The Grammar
and the Substrate

The Capability Escalator argued that every general-purpose technology eventually externalizes the foundations it was built on, passing through five stages: a forcing function, internal capability, a probe by application, externalization, and standardization. The cases were clean — Google → Cloud, Amazon → AWS, IDM → TSMC, Dynamo → Grid. But four ascents that all went the same way is a story, not a law. A law has to survive the cases that contradict it.

So here is the harder question. Is externalization a resource law — a system-neutral consequence of transaction costs, returns to scale, modularity, and the collapsing cost of replication? Or is it a capitalist trap — an artifact of capital accumulation that releases a capability only once the owner has secured a more durable place to extract rent? To decide, we have to go looking for the escalator cases that did not climb: the capabilities that stayed enclosed for decades and collected rent instead of becoming markets. If enclosure and externalization are both common, then "externalization" is not a law but an outcome — and the interesting work is naming the selector variables that decide which path a given capability takes.

Transaction-cost economics. Ronald Coase's "The Nature of the Firm" (1937) asked why firms exist at all rather than pure markets. Because the price mechanism has costs. Oliver Williamson operationalized the answer with asset specificity: when assets are highly specific and co-specialized, the threat of hold-up pushes activity inside the firm; when assets are generic and substitutable, markets govern them cheaply. The lesson for our question is sharp — a firm externalizes its grammar precisely when it already owns the complementary asset that lets it keep extracting value after the grammar is free.

Complementary assets. David Teece's 1986 framework explains who profits from innovation, and it is rarely the inventor. EMI invented the CAT scanner and lost the market; Searle's NutraSweet held only while the patent did. Publishing a capability without owning the complementary assets is a recipe for capture by others. Externalization, then, is not generosity. It is what a firm does once the grammar no longer protects the margin — because something else now does.

Key factors and paradigms. Carlota Perez's Technological Revolutions and Financial Capital (2002) identifies each paradigm's "key factor": a cheap, pervasive input — coal, then cheap steel, then low-cost oil, today's cheap microelectronics — whose plummeting relative cost reorganizes the whole economy around it. When the key factor goes cheap, the paradigm diffuses, and a strong bias toward externalization is introduced at exactly that layer.

Rent and platforms. Mariana Mazzucato's distinction between value creation and value extraction, and Nick Srnicek's Platform Capitalism (2017), supply the trap's vocabulary. A platform externalizes a primitive and then meters it. The grammar is free; the toll booth is not.

Amazon → AWS. The internal forcing function was "undifferentiated heavy lifting" — by Amazon's own oft-repeated estimate, engineers spent the majority of project time rebuilding storage and compute foundations before they could ship anything (a figure leadership has paraphrased as roughly two-thirds, not a measured statistic). S3 launched March 14, 2006 at $0.15/GB-month; EC2 followed in beta that August, developed by a team in Cape Town. The foundations became external products, and the S3 API became a de facto standard cloned by competitors. The grammar escaped.

IDM → TSMC. Morris Chang incorporated TSMC on February 21, 1987 with initial capital of about US$220 million (NT$5.5 billion) — the Executive Yuan held 48.3%, Philips 27.6% (a $58M stake plus process-technology IP licences), and ~24% sat with Taiwanese industrial families (Formosa Plastics took 5%). The barrier was fab cost; the foundry externalized fabrication itself, and the Process Design Kit (PDK) became the grammar that let fabless firms — Nvidia, Qualcomm, Apple — design without owning a fab. TSMC externalized the primitive and metered it.

The Escalator is not nothing. It tracks real efficiency gradients — transaction costs, returns to scale, the collapsing cost of replication. Where the synthesis below departs from the original is not in denying the gradient. It is in denying that the gradient acts alone.

Apple silicon. Apple grew its chip team from roughly 500 engineers to 3,000 by 2024 — and sells the result to no one. The capability is a permanent complement to its devices, never a market. Vertical integration as a permanent moat: the escalator with the down button held.

SAP, Bloomberg, x86. Decades of enclosed grammar, rent extracted the whole time, no externalization in sight. Each owned the complementary asset — the install base, the terminal, the instruction set — and so each kept the grammar enclosed precisely because it could. If the resource law were truly system-neutral, these capabilities should have diffused. They didn't. The owner's grip, not the gradient, decided.

Put the climbers and the refusers side by side and the resource-law story wobbles. The same falling replication costs that "forced" S3 open left x86 enclosed for a generation. A law that fires only sometimes is not a law. It is an outcome with a hidden selector.

The deepest evidence for the trap thesis is that the most consequential commons were not left to any natural tendency. They were created by non-market institutions, often by antitrust coercion.

The 1956 AT&T consent decree. A federal court in New Jersey entered a consent decree obligating the Bell System to license its existing patents royalty-free and future patents at reasonable royalties — roughly 8,600 patents, about 1.3% of all unexpired US patents at the time. The transistor became a commons because a court said so, not because a gradient willed it.

TCP/IP. Kahn and Cerf published the TCP design in 1974; the "flag day" cutover to TCP/IP came January 1, 1983; NSFNET (from 1986) provided the backbone and made TCP/IP mandatory before being privatized in 1995. Commercial internet service did not meaningfully exist until 1989. The protocols were published as royalty-free RFCs and never patented — a commons engineered by DARPA and NSF, not discovered by the market.

Xerox PARC — the inverse lesson. PARC invented the GUI, the mouse, Ethernet, and Smalltalk (the Alto, 1973). Xerox failed to capture any of it; value flowed to Apple and 3Com. PARC shows that representation-layer inventions leak by default — they become commons unless actively enclosed, or unless the inventor owns the complementary assets (Teece again). The state did not need to force PARC's externalization. But where a firm did own the complementary asset, enclosure of the grammar persisted for decades — SAP, Bloomberg, x86. The state had to intervene precisely where the market would not.

The resolution is not a compromise. It is a relocation of the claim. Commons form reliably at the cheap-to-replicate grammar layer — the API, the protocol, the PDK, the representation. They almost never form at the capital-intensive substrate layer — the fab, the wet lab, the terminal, the install base. The gradient is real, but it only externalizes the layer that is cheap to copy. The substrate stays enclosed and rent-bearing because owning it is the whole point.

Read the last two rows and the synthesis pays for itself. Where no non-market actor forced the issue and the owner held the substrate, even the grammar stayed enclosed for decades. The commons is not the default. It is what happens when either competition reaches the cheap layer or an institution reaches in and pries the substrate's grip off the grammar.

A synthesis earns its keep by risking a forecast. Point the grammar/substrate distinction at the next century:

Molecular & matter (synthetic biology, materials). The "parts" registries trend open; the wet-lab substrate stays enclosed and rent-bearing. Prediction: commons appears only at the cheap representational layer, and most decisively when a non-market actor forces it. Falsifier: if a capital-intensive substrate diffuses into a true commons with no state or antitrust push, the synthesis is wrong and the pure resource law is vindicated.

Energy & quantum. Expect open standards and reference designs at the interface, enclosure at the plant and the cryostat. The test is not whether something opens — something always does — but which layer, and who had to push.

The forcing function is the desert vigor — the asabiyyah of a team that must build a primitive just to survive. Internal capability and externalization are the founding of the institution. Standardization, margin compression, and rent are the dynasty's comfortable senescence: the moment the moat becomes a market, the cohesion that built it has already spent itself, and the house must summon fresh asabiyyah to climb again. The escalator is real. So is the descent. And the largest commons in the history of the technology — the transistor, the internet — were not gifts of the gradient. They were taken, by institutions willing to pry the grammar loose from the substrate.

The strategic reading for the compute-commons thesis: own nothing of the fab; refuse to let anyone own the language that describes it. Build at the grammar layer, where the gradient is on your side, and never mistake a metered substrate for a commons.

The limits, stated plainly. Survivorship still lurks — the refusers we named are the famous ones; a fuller census might shift the balance. "Replication cost" is used here as a proxy, not a measured variable, and a serious version would operationalize it. Several figures (the patent count, the TSMC cap table, the AWS time estimate) are drawn from public histories and should be verified against primary sources before any print edition. The grammar/substrate line is a heuristic, not a law — its value is that it is falsifiable, and we have said exactly what would falsify it.