← Upstream archive

Upstream by Tettares

Issue 3 · September 15, 2026 · Physical AI / Robotics

The crown jewels have legs

In physical AI, the most valuable IP is the part nobody filed: the unpatented process knowledge that can walk out the door at 6 p.m.

By Andrew Gerber, Esq., Gerber Law — guest contributor for Tettares. Commentary, not legal advice.

When people ask where the value sits in a humanoid robot, they expect you to point at a wall of patents. The more interesting answer is the people who leave the building at 6 p.m. In physical AI, the most valuable intellectual property is rarely the thing that has been filed and published. It is the unfiled, unpatented process knowledge that lives in a handful of engineers’ heads and a few machines on a factory floor. That makes it enormously valuable and legally fragile at the same time. Trade secrets have no expiry, but they also have no perimeter. They protect you right up until they walk out the door.

The case that made this concrete. On 11 June 2025, Tesla sued a former Optimus engineer, Zhongjie “Jay” Li, and his startup Proception in the U.S. District Court for the Northern District of California, alleging he downloaded confidential data on robotic-hand sensors in his final months and used it to build a competitor Reported · 0.8. Li worked on Optimus from roughly August 2022 to September 2024. Proception was incorporated within days of his departure and, months later, showed dexterous hands Tesla says resemble its own designs Reported · 0.8. The parties later settled, and Proception announced an $11M seed round the same day Reported · 0.8. Two cautions: a complaint is an allegation, not a finding, and a settlement is not an admission. But strip the names away and you have the defining IP pattern of this sector. The crown jewels are portable, and the talent market is on fire.

Why hands, of all things. Because the hand is where robotics is least patentable and most know-how-dependent. The mechanism (tendon routing, sensor placement, the control that makes a grip feel human) is refined empirically, iteration after iteration. That kind of advantage is hard to capture in a public patent claim, so companies keep it as a trade secret. The upside is a moat with no expiry. The downside, as an enforcement matter, is that your protection is only as strong as your offboarding checklist, your device controls, and your ability to prove access and use in court. The Defend Trade Secrets Act gives you a federal cause of action; it does not give you back the knowledge.

Now look at the other end of the stack: the reducer. The strain-wave (harmonic) gear that sits in every precision joint was invented by C.W. Musser in the late 1950s, and those foundational patents expired decades ago Confirmed · context. That is precisely why Chinese entrants like Suzhou Green Harmonic (STAR: 688017) and Leaderdrive can build harmonic reducers at all. Making a harmonic drive is not, by itself, infringement; the core idea is public domain. So how does the incumbent, Harmonic Drive Systems (TSE: 6324), defend a $395M franchise Confirmed · company IR? Not with the expired core patent, but with a thicket of incremental process patents layered over a body of manufacturing trade secrets.

You can see the strategy in the filings. In the Tettares patent dataset, Harmonic Drive Systems' most recent tracked filings are not about the basic gear at all. They are titled things like "Powder lubrication method for strain wave gearing," "Strain wave gearing having fiber seal," "Silk hat-type strain wave gearing," and "Angular transmission error estimation device" (JP 7800877B1, granted) Confirmed · 0.9. That is the tell. When your headline invention is public domain, you patent the lubrication, the sealing, the geometry tweak, and the error compensation, and you keep the actual metallurgy and heat-treat as a secret. The published patents are the visible tenth of the iceberg; they exist partly to signal, and partly to force a would-be copier to design around the last 5% of performance that actually matters.

Where "theft" gets slippery across borders. Two things make enforcement hard in this vertical. First, reverse engineering a legally sold product is generally lawful. It is not trade-secret misappropriation to buy a competitor's reducer and measure it; the line only gets crossed when someone takes confidential information they had a duty to protect. Second, when the alleged taking happens across jurisdictions, a U.S. judgment can be cold comfort. The strategic anxiety underneath the Tesla suit is exactly this: with China now producing the overwhelming majority of the world's humanoids Reported · 0.8, American and Japanese incumbents are getting more aggressive about protecting what they have, not less, because they can feel the know-how gap closing.

A note for investors reading a patent count. Do not confuse a large published portfolio with a durable moat, and do not confuse a thin one with weakness. Analysts have flagged a real disconnect between humanoid startups' patent strength and their valuations Reported · 0.7. Some of the best-funded names are lean on filings precisely because their edge is deliberately unpublished. The provenance question you actually want answered is this: is this company's advantage patented (public, expiring, designable-around) or secret (durable, but only as good as its controls and its people)? Those two profiles fail in completely different ways.

What we're watching upstream

  • The Tesla and Proception docket, and any similar suits. Engineer-mobility cases are the leading indicator of where the real know-how sits.
  • Harmonic Drive Systems' continued process-patent cadence versus Green Harmonic's and Leaderdrive's filings. This is the design-around race in strain-wave manufacturing.
  • Any trade-secret action that crosses between the U.S. and China, or Japan and China, where enforcement is hardest and the stakes are highest.
  • Whether humanoid OEMs shift from patenting to secrecy as their hand and actuator designs mature. It is a quiet but readable signal in filing volume.

The downstream story is "humanoids are coming." The upstream story, the one that should worry every operator and investor in this space, is that the part of the robot worth the most is the part nobody filed. Protect your people and your process, because in physical AI, the crown jewels have legs.


Read Upstream. Fathom the rest. Every fact sourced.

Sources

  • Tesla v. Li / Proception (filed 11 Jun 2025, N.D. Cal.): TechCrunch, Jun 2025 · Fortune, Jun 2025
  • Settlement and $11M seed: The Next Web
  • Strain-wave gear history and core patent long expired: Harmonic Drive strain wave gear technology · example process patent US4934212A "Harmonic drive flexspline manufacture" (Google Patents)
  • Green Harmonic to Optimus reducer (reported): Sina Finance, Feb 2025
  • Humanoid IP-strength versus VC-funding disconnect: PatentVest
  • Trade-secret protection basics: WIPO Guide to Trade Secrets and Innovation
  • Harmonic Drive Systems patent filings (powder lubrication, fiber seal, silk-hat geometry, angular-transmission-error estimation) and FY2026 financials: Tettares dataset (beta), patents and financial_facts tables; patent harvest in progress.

Built from the Tettares dataset (beta). Facts are at pending_review; allegations in active litigation are labeled as such and should be independently verified. Commentary only, not legal or investment advice.

Read Upstream. Fathom the rest. Every fact sourced.

Subscribe, it's free → Explore the data in Fathom →

Built from the Tettares dataset (beta). Facts are at pending-review; rumored items are flagged and should be independently verified before action. Not investment advice.