Blitzscaling as a Description of the AI Race
Blitzscaling, in Hoffman's 2018 book of the same name, is the deliberate prioritization of speed over efficiency in markets where the winner takes most of the value. The argument is counterintuitive only until you understand the math: in markets with strong network effects or steep learning curves, the company that captures the position first reaps disproportionate returns, and the company that captures it second reaps almost nothing. Burning capital to capture the position is rational. Optimizing for unit economics before capturing the position is suicide.
When Hoffman wrote Blitzscaling, the canonical examples were Amazon, Facebook, Uber, and Airbnb. They had each chosen speed at the cost of operational integrity, and they had each won. The book read as a controversial but defensible business doctrine. Six years later, it reads as the operating manual of the foundational model labs. OpenAI, Anthropic, Google DeepMind, xAI, Mistral, Meta — every serious AI lab is blitzscaling. They are spending tens of billions of dollars on compute, ahead of revenue, ahead of clarity about how to capture value, ahead of any reliable map of the market they are creating. They are doing this because the prevailing assumption is that AI is a winner-take-most market and the position will be captured by the lab that reaches the next capability threshold first.
The strange thing is that Hoffman, who articulated this doctrine, has been one of its most thoughtful internal critics in the AI era. He has argued repeatedly that the AI race is not zero-sum, that there will be multiple winners across different niches, and that the metaphors of arms races and Manhattan Projects mislead more than they clarify. This tension between his doctrine and his commentary is instructive. Blitzscaling works when the prize is a market. It is more dangerous when the prize is technological infrastructure that affects every other industry and every nation simultaneously.
The risks of blitzscaling are familiar from the previous generation of technology companies. Move fast and break things, applied to a social network, produced a polluted information environment that is still being cleaned up. Applied to ride-sharing, it produced unsustainable labor models and protracted regulatory battles. Applied to a general-purpose cognitive technology, the breakage surface expands by orders of magnitude. The same speed that wins the market can produce models that have not been adequately red-teamed, infrastructure that has not been adequately secured, and labor disruptions that have not been adequately anticipated.
Hoffman's defense is that the alternative is worse. A precautionary approach, in his framing, cedes the market to actors with weaker safety cultures and lower commitments to democratic values. If OpenAI and Anthropic slow down, the development does not stop — it moves to labs with fewer constraints. The logic is not new; it was used to justify the Manhattan Project. The logic is also not airtight. It depends on the empirical claim that safety can be maintained at full speed, and on the political claim that the right actors are in the lead. Both claims are testable, and both are being tested in real time.
What Hoffman has done, in his AI writing and his investing, is try to add steering to the speed. He pushes for iterative deployment so that the public can shape the technology. He pushes for transparent governance so that the labs are accountable. He pushes for what he calls intelligent regulation — narrow rules targeting specific harms rather than broad rules that ossify the field. Whether this counts as making blitzscaling responsible or as making irresponsibility scalable is going to be the central debate of the next governance cycle. Hoffman has placed his chips on the first interpretation. The chips are not small.
