Across the first four parts I wrote about the deadlines in AI's economy (the K3 weight release, the AI companies' finances, Microsoft's earnings) and the deadlines in geopolitics (Ukraine, Iran, China). Both were cross-sections of the end of the age of concentration. Part five is about what grows after it ends. The conclusion first: the IT industry will shrink. And because the price of software and components has fallen, the materials industries and the manufacturers who assemble gain the advantage.
The Starting Point — The Price Composition of a Product Changed
A product is made of software and hardware. Software here is not only code. Design, procedure, documentation, standards — the entire informational component of a product. The AI revolution is the event that drove the cost of that component toward zero all at once. Apps and websites can now be built by asking a frontier AI. An open model like Kimi K3, whose weights are released on July 27, puts that capability within anyone's reach.
Meanwhile components got cheap too. Through standardization and global sourcing, electronic parts and general-purpose boards have moved close to commodities. So of the four terms in a product's price — software, components, materials, assembly — the first two head toward zero. What remains is materials and assembly.
The Decline of IT — What Disappears Is Not the Work but the Status of Being an Industry
To avoid misunderstanding: the volume of IT work does not disappear. What disappears is the state in which IT is bought and sold as an independent industry.
The precedent is electricity. In the age of electrification, an industry that sold electrification flourished — specialists who installed motors in factories and advised on how to use power. When electrification was complete, that industry vanished. Electricity did not vanish. It became an internal function of each factory. IT now stands at the entrance to the same path. If anyone can build an app, systems revert from a product you buy to an internal function you make. Orders placed with systems integrators, per-seat SaaS, five-year bundled maintenance — these are rentals from an era when software was too expensive to build yourself, and once that premise is gone, the market disappears not through announced cancellations but through the quiet form of no longer being chosen for the next project.
Let me state the trigger of that quiet disappearance precisely. It is that AI made verification nearly free. Defects in a specification, an inflated estimate, the true cost of a SaaS, the hollowness inside a maintenance contract — everything the buying side could not see is visible in a few hours if you let an AI read it. A business standing on information asymmetry cannot stand once the asymmetry is gone. A contradiction is an unpaid invoice, and payment could be deferred only while verification was expensive. From the moment it became free, settlement begins. That is where we are.
What survives on the systems side as an industry is a thin layer: a small number of published standards and kits, and small suppliers who help apply them at each company — not contract work extracting a large sum from one client, but small, continuing payments from many. What is sold is not code. Code is free. What sells is design judgment, standards, operational practice, and trust.
Why Materials and Assembly Win
Three reasons. First, the logic of the residual. If software and components go to zero, the share of materials and assembly in a product's price rises mechanically. Second, the demand effect. When the price of the final product falls, volume increases, and total demand for materials and assembly rises with it. Third, and decisively, the physicality of the barrier to entry. Plants, accumulated formulations and process tuning, sheet metal and thermal design, the fitting of parts, the tacit skills of the shop floor — none of these can be replicated by AI. In a world where the software barrier is gone, only the physical barrier remains a barrier. Demand rises and the wall on the supply side does not fall. The advantage is structural.
Fabless was once the clever strategy because software was scarce. In a world where software is free, it inverts. Today, building the steel box is the harder thing.
What This Means for Japan and Its Regions
This transition favors Japan. The defeat of Japanese manufacturing was not about quality. It was about being unable to produce software, and about cost. The first is resolved by software self-sufficiency through AI; the second by unlocking the structurally low costs of the regions — which until now were sealed off by that single inability to produce software. And the axes that remain, quality and materials, are the domains Japan still holds. Chemicals, specialty steel, carbon fiber, electronic materials — the market-share tables for materials can be read as a roster of this transition's beneficiaries.
Companies out in the countryside, the ones without excess, are going to beat Tokyo companies on cost from here. The only reason they could not before is that they could not produce software.
The Shape of the New Industry
So what is the "new industry"? I think it appears not as a conspicuous invention but as a change of placement. Each organization holds AI in its own box (today, a machine with 128 to 192GB of unified memory), accumulates its own information, builds its own systems, and defends itself. Product development and internal systems building become the same work — once code is a thin layer of open source and AI, what you sell and what you use internally are two faces of the same software. And being able to secure yourself becomes a qualification for doing business, a qualification for selling directly to consumers, and a condition of business continuity. Just as quality assurance once did.
The new industry, in other words, is manufacturing that is self-sufficient in software, earns in the physical, and defends itself. The name is not new. The contents are.
Observation Points
Whether this thesis holds can also be tracked in numbers. On the IT side: not the churn rate of SaaS but the new-adoption rate — the speed at which it stops being chosen for new projects — plus systems integrators' order backlogs and the trajectory of per-seat pricing. On the manufacturing side: the margins of materials and assembly companies rather than components, and the share of small and mid-sized manufacturers doing their own development. And the litmus test running through all of it is the same one named in part three — the day a Sonnet-class model fits in a 128-to-192GB box. On that day the last barrier to software self-sufficiency falls, and every clock in this article speeds up at once.
One analogy to close. This transition resembles the printing press more than the industrial revolution. The press destroyed the copyist's trade but multiplied, by orders of magnitude, the number of readers and writers and the industries of paper and ink. IT as an industry occupies the copyist's position. What disappears is the work of transcription, not the total volume of what gets written. Those who receive the increase are paper and ink — that is, materials, and the ones who assemble.
A disclosure: this piece was developed in dialogue with an Anthropic model (Claude). Anthropic belongs to the IT industry whose contraction is predicted here. Read it with that discounted.
Related
This is the fifth and final part of the series.
- Blog What to Actually Watch When Kimi K3's Weights Drop on July 27 — part one: the checklist for the 27th, and the empty tier
- Blog The Reality of AI Companies, and What Kimi K3 Does to It — part two: revenue, losses, and data-center debt in numbers
- Blog What to Watch in Microsoft's July 29 Earnings — part three: where the shared litmus test (a Sonnet-class model in a box) was identified
- Blog Three Reckonings Are Closing In, All in the Same Week — part four: three deadlines outside AI arriving the same summer
- Blog Three Transitions in Software — the structural map this piece assumes
- Blog Internal Business Systems Are No Longer Something You Outsource — the practical version of "systems revert to an internal function"
- Blog When Fable 5 Returns, Do This First — Verify Every System You Run — the organizational version of "verification becomes free"
References
- Parts one through four of this series (what to watch in the K3 weight release / the reality of AI companies and K3 / Microsoft's July 29 earnings / Ukraine, Iran, China)
- For Moonshot AI's Kimi K3 (released 2026-07-16, weights announced for 2026-07-27), see the references in part one
- Anthropic Just Passed OpenAI in Revenue. While Spending 4x Less to Train Their Models (SaaStr) — https://www.saastr.com/anthropic-just-passed-openai-in-revenue-while-spending-4x-less-to-train-their-models/
- H100 GPU Cost In 2026 (CloudZero) — the commoditization of components and compute — https://www.cloudzero.com/blog/h100-gpu-cost/