Edison Scientific

The lab that runs while you sleep

It is 2 a.m. in a SoMa office and the lights are mostly off. A queue of research jobs is chewing through the night. Each one is reading more papers than a graduate student gets through in a year, and writing more code than most postdocs ship in a quarter. By morning, a biology team in Wilmington will open a report on a target they had not yet thought to ask about. This is Edison Scientific, the company that decided science should not have business hours.

Edison was incorporated in 2025 as a spinout of FutureHouse, a nonprofit research lab co-founded two years earlier by Sam Rodriques and Andrew White. The lab built tools, then watched pharma companies and academic centers ask, gently and then less gently, when they could pay for them. Edison is the answer. The original nonprofit still exists. The for-profit just ships.

Today the company is sixty-one people, headquartered on a quiet block in San Francisco, with a $250 million paper valuation and a flagship product named Kosmos. Most of its users have never met a salesperson. They signed up, ran a job, and came back with their lab.

Science has a throughput problem

The promise of modern biology was that the wet lab would get faster. Sequencing got cheaper. CRISPR got friendlier. Robots replaced pipette tips. And yet a typical drug program still takes a decade, mostly because nobody can read fast enough. The literature doubles every nine years. A motivated graduate student can plough through maybe four hundred papers in a year, if they sleep poorly. Most of the answers a project needs are already published. Nobody has the hours to find them.

Rodriques, a physicist turned bioengineer, spent his postdoc inventing methods for spatial transcriptomics - which is to say, mapping where in a cell each gene gets expressed. He noticed the same thing every researcher notices, then politely pretends not to: most of the work is not at the bench. It is upstream, in the literature, and downstream, in the analysis. The middle part, the bit with the gloves and the gel, is the romantic minority.

What Sam & Andrew bet in 2023

FutureHouse started with a deceptively boring claim: if large language models could read a paper, they could read a thousand. If they could summarize one, they could compare them. From that seed came a small parade of research projects with increasingly ambitious names. ChemCrow in 2022 wired tools into a chemistry agent. PaperQA in 2023 made citations a first-class output instead of a hallucinated afterthought. Aviary in 2024 turned research into a structured environment an agent could navigate. Robin in 2025 produced what was, by the lab's careful telling, the first validated AI-led scientific discoveries.

Robin made it into Nature. That is a sentence most AI companies would tattoo onto a wall. Edison's version is more sober: it is a precondition for the work, not the finish line. The bet was never that AI would replace scientists. It was that scientists, given a competent and tireless lab partner, would attempt the experiments they currently triage away.

A short, suspiciously linear timeline

Kosmos: the AI scientist, on tap

Kosmos is not a chatbot. It is a research agent that runs hundreds of subtasks in parallel - literature searches, data pulls, hypothesis tests, code execution, figure generation - then assembles them into something shaped like a paper. Each claim is back-linked to a source, which is either a deliberate jab at the hallucination crowd or just good manners. Probably both.

Users describe a single overnight Kosmos job as covering what their lab calls "six months of work." That is not a tagline the company invented. It is a number their early customers volunteered. Edison's pricing is correspondingly cheeky: the first six runs are free to academic researchers, then $200 each. Most graduate students spend more on coffee in a week.

What one Kosmos run compresses

Who is already using it

Edison's customer list is a tidy mix of the names you would expect and a few you would not. Incyte, the Wilmington-based oncology and inflammation pharma, runs Kosmos inside its R&D workflows. The Broad Institute and MIT use it in academic settings. The UK Dementia Research Institute applies it to neurodegeneration, where the literature is so vast that no human can hold it in working memory at once.

The seed round read like the same list, rearranged. Spark Capital and Triatomic co-led. Pillar VC and Susa Ventures, already in from FutureHouse days, doubled down. Striker, Hawktail, and Olive joined. Among the angels, Google's chief scientist Jeff Dean and CrowdStrike co-founder Dmitri Alperovitch wrote checks. It is, by any honest reading, a coalition of people who do not need to be convinced that AI can do real work.

What Edison is actually for

Read any AI-for-science deck and you will encounter the phrase "accelerate discovery" enough times to assume it is a legal requirement. Edison says it too. The difference is in what gets accelerated. The company is not promising a new molecule by Tuesday. It is promising that the slow, expensive parts of research - literature review, exploratory analysis, hypothesis generation - become something a team can do on demand, in parallel, without burning out a postdoc.

That is a less sexy claim than "AI cures cancer." It is also a more useful one. The history of scientific tools is mostly a history of compression. Microscopes compressed observation. PCR compressed amplification. Sequencing compressed reading DNA. Each one made experiments cheaper, which made more experiments possible, which made science move. Edison's pitch is that autonomous research agents are next on the list.

What changes if this works

If autonomous research agents become reliable - really reliable, the kind of reliable that does not require a human to babysit the output - the org chart of a biotech changes. Senior scientists spend less time editing literature reviews and more time designing experiments. Small academic labs get access to the same scale of analysis that, until last year, only large pharma could afford. Rare-disease research, perennially underfunded because the unit economics never closed, becomes attemptable. None of this is guaranteed. Some of it is already happening.

There is also a quieter consequence. Most scientific papers, once published, are read by almost no one. An agent that can read all of them flattens the literature in a way no human institution has managed. It also means that the next AlphaFold-shaped breakthrough may not come from a hero team in a marquee lab. It may come from someone running a Kosmos job at 11 p.m. on a Wednesday, asking the right question.

Back to the 2 a.m. queue

It is still 2 a.m. The SoMa office is still mostly dark. The queue is still chewing. The difference, a year from now, is that nobody on the other end of the queue is surprised by what comes out. The morning report is no longer a novelty. It is a meeting input. Edison Scientific's bet is not that it will invent science. It is that it will make the slow parts disappear, quietly, until nobody remembers that research used to wait. That is either the next big platform shift or it isn't. It is, at minimum, the most interesting thing happening on this block of San Francisco at two in the morning.

Tell someone

The receipts, linked

• edisonscientific.com - official site
• LinkedIn - company page
• X / Twitter - @EdisonSci
• platform.edisonscientific.com - run Kosmos
• YouTube - Kosmos x NVIDIA Nemotron demo
• Founders' announcement - "We Raised $70M to Accelerate Science"
• Bloomberg - the seed round
• SiliconANGLE - autonomous AI scientists
• Alzforum - on Kosmos and dementia research
• NVIDIA case study - scaling with Nemotron
• sam-rodriques.com - CEO Sam Rodriques