The Batillus analogy is the image that stays. Massive infrastructure built on the assumption demand grows in a straight line. Demand shifts. The supertanker becomes a monument. That's the AI data centre risk in one historical parallel, and the parallel is uncomfortably precise because the supertankers were also funded by companies who believed the economics were permanent.
The timing question is where I'd push back. The SemiAnalysis data from this week shows the median enterprise spending $136 per employee per year on AI. At that level, nobody is running local models because the cost of your own GPUs, talent, maintenance, and security exceeds the API bill by orders of magnitude. Local makes economic sense when the cloud bill is large enough to justify the fixed cost of running your own stack. For 99% of companies that crossover hasn't happened. It will. But the Karp signal might be two years early, and in venture capital two years early and wrong are hard to tell apart until the third year.
Spot on. My only question is more around timing. There is already a massive inertia of sunk costs and political investment behind the current Big AI Manifest Destiny. And we've seen how the current U.S. administration has approached electrification and renewables.
This is a great post. Thank you.
sewage in, sewage out.
the law of entropy.
The first AI wave made firms tenants of intelligence.
They rented tokens, shipped data outward, accepted opaque pricing, and depended on model owners changing rules from above.
The next wave is about pulling intelligence back into controlled systems.
Not because frontier stops mattering.
Because ownership starts mattering again.
You can't be using much AI, or not in any high value use case. When/if you do, your perfective and analysis will change.
Buying lock, stock, and barrel Karp's biased, transparently self interested commercial pitch doesn't really make sense.
The Batillus analogy is the image that stays. Massive infrastructure built on the assumption demand grows in a straight line. Demand shifts. The supertanker becomes a monument. That's the AI data centre risk in one historical parallel, and the parallel is uncomfortably precise because the supertankers were also funded by companies who believed the economics were permanent.
The timing question is where I'd push back. The SemiAnalysis data from this week shows the median enterprise spending $136 per employee per year on AI. At that level, nobody is running local models because the cost of your own GPUs, talent, maintenance, and security exceeds the API bill by orders of magnitude. Local makes economic sense when the cloud bill is large enough to justify the fixed cost of running your own stack. For 99% of companies that crossover hasn't happened. It will. But the Karp signal might be two years early, and in venture capital two years early and wrong are hard to tell apart until the third year.
Spot on. My only question is more around timing. There is already a massive inertia of sunk costs and political investment behind the current Big AI Manifest Destiny. And we've seen how the current U.S. administration has approached electrification and renewables.
Solid