Factories Are Easy; Learning Is Hard—Why Industrial Policy Keeps Funding the Wrong Layer

Never think about “protectionism versus free trade.” What we call protectionism is simply industrial policy by another name.

The decision to protect certain industries is simply a decision to shift intellectual and material capital towards that industry. The CHIPS act subsidizing the building of new semiconductor “fabs” (manufacturing sites) capacity is all about tariffs and export controls to discourage the Chinese semiconductor industry.

The danger of industrial policy is not that it makes no sense. The danger is that industrial policy makes some sense. Industrial policy is driven by the headlines but skips the footnotes. It caters to the visible—factories, capital expenditures power, and headcounts. But it tends to miss the underlying layers or “platforms”—the set of capabilities that makes the visible, viable and valuable.

Because politicians and other partly educated busy bodies miss the importance of the enabling layers, these layers are more likely to be politicized, regulated, or disrupted. A new fab manufacturing ever more complex chips also needs the newest metrology, inspection, process control, and EDA (electronic design automation).

EDA tools are invisible software layers that enable engineers to design chips with billions of transistors, simulate whether those designs will work, and translate logic into layouts that a fab can physically manufacture.

If fabs are factories and metrology is quality control, EDA is the grammar of silicon—the rules that make complex chips intelligible before they ever exist.

Without the updated EDA, advanced fabs would be mausoleums. Yet EDA is entirely knowledge-based, learned over decades, refined through failure, protected not primarily by patents, but by tacit knowledge and ecosystems.

EDA is driven by the learning curve. Fundamental to the learning curve is that we can never predict what will be learned. And because we can’t predict it, it is all but impossible to subsidize. We can create subsidies to increase the volume of production, which drives the learning curve. But any such subsidies will steer capital and innovation to other, perhaps inferior, products. There is no such thing as a free subsidy.

Governments can fund things that can be blueprinted, because we (even politicians!) already know about them. What cannot be predicted cannot be subsidized. No country has been able to conjure a competitive EDA stack by decree.

Subsidies for artificial intelligence are similarly subject to platform risk. Massive government encouraged investments like Project Stargate focus understandably on compute and power. But compute alone does not produce resilient intelligence systems any more than fabs alone produce competitive semiconductors.

Again, the visible is favored over the critical. Visible are multi-billion-dollar GPU fabrics, the next brilliant generative model, and cap-ex budgets. These are meaningless without critical platform layers:

Orchestration: How workloads are scheduled, paused, scaled, or dropped. Turns computation into continuous service.

Security: Who can invoke a model; what data it can touch.

Identity: Who—or what—is acting right now? Draws the line between delegation and loss of control.

Observability: Ability to see inside a live system while it’s operating, preventing opaque failure and the collapse of trust.

Data integrity: Corrupt inputs produce confident nonsense.

If these critical control layers are weak, AI systems become brittle, failures cascade, and trust evaporates.

Industrial policy supports what is visible, countable, and static. Yet the layers that make AI systems governable are supplied not by builders of concrete and compute, but by firms that trade in understanding, control, and learned restraint—the capabilities industrial policy is least equipped to recognize, fund, or sustain.

Politicians love machines and megawatts. Control layers, by contrast, are software: iterative and intangible; learned, not installed; developed through failure; cheap to begin and expensive to master. Worst of all for political purposes, they do not look like “national power.” It is far easier to boast, “We built a dozen hyperscale data centers,” than to say, “We reduced mean-time-to-understanding,” or, “We improved the auditability of autonomous agents.”

That is why industrial policy so reliably funds the wrong layer. The right layer is not designated in advance, but discovered over time—dynamic rather than static, epistemic rather than material. It is the innovation layer, where learning curves compound and the upside surprise lives, and where policy driven by headlines is least able to follow.