In Part 1 of this series, we made the case that your IP portfolio is probably your most undervalued asset. We walked through why patents matter for fundraising, competitive defense, and scaling — and why founders who ignore IP strategy are leaving real money on the table.
But there is a challenging tradeoff in deciding what to patent.
The patent system is, at its core, a disclosure bargain. You reveal exactly how your invention works, and in return, you get twenty years of exclusivity. That's a good deal when the innovation is easily reverse-engineered, or if you'll need a lot of partnerships that involve technology transfer in order to reach your commercial goals. It's not so good if your margins are based primarily on a lack of competition and you'd like to keep it secret as long as possible.
The Disclosure Bargain: What You're Actually Giving Up
Most founders think about patents only in terms of what they get: exclusivity; investor signaling; and potentially licensing revenue later on.
A patent application becomes a part of the public record. In most jurisdictions, the application is published eighteen months after filing regardless of whether the patent is ultimately granted. Your competitors don't need to reverse-engineer your product anymore. If first mover advantage is key to your strategy, and you don't have high confidence your patent will be granted, filing it becomes a gamble.
For certain innovations, this trade is clearly worth making. Most commonly when composition of matter is involved (because it covers the resulting material no matter how many different methods or trade secrets can be used to make it), and also for a hardware device that can be inexpensively disassembled and analyzed to determine how it was built.
But for other categories of innovation, such as proprietary manufacturing processes, training data pipelines, optimization algorithms, and calibration parameters, all that take place inside the four walls of your corporate offices, the value of patent protection drops. And it can drop further in software fields where the pace of innovation is extraordinarily fast.
To preserve the moat: Would my competitors learn this on their own, or am I about to teach them?
To save time and money: If this patent takes 3 years to be granted, will any of my competitors still care about it by then?
When Trade Secrets Win
Trade secrets protect information that derives its commercial value from being kept confidential. There's no filing requirement, no examination process, no public disclosure, and no expiration date. Protection lasts as long as secrecy is maintained.
For example, the Coca-Cola formula has been protected as a trade secret for over a century. Google's search algorithm has never been patented. Both companies decided, correctly, that the value of perpetual secrecy exceeded the value of time-limited exclusivity.
There are several conditions where trade secret protection is strategically superior to patents.
The innovation is not externally observable. If competitors can't detect your innovation by examining your product, purchasing your service, or observing your manufacturing output, a patent gives away information they wouldn't otherwise have. A "discoverability analysis" should be the first step in any IP strategy session. Although one must keep in mind whether there is active ongoing research in the area that may lead to independent discovery and publishing into the public domain.
The technology lifecycle is shorter than the patent prosecution timeline. Patent prosecution in the United States typically takes two to four years. In fast-moving fields such as AI and software, the commercial life of an innovation may not survive the wait.
Disclosure would compromise a layered advantage. Some innovations derive their value not from a single breakthrough but from a combination of components (e.g., in AI, it could be the particular combination of training data selection, preprocessing techniques, hyperparameter tuning, and model architecture choices). Patenting one layer can reveal enough about the overall system to make the rest more vulnerable.
The § 101 eligibility landscape may be hostile to the innovation type. Since the Supreme Court's Alice Corp. v. CLS Bank International decision in 2014, software and AI-related patents have faced persistent eligibility challenges. The Court established a two-part framework (the "Alice/Mayo test") to determine if a patent claim is (i) directed to an ineligible concept (e.g., an abstract idea, law of nature, or natural phenomenon), and (ii) whether it adds an "inventive concept" to make it eligible. This has impacted AI/ML patents, where courts continue to hold that applying generic machine learning techniques to familiar processes does not meet the patentability threshold — the innovation must demonstrate a concrete improvement to the underlying technology itself. (But the USPTO might be diverging on this — see below re Ex parte Desjardins.)
The Catch: Trade Secrets Have No Defense Against Independent Discovery
There's a fundamental asymmetry founders need to understand before leaning heavily on trade secrets: if a competitor independently arrives at the same innovation through their own R&D, reverse engineering, or even lucky guesswork, you have no legal recourse. Unlike a patent, which grants exclusivity regardless of how a competitor developed their version, a trade secret only protects against misappropriation.
"If nobody stole it, nobody infringed it."
This means trade secrets are strongest when your innovation is genuinely hard to replicate — for example, if it took years of iteration, proprietary data, or domain-specific insight that isn't easily reproduced from first principles or existing research trends. They're weakest when multiple teams in your field are converging on similar approaches and iterating rapidly, which is common in AI and software. Before committing a core innovation to a trade-secret-only strategy, ask yourself honestly: if a well-funded competitor hired ten strong engineers and gave them two years, would they get there on their own? If the answer is yes, a patent might be the better bet, even with the disclosure cost, because at least you'd have grounds to enforce exclusivity.
Case Studies: Trade Secrets in the Wild
Two of the highest-value trade secret disputes in recent memory turned on internal know-how — not patents — and the numbers explain why founders should take secrecy infrastructure seriously.
Waymo v. Uber: The $245 Million LiDAR Lesson (2017–2018)
In early 2016, Anthony Levandowski, a senior engineer on Google's self-driving car project, downloaded approximately 14,000 confidential files related to Waymo's proprietary LiDAR technology before leaving the company. He founded a self-driving truck startup called Otto, which Uber acquired months later for $680 million. Waymo sued, alleging that Uber had knowingly used misappropriated trade secrets to accelerate its own autonomous vehicle program.
The case settled during trial in 2018, with Uber paying Waymo approximately $245 million in equity and agreeing not to use Waymo's confidential technology. Levandowski was later charged federally with 33 counts of trade secret theft and pled guilty to one count.
The investment lesson here is twofold. First, Waymo's LiDAR circuit board designs and sensor calibration methods were genuinely excellent trade secret candidates, because they were deeply technical, not externally observable, and central to competitive advantage. Second, trade secret protection is only as strong as the internal controls that enforce it. One departing engineer with a USB drive created a costly security failure. For investors doing diligence on any company's trade secret strategy, they need to ask: what are your access controls, your confidentiality agreements, and your offboarding procedures?
LG Chem v. SK Innovation: A $1.8 Billion Battery Fight (2019–2021)
The EV battery sector produced what may be the highest-value trade secret settlement in U.S. history. LG Chem (through its subsidiary LG Energy Solution) accused SK Innovation of systematically poaching 77 employees who allegedly brought proprietary know-how about lithium-ion battery manufacturing processes — cell chemistry formulations, electrode coating techniques, and production optimization methods.
The U.S. International Trade Commission sided with LG in 2021, imposing a ten-year ban on SK Innovation importing battery components into the United States. With a $2.6 billion Georgia battery plant under construction and supply contracts with Ford and Volkswagen on the line, SK settled for $1.8 billion in cash and royalties plus a ten-year mutual covenant not to sue.
For deep tech founders, this case illustrates that manufacturing process trade secrets can in some cases turn out to be more valuable than the patents covering the same general battery chemistry. LG's patents were important, but it was the trade secret claims around specific production know-how that generated the billion-dollar settlement.
The AI Industry's Bet on Secrecy (2020–Present)
The most commercially valuable AI companies (e.g. OpenAI, Anthropic) protect their most important innovations primarily through trade secrets rather than patents. Model architectures, training data composition, RLHF reward signals, inference optimization techniques, and system prompts are guarded as confidential information, and not disclosed in patent filings.
Stealth Launches: Shipping Without Publishing
A trade-secret-or-patent decision isn't always binary, and it isn't always permanent. The disclosure clock gives founders a third lever: timing.
Remember the mechanics. A U.S. application publishes roughly eighteen months after its earliest priority date — whether or not it is ever granted. That clock is what turns a filing into a teaching document for competitors. But you can manage it.
Provisionals buy a quiet year. A provisional application establishes a priority date without itself being published. It gives you twelve months to ship product, gather market signal, and decide whether a given innovation is better served by full prosecution or permanent secrecy — all before you've revealed anything.
Non-publication requests keep it dark until grant. In the U.S., you can file a non-publication request and keep an application unpublished all the way through to issuance. The catch is that you must certify you won't pursue the invention in foreign jurisdictions that require eighteen-month publication. In effect, you trade most of your international filing rights for domestic stealth. For a company whose near-term market is the U.S. and whose advantage depends on not tipping its hand, that can be a worthwhile trade.
You can launch and capture a market while your application sits unpublished, deferring the disclosure decision until you actually know how durable the innovation is. Stealth isn't the absence of an IP strategy — it's the deliberate sequencing of one.
The Hybrid Approach: What Smart Founders Do
The best IP strategies are never patent-only or trade-secret-only. Here's the framework we use at Castle Fund and ZeroToIP when evaluating a portfolio company's IP strategy:
Patent the perimeter. File utility patents on the innovations that are externally observable. Any system or apparatus features competitors can see, the system architectures they could reverse-engineer, or the compositions of matter they could analyze, should be considered.
Keep the engine room secret. Protect manufacturing processes, training methodologies, calibration data, optimization parameters, and proprietary datasets as trade secrets. These are the innovations that generate daily competitive advantage that nobody other than your employees should be able to access anyway.
Use provisionals strategically. Provisional patent applications establish priority dates without publication. They give you twelve months to decide whether a given innovation is better served by full patent prosecution or permanent secrecy.
Invest in the infrastructure of secrecy. Trade secrets require active maintenance: robust confidentiality agreements, access controls, employee training, departure protocols, and documentation of what constitutes a trade secret and why. The legal standard requires "reasonable measures" to maintain secrecy — without them, you may have no trade secret to protect at all.
Consider the NVIDIA model. The company has aggressively expanded its patent portfolio over the past several years — thousands of active filings concentrated in GPU and hardware architecture. But CUDA, the computing platform that locks AI workloads into NVIDIA's hardware ecosystem, is protected primarily through a mix of copyrights and trade secrets (its compilers, drivers, specialized libraries, and firmware) rather than patents.
How AI-Powered IP Discovery Might Change Things
LLMs are reshaping the structure of the patent system itself. Here are three dynamics we think will meaningfully shift the trade secret / patent power balance over the next five to ten years.
01 · Two Doors Are Opening — But They Lead to Harder Rooms
Founders building AI-native companies face two distinct patent questions that often get conflated. They shouldn't be, because the regulatory landscape has recently shifted on both — in different ways, for different reasons.
Can you patent the AI itself? This is the § 101 subject matter eligibility question addressed above for Alice. Since Director John Squires took office in September 2025, the USPTO has undergone a pronounced pro-patent recalibration specifically targeting this problem. The Ex parte Desjardins decision was a precedential Appeal Review Panel opinion authored by Squires himself; it vacated a PTAB rejection of machine learning claims and established that examiners should search for a "practical application" showing an improvement to technology, and should avoid dismissing meaningful technical limitations at too high a level of generality. The December 2025 MPEP update codified these instructions. The August 2025 Kim Memo — issued by Deputy Commissioner for Patents Charles Kim — reinforced the same direction: it reminded examiners not to over-extend the "mental process" grouping, instructing that claim limitations encompassing AI in ways that cannot practically be performed in the human mind should not be reflexively treated as abstract. The result: the § 101 gate that was largely shut for AI/ML inventions has opened a crack.
But don't overread the shift. These are USPTO examination postures, and the courts are not bound by them. In Recentive Analytics v. Fox (Fed. Cir. 2025), the Federal Circuit — framing it as a question of first impression — held that claims that do no more than apply established machine learning methods to a new data environment are not patent-eligible, and that iterative training, dynamic updates, and real-time adjustments don't convert functional, results-oriented claims into eligible technology. So the realistic picture is a more permissive patent office sitting on top of a Federal Circuit that remains skeptical. A patent that issues more easily today is not necessarily a patent that survives a § 101 challenge in litigation tomorrow.
Can you patent what the AI produces? This is the inventorship question. The November 2025 Revised Inventorship Guidance (issued November 28, 2025) rescinded the Biden-era 2024 guidance and, with it, the Pannu joint-inventorship factor analysis that the 2024 guidance had layered onto AI-assisted inventions. What survives is the traditional conception test: a natural person must still significantly contribute to the invention's conception — forming the "definite and permanent idea of the complete and operative invention" — and AI systems are now treated as tools equivalent to laboratory equipment, software, or research databases. There is no separate inventorship standard for AI-assisted work. Most practitioners read this as a change in emphasis rather than a change in law; human conception was already the governing rule under Thaler.
For founders, this simplification is welcome. It removes a major source of prosecution uncertainty and aligns the USPTO with the Federal Circuit's holding in Thaler v. Vidal that only natural persons can be inventors. But it also means companies must meticulously document the human inventor's conceptual contributions throughout the development process, particularly when AI tools play a significant role. And courts haven't yet weighed in on where the line falls between "using AI as a tool" and "having AI do the inventing."
02 · § 103 Becomes the Next Battleground
With § 101 eligibility becoming more navigable at the patent office, the substantive fight shifts to § 103 non-obviousness. Patent law evaluates non-obviousness from the perspective of a "person having ordinary skill in the art" (PHOSITA). If AI tools become standard equipment in a researcher's toolkit — and they already are in most technical fields — then what counts as "ordinary skill" ratchets upward. An innovation that would have been non-obvious to a human chemist working alone might be entirely obvious to the same chemist armed with a molecular simulation platform or a protein-folding model like AlphaFold.
Compound this with the prior art problem. AI systems can generate enormous volumes of technical disclosures — research papers, code repositories, design proposals, synthetic datasets — that may qualify as prior art. If AI-generated content is treated as prior art under current law (still an open question), the landscape becomes vastly denser. The gap between what's "known" and what's genuinely new narrows considerably.
For patent applicants, this might mean more rejections at the patent office, but, because of the political goal of the § 101 changes, it might not. Instead, we could face a sharp increase in granted patents, but then courts could diverge and hear more post-grant invalidity challenges on § 103 grounds — much as Recentive shows the Federal Circuit already willing to draw lines the patent office isn't. Trade secrets, which face neither barrier, gain relative strategic value and more predictability in this new environment.
03 · Enablement Requirements Tighten — Forcing More Disclosure or Fewer Patents
The enablement requirement under 35 U.S.C. § 112 demands that a patent specification contain enough detail for a skilled practitioner to reproduce the invention without "undue experimentation." For AI inventions, this is already creating friction — and the pressure is coming from both sides of the Atlantic.
The European Patent Office has established stringent disclosure expectations: its March 2024 examination guidelines (in force 1 March 2024) state that a disclosure is insufficient when "the mathematical methods and the training datasets are disclosed in insufficient detail to reproduce the technical effect over the whole range claimed." The EPO characterizes inadequate AI disclosures as merely "an invitation to a research programme" rather than an enabling disclosure. While U.S. law doesn't formally adopt EPO standards, patent practitioners drafting multilateral applications are increasingly conforming to them, and some commentators believe EPO requirements may presage future USPTO expectations.
And then there's the reciprocal interplay with § 103: if the PHOSITA standard rises (because practitioners now routinely use AI tools), enablement might paradoxically become easier to satisfy with less disclosure — because a more skilled practitioner needs less hand-holding. But that same elevated PHOSITA makes § 103 non-obviousness harder to establish. In the worst-case outcome for patents, what you can patent would actually narrow, while what you must disclose to patent it may increase. This structural squeeze would push IP filers towards trade secrets and away from patents at the margins.
For founders, this is good news. The § 101 door that was largely shut for AI patents is swinging open. But here's the critical nuance: getting through the eligibility gate doesn't mean you've won the patent. It means the examiner now evaluates your claims on the merits — and that's where § 103 obviousness and § 112 enablement become the new chokepoints.
What This Means for Founders Right Now
- 01Conduct a discoverability audit of your core innovations. And test LLMs (locally, or with enterprise-level contracts in the cloud) on your IP portfolio. If AI raises the PHOSITA bar in your field, which of your current patent applications become more vulnerable?
- 02Don't patent by default. The instinct to "file on everything" is expensive and strategically counterproductive. Every patent application is a publication. Have a roadmap and schedule that aligns with your product and financing roadmaps properly. Be deliberate about what you reveal.
- 03Build your trade secret infrastructure early. Confidentiality agreements, access controls, information classification systems, and departure protocols aren't glamorous, but they're the foundation of any trade secret strategy. Without "reasonable measures" to maintain secrecy, the legal protection evaporates.
- 04Think in layers. The strongest IP positions combine patents on observable innovations with trade secrets on internal processes.
Next in the series: we'll explore how IP strategy diverges across sectors — why the patent-versus-trade-secret calculus looks fundamentally different in biotech, energy, and aerospace, and what that means for founders building at the frontier of each.
Castle Fund is a seed-stage venture fund backing deep tech companies with defensible IP moats. If you're building at the frontier of biotech, energy, aerospace, robotics, or semiconductors, we'd like to hear from you.
ZeroToIP helps startups build and execute IP strategies from day one. Learn more.