The Cure May Already Exist: How AI Is Turning Medicine Into a Search Problem

Sometimes the problem is not that medicine has failed to invent the drug.

Sometimes the problem is that nobody knows the drug already exists. That distinction is at the center of one of the more interesting applications of artificial intelligence in medicine.

In January 2024, Joseph Coates was critically ill with POEMS syndrome, a rare multisystem disorder associated with abnormal plasma cells. His heart and kidneys were failing, fluid repeatedly accumulated in his abdomen, and he was on life support. He had become too sick to undergo the stem-cell transplant that might put his underlying disease into remission.

His girlfriend, Tara Theobald, contacted physician-scientist David Fajgenbaum. Using Every Cure's AI-assisted drug-repurposing work, Fajgenbaum examined possible treatments and proposed a three-drug regimen commonly associated with multiple myeloma: dexamethasone, cyclophosphamide and carfilzomib.

The biological jump was not random. POEMS and multiple myeloma both involve abnormal plasma-cell biology. The unusual part was that this particular regimen had not been used for Coates's condition in that circumstance. According to Every Cure, he began improving within a week. Four months later he was healthy enough to receive a stem-cell transplant, and Every Cure's 2025 annual report says he remained in remission.

The AI did not invent the drugs. It changed the cost of searching for the connection.

The story actually begins with Fajgenbaum himself.

Years before Every Cure existed, Fajgenbaum was the patient. As a medical student, he developed idiopathic multicentric Castleman disease, a rare inflammatory and lymphoproliferative disorder capable of causing life-threatening multiorgan dysfunction. Existing therapy did not reliably control his disease.

Fajgenbaum and collaborators investigated his biology and identified abnormal activity in the PI3K/Akt/mTOR pathway. That suggested a possibility: sirolimus, an existing mTOR inhibitor widely used for other purposes. The treatment produced durable remission in Fajgenbaum and other refractory patients in the initial research cohort. Later studies found broader evidence of mTOR activation in the disease, and subsequent cohorts have continued to investigate sirolimus as a repurposed therapy.

One patient searching for one overlooked drug eventually became an attempt to search every drug against every disease.

That is the idea behind Every Cure.

Fajgenbaum's experience raised a deceptively simple question. How many other treatments are already sitting inside the existing pharmaceutical inventory but have never been connected to the disease they might help?

Every Cure was founded in 2022 around that problem. The organization now describes a search space of more than 4,000 existing drugs and roughly 18,000 to 18,500 diseases. That produces approximately 75 million potential drug–disease relationships. A human research organization cannot systematically investigate 75 million possibilities one at a time. A computational system can at least rank them.

This is not simply an LLM recommending drugs.

Every Cure's core MATRIX pipeline is built around biomedical knowledge graphs and machine learning. The platform integrates relationships among drugs, genes, proteins, disease pathways, clinical evidence and other biomedical data.

Its public technical documentation describes three broad stages. First, multiple biomedical knowledge graphs and datasets are ingested and integrated. Second, the graph is transformed into machine-usable features and embeddings. Third, models predict and score potential treatment relationships. Every Cure has since added LLM-based and agentic approaches that can reason through evidence and pressure-test potential matches.

The bottleneck is fragmented knowledge.

Scientific information does not live in one place. One paper may describe a disease pathway. Another may identify a protein target. A third may describe a drug's mechanism. A fourth may contain an obscure case report. A clinical database may hold another piece, and a biobank another. The relevant conclusion may already exist across all of those sources without ever appearing in a single paper. This is where computational systems have an advantage.

AI does not need to know something that humanity does not know in order to create enormous value. Sometimes it only needs to connect what humanity already knows.

That makes the Joseph Coates story less magical—and more important.

POEMS is not biologically unrelated to multiple myeloma. Both involve plasma-cell pathology, and modern POEMS treatment already borrows from therapies used against plasma-cell disorders. The AI did not discover that three completely random medicines could reverse an unrelated condition. It helped surface a plausible treatment path from an adjacent disease that had not been applied to this particular situation. That is exactly the kind of connection a scalable search system should find.

AI still does not make the clinical decision.

A high model score is not medical evidence. It is a reason to look more closely. Every Cure says its scientists and physicians evaluate predicted drug–disease relationships for biological mechanism, existing evidence, clinical feasibility, safety, unmet need and the practical possibility of proving that the treatment works. The organization has publicly described candidates that looked attractive computationally but were rejected during expert review because the real-world biology did not support the hypothesis.

The model narrows the search space. Medicine still has to prove the answer.

The next constraint is validation.

AI can generate candidate relationships much faster than researchers can test them. That moves the bottleneck downstream.

This is already happening beyond one rescue case.

The strongest evidence for Every Cure's model is not Joseph Coates alone. It is whether the organization can repeatedly move computational signals toward clinical practice. One useful example is Rosai-Dorfman disease, a rare inflammatory and histiocytic condition. Every Cure's platform ranked lenalidomide highly as a potential treatment, and the organization combined that signal with existing physician experience and further clinical analysis.

In 2026, updated National Comprehensive Cancer Network guidance elevated lenalidomide plus dexamethasone to a preferred treatment option for Rosai-Dorfman disease regardless of mutation status. That is a different kind of success. Not an emergency rescue—a computational signal moving through evidence review and into clinical guidance.

The economics explain why these opportunities can remain hidden.

Traditional pharmaceutical development is built around the economics of discovering and commercializing new intellectual property. Drug repurposing can create a different incentive structure. If an old generic medicine unexpectedly works for a small rare-disease population, proving that use may create enormous value for patients without creating comparable financial value for a patent owner. There may be no company with enough economic incentive to fund the work. Every Cure says more than 80% of approved drugs are generic, making this incentive problem particularly important.

This is where public capital enters the system.

The market failure is important enough that the U.S. government is funding infrastructure to solve it. In 2024, ARPA-H committed $48.3 million to help Every Cure develop MATRIX. The current ARPA-H award record lists the overall MATRIX program at up to $124 million. In February 2026, Every Cure announced selection for a second phase of up to $76 million, designed to move the computational output downstream: at least 20 prioritized opportunities into preclinical work, and 10 promising opportunities toward clinical trials.

The compute itself is getting dramatically faster.

Google Cloud says an early version of Every Cure's full prediction pipeline required approximately 100 days for one cycle. Using a cloud data architecture and GPU-based computation, that cycle has reportedly fallen to approximately 17 hours. That changes the research process. A 100-day model run is something an organization performs occasionally. A 17-hour run can become iterative: update the data, change the model, test another hypothesis, run it again.

The value of faster compute is not simply that the same answer arrives sooner. It allows the research loop itself to run more times.

The organization is now adding agents.

The next evolution looks increasingly familiar from the broader frontier-AI market. Every Cure said in September 2026 that one of its major technology investments this year is building AI agents to help investigate potential drug–disease relationships. The agents do not eliminate the medical team—every prediction still undergoes expert review. Instead, the agent layer can help collect evidence, reason through mechanisms, challenge assumptions and prioritize which hypotheses deserve scarce human attention. That is the same architectural shift appearing across many knowledge industries: AI moves from generating an answer to performing pieces of a workflow.

This is where the story connects to Navier–Stokes.

The two cases look very different. One involves an unsolved problem in mathematical fluid dynamics; the other involves existing medicines and rare diseases. But both illustrate a larger transition: AI infrastructure is becoming research infrastructure. In the Navier–Stokes case, the system attempts to create knowledge that did not previously exist. In the Every Cure case, the system attempts to recover knowledge that may already exist in fragmented form. One is generative discovery. The other is computational synthesis. Both convert compute into scientific work.

Navier–Stokes asks whether AI can discover something humanity does not know. Every Cure asks whether AI can discover something humanity knows in pieces but has never connected.

This should not be turned into a gigawatt claim.

There is an important infrastructure distinction. Every Cure is not evidence that one biomedical knowledge graph requires a hyperscale AI campus. The organization has not disclosed anything suggesting that its workload looks like frontier-model training. The more interesting point is economic: the value produced by compute can be massively disproportionate to the physical size of the workload. One useful drug–disease connection can matter more than billions of low-value inference responses.

The future value of AI infrastructure will not be measured only by how much compute it produces. It will also be measured by what that compute discovers.

And successful AI can create new physical bottlenecks.

If computational systems become dramatically better at identifying promising therapies, the scarce resources shift: wet-lab capacity, clinical-trial participants, specialist physicians, regulatory review, manufacturing, patient recruitment, and capital willing to fund opportunities that may not offer conventional pharmaceutical economics. That is a recurring pattern in AI. When intelligence becomes cheaper, the physical world becomes more visible as the constraint.

The deepest opportunity may be hiding in plain sight.

Drug discovery traditionally begins with what medicine has not yet invented. Drug repurposing begins with what medicine has already built. That difference creates a remarkable possibility: some future treatments may not require a new molecule, a new manufacturing plant or an entirely new pharmaceutical supply chain. They may require the right existing molecule to be connected to the right disease. The search problem is enormous—but enormous search problems are exactly where computation changes economics.

Bottom line

Joseph Coates is a compelling story. But one dramatic patient outcome is not the larger thesis. The larger thesis began when David Fajgenbaum nearly died, found a new use for an existing medicine, and asked how many similar opportunities medicine had missed. Every Cure is now attempting to answer that question systematically: more than 100 biomedical datasets, thousands of drugs, thousands of diseases, tens of millions of possible relationships, machine-learning models, knowledge graphs, AI agents, human scientists, laboratory validation, clinical trials—and ultimately patients.

It is a reminder that AI's scientific value does not require a machine to independently invent everything from first principles. There is already an extraordinary amount of human knowledge in the world. Much of it is simply too fragmented for any person to search completely.

The cure may already exist. AI's breakthrough may be making the world's medical knowledge searchable enough to find it.