The refinement I'd add is that "data" and "validation" serve very different roles in that phrase, and the gap between them is where most of the value lies. Proprietary data is necessary and increasingly common. Validated data, calibrated to the biology you're actually drugging, is the rare, capital-intensive part, and it's the only thing that survives the stages Anthropic can't touch: tox, efficacy, patient selection.
Which sharpens @Bengüsu's build-vs-partner point nicely. You can rent a model, and soon everyone will. You cannot rent a validated loop for your specific biology, the thing that tells you whether the data predicts the clinic. That is what a small player with real data and tacit knowledge actually owns, and it's the leverage to set terms before the big labs generate their own.
It also makes the China section the sharper story: speed and enrollment scale are a validation advantage as much as a cost one, more shots and faster readouts, and that compounds exactly where model access doesn't.
Interesting perspective. As frontier AI companies expand into life sciences, the real differentiator will be how well they combine powerful models with high-quality biological data, domain expertise, and rigorous validation. Exciting space to watch.
Fully agree with "data and validation are the moat". But the model question is not only about the model but its well-resourced developers (Anthropic, OpenAI, DeepMind etc.) Smaller actors hold the data and the tacit knowledge today, and what I mean by this is life sciences companies, universities and affiliated high-tech research centers, scientists... They will need to answer whether they seek value via AI models on their own or eventually partner with the big AI models, before the big companies get to similar data or expertise. I think many small players, particularly in public orgs, will lack the ML talent to build their own scaffolding, and compute, even though there is more public compute initiatives now. If this is recognized early on, these actors can be more proactive and set their own terms for a long-lasting partnership. Curious to hear your take.
Thank you for your comment. Makes a lot of sense, and it is a known tradeoff between renting AI capabilities, with all the risk involved, or trying to build from the ground up with another set of obvious challenges. I think there is no right answer here, really. It depends on the company skillset, funding, and its vision. The closest to the truth is probably a hybrid approach, with using whatever is available from external AI partners, but also building the core IP with your own resources on your own terms.
Thanks for your insights! I think I'm slightly confused by your statement, "their moat was never the model alone" which implies that Claude Science competes with AI models of the AI-native biotechs you listed. From what I understand, Claude Science solves scientific thinking, analysis, and hypothesis generation whereas the proprietary AI models of the AI-native biotech companies focus on providing high quality hits for a specific target. It seems like Claude Science would instead compete with and replace scientists in these organizations, not the actual company itself, right? Anthropic's drug discovery program would be more likely to partner with these AI-native biotech companies, no?
In my head Claude Science is more of a direct competitor to Edison Scientific and DeepMind Co-Scientist.
Let me know if I'm wrong! I'm very much a noob in the space
Thank you for your question. First of all, you are obvioulsy right saying that Claude Science is a competitor to Edison and alike.
But regarding AI native biotechs, this situation would be true back in 2015-2018, when the first wave of AI drug discovery companies, like Atomwise, were indeed mostly focused on narrow drug design and no one had any broad reasoning capabilities. However, almost all leading AI-native biotechs today such as Insilico Medicine, OWKIN, Recursion, BenchSci, Lantern Pharma, CytoReason, etc, etc, as well as some new entrants like NOETIK, SandboxAQ, etc, are all now in the game of "biological reasoning" and "building end -to end" models for everything. And beyond core drug design tasks, most of these companies are now also offering a wide range of capabilities, which is exactly what Claude is doing. Say, Insilico Medicine's PreciousGPT family of models, or NOETIK's world models of cancer biology, etc. They are essentially centered around the idea of having a general-purpose AI model which then spins all sorts of assets in various verticals (molecules, biomarkers, diagnostics, even clinical-stage insights like patient stratification, etc. If Claude Science can compete on this value proposition, essentially being that end-to-end layer for all sorts of eary biology, and target discovery tasks, it would obviously be a categorical competition, directly with the companies. Say, Insilico claims to have 3500+ users of their AI platforms. People use their platforms for questioning their models about all sorts of things like mechanisms of action, etc, via ChatGPT-style workflow. I imagine this is not cheap. Now, if Claude Science can do the same, it is competition.
But here comes my second part of the argument. Claude Science is JUST a model, they don't really have antying else (yet). The native AI biotechs have been building their wetlab capabilities, internal data acquisition workflows etc etc for years, even a decade. Say, Recursion is able to collect high content screening data (like images of cells in phenotypic cell perturbation experiments) on the order of thousands of experiments per day. So, they generate petabytes of proprietary data directly from wetlab. Claude Science is simply not there yet, their models can only learn from public data or some licensed datasets. So, that's why I say AI natives are usually NOT ONLY models.
The refinement I'd add is that "data" and "validation" serve very different roles in that phrase, and the gap between them is where most of the value lies. Proprietary data is necessary and increasingly common. Validated data, calibrated to the biology you're actually drugging, is the rare, capital-intensive part, and it's the only thing that survives the stages Anthropic can't touch: tox, efficacy, patient selection.
Which sharpens @Bengüsu's build-vs-partner point nicely. You can rent a model, and soon everyone will. You cannot rent a validated loop for your specific biology, the thing that tells you whether the data predicts the clinic. That is what a small player with real data and tacit knowledge actually owns, and it's the leverage to set terms before the big labs generate their own.
It also makes the China section the sharper story: speed and enrollment scale are a validation advantage as much as a cost one, more shots and faster readouts, and that compounds exactly where model access doesn't.
Thank you for your comment, yes, it is a genuinely multi-parameter problem to solve, with many variables affecting the business model and strategy.
Great read. The China angle is gold.
Thank you!
Interesting perspective. As frontier AI companies expand into life sciences, the real differentiator will be how well they combine powerful models with high-quality biological data, domain expertise, and rigorous validation. Exciting space to watch.
Exactly!
Fully agree with "data and validation are the moat". But the model question is not only about the model but its well-resourced developers (Anthropic, OpenAI, DeepMind etc.) Smaller actors hold the data and the tacit knowledge today, and what I mean by this is life sciences companies, universities and affiliated high-tech research centers, scientists... They will need to answer whether they seek value via AI models on their own or eventually partner with the big AI models, before the big companies get to similar data or expertise. I think many small players, particularly in public orgs, will lack the ML talent to build their own scaffolding, and compute, even though there is more public compute initiatives now. If this is recognized early on, these actors can be more proactive and set their own terms for a long-lasting partnership. Curious to hear your take.
Thank you for your comment. Makes a lot of sense, and it is a known tradeoff between renting AI capabilities, with all the risk involved, or trying to build from the ground up with another set of obvious challenges. I think there is no right answer here, really. It depends on the company skillset, funding, and its vision. The closest to the truth is probably a hybrid approach, with using whatever is available from external AI partners, but also building the core IP with your own resources on your own terms.
Great read.
Many thanks!
Thanks for your insights! I think I'm slightly confused by your statement, "their moat was never the model alone" which implies that Claude Science competes with AI models of the AI-native biotechs you listed. From what I understand, Claude Science solves scientific thinking, analysis, and hypothesis generation whereas the proprietary AI models of the AI-native biotech companies focus on providing high quality hits for a specific target. It seems like Claude Science would instead compete with and replace scientists in these organizations, not the actual company itself, right? Anthropic's drug discovery program would be more likely to partner with these AI-native biotech companies, no?
In my head Claude Science is more of a direct competitor to Edison Scientific and DeepMind Co-Scientist.
Let me know if I'm wrong! I'm very much a noob in the space
Thank you for your question. First of all, you are obvioulsy right saying that Claude Science is a competitor to Edison and alike.
But regarding AI native biotechs, this situation would be true back in 2015-2018, when the first wave of AI drug discovery companies, like Atomwise, were indeed mostly focused on narrow drug design and no one had any broad reasoning capabilities. However, almost all leading AI-native biotechs today such as Insilico Medicine, OWKIN, Recursion, BenchSci, Lantern Pharma, CytoReason, etc, etc, as well as some new entrants like NOETIK, SandboxAQ, etc, are all now in the game of "biological reasoning" and "building end -to end" models for everything. And beyond core drug design tasks, most of these companies are now also offering a wide range of capabilities, which is exactly what Claude is doing. Say, Insilico Medicine's PreciousGPT family of models, or NOETIK's world models of cancer biology, etc. They are essentially centered around the idea of having a general-purpose AI model which then spins all sorts of assets in various verticals (molecules, biomarkers, diagnostics, even clinical-stage insights like patient stratification, etc. If Claude Science can compete on this value proposition, essentially being that end-to-end layer for all sorts of eary biology, and target discovery tasks, it would obviously be a categorical competition, directly with the companies. Say, Insilico claims to have 3500+ users of their AI platforms. People use their platforms for questioning their models about all sorts of things like mechanisms of action, etc, via ChatGPT-style workflow. I imagine this is not cheap. Now, if Claude Science can do the same, it is competition.
But here comes my second part of the argument. Claude Science is JUST a model, they don't really have antying else (yet). The native AI biotechs have been building their wetlab capabilities, internal data acquisition workflows etc etc for years, even a decade. Say, Recursion is able to collect high content screening data (like images of cells in phenotypic cell perturbation experiments) on the order of thousands of experiments per day. So, they generate petabytes of proprietary data directly from wetlab. Claude Science is simply not there yet, their models can only learn from public data or some licensed datasets. So, that's why I say AI natives are usually NOT ONLY models.
cool, thanks for the in depth response!
You are welcome!