Selling fear by the token

Andreessen Horowitz recently published a piece arguing that the AI job apocalypse is a complete fantasy. This is good news, knowing that a16z has dozens of billions of dollars invested in AI. Let's look at the numbers.

What the data says

The research consensus is remarkably consistent for a topic generating this much noise.

A recent NBER working paper found that AI adoption has not yet led to meaningful changes in total employment. A Federal Reserve Bank of Atlanta survey found that more than 90% of firms report no employment impact over the last three years. A Census Bureau working paper found that only 5% of AI-using firms report any headcount impact at all, split almost evenly between increases and decreases. The Yale Budget Lab, reviewing the full picture in April 2026, described the labor market as reflecting stability, not major disruption.

Four independent research institutions pointing in the same direction. The AI job apocalypse is not happening. So why does it feel like it is?

The scarecrow is a business model

Running frontier AI models costs an extraordinary amount of money. The compute, the energy, the infrastructure: the capital requirements are unlike anything the technology sector has seen since the early days of cloud. The companies building and selling access to these models need a continuous, high-volume flow of token consumption to justify that investment.

Fear is a fantastic motivator.

If you believe your engineers will be replaced by models, you start asking whether you should be spending on tokens instead of salaries. This is not a hypothetical framing but the kind of calculation being pushed openly in Silicon Valley. Jensen Huang, the CEO of Nvidia, repeatedly said that a $500,000 engineer should be generating $250,000 in token spend, or the math does not work in their favor.

This push to maximize AI consumption regardless of measurable outcome became a trend. Tokenmaxxing is being driven by a combination of AI labs that need the revenue, infrastructure companies that need the utilization, and founders who have built their pitch decks around disruption speed. The incentive structure is straightforward: make the cost of hesitation feel higher than the cost of adoption.

The fact that enterprise AI ROI remains elusive for most organizations does not quiet the message. It intensifies it.

Safety warnings and the capital cycle

Anthropic is a serious company doing serious work on AI safety. That is not in question. But when a lab announces that it cannot release a model such as Mythos because it is too powerful or requires additional safety evaluation, two things happen simultaneously. The public hears responsibility. Investors hear capability, urgency, and moat.

The safety warning and the fundraising round often arrive in the same quarter. That is not a coincidence, it is the structure of the moment. Demonstrating that your model is dangerous enough to require restraint is also the clearest signal you can send that your model is worth backing.

The safety concerns can be genuine and still serve the capital narrative at the same time. Those two things are not mutually exclusive. But the compute dependency — the fact that more powerful models require more infrastructure, more energy, more capital — rarely gets disclosed alongside the warning. The public gets the fear. The cap table gets the context.

What the market rewards

This anxiety does not stay individual. It moves into boardrooms.

Companies that announce AI-driven workforce reductions get a specific kind of attention from analysts. The market has learned to read layoffs framed as AI transformation as a signal of operational seriousness. The reverse is also true: companies that hold headcount steady find themselves explaining why they are not moving fast enough.

This creates a pressure loop. Executives feel compelled to demonstrate AI commitment in ways that are visible and measurable. Reducing headcount is more visible than improving decision quality. The announcement does the work regardless of whether the underlying logic holds.

I wrote about this in March. The Q1 2026 layoffs were not an AI story. They were a pandemic-era overhiring correction dressed in AI language because that framing was easier to sell to investors. The mechanism is the same here, the narrative serves the moment, and the moment serves someone's balance sheet.

When fear gets a face

The scarecrow strategy also has costs that do not show up on a cap table. When the narrative tells millions of people that a handful of technologists in San Francisco are about to make their lives unrecognizable, it does not stay abstract for long. The recent attacks against Sam Altman's home, including a Molotov cocktail attack and a shooting incident, are an extreme consequence of a very deliberately constructed anxiety. Fear, distributed at scale, does not always stay rational.

What is actually shifting

The honest picture is less dramatic and more useful.

AI is changing where in the stack human judgment is required, not eliminating the requirement for it. Repetitive cognitive tasks — data formatting, routine summarization, templated drafting — are being absorbed by models. The tasks becoming more valuable are the ones requiring context, accountability, and the ability to evaluate output from systems that are confident but not always correct.

That shift is real, and matters for how people build skills and careers. But it is not mass unemployment. And treating it as mass unemployment, or allowing others to frame it that way, makes it harder to have an honest conversation about what adaptation actually looks like.

I have argued before that scarcity shifted up the stack the moment LLMs commoditized code production. The same logic applies here. Judgment, architecture, oversight, these are not things a token replaces. They are things a token makes more necessary.

Are you seeing the tokenmaxxing pressure in your organization? And how are you pushing back on it?

Mathieu Flamant

Founder · CTO · mathieuflamant.com