The logic held; the incentives were broken.
Apollo Research has published findings that should unsettle anyone tracking the intersection of artificial intelligence and labor markets. The headline number: AI is compressing wages by approximately $28 billion annually in the United States. Not eliminating jobs. Compressing wages. The distinction matters more than most analysts realize.
I have spent the better part of a decade tracing incentive flows through smart contracts, governance mechanisms, and tokenomic structures. The pattern Apollo identifies in labor markets mirrors what I have observed repeatedly in decentralized finance: the mechanism of extraction shifts before the structure collapses. In DeFi, it was yield subsidies masking liquidity drains. In labor markets, it is wage compression masking productivity gains.
The unemployment rate sits at 3.7-4.0%. The labor market appears healthy. But real wage growth continues to lag productivity growth. The numbers do not reconcile unless you account for what Apollo identifies: AI is not replacing workers in aggregate. It is repricing them.
The Mechanism of Silent Substitution
Let me be precise about what the data shows. Apollo's $28 billion figure represents approximately 0.23% of the roughly $12 trillion annual U.S. wage pool. A small percentage. But the trajectory matters more than the current magnitude.
The mechanism works through what I would call "silent substitution." When a worker equipped with Copilot or ChatGPT becomes 30-50% more productive, the employer's willingness to pay for that role declines proportionally. The job remains. The worker remains. But the market pricing power shifts from labor to capital.
This is not the "AI took my job" narrative. This is the "AI repriced my job" reality.
I traced this pattern before, in a different context. In 2020, I isolated the Compound Finance governance token mechanics and found that yield was largely subsidized by inflationary token emissions rather than organic revenue. The structure appeared stable. The incentives were broken. The same logic applies here: the labor market appears stable because unemployment is low. But the wage mechanism is being quietly rewired.
The $28 Billion Question
The critical question Apollo's research raises is methodological. How was the $28 billion calculated? Model estimation or empirical data? Which industries and job categories were covered?
The confidence level I would assign to this figure is C-minus. Not because the direction is wrong, but because the granularity is insufficient. The number likely captures only direct wage compression effects. It does not account for the hidden costs: the additional hours workers spend learning AI tools, the shift from full-time employment to contract work, the degradation of job quality that does not show up in wage statistics.
Code does not lie, but it can be misled. The same applies to economic data. The $28 billion figure is a floor, not a ceiling.
The Distribution Problem Nobody Wants to Discuss
The wage compression effect is not uniform. This is where the analysis gets uncomfortable.
High-skill workers who effectively leverage AI tools may actually see wage premiums. They become more valuable because they can amplify their output. Low-skill workers whose partial functions get automated face downward wage pressure. The result is a simultaneous expansion of both the "skill premium" and the "low-end squeeze."
This is not a single inequality vector. It is a bifurcation.
I have seen this pattern before in algorithmic systems. In 2021, I spent three months reverse-engineering the bot scripts used in the Bored Ape Yacht Club mint. I identified the specific MEV strategies that allowed insiders to snipe floor prices before public sales. The same dynamic applies here: those with access to the tools extract value from those without.
Bots do not dream, they only scrape. The AI wage compression effect operates on the same principle. Those who control the algorithms extract value from those who do not.
The Entrepreneurship Paradox
Apollo's research highlights that AI lowers the barrier to entrepreneurship. Software development, content creation, and customer service costs have dropped dramatically. The initial capital threshold for starting a business has fallen from "millions" to "hundreds of thousands."
This sounds positive. It is not entirely so.
Lower barriers to entry also mean lower moats. AI-generated code and AI-generated content create a flood of homogeneous projects. The number of new business registrations in the U.S. hit record highs in 2023-2024. But the survival rate is the metric that matters, and that data is less encouraging.
I have seen this dynamic play out in crypto. The proliferation of Layer2 solutions created dozens of networks serving the same small user base. This was not scaling; it was slicing already-scarce liquidity into fragments. The same pattern applies to AI-assisted entrepreneurship: more projects, same demand pool, fragmented attention and capital.
The supply was fixed; the demand was fabricated.
The Algorithmic Pricing Risk
There is a darker dimension to wage compression that Apollo's research only hints at. AI enables what economists call "personalized pricing" in labor markets. Companies can use algorithms to assess each job candidate's reservation wage—the minimum they would accept—and offer precisely that amount.
This is wage discrimination at scale. It is not market-driven; it is algorithm-driven. And it systematically pushes wages downward.
The legal framework for addressing this is murky. Antitrust law has provisions for monopsony—a market with a single buyer. But applying those provisions to algorithmic wage-setting is uncharted territory. The policy response, across the U.S. and EU, remains in the "research" phase. No substantive redistribution mechanisms have been designed.
Transparency is a feature, not a default state. The algorithms determining wage compression are opaque. The data feeding them is proprietary. The accountability mechanisms are nonexistent.
What the Bulls Got Right
I need to acknowledge what the AI optimists got right. The wage compression narrative is not the complete picture.
AI does create genuine productivity gains. The 30-50% efficiency improvements for individual workers are real. The question is whether those gains flow to workers or to capital. Current data suggests they flow disproportionately to capital. U.S. corporate profit margins are at historic highs of approximately 12%, while labor's share of income has declined from 63% in 2000 to roughly 58% today.
But the skill premium is real. Workers who master AI tools are capturing value. The opportunity is not in resisting the technology but in positioning within it.
The entrepreneurship story also has a genuine upside. AI lowers the cost of experimentation. More people can try to build things. The failure rate may be higher, but the raw number of attempts increases. Some of those attempts will succeed.
The Systemic Risk Framework
The second-order effects of AI wage compression are what concern me most. This is where my systemic risk framework comes into play.
If wage compression accelerates while inflation persists, the result is a double squeeze: real wage decline combined with rising living costs. This combination historically precedes social instability. The timeline for technology-driven social backlash is typically 5-10 years. The AI wage compression effect is still early, but if it expands through 2025-2028, the social response could be severe.
The policy response will likely be reactive and abrupt. An AI usage tax. Mandatory redistribution mechanisms. Regulatory interventions that could reshape the entire AI industry overnight.
I have seen this pattern before. In 2022, I modeled the Terra/Luna feedback loop and proved mathematically that the algorithmic stability was a Ponzi structure dependent on infinite growth. I published my critique three days before the total collapse. The accuracy came from cold logic, not intuition.
The same cold logic applies here. The wage compression effect is structural. It will not reverse on its own. The only question is whether the adjustment comes through policy design or through crisis.
The Accountability Gap
The $28 billion figure represents a transfer of wealth from labor to capital. The beneficiaries are corporate shareholders. The costs are borne by workers whose wages stagnate while productivity rises.
The yield was not profit; it was liquidity. The wage compression is not efficiency; it is extraction.
The accountability gap is the core problem. No institution is responsible for monitoring AI's impact on wage distribution. No regulatory framework exists for algorithmic wage-setting. No mechanism ensures that productivity gains are shared with the workers who generate them.
The signals to track are clear. The Employment Cost Index and average hourly earnings data will show whether AI-related wage anomalies appear in the coming quarters. The policy responses from major economies will indicate whether governments understand the stakes. The survival rates of AI-assisted startups will reveal whether the entrepreneurship boom is real or illusory.
The Pre-Mortem
Let me be direct about what I expect to happen. The $28 billion figure will grow. The wage compression effect will accelerate as AI penetration increases from the current 20% of U.S. businesses. The distributional consequences will become more visible. The policy response will lag, as it always does.
The question is not whether AI will reshape labor markets. That is already happening. The question is whether the reshaping will be managed or chaotic.
Algorithmic fairness assumes fair inputs. The inputs to the AI wage compression machine are not fair. They are shaped by access to tools, by data ownership, and by the structural power imbalance between capital and labor.
I have spent my career tracing the gap between how systems are described and how they actually function. The gap between the "AI will create new jobs" narrative and the "AI is repricing existing jobs" reality is exactly the kind of discrepancy that interests me.
The logic held; the incentives were broken. The question now is whether we can fix the incentives before the logic breaks us.