Research in the open. Each week, a deep-researched note on where humanoid robotics, manufacturing economics, reshoring, and factory software actually stand, and what it means for the Moduloa thesis. Once a month, a recap of what mattered. Sourced, dated, and open to correction.
This is Manufacturing's journal, the flagship's weekly research. Each project on the index keeps its own; the Academy journal is the first sibling.
One investigation, four instruments. The thesis states the bet. The register freezes it into dated claims, scored in public. The tracker follows its premise: whether humanoids are becoming ready for factory work. The Journal records the evidence as it arrives.
How these are made: most notes are researched and written by an AI research agent that runs each week under written rules, and the monthly recaps by one that runs on the first of the month. Both publish without a human edit in between. The rules: two independent sources for any claim that is not obvious, no source, no claim, and a note never scores the register. The other notes were written in working sessions with the same tools. Each note carries Sondre's name because the rules, and the responsibility for what is published, are Sondre's. The method page says more.
When AI made code cheap, software hiring tilted to senior and integration roles while the margin went to the makers of memory, turbines and transformers, mostly companies that already owned the capacity; the full argument is in a new thesis companion. Read it →
An AI-run machining hall in Karlsruhe went insolvent while a franchise network that mostly signs up other people's shops reached 16 listed sites and built a flagship of its own: a network does not remove the fixed cost of a factory, it decides who carries it. Read it →
Four notes that each found a piece of the missing framework, a measurement layer, two standards and an honest setup bill, and found every piece one step away from the qualified process and the people who make a cell produce. Read it →
Ford is spending USD 2 billion and well over a year to give one plant a new job, and humanoid robots, physical AI and digital twins each promise to cut that cost: BMW says its twin now checks a new model in three days, but an outside benchmark put the top-scoring model it tested under 50%, and this research found no proven job moving to a factory someone else owns. Read it →
Agility Robotics' merger filing gives the first audited figures from a US humanoid maker: USD 1.78 million of 2025 revenue across nine facilities, a robot that costs about USD 150,000 to build and rents for USD 8,500 a month, and a USD 300 million anchor order from one related party paid partly in warrants. Read it →
A desktop robot cell reported at around USD 5,000 a year ago now starts at USD 11,950 for one arm, installed and configured, and a garage full of cheap equipment cost two founders 90% of their time, which is what BCG's figure of 75% of lifetime cost in setup and reengineering predicts: the hardware was never the barrier. Read it →
A Chinese national standard for robot intelligent control took force on 1 September and its four layers are all inside the machine, while ISO 21423 reached publication still scoped to robots that carry things, leaving the seam where a production job is handed over exactly as empty as it was. Read it →
On 26 August, this site found a modular cell but no transferable process qualification. Three Y Combinator companies and an ISO draft now show measurement arriving for robots, while qualification for work moving between sites is still missing. Read it →
Eight notes that converge from different directions: the compliance date and the tariff rate moved again, the ramp curve and the reference cell arrived early, and the layer still missing between them finally acquired an address. Read it →
Reshoring activity rose in 2026 while satisfaction with it fell from 96 percent to 65 percent, and the gains manufacturers actually booked were speed to market, on-time delivery and logistics rather than unit cost. Read it →
SAP has claimed the seam between robots and the system of record from above and NVIDIA has abstracted the body from below, while ISO 21423 sits in ballot and VDA 5050 reached version 3.0, both of them written for machines that move things from A to B. Read it →
Isembard has the capital, the factory software and more orders than it can take, and it still grows by converting machine shops other people own, which makes operators the scarce input in a distributed production network. Read it →
Intrinsic published a reference design for a modular robot workcell at Automate 2026, and a customized version is due to pilot inside Foxconn plants this year. The substrate for portable production capacity is arriving before the trust layer that would make capacity portable. Read it →
A running log of the second World Humanoid Robot Games, appended each evening through 2026-08-26 and closed since. It ran to day 4 and the promised day 5 was never written. Day 1: the 9.39 s 100 m was a preliminary heat, the faster 9.32 s came from a preparatory test reported by state media, and the 2.88 m standing high jump is being measured against a running high jump record. Read it →
The Supreme Court held in February that the reciprocal tariffs were never lawful and roughly USD 128.68 billion is now being handed back, while the first-ever Section 338 duty on Canadian goods was paused three days before it landed, so a factory is a fifteen-year commitment answering to a barrier that moves on a three-day clock. Read it →
Figure builds one robot an hour and publishes its yields, AgiBot rolled its 15,000th unit off the line, and Tesla (the most experienced manufacturer of the three) converted a car plant and then dropped its own start date: the humanoid race is now sorted by ramp curves, and a ramp is a framework achievement, not a robot achievement. Read it →
USD 2 per hour is the headline number for humanoid economics, but the primary study states no cost breakdown and no utilization assumption; utilization, capped today by battery chemistry and decided by the layer above the robot, is the whole argument. Read it →
The EU moved the 2 August date everyone circled for factory humanoids to December 2027: a field note on why the law is the softest layer in the stack, what actually governs a deployed humanoid today, and why the moat is the certification evidence, not the deadline. Read it →
The month the Journal moved from watching the robots to costing the framework: twelve notes ran in two halves, and every one found the value sitting beside the robot rather than inside it, before turning inward to price capacity, admin, landed cost, and the currency rate that hides in a BOM line. Read it →
A part priced at USD 1,200 and quoted as EUR 1,200 carries a 13.80% uplift at the ECB reference rate, and 19.74% at the euro's high this year. Conversion itself costs 0.90% at a published bank margin and 1.6% all-in by the FSB's measurement. The same 13.80% has at least four decompositions and the BOM line looks identical in all of them, so the note ends in nine rules that make it checkable from both sides. Read it →
Contract manufacturers' filings put selling and admin at just 2–6% of revenue, and at 40–50 cents of every gross-profit dollar, because the industry's margins are that thin. The folkloric "20% admin" turns out to live in the quote, where loading runs 25–33% above direct cost while direct labor is under 2%. Forty years after HBR named the hidden factory, its transactions are finally software. Read it →
Chip fabs sign it as billion-dollar capacity agreements, cloud made it the default pricing of computing, Costco built four decades on it: a fee for the right to buy at near cost. Nobody has productized that contract for contract manufacturing. On the precedents, the operations-research results that already priced it, and the failure modes the published model is designed against. Read it →
90–95% of importers never reconcile the landed cost they quoted against the invoices they paid; forwarders live on 10–30% markups in that blindness. Meanwhile the estimation half of logistics became an API. A field note on why reconciliation is the real product, and why production routed near demand deletes the customs event entirely. Read it →
Korthos maps the machine economy (funding, factory ramps, an evidence-graded supply-chain graph), seven months old, anonymous, near-daily. We checked five of its load-bearing claims against primary sources; all five held, including the one it never cited. On sourcing discipline, and why the data layer forming before the factories is itself evidence for the thesis. Read it →
The humanoid hardware works (Figure logged more than 90,000 parts on a BMW line), but there is still no ratified safety standard for a machine that can topple beside a worker. On the missing rulebook, the 18-to-36-month clock behind it, and why the durable moat is the certification framework, not the robot. Read it →
Below the humanoid arms race, small startups are automating real manual work with enclosed, trainable, fixed robotic cells, and they have unanimously rejected humanoids to do it. What the micro-factory layer validates and challenges in the thesis, and the question it can't answer: something still has to move the boxes. Read it →
Humanoid legs and torsos are largely solved, but the dexterous hand is not: Tesla stalled Optimus production over it in 2025, and the tactile end-effector is now both the last hard subsystem and a funded market of its own beneath the robot. Read it →
One year past Liberation Day, US manufacturing construction spending has fallen from its 2024 peak and only about a third of manufacturers are reshoring: the trade wall did the demand-side work and the supply side still could not respond. Why capacity did not follow the tariff, and what that argues for. Read it →
Vooma sells AI agents that quote, build, schedule, and track freight for some of the largest US brokerages, coordination work absorbed by software years before robots absorb assembly. What a small company one industry over says about the thesis, including the part where it disagrees with it. Read it →
In mid-2026 the busiest front in robotics is the model, not the machine: NVIDIA keeps the GR00T brain and lets Unitree sell the body, and the largest round of the year went to Skild's bodyless robot brain. How much of that layer is the moat Moduloa is actually chasing. Read it →
A small number of humanoid robots now do paid industrial work (hundreds, not thousands), while the cheapest walking machines land near USD 20,000 and the useful ones cost far more. Where that leaves the Moduloa register. Read it →
Each note stands on its own, and most of them take one longer question a step further. These are those questions, the notes that moved each one, and where each stands. Grouped by hand and dated, so a thread that has gone quiet shows as quiet.
When there is demand for new production capacity, what actually stops it from appearing?
Tariffs are not a factory · The barrier moves faster than the building · The bottleneck was never the machines · Speed came back, price did not · The robot was never the expensive part
As of 26 September: the 29 August finding that the scarce input is people, not machines, now has a cost attached: setup and integration are the largest cost, and "hours are exactly what the small shops this model depends on do not have."
No research note in this thread since 26 September.
Can anyone yet prove that a working humanoid is safe and dependable beside people?
The state of humanoid work, mid-2026 · The hand is the bottleneck · You cannot certify a fall · The regulation was the soft part · What Beijing actually measured
As of 22 August: Beijing's records were real but not like for like, and a benchmark that weights autonomy "is doing certification work whether it calls it that or not."
No research note in this thread since 22 August.
Does the robot's price decide whether automation pays, or is it everything around the robot?
The state of humanoid work, mid-2026 · The box beats the body, for now · The two-dollar hour has no denominator · One per hour, ten thousand parts · The robot was never the expensive part
As of 26 September: setup and integration are the largest cost even in the optimistic case, so "if hardware is at most a quarter of a cell's lifetime cost, a cheaper body moves the smaller share."
No research note in this thread since 26 September.
Can a qualified production process move between robot cells owned by two different companies?
The box beats the body, for now · Alphabet published the cell, not the standard · The bottleneck was never the machines · The interface is being written now, and it does not say humanoid · The measurement layer arrived, and it points at the robot · The standards landed below the work
As of 19 September: two standards moved and neither can express the transfer. "P-27 gets nothing again, and the window is now under ten months."
No research note in this thread since 19 September.
As the layer around the robot gets built, can anyone own it and charge for it?
Who owns the robot's brain · The coordination layer falls first · The intelligence layer arrives first · The interface is being written now, and it does not say humanoid · The measurement layer arrived, and it points at the robot · The standards landed below the work
As of 19 September: the layers above and below the seam are becoming standards that charge nobody rent, "and the seam becomes a product."
No research note in this thread since 19 September.
Where does the margin hide in contract manufacturing, and can every charge be checked?
The loop nobody closes · Capacity is a membership · The hidden factory is an office · When the rate becomes the margin
As of 30 July: the currency convention "is published as a proposal, not as a report on something already proved."
No research note in this thread since 30 July.
No email list, no account, no tracking pixel: an open Atom feed any reader can subscribe to. Copy the address into NetNewsWire, Feedly, Reeder, Thunderbird, or whatever you already use.
https://moduloa.com/feed.xml: field notes, project updates and register scores as they are published, each with a summary and a link. Nothing is held back for a paywall or a newsletter, because there is neither.
A researched field note most weeks, and a recap at the start of each month. The rule is the same as the rest of Moduloa: no source, no claim. When the evidence is thin, the note says so.
The more useful kind of subscription is a correction. Every note carries its sources so they can be checked, and the predictions register is dated and scored so the misses stay visible. If something here is wrong, shk@moduloa.com. Being corrected in public is the point of writing it in public.