The bottleneck cascade

In April 2024, Mark Zuckerberg told Dwarkesh Patel that Meta’s biggest constraint in developing AI wasn’t GPUs. It was power. Even with unlimited capital, the lead time on new energy generation was years.

Tyler Cowen made a similar point. When asked why the AI-takeoff may be slower than people expect, he argued that the limiting factors aren’t technical, but human (i.e. cultural and regulatory). For example, even if AI increases the number of high potential drug targets by 10x, the FDA still needs to approve them.

What these examples reveal is a pattern I call the Bottleneck Cascade. Every technological breakthrough makes one input abundant, and in doing so it pushes scarcity somewhere else in the system. 

In other words, progress doesn’t eliminate constraints. It moves them.

History offers countless examples of cascades in action. Here’s a particularly relevant one: in the 19th century, factories had to be built next to rivers to harness water wheels (or else rely on steam engines). The arrival of centralized electricity generation and distribution blew the bottleneck open. Suddenly, you could pipe power into any building in any city, and factories no longer needed to be clustered by rivers.

Electricity didn’t transform manufacturing overnight, however, because cheap electricity created new bottlenecks: the need for transmission infrastructure, new machinery, and above all, new factory designs which took advantage of this strange, new power source.

That’s the pattern consistently seen with general purpose technologies. When a GPT first arrives, productivity often falls. Its real-world utility depends on complementary innovations. In this example, machines with electric motors needed to be invented and “electricity-native” factories needed to be designed before electricity’s full impact could be felt. That took decades. 

And indeed, this is the exact dynamic we’re living through today. AI is the most powerful technology of our lifetimes, but its trajectory will be defined not by raw capability, but by the sequence of bottlenecks it collides with. Two years ago, GPUs were the scarcest resource. As supply ramped, the bottleneck shifted to power. What else is holding AI back? Perhaps talent (how many people in the world can reliably train frontier models…potentially fewer than 1000)? Maybe high-bandwidth memory? What about actual real-world implementation of AI (see the much-discussed MIT report on the failure of enterprise generative AI pilots)?

The cascade continues.

Seeing the world through this lens is useful because it provides a hint as to where value might accrue. Track enough cascades and you’ll realize that the company that breaks a bottleneck is not the one that captures the value. Containers revolutionized shipping, but the container makers didn’t become giants. The winners were the abundance-native firms like Walmart. Walmart won not by making shipping cheaper, but by mastering the scarce capabilities that mattered once it was: logistics, purchasing power, and scale.

The company that unblocks a bottleneck may profit, but the giants are built by those who design for the new abundance. And in a world of abundance, profits accrue to whoever can locate and defend the next true scarcity.

Just like the general purpose technologies that have preceded it, AI will need its own complements to reach the mainstream. New form factors beyond chat boxes. New organizational designs that integrate machine intelligence with human judgment. New UX primitives for delegation and supervision. These are bottlenecks too, and until they’re built, AI’s full potential will remain bottled up.

This sort of analysis is particularly urgent because AI is currently removing bottlenecks all over the place. As a result, the most obvious, consensus ideas are getting funded fast and furious. But the uncomfortable implication of the bottleneck cascade framework is that this is a uniquely dangerous moment for incremental bets. They feel safe, but they’re almost always designed for a world that is already slipping away. And worse, they’re often aimed at categories where AI is collapsing margins, leaving little profit even if adoption takes off.

History is full of cascades, each showing how today’s bottlenecks might unfold. The question isn’t whether AI will face them, but what kind of strange, AI-native firms will emerge once those constraints give way.