Capital is pricing the wrong layer of the stack. $122B flows to compute scale while a single npm misconfiguration proves the moat is lightweight orchestration.

The Tension

The tension is between weight and portability. The compute layer is heavy. Datacenters. Chips. Power plants. It requires massive capital and long construction timelines. It is also where the money is going. The orchestration layer is light. It fits in an npm package. It can be rewritten in an afternoon by motivated engineers.

But the compute bet has its own fragility. Oil surged past $110 per barrel this week. The Strait of Hormuz carries 20% of global oil and LNG. Iran has positioned anti-ship missiles on Abu Musa and the Tunbs. The entire AI infrastructure thesis assumes energy costs that are no longer guaranteed. Oracle did not borrow $50 billion with $110 oil in the model.

Meanwhile, SpaceX filed for IPO at $1.75 trillion. Physical assets: $46 billion. That is 2.6% of the asking price. OpenAI projects $44 billion in cumulative losses before reaching profitability in 2029. The pattern across all three: capital is paying for a future that depends on conditions it cannot control.

For builders, the trade-off is concrete. Do you build on the compute layer, where the money flows but the energy assumptions are unstable? Or do you build on the orchestration layer, where the value is real but the moat is measured in days? Neither answer is comfortable. That is the point.

What This Unlocks

If orchestration is portable and compute is expensive, the winners are the companies that own neither. They are the ones that sit between the two layers and make them useful for specific problems. Vertical applications. Domain-specific coordination. The work of translating raw model capability into something a business can actually use.

The Anthropic leak proves this directly. Claude Code generates significant enterprise revenue, with the majority from enterprise clients. The value is not in the model. It is not in the orchestration code, which is now public. It is in the integration: the specific way the tool fits into enterprise workflows, the trust built over months, the compliance posture. That layer is not in the codebase.

What breaks is the narrative that building bigger models creates durable advantage. If the coordination layer ports in hours, and the compute layer depends on energy prices set by geopolitics you do not control, then the only defensible position is the relationship between the tool and the problem it solves.

For anyone building software right now, this changes the priority stack. Stop investing in model-specific integrations. Build model-agnostic orchestration that can swap providers when pricing shifts. And focus your energy on the part competitors cannot fork: your understanding of the customer's actual problem. Trojanized forks with backdoors are already circulating. The supply chain just got harder to trust. That makes verified, domain-specific tooling more valuable, not less.

Watching Next

Three things I am tracking, each with a specific signal that would change my view.

First, energy pricing in AI contracts. If datacenter operators begin renegotiating power purchase agreements or adding energy price escalation clauses, that confirms the compute layer's cost assumptions are breaking. Watch for Oracle, Microsoft, or Google adjusting their capacity expansion timelines. Any delay is a signal.

Second, orchestration commoditization speed. The 82,000 forks of Claude Code are a live experiment. If three or more production-quality open-source orchestration frameworks emerge within 60 days, the moat thesis for that layer is dead. If they stall at toy implementations, the coordination complexity is higher than the leak suggests.

Third, and this is the one any founder can check: look at your own vendor stack. Count how many layers depend on a single AI provider. If the answer is more than two, you have concentration risk that the market just demonstrated is real. The portability of orchestration means your switching cost should be near zero. If it is not, that is the gap to close this quarter.

Underweighting

I think the genuine underweight in this thesis is the possibility that the market is pricing optionality correctly. A 35x revenue multiple on OpenAI is only irrational if you assume revenue growth plateaus near current levels. If AGI-adjacent capabilities emerge in the 2026 to 2028 window and create entirely new revenue categories, autonomous agents replacing knowledge worker labor at scale, the current multiple could look cheap in retrospect.

The bear case on compute investment requires arguing not just that orchestration is replicable, but that no step-change capability improvement justifies the infrastructure build. I have not made that argument. Today's thesis assumes the current capability ceiling is close to the eventual ceiling. That is the least defensible assumption in the entire stack.

I should also be honest about lens fatigue. This is the fourth consecutive issue focused on system fragility and overvaluation. As I wrote about assumed constants becoming variables, my attention has been trained on what breaks. That is useful. But if inference efficiency improves 10x over 18 months, as historical GPU curves suggest, the energy constraint loosens and the oil signal becomes less structural than I am treating it. The capital might be right. The timing might just be early.

Bottom Line

The market is sending $122 billion to the engine room while a leaked npm package proves the bridge is where the ship gets steered. Capital is not wrong about AI. It is wrong about which layer holds the value. If you build anything this quarter, build it so you can swap every layer beneath your customer relationship without them noticing.