March trade data arrived quietly last week: US imports of large computers hit a pace of $340 billion a year. Not projections. Actual freight crossing actual borders. The number is so large it requires a moment to sit with — more than the GDP of many countries, flowing in as hardware, in a single year, to feed something that mostly lives in software.
On the same day, Anthropic filed confidentially for an IPO. The headlines treated these as separate stories. They are the same story told from opposite ends.
AI is a five-layer system — application on top, then model, then infrastructure, then chip, then energy at the base. Most of the conversation lives in the top two layers. Demos, releases, benchmarks, agent workflows, model evaluations. The excitement is real. But the $340 billion import figure is a signal from the bottom of the stack, and it’s saying something the top-layer conversation tends to skip.
The bottleneck in any layered system almost never lives where the excitement is. It lives one or two levels below where people are looking. When everyone is watching the model race, the constraint is already forming in data centers. When everyone is watching data centers, it’s forming in chip fabs. When everyone is watching chip fabs, it’s forming in power grids. The layer that wins is rarely the one with the best press.
What the IPO Narrative Gets Right — and What It Assumes
Anthropic filing ahead of OpenAI is a real competitive signal. Moving first through the SEC process is a choice, and choices cost something. What it implies about their confidence in near-term revenue trajectory is worth noting.
But IPO narratives are, almost by nature, stories told at the application and model layers. They have to be — those are the layers that generate the revenue lines investors underwrite. A prospectus cannot say “our core asset is a position in a five-layer infrastructure race.” It has to say: here is the product, here is the customer, here is the growth.
The question underneath the Bloomberg skepticism — can these IPOs live up to expectations — is really asking whether the model layer is where durable value accretes in AI, or whether it’s a relay race where the baton keeps passing downward. The honest answer is we don’t know yet. But the $340 billion import number is a vote. And it’s voting for the infrastructure layer.
The Enterprise Knows
Broadcom’s Tanzu Agent Foundations announcement is easy to scroll past. Platform-as-a-service for enterprise AI agents, built on VMware Cloud Foundation. Infrastructure news. Not exciting.
But enterprise technology decisions are slow and expensive to reverse. When a company like Broadcom builds a PaaS layer specifically for AI agents — not AI models, but agents — they’re making a bet that the agent execution environment becomes a durable control point. They’re not racing to build the smartest model. They’re building the substrate that models run on inside the firewall.
This is the old pattern. The companies that win infrastructure bets usually aren’t the ones making the loudest noise at the application layer. They’re the ones laying the plumbing while everyone else is arguing about the fixtures. Enterprise software history is full of this trade. The company that owned the virtualization layer, the company that owned the container orchestration layer — they didn’t win because they had the best demos. They won because they got there early and made switching expensive.
What the Import Numbers Actually Mean
A $340 billion annualized pace of computer imports is not an abstraction. It’s ships. It’s warehouses. It’s power contracts. It’s physical infrastructure being built at a speed that has no precedent in the history of technology deployment.
What that pace is telling you, quietly, is that the physical world has already made its bet. The companies buying those computers — hyperscalers, cloud providers, enterprise data centers — are not waiting for the model layer to stabilize before they invest. They’re treating the demand as certain and building ahead of it. That’s a different kind of conviction than what shows up in a venture valuation.
The AI IPO cycle will generate a lot of analysis about whether model companies can hold margin against open-source alternatives, about whether agents will commoditize the frontier, about whether $150 billion valuations make sense for companies still losing money. All of that is worth thinking about. But the infrastructure layer doesn’t wait for those questions to resolve. It just keeps building.
The layer nobody’s watching tends to be the layer that wins. And right now, 340 billion dollars a year is being very quiet about exactly that.
Leave a Reply