Software is done. Hardware is next.

The clearest measure of what AI can do in software is how long a task it can finish on its own. METR, an independent evaluator, times each task by how long a skilled professional takes, then finds the length at which an AI agent succeeds half the time. That length was about 4 minutes for GPT-4 in March 2023. It was 12 hours for Claude Opus 4.6 in February 2026, and 17.4 hours for an early version of Claude Mythos Preview tested that spring. Since 2023 it has doubled about every 129 days. At 80% reliability the best figure is about 3 hours, which is the number to keep in mind when someone says a task can now be delegated. In May, METR said its task suite can no longer reliably measure beyond 16 hours, and in September it stopped updating its public tracker.

The companies writing the most code say the same thing in their own terms. Google's chief executive said in April that 75% of all new code at Google is AI-generated and approved by engineers, up from 50% the previous fall and from more than a quarter in October 2024. Anthropic says more than 80% of the code merged into its codebase in May 2026 was written by Claude, up from low single digits before February 2025. These are lines of code, counted by companies that sell the tools, and neither is a measure of productivity. Anthropic says so itself: its engineers merged eight times as much code per day as in 2024, a figure it calls almost certainly an overstatement of the real gain.

Price and employment point the same way. Epoch AI estimates that the cost of reaching a fixed level of AI performance has fallen about 13 times a year since 2023, and calls that "faster than for any other transformative technology in history." Its benchmarks are math, science and puzzles, not code. In the Bureau of Labor Statistics' employer survey, computer programmers, the occupation closest to writing code to someone else's specification, fell from 132,740 jobs in May 2022 to 92,230 in May 2025, a drop of 30.5%. The bureau expects a further 7.3% decline by 2035, because companies are expected to use AI to automate repetitive programming tasks. And in February, after Anthropic released plug-ins that let its Claude Cowork agent do legal, sales and data work, Reuters reported Thomson Reuters down nearly 18% in one session, in a wider selloff of data and software stocks.

None of this says software is finished as an activity. It says that producing a working piece of code to a clear specification is getting cheap, quickly. That is the first half of the bet, and the evidence for it is strong.

This is not the first time. In 1954, John Backus later wrote, the programmers attached to a computer center usually cost at least as much as the computer, and programming and debugging could account for three quarters of the cost of running one. That cost was his case for FORTRAN. IBM's first report on it, in November 1954, said it "should virtually eliminate coding and debugging." Backus later judged the team to have been far too optimistic. What followed was not the end of programming but its growth into a mass profession.

The latest count suggests the pattern is still running. The economist James Bessen counts a record 2.5 million US software developers in February 2026, more than 400,000 more than when ChatGPT launched. In his figures, real software output grew about 9.3% a year from 2022 to 2025 while output per developer grew about 6%, so demand outran productivity and more people were hired. Gartner forecasts software spending up 15.5% in 2026. Each time code has become cheaper, the world has wanted more of it.

So the bet has to say what is different this time. Two things are. Every earlier tool raised the level at which a person wrote instructions, but a person still turned intent into a formal language and checked the result. Agents now do most of the turning and much of the checking, over tasks measured in hours. And the labor data have started to split, as section 4 shows: programmers down, developers still up, entry-level hiring almost gone.

Farming is the closer precedent. According to the US Department of Agriculture, farm output grew about 1.46% a year from 1948 to 2021 while the labor used in farming fell about 1.93% a year, and the use of durable equipment and purchased inputs rose. In 1900 about 39% of Americans lived on farms. By 2000 it was about 1%. Farming did not end. It stopped needing most of the people, and the spending moved to machines and inputs. That is the version of done this paper defends: an activity that keeps growing while the labor in it stops being scarce, and the money moves to what it runs on. For AI, what it runs on is chips, memory, power and buildings.

First, the measured gains are smaller than the felt ones. In METR's randomized trial in early 2025, experienced open-source developers working on their own mature projects took 19% longer with AI tools, while believing they had been about 20% faster. A follow-up begun later in 2025 could not produce a reliable estimate, because many developers no longer wanted to work without AI. METR now thinks they are probably faster, by an amount it cannot yet measure. An analysis of about 100,000 developers presented by a Stanford researcher put gains at up to 30 to 40% on simple new code and 0 to 10% on complex existing systems. That matters because most software money is spent on existing systems: US federal agencies report spending about 80% of their IT budgets on operating and maintaining what they already have.

Second, demand for software has been elastic for seventy years and still is. Bessen's record employment is one sign. The Bureau of Labor Statistics still projects software developer jobs to grow 10.2% to 2035, three times the rate for all jobs, although it has cut that from 17.9% in two revisions. Indeed's index of US software job postings has risen 28% from its May 2025 low. Investors do not believe software is dead either. The BVP cloud software index, down 31% for the year at its April low, stood 25% up for 2026 on 9 October. The damage stayed with companies whose product an agent can now do: Adobe and Thomson Reuters were still down about 32% and 21% for the year on 12 October.

Third, cheap code is not yet trusted code. In Veracode's 2026 tests across more than 100 models, AI-written code passed its security checks 56% of the time, unchanged from a year earlier, while nearly all of it was syntactically correct. When maintainers reviewed agent-written changes for METR in 2025, none of the 15 they examined was ready to merge as it stood, including those that passed the project's tests. Google's 2025 DORA research found AI adoption now goes with faster software delivery, and still with less stable delivery.

Taken together, these say the bet is early, not wrong. Code is getting cheaper to produce and more expensive to trust. That moves the scarce work to deciding what to build, fitting it into what already runs and answering for the result. It does not yet shrink software to a small layer.

Inside software, the second half of the bet shows in who is being hired. In the first quarter of 2026, 69.3% of US software development postings on Indeed were for senior roles and 4.5% for entry level, the lowest entry share of any sector Indeed tracks. Senior roles drove 71% of the rebound in postings in the year to May. Stanford's Digital Economy Lab finds that employment of 22 to 25 year olds in the occupations most exposed to AI is 19% below where it would have been had it tracked less exposed work, through fewer hires rather than layoffs, while experienced workers show no such gap. Its authors describe this as a pattern, not a measured effect of AI.

The roles growing fastest sit between a model and a business. At the large tech firms SignalFire tracks, the share of hiring going to forward-deployed engineers, who embed with a customer to make a system work in place, is up 30% since ChatGPT launched, and the share going to AI and machine learning engineers is up 39%, while front-end engineering is down about 25%. In May, Anthropic launched an enterprise AI services firm with Blackstone, Hellman & Friedman and Goldman Sachs to supply that kind of engineer, saying enterprise demand was outpacing any single way of delivering it. Anthropic's own account of where people still lead is "research taste and judgment": choosing which problems matter, which results to trust, and when an approach is a dead end.

That is the bet's architects and integrators, and the data support it as a direction for the work. They do not support it as a direction for the money. Gartner expects IT services spending to grow 5.3% in 2026 against 15.5% for software, and Accenture, the largest integrator, grew revenue 5% in its 2026 fiscal year and guides 3 to 6% for the next. Integration is where the scarce labor is going. On this evidence, it is not where the margin is going. There is also a hole in the model: a profession filling 4.5% of its openings at entry level is not training the seniors it will need.

Follow the money AI is actually spending and it arrives at factories. Microsoft, Alphabet, Amazon and Meta paid about USD 376 billion for property and equipment in 2025 and about USD 295 billion in the first half of 2026 alone, up about 84% on the same half a year earlier, by this research's sum of their filings. Their plans for 2026 add up to roughly USD 720 to 760 billion. When Amazon raised its 2026 plan from USD 200 billion to USD 220 billion in July, its chief executive put the increase down mainly to the higher cost of memory, and said Amazon still would not have enough capacity to meet demand.

The memory makers are where that shows. Micron's quarter to 3 September brought revenue of USD 54.23 billion, against USD 11.32 billion a year earlier, at a gross margin of 86.8%, up from 44.7%. Power is the same story on a slower clock. GE Vernova holds 116 gigawatts of gas turbine backlog and reserved production slots, against output of 20 gigawatts a year: nearly six years of production. Wood Mackenzie put average US lead times for power transformers at 128 weeks in mid-2025, and 143 weeks for the step-up transformers that connect a power plant to the grid. Gartner's fastest-growing line for 2026 is data center systems, up 62.5%.

Economics predicts this. Clayton Christensen called it the conservation of attractive profits: when one stage of a product becomes modular and cheap, profit reappears at the adjacent stage that is still hard. The Stanford economist Charles Jones puts numbers on the same idea. Software spending is about 2% of US GDP, so even infinite amounts of what software does today would raise GDP by only about 2%. In his model, automation raises the price of the remaining weak links, the tasks that stay slow, and making goods is among the tasks he says still need people. There is evidence for the mechanism from the last time experimentation got cheap. After Amazon Web Services cut the cost of trying a software idea, venture money moved toward software and away from complex physical technology, in research by Michael Ewens, Ramana Nanda and Matthew Rhodes-Kropf. Advantage competes away fastest where experiments are cheapest. Today that is software, for everyone.

Hardware is not uniformly the next frontier, and the evidence against a broad reading is strong. Investors have not moved: AI took 82.7% of US venture dollars in 2026 to the end of September, according to PitchBook and the NVCA. Where money has gone into physical AI, prices run ahead of revenue. Figure AI was valued at USD 39 billion in September 2025 without disclosing revenue, and Unitree, a profitable Chinese robot maker, listed in Shanghai in August at about 36 times its 2025 revenue and opened 629% higher. US factories ran at 75.7% of capacity in August, below their long-run average of 78.2%. Spending on new US manufacturing buildings in August was 19% lower than a year earlier, and Kearney's reshoring index is still negative.

The scarcity margins of 2026 went to companies whose capacity took decades to build: memory fabs, chip foundries, turbine lines. A newcomer does not build a fab to join them. Design is getting cheaper too, which pushes the bottleneck further toward production. Quilter says its AI laid out an 843-part Linux computer in 38.5 hours of human time, against 428 hours quoted for doing it by hand. A board house still had to make it.

For a small or mid-size company with a product to build, what is scarce is not machines. It is qualified, scheduled, trustworthy capacity at small and changing volumes. Protolabs counted 6.1% fewer customers in 2025, which it put down partly to a strategy of winning larger orders. Xometry, a marketplace that routes jobs to about 5,000 suppliers, grew marketplace revenue 45% in the second quarter of 2026, and Siemens invested USD 50 million in it. MIT's study of 150 production start-ups found that many moved production abroad when they reached commercial scale, because that is where the capital, the production capabilities and the lead customers were. In regulated products the clock is slower still: most EU notified bodies take 13 to 24 months to issue a new certificate under the medical device regulation.

That is a narrower claim than hardware is next, and it is the one this site is built on.

The bet has an uncomfortable corollary for Moduloa. Its factory OS, its quoting and its routing are software. If code is cheap, a competitor can rebuild them in months, and Joel Spolsky's old observation applies: software commoditizes hardware far more easily than hardware commoditizes software. In Christensen's account of the personal computer, the profit went to the operating system and the processor, not to the company assembling the machines. Moduloa cannot claim that nobody else could build what it builds, and a Moduloa that only sold hours on machines, or only routed orders, would be on the losing side of this.

What cheap code cannot copy is knowledge that has to be gathered, and the people who gather it. Manufacturing know-how sits scattered across thousands of small shops and in the heads of the people who run them. Hadrian's founder, Chris Power, has estimated that about 3,000 small US machine shops make the precision parts for aerospace and defense, with about USD 40 billion a year in revenue between them. Xometry, which routes work to shops like these, describes its market as massive, highly fragmented and regional. And the knowledge is leaving: Deloitte and the Manufacturing Institute project that 1.9 million of the 3.8 million US manufacturing jobs needed by 2033 could go unfilled. Hadrian's answer, as this site's thesis benchmarks it, is to build and run its own software-driven factories. Moduloa's answer is to gather the shops, the people and what they know into one shared framework, and to build it in the open.

So the foundation starts now, and it starts open. The research and documents here are published under CC BY 4.0 so anyone can use and correct them, the register of who is on board is public, and the first contributing company is aboard to build the factory OS data model as one versioned repository. Moduloa builds with AI too, and says so: the research behind this paper was gathered and checked with AI agents, against primary sources. If code is cheap for everyone, the advantage is not in hiding it. It is in being the place where the most people bring what they know, and where every production run adds to a record anyone can audit. Each contributor makes the framework more useful to the next one, and that kind of advantage accumulates instead of competing away.

The register already carries this bet. P-18 says the durable moat proves to be the framework, meaning standards, data, certification and training, not the robots. P-14 adds the people: a new profession, the robot-era production technologist, working across machines, robots, fixtures, software, quality and data, the physical counterpart to the system integrators of software. This journal found that integration, not the robot, is the expensive part of automation, which is why those people matter. The software labor data carry a warning for P-14: an industry that stops hiring juniors runs out of seniors.

Three cautions belong here. Today the foundation is one founder, one contributing company and a body of published research, and the advantage described above exists only if many more people come. The International Federation of Robotics says most humanoids remain prototypes or pilots, and that conventional industrial robots are economically superior for production today, so the humanoid floor Moduloa designs for is a target, not a present fact. And others see the same seam: Y Combinator's requests for startups this fall include new operating systems for the physical world that manage robot and human labor together. Today's field note makes the balance-sheet version of the point: a distributed network does not remove the fixed cost of a factory, it decides who carries it.

The capability curve stalls. If the task length frontier agents complete at 80% reliability is still about 3 hours at the end of 2027, writing code has not become cheap enough to stop being scarce. METR no longer updates its tracker, so this needs whatever measurement replaces it.

Entry-level software hiring comes back. If the entry share of software postings climbs well above today's 4.5%, and computer programmer jobs stop falling, the old pattern has reasserted itself and cheaper code is simply making more programmers.

Application software keeps its margins. If the companies hit hardest in 2026 recover and seat-based pricing holds, the market has decided that agents complement application software rather than replace it.

The hardware margins were a cycle. Memory is the classic boom and bust industry. If Micron's gross margin falls back toward its 2025 level of about 40% as new fabs arrive, and turbine and transformer lead times return to normal, the money that moved to hardware in 2026 was a shortage, not a shift.

Production itself becomes a commodity. If robots make any cell able to do any job, capacity gets sold by the hour, as David Cahn of Sequoia argues GPU computing already is, and the margin moves on: to energy, land, components, capital or whoever owns the customer, not to a framework.

Certification cannot travel. If regulators and customers keep tying an approval to one manufacturer and one site, a network of certified, portable capacity has nothing to sell. P-27 and P-10 in the register test this directly.

Nothing here scores a claim. The register is reviewed on its own schedule, under its own sourcing rules.

Software is not finished. Writing it is. In 2026 code stopped being the scarce input, and the evidence for that is strong. What stays scarce around it is judgment, integration and trust. The jobs are moving there, but so far the money is not. The money moved to physical capacity, and mostly to the companies that already owned it.

Moduloa's bet sits in the gap that leaves: access to qualified production capacity for everyone who does not own a fab, through a framework whose value is not its code but what gathers around it. That means the knowledge contributors bring, the production record, certification that travels between sites, and the people who make a cell produce, built in the open so more can add to it. Hardware is next only for whoever can get it made. None of this exists yet. Moduloa is at Stage 0, with no hub and no customer, and the six tests above are how it will know if it is wrong. The full plan is in the working thesis, and what is being built is on Manufacturing.

METR: Measuring AI Ability to Complete Long Software Tasks ·METR: time horizon benchmark results, version 1.1 (data file) ·METR: Frontier Risk Report, February to March 2026 ·METR: Time horizons ·Google: Sundar Pichai shares news from Google Cloud Next 2026 ·Semafor: Google CEO says 75% of company's new code is AI-generated ·Google: Alphabet Q3 2024 earnings remarks ·Anthropic Institute: When AI builds itself ·Epoch AI: The plunging price of thought ·BLS: Occupational Employment and Wage Statistics, national estimates 2022 to 2025 ·BLS Occupational Outlook Handbook: Computer Programmers ·BLS: Occupational projections, 2025 to 2035 ·Reuters: Anthropic's new AI tools deepen selloff in data analytics and software stocks, investors say ·John Backus, ACM History of Programming Languages, 1978: The History of FORTRAN I, II, and III ·James Bessen, TPRI: Why AI hasn't killed software developer jobs ·Gartner: Worldwide IT spending to grow 14.2% in 2026, totaling USD 6.37 trillion ·USDA ERS: U.S. Agriculture Production Grew Steadily From 1948 to 2021 as Productivity Increased ·USDA ERS Amber Waves: The 20th century transformation of U.S. agriculture, data feature ·METR: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity ·Reuters: AI slows down some experienced software developers, study finds ·METR: We are Changing our Developer Productivity Experiment Design ·METR: Research Update, Algorithmic vs. Holistic Evaluation ·Yegor Denisov-Blanch (Stanford), AI Engineer World's Fair: Does AI Actually Boost Developer Productivity? ·GAO-25-107795: Agencies Need to Plan for Modernizing Critical Decades-Old Legacy Systems ·Indeed Hiring Lab via FRED: Software development job postings in the United States ·FRED: BVP Nasdaq Emerging Cloud Index ·Reuters: US software stocks scale fresh 2026 highs as AI disruption worries fade ·Veracode: LLMs Are Getting Smarter, But Not Safer, 2026 GenAI Code Security Report ·Google Cloud: Announcing the 2025 DORA Report ·Indeed Hiring Lab: The Labor Market Is Tilting Toward Seniority ·Indeed Hiring Lab: AI and Job Postings, From Destruction to Creation? ·Stanford Digital Economy Lab: No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19% ·SignalFire: State of Talent Report 2026 ·Blackstone: Anthropic partners with Blackstone, Hellman & Friedman and Goldman Sachs to launch enterprise AI services firm ·Accenture: fourth quarter and full year fiscal 2026 results, Form 8-K ·SEC EDGAR: Alphabet, payments to acquire property, plant and equipment (one of the four series summed) ·Meta: second quarter 2026 results, Form 8-K ·Fortune: Andy Jassy said Amazon will spend USD 220 billion this year, and still won't have enough capacity to meet demand ·CFO Dive: Microsoft holds line on AI spending plans ·W.Media: AI infrastructure demand pushes Alphabet's 2026 capex guidance to USD 205 billion ·Micron: fourth quarter and full year fiscal 2026 results, Form 8-K ·GE Vernova: second quarter 2026 results, Form 8-K ·Turbomachinery Magazine: GE Vernova gas turbine backlog hits 116 GW as power orders more than double ·Wood Mackenzie and American Clean Power: Making the connection, executive summary ·EEPower: Transformer supply chain woes persist as energy demand grows ·Harvard Business Review: Breakthrough Ideas for 2004, including Christensen on the law of conservation of attractive profits ·Christensen, Raynor and Verlinden, HBR: Skate to Where the Money Will Be ·Charles I. Jones, NBER w34779: A.I. and Our Economic Future ·Timothy Taylor: Growth Effects of AI, Modest for a Decade or Two, At Least ·Ewens, Nanda and Rhodes-Kropf, NBER w24523: Cost of Experimentation and the Evolution of Venture Capital ·PitchBook and NVCA: Q3 2026 Venture Monitor ·Figure: Series C ·Bloomberg Tax: Robotics Startup Figure AI Valued at USD 39 Billion in New Funding ·Reuters via New Straits Times: Unitree soars in Shanghai debut ·Shanghai Stock Exchange, China Daily: Unitree prospectus figures ·Federal Reserve G.17: Industrial Production and Capacity Utilization ·US Census Bureau: Monthly Construction Spending, August 2026 ·Kearney: 2026 Reshoring Index ·Quilter: Project Speedrun ·VentureBeat: Quilter's AI just designed an 843-part Linux computer that booted on the first try ·Proto Labs: Form 10-K for 2025 ·Xometry via CIMdata: Xometry Reports Record Second Quarter 2026 Results ·Xometry: Form 10-K for 2025 ·Reynolds, Samel and Lawrence, MIT Press 2014: Learning by Building ·European Commission: 20th notified body survey on MDR and IVDR certification ·Joel Spolsky: Strategy Letter V ·CNBC: Machine-parts start-up Hadrian raises USD 90 million as it seeks to shake up the aerospace supply chain ·The Manufacturing Institute and Deloitte: Manufacturers Need as Many as 3.8 Million New Employees by 2033 ·IFR: Five Million Robots now Operate in Factories Globally ·IFR: World Robotics 2026, Industrial Robots, Executive Summary ·Y Combinator: Requests for Startups ·David Cahn, Sequoia Capital: AI's USD 600B Question ·Moduloa: Working thesis v0.2 ·Moduloa: The robot was never the expensive part ·Moduloa: Someone has to own the hall ·Moduloa: The work went to the integrators. The money went to the factories.

What these figures are, and are not. METR's time horizons come from its published data file and reports; METR says its suite cannot reliably measure beyond 16 hours, and that a 50% horizon is not a safe delegation threshold. The shares of AI-written code are the companies' own counts of lines, not measures of productivity, and come from companies that sell AI tools. Epoch AI's price index is built on math, science and puzzle benchmarks, not code. The occupational counts are BLS employer-survey May estimates, which pool three years of samples and lag. The Stanford, Indeed and SignalFire findings are descriptive, and the Stanford authors caution against reading them as caused by AI. Bessen's productivity and output figures are his own assembled series, which he calls suggestive. The four companies' equipment spending is this research's sum of cash paid for property and equipment in their filings, excluding finance leases; their 2026 plans are guidance on differing bases and are not comparable to it as a growth rate. The Adobe and Thomson Reuters year-to-date moves and the cloud index change are computed from daily closing values. Unitree's revenue multiple is computed from its offer valuation and its prospectus revenue. The Quilter hours and every design-tool gain cited are vendor figures. Christensen's personal computer account and Spolsky's asymmetry are arguments, not measurements. The Backus account was written by him in 1978. The count of small US machine shops and their revenue is Chris Power's own estimate as reported by CNBC in 2022, and the unfilled-jobs figure is a projection, not a measurement. Absence findings, such as no rigorous study of how contract manufacturers rank small customers, are negative results from research completed on 12 October 2026, not proof that none exists.