The market is watching the wrong metric. Over the past seven days, the narrative has been dominated by price action and ETF flows, but the signal that matters is a quiet economic finding: AI is compressing wages by an estimated $28 billion annually. This is not a prediction. This is a live audit of the labor market, and the results are being written into every future earnings report.
Apollo Research has published a finding that cuts through the noise. The $28 billion figure represents the current annual impact of AI on wage suppression. The number is small relative to the $12 trillion US wage pool—roughly 0.23%—but the trajectory is the story. We are not looking at a system failure. We are looking at a system that is functioning exactly as designed. Logic dissolves when code meets human greed, and this time the code is running on the labor market.
Context
The traditional doomsday narrative around AI has always been about job displacement. The media cycle loves a headline about mass unemployment. The reality is far more insidious. Unemployment in the US remains low, hovering around 3.7% to 4.0%. The jobs are still there. What is changing is the price tag attached to them.
This is the shift from explicit replacement to implicit devaluation. AI tools like Copilot and ChatGPT increase individual output by an estimated 30% to 50%. When output per worker rises but total demand stays flat, the employer's willingness to pay for that worker drops. The role survives. The salary does not. Trust is a vulnerability we audit, not a virtue, and the market is auditing labor costs in real-time.
Apollo's data aligns with what I have observed in my own audit work over the past three years. When I look at protocol teams and tech startups, the headcount is not shrinking as fast as the burn rate. The composition is changing. The leverage has shifted from the employee to the employer. The bridge was never built, only imagined, and the bridge between labor productivity and labor compensation is collapsing.
Core
Let me deconstruct the $28 billion figure from a first-principles perspective. The number is not a single line item. It is an aggregate of multiple compression vectors across the US economy. Based on my experience modeling interest rate curves and liquidation engines, I recognize this pattern. It is a systemic risk that is being priced into the market in a non-linear fashion.
The first vector is direct substitution. In sectors like customer service, content creation, and basic software development, AI agents can now perform tasks that previously required a human employee. This is not about firing the entire team. It is about not hiring the second or third person. The marginal hire has been replaced by an API call. Every summer has a winter of truth, and the winter for low-margin labor is already here.
The second vector is productivity arbitrage. When a worker uses AI tools, their output increases. The employer captures the majority of this surplus. This is a transfer of economic rent from labor to capital. The corporate profit margin is at a historical high of around 12%, while the labor income share has dropped from 63% in 2000 to roughly 58% today. AI is accelerating this trend. It is not creating the inequality; it is making the existing inequality more efficient.
The third vector is the most dangerous and the least discussed: the entrepreneurial illusion. Apollo's research suggests that AI lowers the barrier to entry for startups. The initial capital requirement drops from the million-dollar range to the hundred-thousand-dollar range. On the surface, this looks like a democratization of opportunity. The data on new business registrations in 2023 and 2024 supports this narrative—they are at record highs.
But this is where the audit gets uncomfortable. AI lowers the barrier to entry, but it also lowers the moat. When everyone has access to the same AI tools, the differentiation disappears. We are seeing a surge in homogeneous startups that are all building the same AI wrapper on the same APIs. The result is not a thriving ecosystem. It is a bubble of marginal ventures with high failure rates. The entrepreneurial path is becoming a new form of self-exploitation, where founders work longer hours for a lower probability of success.
I built a simple Python model to stress-test this scenario. If the startup failure rate increases by 10% while the creation rate increases by 20%, the net effect on job creation is negative. The aggregate wage pool shrinks even as the number of "founders" grows. Complexity is just laziness wearing a mask, and the complexity of the modern labor market is hiding a simple truth: AI is transferring pricing power from the individual to the institution.
There is also a hidden dimension in the Apollo data that deserves attention. The $28 billion likely underestimates the total impact. It probably captures direct wage compression but misses the hidden costs. Workers are spending unpaid hours learning new AI tools. Full-time roles are being converted to contract or gig positions to avoid benefits and long-term commitments. The quality of employment is declining even when the quantity remains stable.
The distribution of this compression is not uniform. High-skill workers who can leverage AI are seeing a premium. They are the ones who can command higher salaries because they are now twice as productive. Low-skill workers are bearing the brunt of the compression. They are the ones whose tasks are being partially automated, and their bargaining power is evaporating. This is not a single inequality curve. This is a barbell, with the middle class being squeezed from both sides.
Contrarian
Now let me play the contrarian role, because the bulls are not entirely wrong. The AI apocalypse narrative is lazy. The reality is that AI is creating value, and some workers are capturing it.
The productivity gains are real. The 30% to 50% output increase is not fiction. For the first time in a decade, we are seeing a potential reversal of the productivity stagnation that has plagued developed economies since the 2008 financial crisis. This is not a zero-sum game. There is a surplus being created, and the question is purely about distribution.
I have to acknowledge that my own analysis has a blind spot here. I focus on the compression vectors, but I may be underestimating the creation vectors. AI is enabling new categories of work that did not exist before. Prompt engineering, AI model tuning, and data curation are new job categories that are emerging. These roles may not yet compensate for the compression, but they are growing.
The policy response is also not entirely absent. There is talk of universal basic income, retraining subsidies, and even AI usage taxes. These are crude instruments, but they represent a recognition that the problem exists. History shows that social backlash to technological shocks typically lags by 5 to 10 years. We are still in the early phase. The window for a measured policy response is open, though it is closing.
The counter-argument to my entire thesis is that the $28 billion is a rounding error. In a $28 trillion economy, this is 0.1%. The impact is real but not yet systemically significant. The market is right to focus on the growth potential of AI rather than the current distributional effects. I am measuring the first drop of rain and calling it a flood. The flood may never come.
But this is where the risk assessment diverges. The $28 billion is not a static number. It is a growth rate. If this figure doubles every 18 months, which is consistent with AI adoption rates, we are looking at a $100 billion+ annual impact within three years. At that level, the impact on consumer spending and aggregate demand becomes macro-relevant. Silence in the blockchain is louder than the hack, and the silence in the wage data is becoming deafening.
Takeaway
The $28 billion wage compression figure is not a headline. It is a diagnostic signal. It tells us that AI is no longer a future risk. It is a current variable that is being priced into the labor market. The question for the next 12 to 24 months is not whether AI will disrupt the workforce. It is whether the disruption will be a controlled burn or a wildfire.
I am watching the Employment Cost Index and average hourly earnings data with the same intensity that I used to watch liquidation engines. The signal will not come from a single dramatic event. It will come from a series of small, compounding data points that reveal the direction of the transfer. Interoperability is the illusion of safety, and the interoperability between AI and labor is the new attack surface.
The market needs to start treating wage compression as a first-order variable in its models. This is not a labor market issue. It is a macro risk factor that will eventually hit consumer spending, corporate earnings, and by extension, the risk assets that we all track. The code is running. The question is who is auditing the output.