AI Becomes Too Big To Fail — 2026-10-13

Let's start with markets, because markets are often where the abstraction becomes visible.

On October thirteenth, WSJ What's News released two episodes that almost form a single sentence.

Rising borrowing costs threaten the Wall Street party, and IMF warns AI's riches are increasingly concentrated.

The first is the cost of capital side.

The second is the distribution side.

That pairing matters.

AI is no longer just a high growth software category where investors can debate multiples.

Um, it is a capital intensive build out that touches public markets, credit, sovereign strategy, data centers, energy, labor, and industrial policy.

The same day, Prof G Markets asked, Has NVIDIA become too big to fail? Ed Elson and Liz Hoffman discussed NVIDIA's AI holding company strategy and the idea that NVIDIA has become too essential to the AI system to fail in any ordinary way.

That's a remarkable phrase.

Too big to fail used to belong to banks, then maybe cloud platforms.

Now we're testing whether an AI chip company can become systemic infrastructure, and you can see why.

If every frontier lab, hyperscaler, enterprise AI platform, robotics company, and defense contractor is downstream of the same accelerators, memory supply chains, networking gear, and developer ecosystem, then NVIDIA is not merely selling picks and shovels.

It is helping define the terrain where everyone else digs.

The October thirteenth Innermost Loop newsletter framed this with a useful line.

The intelligence explosion comes with an invoice.

It cited estimates that a one gigawatt data center in Abilene costs about thirty billion dollars, that memory costs are up fivefold in a year, and that Broadcom's Hock Tan and Intel's Lip Butan both increasingly see capital, not just chips or power, as the binding bottleneck.

It also noted that NTT data is reportedly in talks to build Anthropic a five hundred megawatt data center in Osaka.

That is a very different AI story from the twenty twenty two chatbot moment.

The scarce thing is not only intelligence.

It is financing capacity, grid access, power purchase agreements, HBM supply, land permitting, interconnects, cooling, and enough revenue per megawatt to make the math work.

This is where Gavin Baker's October thirteenth posts on X are a useful counterweight.

Baker argued that an OpenAI Anthropic oligopoly at the model layer would be negative for nearly every company except those two, especially the token supply chain, because they could become monopsony buyers for hyperscalers, neo clouds, and data centers.

But he also argued that everyone can win if open source models, frontier labs, infrastructure, applications, and enterprises all coexist.

His more interesting point was that open weight models are increasingly being post trained in America on enterprise private data to create specialized models, which can then be combined with frontier models behind a router.

Let's start with markets, because markets are often where the abstraction becomes visible.

On October thirteenth, WSJ What's News released two episodes that almost form a single sentence.

Rising borrowing costs threaten the Wall Street party, and IMF warns AI's riches are increasingly concentrated.

The first is the cost of capital side.

The second is the distribution side.

That pairing matters.

AI is no longer just a high growth software category where investors can debate multiples.

Um, it is a capital intensive build out that touches public markets, credit, sovereign strategy, data centers, energy, labor, and industrial policy.

The same day, Prof G Markets asked, Has NVIDIA become too big to fail? Ed Elson and Liz Hoffman discussed NVIDIA's AI holding company strategy and the idea that NVIDIA has become too essential to the AI system to fail in any ordinary way.

That's a remarkable phrase.

Too big to fail used to belong to banks, then maybe cloud platforms.

Now we're testing whether an AI chip company can become systemic infrastructure, and you can see why.

If every frontier lab, hyperscaler, enterprise AI platform, robotics company, and defense contractor is downstream of the same accelerators, memory supply chains, networking gear, and developer ecosystem, then NVIDIA is not merely selling picks and shovels.

It is helping define the terrain where everyone else digs.

The October thirteenth Innermost Loop newsletter framed this with a useful line.

The intelligence explosion comes with an invoice.

It cited estimates that a one gigawatt data center in Abilene costs about thirty billion dollars, that memory costs are up fivefold in a year, and that Broadcom's Hock Tan and Intel's Lip Butan both increasingly see capital, not just chips or power, as the binding bottleneck.

It also noted that NTT data is reportedly in talks to build Anthropic a five hundred megawatt data center in Osaka.

That is a very different AI story from the twenty twenty two chatbot moment.

The scarce thing is not only intelligence.

It is financing capacity, grid access, power purchase agreements, HBM supply, land permitting, interconnects, cooling, and enough revenue per megawatt to make the math work.

This is where Gavin Baker's October thirteenth posts on X are a useful counterweight.

Baker argued that an OpenAI Anthropic oligopoly at the model layer would be negative for nearly every company except those two, especially the token supply chain, because they could become monopsony buyers for hyperscalers, neo clouds, and data centers.

But he also argued that everyone can win if open source models, frontier labs, infrastructure, applications, and enterprises all coexist.

His more interesting point was that open weight models are increasingly being post trained in America on enterprise private data to create specialized models, which can then be combined with frontier models behind a router.