For a hundred years the moat in this industry was a molecule and a patent on it. You discovered a compound, claimed it as composition of matter, and held a twenty-year wall no competitor could cross — the single most reliable moat in technology, and the reason capital has always trusted biology. But the AI-native biotech this whole series has been describing — the closed-loop factory that gets better at inventing drugs every cycle — does not keep its value in the molecule. The molecule is a receipt. The value is in the engine that produced it: the model weights, the training data, the phenomap, the loop. And that engine sits on legal ground with no floor and no walls. The law is very good at protecting a molecule. It does not yet know how to protect a factory that invents molecules. Everywhere on Earth that ground is unsettled — even the United States built a rule for it and then tore the rule up. But nowhere is it thinner than in the one country this thesis keeps telling to go build such a factory. India can patent the drug its engine designs. It cannot, today, cleanly protect the engine — and the engine is the entire point.
The moat moved, and the law stayed put
Recall where the value went. The argument of the platform essays is that in AI-native biotech the balance-sheet asset stopped being the pipeline and became the platform — the recursive engine whose every experiment improves every future one. Recursion's real asset is not any single molecule in its clinic; it is the phenomap, the learned map of cellular states its 2.2-million-experiments-a-week factory has built. Isomorphic's real asset is not a binder; it is the AlphaFold-descended model that designed it. The molecule is downstream. The factory is the company.
Now ask the question the whole platform thesis quietly rests on: who legally owns the output of a factory that invents things? For a molecule, the answer is a century old — composition-of-matter patent, human inventor, twenty years. For the factory itself — the weights, the data, the closed loop — the answer is a shrug. Intellectual property law was built around a human inventor conceiving a specific invention. It has no clean category for an engine that generates inventions on its own, and it has no settled way to protect the engine as distinct from the things the engine makes. The moat migrated upstream to a layer the law never designed a wall for.
The American wobble
Watch the most developed patent system on Earth try to build that wall, and flinch.
In 2022 the Federal Circuit ruled, in Thaler v. Vidal, that an artificial intelligence cannot be named as an inventor on a US patent — the statute means a human being, full stop. That did not kill AI-assisted patents; it forced a workaround. Name the human scientists who selected, validated, and characterised the AI's output as the inventors, and proceed. Boring and functional. Then, in February 2024, the patent office tried to make the workaround rigorous: a formal guidance specifying, through a multi-factor test, exactly how much human contribution was enough to count. For twenty months that guidance was the map every AI-native biotech filed against.
And then, in November 2025, the office rescinded it — withdrew the framework in its entirety and replaced it with something conspicuously thinner, a posture commentators quickly nicknamed "don't ask, don't tell." The most sophisticated IP regime in the world spent two years constructing examiner guidance for AI-assisted invention and then un-built it. Read that sequence as what it is: not a settled doctrine but a floor being laid, tested, and pulled up again. Even where the patent system is deepest, nobody yet knows how to police the boundary between what the human did and what the machine did — so the rule keeps moving. That is the floor wobbling. Now go somewhere the floor was never poured.
Two fallbacks, and why they hold in America
If you cannot cleanly patent the engine, you have two ways to protect it, and the serious platforms use both.
The first is to patent the molecule and keep the engine as a trade secret. Recursion does not try to patent its phenomap; it treats it as a trade secret and patents the compounds that come out of it under human-named inventors. This is the dominant strategy, and it works — in the United States — because the United States has a federal trade-secret statute with real teeth, the Defend Trade Secrets Act of 2016, that gives the secret a floor made of law: injunctions, damages, federal court. The patent on the molecule may rest on a wobbling inventorship rule, but the secret on the engine rests on a statute. Wobbly upstairs, solid in the basement.
The second is simply the composition-of-matter route done carefully — name the chemist who made the significant human contribution, and the molecule is safe regardless of how the model helped. Also functional, also dependent on a legal system that will back the human-inventor fiction when challenged.
Both fallbacks assume the same thing: a legal environment where patents are enforceable and, crucially, where a trade secret is a statutory asset and not merely a private promise. Change that assumption and the whole protective structure changes. India changes that assumption on every axis.
India, thinner on every axis
Run the identical analysis in India — the country the platform essays keep nominating to build the AI-native factory no incumbent has won — and it is weaker at every layer that matters.
Start with the patent on the engine, because India is structurally more hostile to it than America is. India shares the natural-person baseline: the Indian Patent Office refused the same DABUS application the US and Europe refused, holding that an inventor must be a person. Fine — that is the global consensus and not the problem. The problem is one rung down. Section 3(k) of India's Patents Act expressly bars from patentability "a mathematical or business method or a computer programme per se or algorithms." The molecule an AI designs is still patentable, because it is claimed as composition of matter on a different track entirely. But the model — the weights, the training method, the closed-loop algorithm that is the actual moat of a Recursive Discovery Factory — runs straight into §3(k). There is a narrow escape, codified in India's 2025 computer-related-inventions guidelines, for software that produces a concrete "technical effect." It is narrow, and it was written for embedded systems, not for a discovery engine. Where American law merely has no clean category for patenting the engine, Indian law has an express exclusion pointed at exactly the engine's substance. That is worse, not equal.
Now the trade-secret fallback — and here is the sharpest fact in the essay. The strategy that saves the American platform, keeping the engine as a statutorily protected trade secret, has no statute to stand on in India. India has no dedicated trade-secret law at all. It protects secrets only through the ordinary law of contract — non-disclosure agreements — and an equitable action for breach of confidence inherited from English common law. There is no Indian equivalent of the Defend Trade Secrets Act, no equivalent of the EU Trade Secrets Directive. India's own 22nd Law Commission recognised the gap and, in March 2024, recommended a dedicated statute and even drafted a Protection of Trade Secrets Bill to fill it — and that bill sits unpassed. So the wall an Indian RDF would put around its phenomap is not a statutory wall. It is a stack of contracts. Contracts can be strong — India's entire IT-services industry runs on exactly this, confidentiality enforced by agreement — but a contract binds only the person who signed it, and a trade-secret statute binds the world. The American platform's engine sits behind a law. The Indian platform's engine sits behind an NDA.
And the guidance layer, where the state could at least reduce the uncertainty, is a set of recommendations without rules. There is no Indian patent-office examiner guidance on AI inventorship — a parliamentary committee flagged the Act as "not well equipped" for AI inventorship back in 2021, and the 2026 DABUS order settled the narrow personhood question, but no examiner has a rulebook for the harder, everyday question of how much machine contribution voids a human-named claim. On the regulatory side the same shape repeats: India has no drug-development-AI guidance to match the FDA's, only a 2025 draft covering AI in medical-device software — a different track from drug discovery. Recommendation, draft, flag, order — but no floor and no rulebook, precisely where the money now concentrates.
Why this bites India's own thesis
Put the pieces together and the problem is not generic; it is aimed, with unusual precision, at exactly the future this publication keeps prescribing to India.
The therapeutics thesis argues that India's one genuinely open frontier is AI-native platform biotech — the Recursive Discovery Factory that no incumbent has won in any geography, the shape whose value compounds even when a partnership is delayed. That is the essay's most optimistic claim about India. This essay is its most uncomfortable footnote. Because an RDF's entire value is the engine — the model, the data, the loop — and the engine is precisely the asset Indian law protects least. India can patent the molecule the factory produces; it struggles to patent the factory, and it cannot statutorily seal the factory as a trade secret. The country being told to build the AI-native discovery engine has the flimsiest legal foundation on Earth for owning one. This is not the familiar India capital gap, the thin cheque-book that recurs across the series. It is an IP gap, and it is structural: the moat India is being urged to build is made of the one material its law does not yet recognise as property.
Where it breaks
The honest bounds, because "no floor, no walls" is a rhetorical compression and the truth is more specific.
India's thinness is a statutory and guidance thinness, not an absence of all protection. Molecules remain fully patentable; the composition-of-matter route works exactly as it does in the West. Contracts and the breach-of-confidence action do bind, and they bind hard enough that a large services economy trusts them daily. A device-side AI framework now exists. The trade-secret bill is drafted. So the correct statement is not that an Indian RDF is defenceless — it is that its defences are contractual and fragile where an American RDF's are statutory and firm. Walls made of agreement, not walls made of law.
The floor is wobbling everywhere, not only in India. The American rescission of November 2025 is the whole industry's problem, not India's alone; every AI-native biotech on Earth is building on unsettled inventorship ground. India is the worst-positioned, not the only-exposed, and the essay should not pretend the West has solved what it has merely papered over.
And the gap is legislatable — which is the most important bound, because it points at the fix. Nothing here requires India to out-invent anyone. It requires the state to pass a trade-secret statute it has already drafted, issue examiner guidance on AI inventorship it has already been advised to write, and extend its AI-regulatory framework from devices to drug development. Each of those is a stroke of the pen, not a scientific breakthrough. The gap is real, and it is also the cheapest gap in the entire thesis to close.
Close
The moat moved off the molecule and onto the factory that makes molecules, and the law did not follow. Everywhere the floor beneath that new moat is unsettled — the United States poured one, tested it, and pulled it up within two years. In India the floor was never poured and the walls were never built: the molecule is protected, the engine that invents molecules is not; the patent route runs into an express software exclusion, and the trade-secret route runs into the absence of a trade-secret law. The publication keeps telling India that its frontier is the AI-native factory no one else has won. It should say the quiet part alongside it. India is being asked to own the one asset its legal system protects least — and the single most valuable thing the state could do for the whole of frontier biotech is not another biofoundry subsidy or another mission document, but two pieces of legislation that already sit drafted and unpassed: a trade-secret statute and an AI-inventorship rule. Build the factory, by all means. But a factory needs a floor to stand on, and right now India is offering its most important industry a moat with none.
The IP-layer flagship of the Atoms and Cells therapeutics thesis, spun out of the Index's Section 4 and applying across every AI-native platform in the series. US chain (consistent with the Index): Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022) — a natural person must be the inventor; USPTO Inventorship Guidance for AI-Assisted Inventions, 89 FR 10043 (13 February 2024) — the Pannu significant-contribution test; Revised Inventorship Guidance, 90 FR (28 November 2025, Director Squires) — rescinding the 2024 guidance in full and withdrawing the Pannu framework. India, verified this pass: the Indian Patent Office refused Thaler's DABUS application (No. 202017019068, filed 2021; order surfaced early 2026) on natural-person grounds under Sections 2/6/7 and the General Clauses Act 1897; Section 3(k) of the Patents Act 1970 bars "a mathematical or business method or a computer programme per se or algorithms," with only a narrow "technical effect" escape codified in the CRI Guidelines 2025 (effective 29 July 2025) — so the molecule is patentable as composition of matter while the model/algorithm faces the §3(k) wall; India has no dedicated trade-secret statute (protection via the Contract Act 1872 §27, the Specific Relief Act 1963, and equitable breach of confidence only), against the US Defend Trade Secrets Act 2016 and the EU Trade Secrets Directive 2016/943, with the 22nd Law Commission's Report on Trade Secrets and Economic Espionage (5 March 2024) recommending a sui generis law and appending a still-unpassed draft Protection of Trade Secrets Bill 2024; there is no IPO examiner guidance on AI inventorship (only the Parliamentary Standing Committee on Commerce's 161st Report, tabled 23 July 2021, flagging the Act as "not well equipped," plus the 2026 DABUS order); and India has no CDSCO drug-development-AI guidance to match the FDA's January 2025 draft, only a CDSCO Medical Device Software (SaMD) draft guidance of 21 October 2025. The bound stated in the text: India's thinness is statutory and guidance thinness, not an absence of all protection — molecules remain patentable, contracts and breach-of-confidence bind, a device-side framework exists, and the missing laws are drafted. The thesis in one line: the moat moved to the factory, no regime protects that layer cleanly, India's protection is thinnest of all (an express software-patent exclusion plus no trade-secret statute), and the cheapest thing India could do for its frontier-biotech future is pass the trade-secret bill and issue the AI-inventorship guidance already sitting on the shelf.