Is the Prusa MK4 actually self-replicating?
The honest answer in five minutes. Yes, but not in the sense laypeople mean — and not in the sense that should let any humanoid company use the word about themselves.
The question, stated precisely
“Self-replicating” is the most overloaded word in robotics. We split it into three numbers and a class.1 Until you commit to which one you mean, every argument about the MK4 is people talking past each other.
- P% — the fraction of discrete parts in the bill-of-materials the machine can produce itself. Cheap to compute. Easiest to game: every printed cable clip pushes the number up.
- M% — the fraction of finished mass the machine produces itself. Penalises decoration. Rewards machines that print their own frame.
- C%— the fraction of BoM cost the machine can cancel, valued at arm's-length market prices. The honest measure, because plastic is cheap and integrated circuits are not.
- Class — 1 (autotrophic, no imports), 2 (autotrophic-lite, only specialists imported), 3 (partial kinematic, prints a defined subset from standardised feedstock), 4 (assembly-only), 5 (swarm).
A machine's class is the highest it can honestly claim. Its three numbers tell you whether the claim is interesting or cosmetic. If you only ever cite one of P%, M%, C%, you are either a marketer or making a mistake.
What the MK4 actually self-produces
Prusa runs a print farm in the Holesovice district of Prague.3 Hundreds of MK3 and MK4 machines run twenty-four hours a day, and a meaningful fraction of what they print is plastic parts for new MK4 kits. This is not a marketing diagram. It is a working factory whose output includes the next generation of its own machines. By any honest definition of the word, that is self-replication of a kind.
The parts a MK4 prints for another MK4:
- X, Y, and Z carriages
- Frame end-caps, idler tensioners, motor mounts
- Fan ducts and shrouds
- Cable guides and strain reliefs
- Spool holders, knobs, display housings
- Tool-changer caddies and accessory brackets
That is roughly forty to sixty distinct printed parts per machine, most weighing between five and eighty grams, all produced from a single feedstock — PETG filament — that the machine itself melts and lays down. It is genuinely the highest-volume Class-3 operation on Earth.3
What it imports — and why those parts dominate cost
Everything else comes from a normal global supply chain. The imports fall into three buckets, and each bucket is harder to self-produce than the last.
Commodity mechanicals
Aluminium extrusions, threaded rods, GT2 belts, pulleys, M3/M4 fasteners. Cheap, abundant, made by everyone. The MK4 buys these from the same wholesalers feeding every other 3D printer vendor on the planet. They are commodities precisely because no one wants to make them in their garage.
Specialist mechanicals
Linear rails, MGN/MGW carriages, hardened steel rods, sealed ball bearings. These need micron-scale tolerances and hardened-steel grinding that is structurally outside what FDM-printed plastic can replace.7 Bearings alone are flagged in our gap analysis as ~0% self-produceable at spec, even at twenty-year horizons.
Active subsystems
Buddy 32-bit control board (STM32 microcontroller, stepper drivers, custom PCB), HMS color display, NEMA 17 stepper motors with neodymium magnets and precision laminations, magnetic heated bed, 240W mean-well-class power supply, V6/Revo E3D-derived hotend, optional Nextruder. The motors are the single largest line item.5 The control board contains ICs that no FDM printer can fabricate.6
A few of these imports are Prusa-bespoke: the Nextruder, the Buddy board firmware stack, the loadcell-equipped hotend assembly. Most are commodity parts shared with the wider 3D-printer industry. None of them is printable on the machine itself.
The numbers
From our research base1:
Prusa MK4 — Class 3 (Partial Kinematic), $799 kit
The drop from 25-35% to 2-5% as you cross from part count to cost is the entire story. Printed PETG is roughly thirty US cents per cubic centimetre at retail; the heated bed alone is over a hundred dollars; the Buddy board with its STM32 is over two hundred. The machine prints what is cheap and buys what is expensive, by necessity. Our Class-3 economics summary makes this general: today's shipping Class-3 machines produce roughly 25-40% of parts by count but only 2-7% by cost; the remaining 93-98% is bought motors, electronics, bearings, extrusions, and power.8
That is not a problem with the MK4. It is the shape of the Class-3 frontier in 2026.
The honest verdict
Yes, the Prusa MK4 is self-replicating in the technical sense: it is a shipping Class-3 (Partial Kinematic) machine that produces a non-trivial fraction of the parts in copies of itself, from a single standardised feedstock, at scale, profitably.
No, the Prusa MK4 is not self-replicating in the sense laypeople mean when they hear the phrase. By cost, it cancels two to five percent of its bill of materials. The remaining ninety-five-plus percent is a normal industrial supply chain — motors, electronics, bearings, extrusions, batteries-via-PSU, fasteners — every one of which Prusa buys from someone else.2
Both statements are simultaneously true. C% is the line that matters, because cost is what self-replication would have to cancel for the word to translate into a real economic claim — a self-replicator that gets cheaper with every generation, that is cost-bounded by raw feedstock rather than by global supply, that could plausibly bootstrap somewhere with no industrial base. None of that is the MK4. The MK4 is a Class-3 machine doing Class-3-ceiling-grade work, and the Class-3 ceiling in 2026 is five percent.
Compared to the original RepRap Mendel
Adrian Bowyer's RepRap Mendel (2009) is the historical anchor for any Class-3 conversation.18 It scores higher than the MK4 on every dimension of self-replication.4
| Dimension | Prusa MK4 | RepRap Mendel |
|---|---|---|
| P% | 25-35% | 40-55% |
| M% | 6-12% | 10-20% |
| C% | 2-5% | 3-8% |
| Build time | 40-60 h | 60-100 h |
On its own terms the Mendel is the purer Class-3 machine. Its frame topology is more than half printed plastic by part count. Its cost fraction is meaningfully higher.
And yet the Mendel is obsolete. Its electronics are pre-32-bit (RAMPS on Arduino Mega). Its build is fussy and slow. Its print quality at modern speeds is not competitive. The MK4 is strictly worse as a self-replicator and strictly better as a business: it ships in the hundreds of thousands; the Mendel never escaped the tens of thousands. “Purer” does not mean “more useful”. The lesson is that the open-BoM advantage is real but it is not a print advantage — it is a commodity-substitution advantage that buys roughly twenty percent on cost.17
What would push it to Class 2?
Class 2 (Autotrophic-Lite) means producing all structural parts and most non-specialist subsystems from raw or near-raw feedstock, importing only the genuinely specialist items — ICs, rare-earth magnets, specialist chemicals. Nobody ships at Class 2 today. Three concrete moves would close the gap, each with a current technology readiness level worth taking seriously.
1. Printable electronics for the Buddy board
The Nano Dimension DragonFly IV produces multi-layer PCBs including embedded conductors, dielectrics, and some passives.9It cannot fabricate the active components — the STM32, the gate drivers, the regulators. Voltera's desktop V-One and NOVA fill the lower-end niche for single-side and flex circuits.10 The eFlesh tactile sensor — a $5-BOM, hobbyist-printable magnetic sensor — proves that component-level fab is reaching consumer scale for at least one sensor class.11 TRL 4-5 for boards, 2-3 for active components.
2. Printable steppers and their gearboxes
Our gap analysis puts electric motors at ~0% self-produceable today — wire drawing, precision laminations, and NdFeB magnets are all chemistry- or capex-hard.5 But the gearbox half of the actuator is moving. An open-source mini 6-DOF arm published in 2026 demonstrates 3D-printed split-ring planetary gearboxes and an inverted-belt differential wrist, using commodity TMC5150 drivers and AS5048a encoders.12 That is the highest P%/M% of any product in our index, and proof that the gear-train portion of the actuator is now hobbyist-fabricable. The motor itself remains imported. TRL 6 for printed gearboxes; TRL 1-2 for printed motors.
3. Printable bearings and rails via metal binder jetting
Sealed ball bearings need ball roundness under five micrometres in hardened steel.7No FDM machine will ever do that. But Desktop Metal's Shop System produces structural metal parts in steel, copper, and stainless at roughly one-percent dimensional tolerance,13 which is enough for non-load-bearing brackets, idler axles, and motor mounts that are currently bought as machined aluminium. A second-generation MK4 with a shop-scale binder-jet sister machine in the same room would shrink the imported-metal line item materially. TRL 8 at shop scale, TRL 4 at desktop.
None of these three is hypothetical. All three are shipping products with prices and serial numbers. Combine them and a credible Class-2 desktop-scale fab moves from twenty-year speculation to ten-year roadmap. That is the moves we are tracking.
Why this matters for humanoid robotics
A 3D printer has on the order of a hundred fifty distinct parts. A humanoid robot has tens of thousands. The MK4 is one of the highest P% machines on Earth and it is still ninety-five percent imported by cost. Anyone telling you a humanoid is “self-replicating” is either using the word differently than we do or selling you something.
Tesla Optimus is the cleanest example. Within Gigafactory walls, Tesla can produce roughly ninety-five percent of Optimus by part count and ninety percent by mass — they have the Giga Press, the FSD silicon, the in-house actuator design.14 But their C% from the perspective of anyone outside Tesla is approximately five percent. The actuators are designed by Tesla and built by Sanhua and Tuopu. The harmonic drives come from Green Harmonic in Suzhou. The cells come from CATL. None of those parts is available to a third party. That is Class 4 (Assembly-Only) with a vertical-integration story bolted on, not self-replication.
If the entire 3D-printer industry needed forty years to push C% from zero to five percent on a hundred-fifty-part machine, the humanoid industry is not about to leapfrog on a thirty-thousand- part machine. The people repeating that they will are repeating a press release, not analysing a supply chain.
The opportunity
We are publishing the Self-Replication Index because nobody else has — there is no consistent, cited, numeric measure of how self-replicating any given machine actually is.15This essay is the first hero entry. Four further moves follow from it directly:
- Score fifty machines against P%/M%/C% so the index has reference value.
- Fund a public BoM teardown of the Unitree G1 — the first honest C% number for a humanoid will become the most-cited chart in the industry.16
- Close the electronics gap by investing alongside Nano Dimension and Voltera, the only two companies in striking distance.9,10
- Build a Class-3-native humanoid kit — make the C% argument ourselves, instead of measuring others making it.
See the full taxonomy and product rankings on /self-replication, the Tesla teardown on /companies/tesla-optimus, and our actuator-supply analysis on /actuators.