Start with the deaths.
Victor printed a Pentaborg. Two parts, grey PLA, screwed together, about the size of a plum. The question is whether that becomes a shippable robot, and for whom. The honest way in is backwards: the social robot category has killed companies with a hundred times his funding. Nothing gets recommended here until we know exactly how they died.
01 The graveyard
Every one of these died the same death, wearing different clothes
Not one died of bad engineering. Jibo and Anki built genuinely beautiful things. They died because the object cost more than the job it did was worth to the person paying, and because consumer robotics needs each funding round to be several times the last. Anki is the frightening one: it sold 6.5 million devices and still shut down in a single day.
What it forbids: treating unit sales as proof of viability. Selling well and being solvent are different things. A character robot's cost base is a games studio wearing a factory, and five distinct Pentaborg personalities is exactly that kind of content cost.
Sources: The Robot Report, 2019; IEEE Spectrum, 2019.
What it forbids: any design where Victor's ability to keep paying a hosting bill in 2031 is a product feature. The more emotionally effective the product, the larger the liability on shutdown. This kills the 2023 ideation note's ESP32-with-camera-and-WiFi concept outright.
Sources: Axios, Dec 2024; OECD.AI incident record 2024-12-09-ab89; AppleInsider, Dec 2024. Funding total is disputed: CB Insights records USD 68.8m over 9 rounds, another aggregator USD 86.3m over 7. No primary filing was opened, so treat the figure as unresolved.
What it forbids: a talking Pentaborg that answers questions is a worse ChatGPT with a battery. Any surviving version has to do something where being a physical object in a room with people is structurally the point.
Sources: TechCrunch, 2019; The Robot Report, 2019.
What it forbids: the belief that sufficient emotional attachment eventually fixes the economics. It does not. It never has.
Sources: Japan Times / AFP, 26 June 2026; TechXplore, 2026.
What it permits, oddly: this is the entry that makes a small ambition rational. The category cannot clear a multinational's hurdle rate. A product selling 400 units a year at real margin is a failure for Bosch and a meaningful outcome for one person in Mayo.
Sources: Engadget, 2018; The Robot Report, 2018.
- Reach Robotics, MekaMon (Bristol, closed Sept 2019). AR battle robot at around GBP 300. Its CEO: "the consumer robotics sector is an inherently challenging space". Novelty exhaustion: AR battling is a two-week experience on a cost base built for years. Note the escape hatch it reached for was education, the same direction Sphero took.
- Keecker (France, closed 2019, roughly USD 8m raised). A rolling home projector-robot. BOM cost with no defensible job: a projector on wheels competes with an EUR 80 projector on a table.
- Misty Robotics (Sphero spin-out, acquired by Furhat Jan 2022). "Developer platform" is not a consumer market. Misty sold a robot to people who would build the app that made it worth owning. Nobody did, at scale.
- Amazon Astro for Business (killed Sept 2024, under a year after launch). Even trillion-dollar balance sheets brick mobile robots when the unit economics do not close. Amazon at least offered refunds.
Sources: TechCrunch 2019 and 2022; The Robot Report 2019; TechCrunch July 2024.
02 The survivors
The one number that reframes the whole venture
Tonies did EUR 630m of revenue in FY2025. The Toniebox itself was EUR 161m of it. Roughly 74% of the money is the content figurines people choose to buy afterwards. There is no LLM, no cloud personality, no subscription lock, and the box works offline once content is loaded.
The 5i framework is natively a content system. Twenty-five specialisations is twenty-five blades for one razor.
FY2025 group revenue EUR 630m, up 31% (36% constant currency). North America EUR 276m (+40%), DACH EUR 214m (+16%), rest of world EUR 141m (+68%). Adjusted EBITDA margin 8.6%, above guidance. FY2026 guidance is above EUR 760m at 9 to 11% EBITDA.
Hardware is roughly 26% of revenue. The box is a razor and the figurines are the blades. "A companion" killed six companies on this page. "Plays the story on the figurine you put on it" built a EUR 630m business.
Source: tonies SE FY2025 results release. Every sub-figure was independently re-checked. One related claim, "over 9.5m Tonieboxes sold since 2016", could not be confirmed in the release text and is marked unverified.
Tokyo. Qoobo, the headless cat-tail cushion, took 5,000-plus preorders. Nekojita FuFu is a JPY 3,800 (around USD 25) robot that blows on your hot drink to cool it. Mirumi is around USD 70 and peeks shyly out of your bag. Each is crowdfunded before tooling, has no cloud, no account, and no ongoing service obligation.
This is the closest thing in the entire research set to a shape one person can actually operate. Small, cheap, opinionated, finished. Ship it, and it stays shipped.
Sources: Dezeen, Jan 2025; TechCrunch, Jan 2025.
- Miko (Mumbai). Revenue INR 358 crore in FY24, roughly EUR 39m, up 58%. From USD 199 in the US plus a USD 99/year content subscription. The counter-example to Jibo: same category, a quarter of the price, content library rather than personality as the value. A reported USD 155m Series D at USD 550m post-money was covered in Aug 2025 as being raised by board resolution, not confirmed closed.
- LOVOT (Groove X, Tokyo). Fewer than 20,000 units since 2019 at around JPY 349,000. Net loss JPY 2.7bn in FY2023, cumulative losses JPY 8.5bn. But 90% daily active users past 1,000 days and 60+ minutes of interaction a day. Read that carefully: the best retention in the category, still deeply loss-making. This is the strongest single argument against building an emotional companion. Single-source (KrASIA); Groove X is private and files nothing.
- Petoi (US). 30,000+ robots to 60+ countries, used in K-12, camps and university labs including Carnegie Mellon. Open source, no cloud, no personality promise. Nothing to brick.
- Tamagotchi (Bandai). 100 million units since 1996, milestone confirmed Aug 2025. A sub-USD 30 object with no motors, no cloud, no LLM, that produces care-and-attachment through software alone. The cheapest way ever found to create the emotional effect this whole category chases.
- UBTECH (HK: 9880). 2025 revenue above RMB 2bn (+53.3%), net loss RMB 789.8m, narrowed 32%. Humanoid segment revenue up 2,203.7% to RMB 821m; 1,079 humanoid units sold across the segment. It abandoned consumer and went industrial B2B. It started with consumer humanoids and toy kits; it now sells factory humanoids to carmakers. Consumer was the wrong end.
- Sphero. Deliberately exited Disney and Star Wars licensing to sell its own robots into schools. The licensing cliff was the danger and it jumped before it. BB-8 sold on a film release; a film release expires. Curriculum does not. Revenue figures circulating for Sphero are aggregator estimates and should not be quoted.
03 The five laws the graveyard writes
Every survivor has at least four of these five. Every corpse lacked at least three. This is the test any Pentaborgs product has to pass before it deserves a single euro.
Toniebox plays loaded content offline. Tamagotchi has no server. Petoi is open source. Yukai's objects have no account. Jibo, Moxie and Astro all bricked.
Not "we will open-source it if we fail". Designed from day one so the core value survives fiveinnolabs ceasing to exist. For a solo founder this is not a nice-to-have, it is the price of being taken seriously by anyone who read about Moxie.
Tonies: EUR 161m hardware inside EUR 630m total. Miko: USD 99/year for a library you opt into. The dead either had no recurring revenue at all (Jibo, Kuri, MekaMon) or made the subscription a hostage arrangement (Moxie).
The 5i asset is unusually well shaped for this. Five types, 25 specialisations. A magnetic five-piece set where the innovator types physically join is a razor with a collection mechanic built into the framework, and it needs no electronics at all.
Sphero, Petoi, Elephant Robotics and Wonder Workshop all sell to schools on budget cycles. Every dead company in section 01 sold a companion to a curious individual.
Victor already stands inside one of these channels. His EU AI Act Article 4 corporate training clients buy on a training budget, from someone who will run the workshop again next quarter. He does not have to build the channel, which is the single rarest asset on this page.
Dead: Jibo USD 899, Kuri USD 700, Moxie USD 799, Keecker. Alive: Miko USD 199, Ropet USD 299, Tamagotchi under USD 30, Yukai USD 25 to 70. LOVOT at roughly USD 3,100 survives only on venture capital and loses money doing it.
Note what this does to the maths: a 50mm figure priced under EUR 250 cannot carry servos, a speaker chamber and a certification file. The price cap and the ambition ladder are the same constraint seen from two sides.
If the answer to "what does it do" needs a second sentence, the category history says it dies. The test is whether a buyer can repeat the product's function to a colleague accurately, having heard it once.
04 The 2025-26 AI toy wave, and the trap with Victor's name on it
By October 2025 more than 1,500 Chinese companies had entered the AI companion toy category. Then US PIRG's Trouble in Toyland 2025 tested them. FoloToy's Kumma bear, running GPT-4o, told children where to find knives, pills and matches and how to light them, and discussed sexual content.
In order, what followed: FoloToy suspended sales, then returned the product to market days later. OpenAI suspended the developer, which is the critical commercial fact: your model provider can terminate your product with one enforcement action, at any time, with no notice. The US CPSC told Congress it is "neither equipped nor authorized" to evaluate emotional or psychological harm. Maryland enacted an AI Toy Safety Act. A bipartisan federal bill (S.5171) cleared Senate Commerce Committee, though note it is a study-and-report bill, not a regulatory statute. The FTC opened an inquiry.
Regulatory cost in this category is now rising faster than differentiation is available. Separately: every market-size figure for smart toys should be discarded. Two vendors describe the same 2025 category as USD 8.55bn and USD 42.15bn. A 5x disagreement on the current year means nobody knows.
Sources: US PIRG Trouble in Toyland 2025; CNN, 19 Nov 2025; CPSC written response to Congress; Crowell & Moring on the Maryland Act; Senate Commerce Committee, 5 Aug 2026.
The one failure mode Victor is most likely to walk into
Building an LLM-powered Pentaborg companion, because he owns the character IP and the AI expertise, and the two feel like they obviously combine.
That is precisely the trap. It would need an ESP32 with mic, speaker and WiFi, a model API, a per-device inference cost, a privacy posture, and, because his characters are appealing, a real chance children end up talking to it. That single decision imports every failure mechanism on this page at once: cloud dependency, a model provider who can switch him off, per-unit variable cost against a one-time sale, and an emotionally bonded user with a claim on him if he stops.
The specific version for him is worse than generic. A compliance educator whose own AI product causes a child-safety or data incident loses the training business, which is the actual revenue, to protect a hardware side-bet that was never going to reach EUR 2.5k MRR. Embodied had USD 69m or more and Intel, Amazon, Sony and Toyota behind it, and it still ended with no refunds paid and a permanent incident record. The asymmetry runs entirely against him.
05 What we actually hold
- The file is a Blender mesh, not CAD.
3Dpentaborg.stl, 2.36 MB, 47,172 triangles, header "Exported from Blender-3.0.0", dated 12 September 2022. No.step,.f3dor.3mfexists anywhere in the asset tree. - No supplier can quote from it. A mesh has no units, no tolerances, no wall thickness, no draft angles and no parting line. Every factory's first email will ask for a manufacturable file. A parametric CAD rebuild is line item one of any hardware path, not a detail to sort later.
- The proportions are the constraint. Bounding box 2.66 x 1.93 x 4.46 in Blender units: the figure is 1.68 times taller than it is wide and 2.31 times taller than it is deep. Scaled to a 100 mm figure that is a 60 x 43 mm footprint. Whether a battery, a speaker and a servo fit inside that envelope is a real engineering question, and it decides whether the design language has to change.
- The print itself is hobby FDM. Grey PLA, roughly 0.2 mm layers, visible seam stringing on the torso, support-free angular geometry on the head. That last part is genuinely good news: a faceted form with no undercuts is friendlier to both printing and moulding than a rounded organic one.
Irish application 267597, "Pentaborgs", filed 1 September 2022 in Class 28 (toys) and Class 41 (education and training). Examined, accepted, and advertised in Journal 2476 on 9 November 2022. Then withdrawn on 28 June 2023 under Section 45(2) of the Trade Marks Act 1996, which is what happens when the registration fee goes unpaid after acceptance.
Two consequences. First, there is no live mark today, so any plan that assumes brand protection is assuming something that does not exist. Second, the 2022 examination went through cleanly, which is decent evidence the word mark is registrable. Refiling costs EUR 70 per class plus EUR 177 on acceptance, so roughly EUR 317 for two classes across seven to ten months.
Verified against the IPOI eRegister on 27 July 2026. The Section 45(2) cause is inferred from the statute and the dates, not stated on the record.
A research run on 27 July 2026 mapped 137 uses for the dormant Pentaborgs asset. The highest-scoring one was not a product at all: a 45-minute facilitation instrument run at leadership offsites, sold to the EU AI Act Article 4 clients Victor already has. Zero manufacture, zero certification, zero inventory. A pilot email for it is written and sitting unsent at c1-pilot-email.md.
That matters for intellectual honesty. Any robotics recommendation this run produces has to beat a zero-capital option that could start this week, and beating it is not assumed. If the answer at the end is "the instrument still wins", that is a real finding and it will be stated plainly.
06 Landing next
Eleven agents returned, three verifiers tore into them
Eight research streams and three cold verifiers have finished. The verifiers found real defects and they have been applied above: a disputed funding figure de-verified, a segment sales total wrongly attributed to one product model, an unconfirmed Toniebox unit count flagged, and a federal bill correctly downgraded from statute to study.
The supplier scan, the eight costed niches, the regulatory gate map and the three-tier unit economics are written and audited. They land on this page next, once their own defects are repaired. Two of them are serious: a break-even calculation that double-counted EUR 7.81 a unit, and a connected-device compliance case built on an EU regulation that gets repealed in December 2027.