The Correct Economics of AI

A Systemic Analysis of Consumer-Side Value Fracture, Token Supply Mismatch, and Industrial Flywheel Reconstruction

This paper works backward from the consumer side’s capacity to realize value, tracing the structural crisis across the entire AI value chain. It argues that the fundamental contradiction in today’s AI economy lies neither in supply-side technological capability nor in the middle layer of model competition, but in the demand-side breakdown of the value feedback loop. Drawing on semi-annual industry data cross-comparisons from 2023 to 2026, it proposes a “Token Triage” architecture and an “outcome-oriented” management paradigm as flywheel repair pathways.

IThe Duopoly and the Hardware Layer’s Hard Cash

As of April 2026, AI industry competition has converged into a duopoly. Anthropic’s annualized revenue has surpassed $30B, while OpenAI stands at approximately $24–25B. Anthropic doubled from $14B in February to $30B in April over just eight weeks — Meritech Capital analysts noted that after reviewing IPO trajectories of over 200 publicly listed software companies, they had never seen such a growth rate. The critical structural difference lies in customer composition: 80% of Anthropic’s revenue comes from enterprise clients, while approximately 70% of OpenAI’s comes from ChatGPT consumer subscriptions. Enterprise revenue has higher retention, stronger expansion economics, and lower churn.

However, the entities capturing “hard cash” are not the AI model companies but the hardware supply chain.

SK Hynix Q1 revenue was KRW 52.58 trillion ($38.6B), with operating profit of KRW 37.61 trillion ($27.6B). Financial analysts forecast Q2 operating profit to exceed KRW 60 trillion, and Q3 to exceed KRW 70 trillion. The HBM market is projected to grow from $35B in 2025 to $58B in 2026, potentially surpassing $100B by 2028.

IIThree-Layer Funnel Fracture: The Crisis Is at the Consumer End

The industry value chain can be decomposed into a three-layer funnel, with fracture points propagating from the bottom up:

Layer 1 is validated. Layer 2 survives on fundraising — OpenAI completed $122B in funding, Anthropic $30B. The true life-or-death choke point is Layer 3: after users consume tokens, is the output a “digital product” or “digital waste”?

2.1 Structural Inversion of Unit Economics

The entire AI industry burns approximately $400B annually against revenue of only $50–60B. At the micro level: GitHub Copilot loses over $20 per user per month; heavy users incur $80 in compute costs against a $10 subscription; Anthropic users consume approximately $8 in compute resources for every $1 in subscription revenue. OpenAI is projected to lose $14B in 2026 and does not expect positive free cash flow until 2029. Anthropic is projected to reach positive cash flow in 2027 — though training costs through 2030 will still require approximately $30B (OpenAI approximately $125B over the same period).

U.S. consumer spending on AI services totals approximately $12B annually, compared to $527B in infrastructure investment — a ratio of 2.3%. Global AI users exceed 1.35 billion, but the vast majority are free users. Only about 5% of ChatGPT’s 900 million weekly active users convert to paid subscribers.

2.2 Six-Cycle Horizontal Data Comparison

2.3 Evidence of Consumer-Side Retention Collapse

ChatGPT Plus has a 6-month paid retention rate of 71% (industry highest), Claude Pro 62%, Gemini Advanced 60%, and Character.AI only 47%. These figures may appear adequate, but the median net revenue retention (NRR) for AI-native SaaS is only 27–40% — compared to a median NRR of 82% for traditional B2B SaaS. AI products cannot reach even half that benchmark. 41% of consumers are experiencing subscription fatigue, with 44% of cancellations occurring within the first 90 days.

The most fatal signal: OpenAI itself projects that ChatGPT Plus ($20/month) subscribers will plummet 80% from 44 million in 2025 to 9 million in 2026, pivoting to an $8 Go tier targeting 112 million users. Plus growth rate has already decelerated from approximately 3×/year in 2024 to approximately 1.15–1.2× in 2026. ChartMogul has termed this the “AI tourist effect” — users sign up out of curiosity, try the product briefly, then churn.

IIIThe Value Spectrum: Defining “Digital Waste”

“Digital waste” is this paper’s core concept, but it should not be understood as a binary classification. The value of token output is distributed across a four-layer spectrum:

The structural problem of the current industry is that the vast majority of token consumption occurs at the L0 and L1 layers — layers that either cannot be monetized or have monetization paths that are too long and too weak. Only 10–15% reaches the L2 direct substitution layer, and fewer than 3% reach the L3 value creation layer. The Deloitte 2026 report confirms: 66% of enterprises report efficiency and productivity gains (L1–L2), but only 20% have achieved revenue growth (L3), and 74% merely “hope” to grow revenue through AI in the future.

HBR’s April 2026 article precisely named this trap: the “micro-productivity trap” — task-level AI efficiency gains that fail to translate into enterprise-level value. The collapse of the creator economy is the consumer-side mirror case: ChatGPT’s free tier can generate the entire content of a $297 course in 10 minutes, causing course completion rates to drop below 5% and refund rates to climb to 22%. When AI output itself loses scarcity, token consumption at the L0–L1 layer becomes “digital waste” in economic terms — not because the output is useless, but because it cannot sustain the continuity of subscription payments.

IVThe Dual Squeeze on Token Costs

If consumer-side value is already trending toward zero, the trajectory of token costs becomes the life-or-death variable for the flywheel. This paper’s finding is: The true cost of tokens is rising — even though list prices appear unchanged.

4.1 Three Layers of Hidden Cost Inflation

Layer 1: Tokenizer-generation inflation. After Claude Opus 4.7 adopted a new tokenizer, the same text consumes up to 35% more tokens. Price lists remain unchanged, but actual costs rose 12–27%. The critical detail: for medium-length prompts (10K–25K tokens), caching absorbs only 9% of the additional tokens, meaning the price increase is almost fully passed through to users; only prompts exceeding 128K tokens achieve 93% cache absorption. The overwhelming majority of ordinary consumer use cases involve precisely short-to-medium prompts.

Layer 2: Rapidly widening frontier-to-budget price gap. As of May 2026: Gemini Flash-Lite costs $0.10 per million input tokens, while frontier models cost $5.00 — a 50× gap. Over the past three years, small model prices have dropped 99.7%, but frontier model pricing remains firm. Consumers face a brutal choice: using cheap models produces lower-quality output that is even harder to monetize (directly generating L0-layer “digital waste”); using frontier models costs too much to ever recoup.

Layer 3: Transition from subscription to consumption pricing. Cursor replaced fixed request quotas with credit pools in June 2025; the Pro plan at $20/month covers only about 225 Claude Sonnet requests. One developer generated $350 in overages in a single week. Another developer consumed 10 billion tokens over 8 months of daily use, exceeding $15,000 at API pricing ($3/$15 per million tokens). The “buffer” of flat subscriptions is being pulled away.

VTokenmaxxing: The Institutional Disaster of Process-Oriented Management

While the consumer-side flywheel fractures, an even more absurd phenomenon has emerged on the enterprise side — Tokenmaxxing: using token consumption volume as a productivity proxy metric, where higher consumption is equated with higher productivity.

5.1 Cross-Industry Case Matrix

Meta (“Claudeonomics” Incident): A Meta employee created a “Claudeonomics” leaderboard on the company intranet, tracking token consumption across 85,000+ employees and displaying the top 250 power users. In 30 days, the entire company burned through 60 trillion tokens, estimated at approximately $900M at Claude Opus public pricing. The #1 user consumed 281 billion tokens in 30 days — averaging 9.36 billion per day — costing an estimated $1.4M for that single individual. The leaderboard featured gamified titles: “Token Legend,” “Session Immortal,” “Cache Wizard,” “Model Connoisseur.” Some employees ran AI Agents idle for hours solely to climb the rankings. Neither Zuckerberg nor CTO Bosworth made the top 250. The leaderboard was shut down 48 hours after the story leaked, citing “data exposure.” Meta has incorporated “AI-driven impact” as a core metric in its 2026 performance reviews.

Uber (Budget Exhaustion Incident): In December 2025, Uber opened access to Claude Code; adoption surged from 32% to 84% by March 2026 (across 5,000 engineers). 95% of engineers use AI tools monthly, 70% of committed code is AI-generated, and 1,800 AI-written code changes ship weekly. AI-related costs have grown approximately 6× since 2024. An internal leaderboard ranks engineers by AI usage volume. CTO Praveen Neppalli Naga admitted: the $3.4B R&D budget was exhausted in four months — “I have to go back to the drawing board and replan.” Individual engineer monthly API costs range from $500 to $2,000.

Other enterprises: Disney’s streaming technology division deployed a token-tracking dashboard; one employee made 460,000 Claude API calls in 9 working days — 51,000 per day — a rate achievable only with autonomous agents running idle in the background. Visa consumed approximately 1.9 trillion tokens in March, doubling its February volume. Microsoft has maintained a similar internal token dashboard since January 2026, with engineers acknowledging deliberate usage inflation.

5.2 Jensen Huang’s Amplification Effect

At GTC 2026’s All-In Podcast, NVIDIA CEO Jensen Huang proposed that an engineer earning $500K/year should consume no less than $250K in tokens annually; otherwise, he would be “deeply alarmed.” If they spent only $5,000, he would “go ape.” Asked whether NVIDIA’s 42,000 employees are spending a $2B token budget, he said “we’re trying to.” He compared not using AI to “a chip designer saying they want to use paper and pencil.” Tokens are becoming Silicon Valley’s “fourth compensation component” — alongside base salary, bonus, and equity.

5.3 The 10× Token Gap between Senior and Junior Developers

A senior developer crafting precise prompts obtains a production-grade solution with 8,000 tokens; a junior developer pasting 800 lines of chaotic code with “fix this” burns through 80,000 tokens — with worse output quality. Root cause: junior developers lack systematic experience and logical structure; their prompts contain highly disordered relationships, forcing the AI to consume massive tokens on analysis, validation, and trial-and-error. Developers who switch to precise prompting report 30–50% token reduction per task.

5.4 Counter-Evidence from 22,000 Developers

Faros.ai analyzed two years of telemetry data from 22,000 developers across 4,000 teams:

Speed went up; quality collapsed. Tokenmaxxing is a textbook case of Goodhart’s Law: “When a measure becomes a target, it ceases to be a good measure.” Salesforce launched AWU (Agentic Work Units) as a counterweight, measuring output and impact rather than token consumption. Appian’s CEO called Tokenmaxxing “the Soviet Union evaluating chandeliers by weight.” Uber’s lesson ultimately validated the root cause: the team that designed the leaderboard to drive adoption was not the same team managing the AI budget — this organizational disconnect is more lethal than any pricing model.

29% of employees admit to deliberately sabotaging their company’s AI strategy (rising to 44% among Gen Z). 76% of executives view employee resistance as a serious threat. But 75% of executives simultaneously admit their company’s AI strategy is “more about optics than substance.” When executives push AI adoption mandates without genuine strategy, resistance is the logical response.

VIThe Solution: Token Triage Architecture

Based on the analysis above, this paper proposes Token Triage as the core architecture for structural industry repair — restructuring the token supply according to the value attributes of demand, achieving an “oil-water separation.”

6.1 Technical Feasibility and Physical Constraints

Capability side: Sub-billion-parameter models can now handle many practical tasks. Llama 3.2 (1B/3B), Gemma 3 (as low as 270M), Phi-4 mini (3.8B), and Qwen2.5 (0.5B–1.5B) all target efficient on-device deployment. Meta’s ExecuTorch runtime is only 50KB, supporting 12+ hardware backends (Apple, Qualcomm, Arm, MediaTek); over 80% of mainstream Edge LLMs on HuggingFace work out of the box, already serving billions of users across Instagram/WhatsApp/Messenger. Google just launched LiteRT-LM, a production-grade framework specifically for deploying LLMs on edge devices. Over 42% of developers are already running LLMs locally. 4-bit quantization achieves 4× memory compression. Test-time compute allows Llama 3.2 1B to outperform 8B models with search strategy assistance. Gartner predicts that by 2027, organizations will use small task-specific models 3× more frequently than general-purpose LLMs.

Constraint side (absent from V1): Mobile device memory bandwidth runs at 50–90 GB/s, while data center GPUs operate at 2–3 TB/s — a 30–50× gap. LLM inference is memory-bandwidth-bound: every generated token requires streaming the full model weights. Available device RAM is typically under 4GB (shared with the OS and other services), limiting maximum model size. MoE (Mixture of Experts) architectures remain difficult at the edge: although computation is sparse, all experts must still be loaded into memory. This means local SLMs can cover only L0–L1 scenarios and some simple L2 scenarios, and cannot replace frontier models’ deep reasoning — which precisely validates the necessity of triage: let local devices handle what they can, and route requests exceeding their capacity to the cloud.

6.2 Three Fracture Points in Flywheel Repair

Fracture Point 1 (Consumer Payment Collapse): Once social demand is triaged to local devices, consumers no longer need a $20/month subscription for everyday chatting. The retention death spiral of AI-native SaaS with NRR of only 27–40% disappears — because those users should never have been on the cloud in the first place.

Fracture Point 2 (Unit Economics Inversion): Low-value users who consume $8 in compute for every $1 in subscription revenue are triaged to local devices. Only high-value B2B clients remain on the cloud — with high ARPU, well-defined needs, and monetizable output. The unit economics of AI companies instantly improve.

Fracture Point 3 (Training Data Flywheel): Interaction data from B2B users is of far higher quality than consumer-side casual chat. After the oil-water separation, cloud models receive cleaner, more specialized training signals, and model evolution becomes more focused — enabling the positive flywheel to actually start turning.

6.3 Political Economy Resistance (Absent from V1)

Token Triage is the optimal solution for industrial efficiency, but it faces three layers of resistance:

Valuation narrative conflict: ChatGPT’s 900 million weekly active users are the core story behind OpenAI’s $852B valuation. If they acknowledged that most users are “AI tourists” and actively triaged them to local devices, user numbers would plunge and the valuation thesis would collapse. Every AI company has an incentive to keep all users on the cloud, even if serving them is loss-making.

Reverse incentive from hardware suppliers: Jensen Huang’s $250K/person token budget advocacy is the counterexample — GPU suppliers’ interests lie in maximizing cloud token consumption, not in optimizing token allocation. When the shovel seller dominates the narrative, triage proposals are naturally suppressed.

The double-edged sword of open source: Open-source models like Llama, Qwen, and DeepSeek form the technical foundation for consumer-side localization. But Meta’s strategic motive for releasing Llama is to erode competitors’ API revenue — if on-device deployment truly replaces cloud at scale, it would validate Meta’s strategy while undermining the business models of Anthropic and OpenAI. There is an incentive misalignment between the beneficiaries and the promoters of triage.

Who has the incentive to act first? Apple is the natural candidate — its business model is based on device sales, not cloud tokens, and on-device AI directly enhances hardware value propositions. Apple Intelligence is already on this trajectory. Conversely, pure API companies (Anthropic, OpenAI) would need extraordinary strategic courage to proactively triage — it amounts to admitting that a portion of their revenue is “unhealthy.”

VIICounter-Arguments and Responses

An honest thought paper must first present the strongest opposing arguments, then explain why its own framework has greater explanatory power.

7.1 The Optimist’s Three Core Arguments

Argument 1: “Leading companies have already proven ROI.” BCG data shows “Visionary players” achieving 1.7× revenue growth and 3.6× three-year TSR. Companies that have scaled from pilot to production average 1.7× ROI, with supply chain and finance domains seeing 26–31% cost savings. IBM achieved $3.5B in AI-driven cost savings. Coding assistants show 376% three-year ROI with a payback period under 6 months.

Argument 2: “AI companies have real revenue.” JPMorgan argues AI does not meet classic bubble criteria. Jerome Powell distinguishes AI from the dot-com era — current AI companies have real revenue and real customers. Anthropic at $30B ARR, OpenAI at $24B — incomparable to the dot-com era.

Argument 3: “Early stages always look like this — patience is needed.” Standard payback periods for enterprise technology are 7–12 months; AI requires 2–4 years. This is the normal cadence of emerging technology adoption, and premature pessimism is unwarranted.

7.2 This Paper’s Responses

To Argument 1: This paper fully concurs — and this is precisely what proves its core thesis. Every single one of these success cases is concentrated in highly structured B2B scenarios: financial risk management, supply chain optimization, coding assistance, document automation. Without exception, they belong to the L2–L3 layers of the value spectrum. Not a single case demonstrates that L0–L1 layer token consumption can generate ROI. This is exactly the logical foundation of Token Triage — concentrating cloud resources on validated high-value scenarios rather than spreading them thinly across all users.

To Argument 2: AI companies do indeed have real revenue. But 70% of OpenAI’s revenue comes from consumer subscriptions, and that cohort is churning at an 80% rate. 80% of Anthropic’s revenue comes from enterprise clients — which is precisely why Anthropic is overtaking OpenAI: its revenue structure naturally more closely resembles the ideal state post-Token-Triage. Real revenue does not equal sustainable revenue.

To Argument 3: “Patience is needed” presupposes that the flywheel is slowly turning. But when 53% of investors expect returns within 6 months, 98% of boards demand ROI proof, and 42% of enterprises are already abandoning AI projects — the patience window the market is granting is closing. AI will not disappear, but the patience of those paying for AI is finite. The question is not “can AI eventually generate value” (it can), but “can the flywheel start turning before patience runs out.”

VIIIConclusion: The Paradigm Shift from Process to Outcome

The fundamental contradiction of the current AI industry can be precisely stated as: When consumer users cannot use token output to create differentiated, monetizable products, and when enterprises cannot convert AI into quantifiable financial returns, the entire value chain’s investment loses its positive-feedback anchor point. Without an anchor point, the flywheel is idling and burning fuel.

How long the idle can persist depends on two variables: first, the inertia of hardware investment cycles — committed GPU orders and data center construction will not stop immediately; second, the capital market’s patience — 53% of investors expect returns within 6 months, and 98% of boards demand ROI proof. When the crossover of these two curves arrives — hardware order inertia weakens while investor patience expires — the industry will face a genuine correction.

The Token Triage architecture and outcome-oriented paradigm shift proposed in this paper, while the optimal solution for industrial efficiency, is the hardest to execute under current capital market incentive structures — it requires companies to sacrifice short-term valuation narratives for long-term health. The validated 29% of successful enterprises share four characteristics: AI directly linked to revenue outcomes, governance preceding scale, business teams owning AI workflows, and the entire initiative treated as organizational redesign. The common essence of these four characteristics is “outcome orientation.”