14162
- LeCun, Y. (2022, June 27). A path towards autonomous machine intelligence (Version 0.9.2). Read source
- Lee, H. P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. CHI 2025, Article 1121, 1–22. Read source
- Marx, P. (2026, October 20). Hyperscale has arrived. Disconnect. Read source
- Merchant, B. (2026, November 3). 20% of AI layoffs have failed. The Tech Report. Read source
- METR. (2025, July 10). Measuring the impact of early 2025 AI on experienced open source developer productivity. Read source
- Muir, M. (2026, November 4). Big tech’s maga alliance is over. The Tech Report. Read source
- NVIDIA. (2025, September 22). OpenAI and NVIDIA announce strategic partnership to deploy 10 gigawatts of NVIDIA systems. Read source
- Patel, D., & Nishball, D. (2024, February 21). Groq inference tokenomics: Speed, but at what cost? SemiAnalysis. Read source
- Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631(8022), 755–759. https://doi.org/10.1038/s41586-024-07566-y
- Wolff, R. (2026, March 31). Economic update: Criticizing pro capitalist ideology. Democracy at Work. Read source
- Zitron, E. (2026, October). Exit liquidity and running dry. Where’s Your Ed At. Read source