A remark about where Isomorphic Labs keeps its models exposes a structural blind spot in this week’s pacing debate: the throttle everyone is arguing about sits on only one valve in a system that has already developed several others.
Isomorphic Labs — Key Markers
2021
Founded as a DeepMind spinoff by Demis Hassabis; Google-backed from inception, focused on AI-driven drug discovery.
Active collaborations
Partnerships established with Novartis, Eli Lilly and Johnson & Johnson across oncology and immunology programmes.
May 2026
$2.1 billion raised in latest funding round, led by Thrive Capital. Valuation undisclosed.
15 Sept 2026
Chris Butler (drug discovery lead) speaks at the Semafor Future of Health Forum; remarks on internal model architecture reported by Semafor’s J.D. Capelouto.
Reporting by Semafor’s J.D. Capelouto from Tuesday’s Future of Health Forum captured a remark that is easy to misread. Chris Butler, who leads drug discovery at Isomorphic Labs, described the architecture of his company’s AI work rather than staking out a position on industry governance. “We’re able to still progress regardless of what decisions are made elsewhere,” he said — not as a defiant claim, but as a description of operational reality. His models do not live outside the company; they live inside it. “All of our AI models are locked down and locked down in-house. So, really, for us, it’s about, how do we use the models that we’ve built internally to help to accelerate the drug discovery process.”
Butler did not oppose pacing. He did not criticise Dario Amodei or any safety advocate. He did not say Isomorphic refuses to slow down. He described where his models live — a fact about infrastructure, not a policy argument. On the therapeutic side, he confirmed the company is moving smoothly through preclinical development in oncology and immunology alongside its pharma collaborators, and stressed the regulatory dimension: “It is really important that we continue to see that progress toward the clinic.” He declined to give clinical-trial timeline specifics, and none are offered here.
Isomorphic is the Google-backed drug company founded in 2021 as a DeepMind spinoff by Demis Hassabis. Its $2.1 billion May raise, led by Thrive Capital at an undisclosed valuation, makes it one of the best-capitalised AI-drug-discovery companies in existence. Its collaborators — Novartis, Eli Lilly and Johnson & Johnson — are among the largest pharmaceutical firms on earth. The company is private; Alphabet, Novartis, Eli Lilly and Johnson & Johnson are publicly listed. Nothing here is investment advice, medical advice, a view on any security or a recommendation.
The key insight: Butler’s remark is a statement about infrastructure topology, not governance preference. Its analytical weight lies not in what Isomorphic intends to do, but in what it reveals about the structural limits of every pacing mechanism debated this week — none of which touches capability already trained, held and deployed inside a vertical.
A release-cadence agreement binds whoever is about to ship a frontier model. It has no grip on a company running models it already built and holds internally — which is a claim about coverage, not about whether pacing is a good idea.
The Structural Read
The analysis below is this publication’s own, not anything said or implied at the forum.
Every coordination mechanism proposed in the pacing debate this week — escalating capability tiers, embedded evaluator commitments, a narrow antitrust waiver to enable joint slowdowns, a more conservative release cadence — operates on the same variable: the rate at which new frontier capability is created and released. That is a flow instrument. What Butler described, without intending to make a policy argument, is a stock problem.
A company running models it built and holds internally is not downstream of anyone’s release decision. An agreement struck among frontier laboratories changes nothing about what that company can do tomorrow morning. The coverage is structural, not incidental: any firm that has already trained a capable domain model and keeps it in-house sits entirely outside the perimeter that release-cadence agreements can draw. Pacing is a supply-side instrument arriving after a considerable amount of supply has already shipped — and that is a claim about coverage, not a claim about whether restraint is desirable.
The relationship also compounds in the wrong direction. The more successfully AI capability diffuses into verticals — drug discovery, legal research, materials science, financial modelling — the smaller the share of total deployed capability that any frontier agreement can reach. A coordination mechanism optimised for 2023’s supply chain is being applied to 2026’s deployment map.
Permission Layer — Coverage Gap
A permission mechanism only governs what passes through its gate. Release-cadence agreements gate the frontier laboratory’s output. They do not gate the capability a vertical company trained last year and runs on its own infrastructure today. The gate is real; the coverage is partial — and the gap grows with every successful diffusion event.
The Hassabis connection deserves careful handling, because the obvious reading of it is wrong. Isomorphic was founded as a DeepMind spinoff by Demis Hassabis, who days earlier was among those publicly backing the direction of Amodei’s argument. That is not a contradiction and no hypocrisy is alleged — against Hassabis, against Isomorphic, against DeepMind or against Alphabet. Hassabis does not speak for Isomorphic; Butler does not speak for Hassabis. Supporting restraint on frontier general-purpose systems is entirely consistent with continuing domain-specific drug-discovery work, which is a different activity with a different risk profile. The genuinely interesting point is subtler and more structurally damaging to the proposal than any personal inconsistency would be.
The pacing conversation has been conducted throughout as though “AI capability” were a single object with one throttle attached to it. In practice it is many objects, held by many organisations, at many stages of diffusion — and the throttle everyone has been arguing about sits on only one of them. That is the observation Butler’s remark makes available, even though he was not making it.
There is also a cost-benefit asymmetry that explains why application companies will not volunteer for the same restraint as frontier labs, regardless of their views on safety in the abstract. For a frontier laboratory, the marginal benefit of the next capability increment is diffuse across a general-purpose user base, and the risk scenarios are heavily theorised. For a preclinical oncology and immunology company working alongside Novartis, Eli Lilly and Johnson & Johnson, the benefit of the next model improvement is specific, nameable and ultimately measured in patients. A blanket norm therefore asks verticals to price forgone therapeutics against a risk profile that largely belongs to someone else’s system.
That framing does not assert that domain-specific models are safe, low-risk or exempt from scrutiny. Risk profiles differ by domain, and the hazards specific to AI in drug discovery — including biosecurity and misuse vectors — are serious and distinct. Butler did not discuss those topics at the forum; the Semafor reporting did not address them; and this article does not adjudicate them. The claim is only that the risk architecture differs, and that a governance instrument designed for one risk architecture provides incomplete coverage of another.
What a Mechanism With General Coverage Would Have to Look Like
If capability already held in-house sits beyond the reach of release-rate agreements, then instruments binding at the point of deployment rather than release are the only ones that touch the whole field. An audit-and-certification standard attaches to a deployed agent — wherever it runs and whoever trained it. A disclosure obligation attaches to a listed company’s public statements about what its systems do and what they have been validated against. A procurement requirement attaches to a purchase, reaching any supplier regardless of whether it participated in a frontier-lab coordination agreement.
Each of those instruments reaches a firm running internal models it trained itself. A release-cadence agreement among frontier laboratories reaches none of them. This is not a prediction that any deployment-binding mechanism will be adopted or will work — several were publicly refused or remain unproven within the last four days. It is an observation that the pacing debate has been optimising a throttle on one valve in a system that has already developed several other flows.
Three Implications
IMPLICATION 1 — The Coverage Problem Compounds Over Time
Every domain in which a capable AI model is trained and internalised reduces the share of total deployed capability that frontier-lab agreements can reach. This is not a temporary gap to be closed by faster policy iteration; it is a structural feature of how capability diffuses. Governance frameworks that do not account for already-distributed stock will become less comprehensive with each passing quarter, not more.
IMPLICATION 2 — Deployment Is the More Durable Regulatory Surface
Audit-and-certification standards, disclosure obligations tied to listed-company filings and procurement requirements all attach at the point of use rather than the point of release. They are harder to design, require domain-specific expertise and face their own political obstacles — but they are the only instruments with structural coverage of the entire deployed field, including verticals running their own internally trained models. The policy conversation has been gravitating toward this without quite arriving at it.
IMPLICATION 3 — The Single-Throttle Model of AI Governance Is Already Obsolete
The framing that treating “AI capability” as a single object with one throttle was always a simplification; it is now actively misleading. Capability exists at the frontier lab level, the large-enterprise level, the startup level and the domain-specialist level — each with a different risk profile, a different diffusion path and a different relationship to any proposed coordination mechanism. Governance architecture that doesn’t map this topology will keep solving for the wrong variable.
Business Engineer Framework
The Permission Layer — and Why It Only Governs What Passes Through the Gate
The Permission Layer framework maps where government, regulation and coordination mechanisms sit in the AI stack — and, critically, which layers they can and cannot reach. The Isomorphic moment is a textbook Permission Layer case: a gate exists, it is real, and its coverage is structurally partial. Understanding the topology of where the gate sits relative to deployed capability is the first step toward designing instruments that actually bind. The full Map of AI framework plots 200+ companies across nine layers and shows exactly where each type of instrument attaches.
Chris Butler’s remark at the Semafor Future of Health Forum was a description of infrastructure, not a governance argument — but it inadvertently named the central structural problem with every pacing proposal debated this week. Pacing regulates the flow of new frontier capability; it has no purchase on the stock of capability already trained, distributed and running inside verticals. Isomorphic Labs is one data point; the dynamic is general. A governance framework that optimises one valve while several others run freely is not governing the field — it is governing the part of the field that was easiest to reach when the conversation started.
Sources: Semafor — J.D. Capelouto, 15 September 2026. Structural analysis is this publication’s own. Company facts sourced from Semafor reporting and publicly available corporate records. Nothing in this article is investment advice, medical advice, a view on any security or a recommendation.
Chris Butler did not argue against pacing, did not criticise Dario Amodei or anyone advocating restraint, and did not say Isomorphic Labs refuses to slow down. He described the fact that the company’s models are built and held internally. Nothing here should be read as positioning him, Isomorphic Labs or its investors as opponents of AI safety measures. The distinction between the flow of newly released capability and the stock of capability already held is this publication’s analysis. Neither Butler nor the reporting drew it, and it should not be attributed to them. Demis Hassabis founded Isomorphic Labs and was among those publicly supporting the direction of the pacing argument. Those positions are not in conflict: he does not speak for Isomorphic, Butler does not speak for him, and supporting restraint on frontier general-purpose systems is consistent with continuing domain-specific drug-discovery work. No hypocrisy or inconsistency is alleged against him, Isomorphic Labs, DeepMind or Alphabet. The valuation attached to the May funding round was not disclosed, and none is stated or estimated here. Butler declined to give clinical-trial timeline specifics, and no timeline is offered. No claim is made about any drug, trial, regulatory approval or patient outcome, and nothing here suggests any therapy is close to market. This is not medical advice. Nothing here asserts that domain-specific models are safe, low-risk or exempt from scrutiny; the observation is that risk profiles differ by domain. Butler did not discuss biosecurity or misuse, the reporting did not address either, and this article does not adjudicate them. No prediction is offered about which governance mechanism, if any, is adopted or succeeds. Isomorphic Labs is a private, Google-backed company; Alphabet, Novartis, Eli Lilly and Johnson & Johnson are publicly listed; Anthropic and OpenAI are private. No claim is made about any share price or market effect. This is businessanalysis, not investment advice, no view is expressed on any security, and no recommendation is made.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
