Crawl, walk, run.

Understanding the brain is an ambitious vision so we will sequence our work in order to offer products that create value along the way.

The best analogy for our approach is SpaceX.

1. 1
   Crawl · Cargo

   An Alzheimer's drug readout at synapse resolution

   Tell a drug program whether its compound saved synapses in the mouse, before it spends hundreds of millions of dollars on a Phase 3.

   The problem

   Alzheimer's is a disease of lost synapses: synapse density tracks cognitive decline far more closely than plaques or tangles do. Yet drug programs are judged on plaque load, bulk stains and behavior in mice, a record the NIH's MODEL-AD consortium sums up as hundreds of interventions that improved cognition in mice and models that have not predicted a therapy's effect in trials. In people the decisive test is still a Phase 3 of about 1,800 patients over 18 months, in a program modeled at $462 million per agent through Phase 3.

   Our solution

   We count every synapse the detectors find in a block of tissue, in 3D, and place each one relative to the pathology. In the public 3xTg mouse volume, synapse density within 10 micrometers of a plaque is 43 to 62 percent below plaque-free tissue in the same layer, depending on which of the preprint's two synapse detectors is used, while the loss averaged over the whole block is 6 percent, which is all a bulk stain sees. The product is a blinded study in the sponsor's own model, with these targets: eight animals per arm, treated, untreated and non-transgenic, one block of hippocampus each, the analysis fixed in advance, delivered in about six months and priced per animal, reporting synapse density by layer and by distance from plaques, tau and microglia before the decision to go to the clinic.

   Why it is better than what exists

   Today the question a program can answer before the clinic is whether a compound cleared plaques and moved a behavior score; in people, synapses are read only by SV2A PET, an indirect marker, and the program's fate is decided by cognition scores at the end of Phase 3, about 7.6 years and $462 million per agent after Phase 1 begins. Our study answers it in the mouse before Phase 1, with a placed count whose local deficit was seven to ten times the block average in our reanalysis of one plaque in one animal, in the same animals the program already runs. What that is worth is the clinical years and dollars not spent on compounds that fail the synapse test, up to 7.6 years and $462 million per agent, and the first paid study, which also runs array tomography on the adjacent block as the head-to-head, is the test of whether the readout earns it.

   The market if it works

   The buyer is a program deciding whether a compound goes forward: 158 drugs are in 192 Alzheimer's trials this year, with more behind them in preclinical work. Pharma has paid for human brain data inside target-discovery alliances, Merck $25 million upfront to Cerevance and GSK €33.5 million to Muna, and behind the programs is a disease of 7.4 million Americans, $409 billion a year in dementia care and Leqembi at about $555 million in sales in its third year. The pool grows as people outlive heart disease: a third of Americans 85 and older have Alzheimer's dementia, dementia is already the leading cause of death in the United Kingdom and Australia, and Japan, with a third of America's age-standardized heart-disease death rate, expects 5.84 million people with dementia by 2040.

   Terry et al., Ann Neurol 1991 · Sukoff Rizzo et al., MODEL-AD, 2020 · Territo, MODEL-AD, Alzheimer's & Dementia 2024 (AAIC abstract) · Mecca et al., SV2A PET, 2020 · Cummings et al., costs, 2022 · Scott et al., Ann N Y Acad Sci 2014 · van Dyck et al., Clarity AD, NEJM 2023 · Cummings et al., pipeline 2026 · Han et al., bioRxiv 2023 · Koffie et al., PNAS 2009 · Alzheimer's Association, 2026 Facts and Figures · Alzheimer's Association, 2025 Facts and Figures · Eisai FY2025 Leqembi sales (Fierce Pharma) · Alzheimer's Society, UK 2024 deaths · Australian Bureau of Statistics, 2024 · GBD 2021 ischemic heart disease, G20 · Japan Times, dementia estimate 2040

2. 2
   Walk · Starlink

   A readable map of the human brain, one disease region at a time

   The same measurement in the human regions each brain disease destroys, read with the tools we build, and sold to everyone who needs them.

   The problem

   Neurology drugs are approved at 5.9 percent from Phase 1 against 7.9 percent for drugs overall, and the animal results that put them in the clinic are rarely checked against human synapses, because the human data are thin. The largest dense human volume is a cubic millimeter of cortex from one surgical sample; the Cajal group's Alzheimer's samples are five cases and five controls per study. A human disease region mapped densely enough to read which cells and pathways lose their synapses does not yet exist.

   Our solution

   Map the regions each disease destroys at synapse resolution, from banked tissue with the donor's history attached and from living surgical tissue: hippocampus and entorhinal cortex for Alzheimer's first, then substantia nigra, motor cortex and spinal cord, prefrontal cortex. Banked tissue answers whether people who had plaques and stayed sharp kept their synapses; living cortical slices, which survive two to three weeks in culture, give a compound's acute effect on human synapses against a vehicle slice from the same patient. The software we build to read our own maps, starting with our proofreading model, goes to everyone producing maps.

   Why it is better than what exists

   The human brain data pharma has paid for are gene-expression and cell maps; a synapse-level map of a human disease region has not been sold. The mouse stays in the loop for the dosing experiments a person cannot be given, and the human map is what the mouse result gets checked against.

   The market if it works

   3.4 billion people live with a condition of the nervous system, the leading cause of ill health worldwide by disability-adjusted life years (the count includes headache and stroke). Pharma pays upfront for discovery platforms it can make decisions with: Roche paid Recursion $150 million upfront for up to 40 programs. The incretin drugs, GLP-1s among them, are forecast at $200 billion a year by 2030, the scale of one treatable disease class.

   BIO, clinical development success rates 2011 to 2020 · Shapson-Coe et al., Science 2024 · Montero-Crespo et al., Brain 2021 · Domínguez-Álvaro et al., Acta Neuropathol Commun 2018 · Schwarz et al., Sci Rep 2017 · Schwarz et al., eLife 2019 · GBD 2021 · J.P. Morgan Research, February 2026 · Recursion and Roche, 2021 · MICrONS Consortium, Nature 2025

3. 3
   Run · Mars

   A brain that runs

   A model of a mapped brain that predicts what it will do, and what a drug does to it.

   The problem

   We cannot yet predict what a brain will do from its structure. Drugs for the brain are found by trial and error because the models of the system they act on capture only parts of it, and the trial is the experiment.

   Our solution

   A simulation of a mapped brain, starting with the fly, then the MICrONS cubic millimeter, the one volume where activity and wiring come from the same neurons, then a whole mouse brain, which we intend to have mapped by 2033, checked against recorded samples of its activity and its behavior. The methods now reading artificial networks (Why now) apply here, because a simulated brain has activity to read. The first version is a research instrument; the later ones are where brain drugs get designed.

   Why the cost of a map will fall

   Costs collapse when the tool changes: Falcon 9 cut the cost of a kilogram to orbit about 20x below the Shuttle, and a human genome fell about 570,000-fold in nominal dollars, about a million-fold in 2025 dollars, in the twenty-two years from the $300 million draft of 2000 to $525 in 2022, most of it after the sequencer changed in 2008. A cubic millimeter of brain map has fallen about 10x since the first measured tracing rate in 2013, because its cost is people tracing by hand, and wages do not fall. We are changing the instrument, from electrons to light, and the labor, from people to models, and our target is the next 100x: a mouse brain under $100 million, the map a brain that runs is built on (Mapping, above).

   The market if it works

   Where drugs get designed, and, much further out, the mind itself: if a human mind can one day be reproduced outside biology, that is a path to extending human life indefinitely. The odds of success for this long-term vision are low and the journey will be long. But we are pursuing it anyway, because at a minimum it will further our understanding of the brain, and if we are successful it will be a tremendous benefit to humanity.

   Dorkenwald et al., Nature 2024 · MICrONS Consortium, Nature 2025 · Anthropic, 2025 · Herculano-Houzel et al., PNAS 2006 · Azevedo et al., J Comp Neurol 2009

   A kilogram to low Earth orbitA human genomeA cubic millimeter of brain mapFilled: spent or measured. Hollow: projection or goal. Bar: published range.