Friday charts: How much intelligence is too much?

Nvidia co-founder Jensen Huang believes that investing in AI infrastructure is a reinforcing feedback loop: “More compute creates better AI; better AI creates more usage; more usage creates more revenue; and more revenue drives more compute,” he posted this week on his new X account.

Huang sells semiconductors, so this might be a sales pitch. But it might also give him a unique perspective on AI. In the Moore’s Law era of rapidly improving semiconductors (which Huang has personally done so much to perpetuate), more investment led to better and cheaper chips, which made the chips more useful, which sold more chips, which funded more investment, which led to better chips.

Investment in large language models might work the same way.

But there are some big counterexamples, too — like agriculture and manufacturing.

The reason hardly any of us are farmers anymore is because investment in technology like tractors and fertilizer expanded the supply of food far faster than we could eat it (much as we’ve tried).

More recently, it’s also the primary reason why fewer and fewer of us work in factories. “Faster productivity growth interacting with unresponsive demand has been the dominant force behind the declining share of employment in manufacturing in the United States and other industrial economies,” one study found.

No matter how cheap and abundant manufactured goods become, there are only so many cars, refrigerators, TVs, and couches we can use (try as we might).

Now the same thing is happening in China, which is producing so many cars for export, there are not enough ships to carry them abroad. As manufacturing productivity outpaces consumption, Chinese workers are shifting into the gig economy, much as American workers once shifted from farming to manufacturing, and then from manufacturing to services.

Could the same happen in this new industrial revolution of AI? There might similarly be a limit to how much AI we can consume, however good and cheap it gets.

Intelligence is certainly different from manufactured goods. As televisions got cheaper, we bought more of them. But once there’s a TV in every room, we’re done. 

AI might be different. There’s no obvious limit to how much intelligence we might use.

In theory. In practice, there’s some early evidence pointing the other way.

The world’s most advanced LLM, Anthropic’s Fable 5, has proven less popular than expected: It accounts for only 11% of the tokens Anthropic sells, with customers continuing to prefer Anthropic’s less expensive Opus models.

Some take this as a sign that there’s a ceiling on how much intelligence we can use. “With Fable 5, we’ve found a new upper bound for how much businesses are willing to spend on AI,” Ara Kharazian says.

Others question the usefulness of the intelligence we’re already using. 

Sharon Goldman reports there’s a growing realization among companies using LLMs to write code: “Generating more code doesn’t necessarily mean producing more value.”

“We found that generated code requires a lot of work,” a VP of engineering told her. AI-written code may look plausible and compile successfully, but problems arise when it hasn’t been thoroughly vetted by human programmers. “Without that foundation, automation just accelerates the mess,” the VP explained.

Similarly, a Harvard Business Review study found that agentic AI might make more work than it saves.

“What looks like higher productivity in the short run can mask silent workload creep and growing cognitive strain as employees juggle multiple AI-enabled workflows,” the authors warned. 

"You had thought that maybe, ‘Oh, because you could be more productive with AI, then you save some time, you can work less,’” one software engineer told them. “But then really, you don’t work less. You just work the same amount, or even more."

That doesn’t disprove the idea that AI could create a self-reinforcing cycle of better models, more usage, and more investment. 

But that cycle only works if better and cheaper AI creates enough value to make us want still more of it. Jensen’s flywheel keeps spinning if each round of intelligence creates enough demand for the next. We don’t yet know for sure that it will.

There may be no obvious limit to intelligence, but there might still be a limit to what we’re willing to pay for it.

Let’s check the charts.

Moore’s Law for compute?

Epoch.ai calculates that the amount of compute you can buy with $1 has been rising at a rate of 49% per year — a doubling every 21 months, if it continues.

Demand signal?

In July, 43.5% of Ramp’s US customers paid for intelligence manufactured by Anthropic. I’m not sure how representative of the country that is, but it seems like a lot. 

Supply signal?

Anthropic’s newest model, Fable 5, accounts for only 11% of the dollars Ramp sees being spent on Anthropic models (the purple bars at the bottom). Customers seem happier with the cheaper Opus models (in black and orange).

Counterproductive so far?

Data from Bank of America shows that the AI investment boom is adding to inflation due to 1) rising input costs for electronics (memory prices, mostly, I assume) and 2) a wealth effect supporting consumer demand (because stocks are going up, I’m guessing). The hoped-for disinflationary effects of productivity are not yet visible in the data.  

The cost line:

In total, customers are on pace to spend about $115 billion buying intelligence from Anthropic and OpenAI this year.

The new corporate divide:

Data from OpenAI shows that the top 10% of its customers use 17x more compute than the average user. The study also finds that the biggest corporate users also have the most productive employees. How to interpret a study from OpenAI that finds OpenAI makes its customers more productive is left for the reader as an exercise.

The corporate divide (2):

Ramp reports that in July, the top 1% of businesses spent $7,400 per employee on AI. The median firm spent just $11.95.

Not productive enough?

Torsten Slok notes the AI boom is being funded by investors rather than earned from customers. Semiconductor makers like Nvidia have a profit margin of 41%. Model providers like OpenAI have an operating loss of 59%. “Capital can bridge the gap for a while, but not indefinitely,” Slok says. 

More code, fewer users.

A study finds that agentic coding has led to an explosion of apps in Apple’s App Store, but few of them are being used. The authors speculate this is either because the apps are of lower quality or because the flood of supply stops them from being discovered.

A lot more code. A little more software.

The same paper finds that agentic coding has increased the amount of code written by 290%, but the number of finished products by only 30%. “In software,” the authors conclude, “the binding constraint appears to be shifting from writing code to reviewing, integrating, and ultimately distributing it.”

The supply of intelligence may be rising faster than our ability to productively use it.

Have a great weekend, productive readers.

— Byron Gilliam