AI-discovered drugs pass Phase I trials at nearly double the industry rate. They pass Phase II at exactly the industry rate. The split is not a failure of current models. It is a boundary — between problems that yield to search and problems that require contact with reality. The same boundary runs through software, finance, writing, and the $650 billion infrastructure bet.
One hundred and seventy-three AI drug programs are now in clinical or preclinical stages. The data is starting to resolve. Not conclusively — no AI-discovered drug has received FDA approval — but clearly enough to see the shape of what AI does and does not change about discovering medicine.
Phase I clinical trials test whether a molecule is safe and behaves as the physics predicts. Does it absorb? Does it distribute? Does it metabolize on schedule? These are structural questions — the molecule either matches the model or it does not. AI-discovered drugs pass Phase I at eighty to ninety percent. The industry average is forty to sixty-five percent. AI roughly doubles the structural hit rate.
Phase II clinical trials test whether binding to the target actually treats the disease. Does the interaction produce the therapeutic effect? Does the patient improve? These are functional questions — the answer depends on biological networks, compensatory mechanisms, patient variation, and emergent properties that cannot be computed from the molecular structure alone. AI-discovered drugs pass Phase II at approximately forty percent. The industry average is approximately forty percent.
Below a certain threshold of complexity, AI is transformative. Above it, AI is average.
The First Proof
Rentosertib is the first AI-designed drug to demonstrate dose-dependent clinical efficacy. Developed by Insilico Medicine for idiopathic pulmonary fibrosis, it reached human trials in eighteen months — the industry average is four years. The target is TNIK, a well-characterized kinase with a thoroughly mapped structure-function relationship.
Kinases are among the most structurally predictable drug targets in pharmacology. The distance between knowing a kinase's shape and predicting what happens when you bind to it is shorter than for almost any other target class. Rentosertib's efficacy proof comes from the easiest territory — the place where structure and function are closest together.
This is not a failure of AI drug discovery. It is an ordering principle. The first successes will cluster where the gap between structural prediction and functional outcome is smallest. Kinases first. Then ion channels and GPCRs. Then complex multi-pathway disease mechanisms last. The sequence is the boundary made economic.
The Boundary That Does Not Move
In February, Isomorphic Labs released IsoDDE — an engine that doubles AlphaFold 3's accuracy on protein-ligand structure prediction. It predicts binding affinity matching physics-based gold standard methods at a fraction of the cost. It discovers cryptic binding sites that only appear during molecular interaction. This is genuinely impressive work.
It is also entirely below the boundary. IsoDDE computes conditional structure from molecular physics — what shape does the protein take when this ligand approaches? That question is deterministic. The answer exists in the physics. Better models find it faster and cheaper.
The question IsoDDE does not answer — the question no structural model answers — is what happens to the disease when you bind there. Therapeutic function depends on network effects, compensatory pathways, immune response, patient biology. These properties emerge from context-dependent interaction. They cannot be computed from molecular structure for the same reason you cannot compute the meaning of a sentence from the shapes of its letters.
The drug discovery community has articulated this independently, without the formalism. Their summary of 2026: the most important question is not whether AI can accelerate preclinical timelines — we know it can — but whether it can improve clinical success rates. Speed below the boundary is proven. Accuracy above it is unproven. They are living inside the boundary and describing what they see.
Here is the distinction that matters: this is a complexity boundary, not a capability boundary. Better AI makes the region below the boundary larger and cheaper. It does not move the boundary itself. IsoDDE is Moore's Law for structural biology — faster and cheaper within the computable, not past it.
The Same Boundary Everywhere
Drug discovery provides the cleanest empirical test because the Phase I and Phase II split maps directly onto below and above. But the same topology appears across every domain where AI is deployed at scale.
Software. AI-generated code compiles. It passes syntax checks. It matches structural patterns in the training data. Whether the code is the right code — whether it solves the business problem it was written for — requires deployment and user response. Structural correctness is below the boundary. Fitness for purpose is above it.
Finance. Quantitative execution — high-frequency trading, order routing, derivative pricing — excels. These operations are deterministic from market microstructure. Alpha generation — identifying which assets will outperform — is at parity or worse. It requires contact with future market states that cannot be computed from present data.
Writing. Structural output is massively accelerated. Style, syntax, volume, formatting — all below the boundary. Truth, meaning, originality — all above it. A language model can produce fifteen well-structured articles in a day. Whether they are worth reading requires a reader.
Prediction markets. A recent benchmark showed that giving language models more internal reasoning time worsened their probability calibration. More computation below the boundary did not help above it. What improved calibration was web search — external information from outside the model. The boundary again: internal search is insufficient. External contact is required.
Closed games confirm the pattern by its absence. Every property of a Go position is computable from the rules. There is no region above the boundary. AlphaGo's dominance is total precisely because the game is entirely structural. When there is no above-boundary region, there is no boundary — and AI wins completely.
The universal pattern: AI produces speed below the boundary and parity above it.
What the Boundary Means for the Infrastructure Bet
The AI infrastructure buildout does not need to cross the boundary to be profitable. The region below it is enormous: structure prediction, code generation, content production, data processing, execution optimization. Making all of this ten times faster and a hundred times cheaper transforms economics even without touching the region above.
But the narrative driving the investment — that sufficient compute eventually produces artificial general intelligence, systems that can do anything — is a claim that the boundary will dissolve. The drug discovery data suggests otherwise. One hundred and seventy-three programs. Dramatic structural acceleration. Clinical efficacy at industry baseline. The searchable space has expanded. The unsearchable space is unchanged.
This is not an argument that the six hundred and fifty billion dollars is wasted. It is an argument about what it buys. Below the boundary, AI is a revolution — compressing years into months, multiplying throughput, reducing costs by orders of magnitude. Above the boundary, AI is a tool — useful, but not transformative. Not yet. And the data does not show the boundary moving.
The question worth six hundred and fifty billion dollars is not whether AI can search faster. It demonstrably can. The question is whether search, no matter how fast, can substitute for contact with reality. The Phase I and Phase II numbers are the clearest answer available: within the region where structure determines outcome, AI is extraordinary. At the boundary where function emerges from context, AI is ordinary.
The boundary is not a ceiling on AI's value. It is a map of where the value concentrates — and a warning about where the narrative outruns the evidence.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)