AI Costs More Than Human Labor—For Now
In recent years, his research has extended to the economic and fiscal implications of technological change, including the interaction between artificial intelligence, demographic shifts, and public finance sustainability.
He holds a PhD in Mathematical Finance from Boston University, and previously earned an MSc in Finance and Economics from the London School of Economics. He completed his undergraduate studies in Economics at Seoul National University under the Korea Foundation for Advanced Studies scholarship program.
He regularly contributes analytical essays on the broader socioeconomic implications of AI to The Economy Review.
Authored on
Modified
AI is moving from a low-cost software experiment to a metered production input whose bill rises with use Cheaper tokens do not guarantee lower spending: agents, context, tools, and repeated inference can expand consumption faster than unit prices fall The automation threshold is crossed when AI becomes cheaper, more predictable, and sufficiently reliable at the task level—not merely when a model can perform the task
A follow-up to Fortune’s 2026 interview on the cost of AI compute and human labor
The automation threshold is crossed when AI becomes cheaper, more predictable, and sufficiently reliable at the task level—not merely when a model can perform the task. What the Fortune interview reveals about token economics, enterprise budgeting, and the next automation threshold.
The paradox behind the Fortune interview
The interview began with a contradiction that many executives now recognize. Technology firms are reducing headcount, committing extraordinary capital to AI, and still discovering that compute can cost more than the employees using it. Nvidia executive Bryan Catanzaro described compute costs on his team as exceeding employee costs; Fortune also reported that major technology firms had announced about \$740 billion in capital expenditures while more than 118,000 technology layoffs had been recorded in 2026 at the time of publication.
This is not evidence that AI has failed. It is evidence that enterprise AI has entered the difficult stage between demonstration and industrialization. A chatbot can appear inexpensive at the interface while the production system behind it consumes data, retrieval, orchestration, monitoring, security, and human review. In the Fortune interview, I described the moment as a “short-term mismatch”: firms are being asked to show AI-driven productivity before AI has become a universal low-cost substitute for labor.
The current contradiction is real: firms can lay off workers, spend more on AI, and still find that human labor remains cheaper for many production tasks.
A short-term mismatch, not a permanent verdict
The present cost disadvantage of AI reflects three transitional conditions. First, infrastructure remains scarce and expensive. Chips, high-bandwidth memory, advanced packaging, power, cooling, and data-center capacity must expand before low model prices become low production costs. Second, many organizations are running premium models on tasks that do not require premium capability. Third, enterprise workflows still carry a large human backstop because models are unreliable at the edges and accountability cannot be delegated to a probability distribution.
These conditions will not remain fixed. Hardware efficiency will improve, model routing will become more sophisticated, and firms will learn which tasks justify frontier models. Pricing will also migrate away from simplistic flat subscriptions toward usage, capacity, and outcome-based contracts. The important conclusion is therefore temporary but consequential: AI labor may cost more than human labor today in selected workflows, yet the cost curve is moving quickly enough that executives should redesign work before the economic threshold arrives.
Cheaper tokens can produce larger budgets
The most important correction to conventional software thinking is that a lower unit price does not necessarily reduce total expenditure. Bain reports that average token cost fell by half between December 2024 and December 2025 while tokens consumed rose 4.5 times. Organizations upgraded to frontier models, assigned agents more complex multistep work, and expanded use once teams discovered new workflows. The result was not a clean cost saving. The models became cheaper per token while the bill remained stubbornly high.
This is a form of the Jevons paradox: efficiency lowers the cost of an individual unit and thereby makes more consumption economically attractive. In AI, the effect is amplified by agentic systems. One visible answer may require multiple internal searches, tool calls, context reloads, verification passes, and retries. The correct question is not whether token prices are falling. It is whether cost per accepted business output is falling after the full workflow is counted.
The bill extends beyond model inference
Token prices are visible because providers publish them. The larger cost stack is often hidden across budgets and teams. Data engineers prepare and govern information. Security teams control access. Product teams integrate models into systems. Reviewers validate output. Operators handle exceptions. When an agent fails, the organization pays again through rework, delayed service, customer remediation, or compliance exposure.
This is why a low API price can coexist with poor automation economics. A technically impressive system can still be more expensive than the human process if it requires frequent escalation or produces outputs that cannot be accepted without substantial review. BCG similarly argues that much of the value from AI depends on process redesign rather than the algorithm alone, and that efficiency gains must be translated deliberately into P&L impact.
Why layoffs can precede full automation
The coexistence of high AI costs and layoffs is easier to understand once jobs are separated from tasks. A firm does not need an autonomous system that replaces an entire occupation before it can reduce hiring. If AI allows a smaller team to draft, search, summarize, test, and coordinate more work, the company can consolidate roles or leave vacancies unfilled while retaining humans for judgment and exceptions.
The current technology-sector cycle also includes factors that are not pure AI substitution. Many firms over-hired during the pandemic, when online demand and remote operating models appeared more permanent than they proved to be. At the same time, companies making enormous AI investments face investor pressure to demonstrate margin improvement. Headcount is one of the fastest and most visible costs to adjust. The result is an AI-shaped restructuring cycle that can occur before AI is cheaper than people in a complete workflow.
| Bit Tech Capex | Layoffs | Mismatch |
|---|---|---|
| $740B | 118,000+ | Short-term mismatch |
| Big Tech capital expenditure announced in 2026 at the time of the Fortune article | Technology layoffs recorded in 2026 at the time of publication | AI spending and labor reductions can rise together before full substitution |
The replacement threshold is task-specific
The decisive comparison is not AI versus humans in general. It is AI versus the human process for a specific task, at a specific volume, under a specific risk standard. MIT’s 2024 analysis of computer-vision tasks estimated that only about 23 percent of wages paid for those tasks were economically attractive to automate under then-current cost assumptions. Technical feasibility was broader than economic viability.
The threshold moves with volume, reliability, regulation, and exception frequency. High-volume, standardized, low-liability tasks can cross early. Low-volume work with tacit context or costly errors may remain human-led long after a model can produce plausible output. For executives, this means that automation portfolios should be built task by task, not announced job category by job category.
Three economic states of AI labor
| State | Operating model | Human role | Labor effect |
|---|---|---|---|
| Augment | AI assists a worker who still owns the process | Higher quality, faster cycles, limited headcount effect | Human remains primary operator |
| Compress | A smaller AI-leveraged team absorbs work previously spread across more roles | Hiring restraint, role consolidation, higher output expectations | Human owns exceptions and accountability |
| Substitute | AI performs the workflow end to end at lower accepted-output cost | Direct labor displacement becomes economically rational | Human oversight becomes periodic or regulatory |
The cost curve is moving
The present advantage of human flexibility should not be mistaken for a permanent shield. Stanford’s 2025 AI Index reported that the cost of querying a model at GPT-3.5-equivalent performance fell from about \$20 per million tokens in November 2022 to \$0.07 by October 2024. Fixed-performance inference therefore became more than 280 times cheaper in less than two years.
Physical capacity is expanding at the same time. The International Energy Agency projects global data-center electricity consumption to rise from roughly 415 TWh in 2024 to around 945 TWh by 2030, with AI the largest driver of the increase. This build-out is expensive in the near term, but it expands the infrastructure available for inference and can lower cost per useful task as supply, hardware, and workflow efficiency improve.
From tokenmaxxing to ROI-maxxing
The first phase of enterprise AI rewarded visible adoption: seats activated, prompts submitted, agents launched, and token consumption increased. Those metrics are now becoming liabilities because they measure activity rather than value. The next phase requires what might be called ROI-maxxing: using the least expensive combination of model, workflow, and human oversight that reliably produces an accepted outcome.
The essential management discipline is to instrument AI consumption at the workflow level. A company should know what it costs to resolve a customer issue, draft an accepted contract, produce a usable code change, or complete a compliance review. Without that denominator, model routing and budget controls are guesses.
| Metric | Executive interpretation |
|---|---|
| Cost per accepted output | Tokens + tools + integration + review + rework, divided by accepted business outputs |
| Human review minutes | Time spent checking, correcting, escalating, and signing off each output |
| Exception rate | Share of cases that leave the automated path and require human intervention |
| Model-routing efficiency | Share of tasks handled by the smallest model that meets the quality threshold |
| Value per million tokens | Revenue protected, cost avoided, or cycle time saved for each unit of consumption |
| P&L realization | Whether measured productivity becomes lower cost, greater capacity, or better customer outcomes |
A 90-day executive agenda
| Steps | Objectives | Tasks |
|---|---|---|
| 1 | Instrument one workflow end to end | Measure tokens, tool calls, integration expense, review time, exception handling, and accepted outputs |
| 2 | Establish the human baseline | Document current labor cost, cycle time, error rate, and service quality before comparing AI economics |
| 3 | Route models by task | Reserve frontier models for work that justifies the premium; use smaller or specialized models for volume |
| 4 | Cap and observe agent loops | Set stopping rules, retry limits, and alerts before autonomous workflows create invisible consumption |
| 5 | Name an exception owner | Define who intervenes when the automated path fails and who carries accountability for the outcome |
| 6 | Translate productivity into P&L | Decide in advance whether saved time becomes lower headcount, greater capacity, faster growth, or better customer experience |
Conclusion: predictability is the real tipping point
The intermediate labor market will reward people who can supervise AI inside a real process. Prompting alone is not a durable advantage. Workers need domain context, evaluation skill, data literacy, exception judgment, and an understanding of how model use changes cost and risk. The most valuable employee may be the person who knows when a smaller model is sufficient, when a human should intervene, and how to verify that an output can be accepted.
This creates a divide between ordinary human labor and AI-leveraged human labor before it creates a clean divide between humans and machines. The immediate risk is labor compression: fewer employees are expected to produce more. The longer-term risk is substitution as cost, reliability, and infrastructure improve. The appropriate response is neither reassurance nor panic. It is deliberate role redesign and credible training before the threshold moves.
AI will not become a universal labor substitute simply because models become more capable. The tipping point arrives when AI can complete a task at lower accepted-output cost, with sufficiently low variance, manageable exceptions, and accountability the organization can defend. That is why the phrase from the Fortune interview matters: AI must become not only cheaper than humans, but cheaper and more predictable at scale.
The strategic question is no longer “How many people use AI?” It is “Which workflows produce more accepted value per dollar of compute and per hour of human oversight?
Related reading from The Economy
- AI Labor Costs Are the New Test of the Automation Economy
- The AI Labor Divide: Who Wins, Who Survives, and Who Falls Away
- “Token Prices Are Falling, but AI Costs Keep Rising”
- Leverage Builds Behind the AI Investment Frenzy
- “Breaking Through the Limits of Process Shrinkage”
Sources and further reading
1. “The cost of compute is far beyond the costs of the employee”: Nvidia executive says right now AI is more expensive than paying human workers. Fortune, 2026. (Yahoo Finance version: ‘The cost of compute is far beyond the costs of the employees’: Nvidia executive says right now AI is more expensive than paying human workers)
2. How Token Economics Will Change Opex. Bain & Company, 2026.
3. Four Ways to Create a Lasting Cost Advantage from AI. Fortune / BCG, 2026.
4. Beyond AI Exposure: Which Tasks Are Cost-Effective to Automate with Computer Vision?. MIT CSAIL / MIT Initiative on the Digital Economy, 2024.
5. Artificial Intelligence Index Report 2025 - Research and Development. Stanford Institute for Human-Centered AI, 2025.
6. Energy and AI - Executive Summary. International Energy Agency, 2025.
7. Tokens Are Getting Cheaper, but Companies Are Spending Even More on AI as a Result. Fortune, 2026.
8. Your AI Budget Is Growing. Your Returns Aren’t. Here’s Why.. Bain & Company, 2026.
In recent years, his research has extended to the economic and fiscal implications of technological change, including the interaction between artificial intelligence, demographic shifts, and public finance sustainability.
He holds a PhD in Mathematical Finance from Boston University, and previously earned an MSc in Finance and Economics from the London School of Economics. He completed his undergraduate studies in Economics at Seoul National University under the Korea Foundation for Advanced Studies scholarship program.
He regularly contributes analytical essays on the broader socioeconomic implications of AI to The Economy Review.
