Every general-purpose technology eventually sells the foundations it was built on. Once you have enough cases, the sequence stops looking like a story and starts looking like a law — and a law can be pointed at the next century.
A reader watching Anthropic ship Claude Design — a whole product, with tutorials, arriving right behind Cowork's move on enterprise knowledge work — asked the natural question: why pick a fight on the designers' turf at all? The honest answer is that it isn't a fight. It's a survey expedition. And the only way to see that clearly is to find the closest historical example, not as a one-to-one map but as a foothold: a place where we already watched the whole thing happen and recorded the dates.
The example almost everyone reaches for is Google. It is the right one — provided we tell it accurately, because the accuracy is where the pattern lives.
Search at planetary scale was a problem nobody had a finished answer to in 1998, so Google had to invent the foundations underneath it. Those foundations were published as papers, and the dates matter: the Google File System in 2003, MapReduce in 2004, Bigtable in 2006. None of these was a product. Each was internal plumbing built to be more general than the thing that motivated it — which is the first quiet tell of the pattern.
Then came the proof. Gmail (2004) and the browser-based productivity that became Google Docs (2006) were not just applications; they were evidence that the infrastructure generalized — that the same plumbing serving search could serve arbitrary internet-scale software. Only after that proof did Google productize the basement it had been living in: App Engine in 2008, Compute Engine in 2012, and the assembled Google Cloud Platform. The cost center became the product.
The consumer apps proved the infrastructure worked. The platform was the act of selling that infrastructure to everyone else.
The final move is the one most people forget, and it is the most important. The papers escaped. Hadoop and HBase are open-source reconstructions of MapReduce and Bigtable, and they seeded much of the cloud ecosystem that followed. What began as one firm's secret became the public grammar of an industry.
Here is where the LLM labs depart from the script — and the departure is the most revealing thing about them. Google externalized its raw primitive last. The labs externalized theirs first. The breakthrough — cognition sold by the token — was itself the immediately sellable artifact, so the API came before the apps rather than after them.
Which means Code, Cowork, and Design are not the destination. They are Gmail and Docs: probes proving a deeper capability generalizes. But the capability being mapped is no longer cognition, which is already sold. It is agency — the handoff of intent. Code maps delegate a coding goal. Cowork maps delegate knowledge work. Design maps iterate on a conceptual artifact through dialogue. Each first-party app is a cartographer planting a flag on a different axis of a territory that has not yet been cleanly externalized.
The journey is not the product. The map of the journey is the product — and it externalizes next as agent SDKs, computer-use APIs, and open protocols like MCP.
Design feels like a frontal assault on a tangential audience only if you mistake the probe for the prize. It is reconnaissance on the generative-iteration frontier, run before that frontier is sold as a platform.
Strip the specifics away and the same five movements appear, in order, every time. This is the escalator. Hover or tap a step.
A single hard, high-value vertical product demands a new primitive of composition that does not yet exist.
The firm builds infrastructure deliberately more general than the product that motivated it.
First-party apps prove the primitive generalizes — and discover the abstractions outsiders will need.
The internal capability becomes the product. The moat becomes a market; the firm's boundary moves.
Protocols and open source crystallize. Margins compress, and the unit of composition rises one altitude.
One case is an anecdote. The reason this reads as something closer to science is that the same five stages recur in industries that share no people, no decade, and no underlying physics. Below: four externalizations, read against the identical scaffold. Select a column to isolate it.
| Stage | Google → Cloud | Amazon → AWS | Foundry / Silicon | Electricity / Grid |
|---|---|---|---|---|
| 01Forcing function | Web search at planetary scale, with no off-the-shelf answer. | Surviving holiday-peak retail load without buckling. | The soaring cost of owning a fab made integrated manufacturing untenable for new entrants. | Factories needed reliable power and built isolated on-site dynamos to get it. |
| 02Internal capability | GFS '03, MapReduce '04, Bigtable '06 — plumbing built to generalize. | A service-oriented internal platform of storage and compute primitives. | Process mastery, historically locked inside integrated device makers like Intel and TI. | Generation expertise and the central station — Edison's Pearl Street, 1882. |
| 03Probe by application | Gmail '04 and Docs '06 prove internet-scale software runs on it. | Amazon.com itself runs on the primitives — the infra eats its own cooking. | TSMC '87 proves fabrication can be decoupled from design entirely. | Insull's Chicago Edison proves load diversity and metering beat self-generation. |
| 04Externalization | App Engine '08 → Compute Engine '12 → GCP, sold to all. | S3 and EC2 '06 — the plumbing sold to outside developers. | The foundry + Process Design Kit sell fabrication as an interface to atoms. | Power sold as a metered utility delivered over the grid. |
| 05Standardization | Hadoop and HBase clone the papers; cloud primitives commoditize. | "Cloud" becomes a category; the S3 API becomes a de facto standard. | The fabless model + shared EDA/PDK toolchains become the industry's grammar. | AC standardization and interconnection make power ambient and universal. |
Three of these had nothing to learn from the others. Insull was metering electricity in Chicago decades before a transistor existed; Morris Chang's pure-play foundry predates the public web. When unrelated systems converge on the same five-stage structure under their own local pressures, you are no longer looking at imitation. You are looking at a convergent solution to a recurring problem — which is exactly the signature of a real underlying mechanism rather than a fashion. Deep technology becomes legible the moment you have enough independent data points for the pattern to survive the removal of any one of them.
Each turn of the escalator does the same structural thing: it industrializes a layer of what Fred Brooks called accidental complexity into reusable substrate, freeing the next generation to wrestle only with the essential complexity of a higher altitude. The result is a single ladder — the rising unit of composition that builders manipulate. Tap any rung.
The hard-won internal capability of one generation becomes the cheap API of the next. The thing you sell at the top of one cycle is the thing you take for granted at the bottom of the next.
Each externalization buries a layer of accidental complexity so completely that builders forget it was ever a problem — and immediately start composing at the level above it.
If the mechanism is real, it predicts the same way a physical law does: nobody wins by selling the raw physics. Google never sold disks; Anthropic never sold matrix multiplications. The sellable thing is always the abstraction that hides the mess. Read against three frontiers, the escalator tells you where the value will crystallize.
The MapReduce of quantum will be a programming model that hides decoherence — not a faster qubit.
Some firm solves one internal problem first — molecular simulation, logistics, optimizing its own AI training — and builds the error-correction stack to do it.
Error-corrected logical qubits plus a hybrid classical-quantum compiler. The buyer never touches decoherence, exactly as the cloud buyer never touches a disk controller.
A published logical-qubit programming model that open source promptly clones — the Hadoop-of-quantum. That is Stage 05 arriving.
Whoever owns the logical-qubit cloud becomes the GCP of its altitude — unless the model layer externalizes as a protocol instead.
The product is never the nanobot. It is the foundry plus the inverse-design compiler — Matter-as-a-Service.
First-party pharma and materials products become the probe-apps — Stage 03 — proving a closed design-fabricate-verify loop before the loop itself is sold.
A foundry that compiles intent — "a material with these properties" — to structure, fabricates it, and verifies it. TSMC externalized fabrication for chips; this does it for atoms.
A shared inverse-design representation — the Process Design Kit for molecules — that decouples the designer from the fab. Fabless, for matter.
The owner of the foundry and the PDK, precisely as the fabless era handed power to whoever held the process — unless the kit is published as a commons.
Cheap energy is not a rung on the ladder. It is Carlota Perez's key factor — the abundant cheap input that flips a whole paradigm from installation into deployment.
Cheap iron, cheap oil, and cheap microchips each defined an era not as a product but as the input that made everything above them affordable. Cheap energy is the next such factor.
It is what turns quantum cooling, molecular fabrication, and trillion-token inference from boutique into ubiquitous. Without it the rungs above stay rich-world luxuries.
Whoever solves energy for their own crown-jewel clusters externalizes the capability that mattered: siting-and-grid-orchestration as a platform — the seam the compute-commons thesis circles.
Insull did exactly this with electricity a century ago: solve generation for yourself, then sell the orchestration as a metered utility to everyone.
The kicker is recursion, and it is why the next century compounds faster than this one. Each layer is built with the externalized output of the layer below: AI designs the quantum error-correction codes and the nanostructures; quantum accelerates the AI training and the molecular search; nanotech fabricates cheaper energy harvesters and faster chips; cheap energy makes all of it affordable. A flywheel of flywheels, climbing toward a rung where the thing sold is no longer a token or a service but an outcome with a proof attached.
Here is the one thing the mechanism is silent about, and it is the thing that matters most. The escalator is structurally neutral about who ends up owning the externalized layer. By default it concentrates: whoever owns the logical-qubit cloud, the matter foundry, the energy fabric becomes the AWS of their altitude. Clayton Christensen's conservation of attractive profits is just the escalator's accountant — as one layer commoditizes, value integrates into whoever sits one rung above it. That is precisely why a model lab climbs into Design and Cowork as the API beneath it commoditizes.
But the web is the counter-example written into the ladder itself. HTTP and TCP/IP prove an externalized layer can arrive as a protocol-commons rather than a private platform. The pattern guarantees externalization is coming at every altitude. It does not guarantee whether the foundry, the qubit cloud, and the energy fabric arrive as rent-extracting platforms or as commons. That is not physics. That is a choice, and it has an expiry date.
Capability is forged under the asabiyyah of solving a founding problem, institutionalizes into platform, then ossifies into rent.
Ibn Khaldun would recognize the timing exactly. The window to install a commons is during externalization — after the capability is proven but before the incumbent's gravity sets. Once the foundry is a fortress, the argument is over. This is the whole wager of a compute commons: not to resist the escalator, which is futile, but to ride it deliberately onto the one rung where the protocol layer can still be claimed for the public before it is enclosed.
So when the next firm ships a strangely specific first-party app on the edge of a capability it hasn't named yet, do not read it as a land grab on someone's turf. Read it as a survey stake. The territory is being mapped. The only open question — the only one the pattern leaves to us — is who will be allowed to live there.