The Supply Chain ‘Ain’t’ What It Used to Be

The big goal of industrial policy advocates these days is to “reshore” the supply chain, for instance, by moving semiconductor manufacturing back to the United States.

What the industrial policy-types miss is that in our economy, in which all wealth is knowledge and all growth is learning, the real supply chain is equally immaterial. Today, the supply chain is the “experience curve.”  And experience curves, unlike factories, cannot be bought, or at least not budgeted.

When Michale Polanyi famously declared “we know more than we can tell,” he was replying to his experience of fascist and socialist central planning. Polanyi argued that successful human endeavors depended on a vast amount of “tacit” knowledge: intuitive, unarticulated, and based on personal experience, tradition, and practice, rather than articulable facts. He rejected the positivist claim that what cannot be stated clearly cannot be knowledge.

But planners must be positivists. What cannot be stated clearly cannot be put into the plan. What cannot be stated clearly cannot be made into an order or directive—or a government budget allocation.

The experience curve lives in the minds and even the bodies of those who have the experience. First comes discovery, then comes explanation. Formalization into rules comes last, and maybe never. Much of the most important knowledge in a well-functioning civilization cannot be centralized, specified, or planned without being damaged.

This becomes truer as tasks become more complex, incorporating more knowledge.

Project Stargate proposes to solve our AI challenges by building $500 billion worth of data centers. That can be budgeted because we know the cost of the bricks and mortar, the cooling systems, and the graphics processing unit (GPU) fabrics.

But progress in artificial intelligence does not come from bricks, mortar, and machines. Quite the reverse. Progress is measured by how much material is subtracted. More efficient models use less raw compute power. Capability advances less through brute force than through accumulated organizational learning—architecture, data judgment, training craft, and failure analysis—none of which can be mandated, subsidized, or scaled by decree.

Improvements in methods for training large models flow from accumulated know-how about how systems respond to scale, feedback, and failure, that cannot be specified in advance or subsidized directly. This is craft knowledge. It lives inside teams. It is similar with data management. Experience teaches teams what to keep, what to discard, and how to reshape training sets, over time reducing needed computational power rather than increasing it.

Research on experience curves show that, for every doubling of accumulated production, costs fall by somewhere between 15% and 30%, depending on the maturity of the industry. In that sense, it is possible to subsidize experience curves by subsidizing increased production. This does not automatically make it wise to do so. The accumulation of experience will always take time, disappointing those who believe bigger budgets and more factories can solve everything.

Consider the huge wafer fabs being built in Arizona by Taiwan Semiconductor (NYSE: TSM) and heavily subsidized under the CHIPS act. The Arizona fab uses the same fantastically expensive equipment as fabs in Taiwan, including ASML photolithography machines costing hundreds of millions. It uses the same process “recipes” to build at advanced nodes.

Yet, currently, finished wafers produced in the Arizona plant cost between 30% and 50% more than wafers produced in Taiwan. Yields, the percentage of good chips from a wafer, have been slow to increase.

To address these challenges, TSMC temporarily relocated hundreds of experienced workers from Taiwan to Arizona—a direct concession of the experience curve as supply chain. For what the Arizona fab was missing, and will continue to miss for some time, was not machines, but the tacit knowledge carried in the minds and bodies of experienced teams.

This does not mean the Arizona fab will fail. It means that even when the same company deploys the same tools and the same processes, manufacturing performance depends on accumulated local learning—teams that have worked together through years of iteration, error, and yield improvement.

It means that you can relocate the material components of a supply chain on a schedule. But experience comes in its own time. Time has a cost. The industrial policy folks ignore that cost at their peril and ours.