A New Drug Burns $2 Billion and 10 Years. Benchling's CEO Wants AI to Cut That to Three

Sarah Guo introduces her No Priors guest as Saji: Sajith Wickramasekara, cofounder and CEO of Benchling, the system of record for biotech R&D. He is a software engineer who walked into a biology lab, found paper notebooks and desktop spreadsheets, and decided—naively, he now says—that building good tools for scientists could not be that hard. Thirteen years later the product helps scientists design molecules, plan and run experiments, organize and analyze the data, and share it with colleagues.

The company now works with about 1,300 biotech and pharma companies and scientists at more than 7,000 academic institutions. Household names on the roster include Moderna, Sanofi, Eli Lilly, and Regeneron, plus AI-native biotechs such as Isomorphic Labs and Xaira. From that seat Wickramasekara can see the whole sector's cycle, and where AI might actually bite. The data itself is not just gene sequences and yes/no assays. Making a medicine, he says, has 9,999 steps after you come up with a molecule.

You find a biologically meaningful target, design and optimize a molecule, test it in dishes, cell lines, and animals, then take it into successively larger human trials—all while inventing a manufacturing process that is economic, safe, and high-quality, and navigating regulators. Seven to ten years later you might have a commercial drug, and even then the work is not over. Benchling's job is to pull the heterogeneous lab data—molecules, the work that created them, tests, animal readouts, fermenter scale-up—into one searchable place so scientists can decide.

Guo asks him to characterize the biotech macro. He is not a macro specialist, but he is talking to all of his customers. The last couple of years, he says, are the equivalent of the dot-com bust for biotech. COVID was the peak: mRNA reopened the world, generalist money poured in, and exuberance followed. The bust was not only that hangover. Interest rates, tariffs, regulatory uncertainty, China, and scientific technologies that are still important but have not become commercially successful as fast as people wanted all stacked up.

The technologies that got people hyped were new modalities: gene editing, cell and gene therapies, RNA. These are real categories, with approved medicines already in patients—reprogrammed immune systems, gene therapies, mRNA. Investors and companies put a lot of money in, and some of those categories are now in the trough of disillusionment, having taken longer and cost more than expected. In 2021 every biotech was told to build a platform that would cure a bunch of diseases, with hundreds of millions of dollars and free capital. Companies taking a technology risk picked smaller, simpler proof-of-concept diseases. Then investors changed their minds faster than companies could: platforms were no longer valuable, and the proof of concept suddenly defined them.

China is the piece he weights most. If the last decade was biologics and new modalities, the next is speed and cost. People want more drugs, cheaper, and China is very good at things related to speed and cost. Chinese biotechs have been creating molecules and taking them into early clinical development in China, fast and cheap, even in the new modalities. The biggest companies in the world—Merck, Pfizer, Lilly—have gone to China to buy molecules they historically would have bought from American biotechs.

The in-market example people might recognize is Carvykti, a Johnson & Johnson cancer immunotherapy, developed with China's Legend Biotech, that reprograms the immune system against multiple myeloma. When Legend presented the data, a lot of people assumed it would not replicate; J&J saw the promise. The medicine is now commercially successful and widely used by those cancer patients in the US. Western reaction is a mixed bag: some feel inspired that American biotech has to get faster and cheaper, some reach for nationalist skepticism about regulatory and ethical standards and whether the data will hold. Wickramasekara's read is blunt. China as a major biotech player is here to stay.

Wickramasekara wishes people would appreciate that medicines are magic. Prescription drugs are 9 percent of US healthcare spending, and they are the part of the system with the best ROI, because they go generic. A statin he takes today costs almost nothing; twenty years ago it was expensive. The rest of healthcare is labor-dependent, and labor gets more expensive. We should want more drugs, and we should get to stockpile them. Instead it takes more than $2 billion and about 10 years to bring a medicine to market, and most of those medicines fail very late. You get seven to ten years in, you have spent hundreds of millions, and the trial fails because the medicine is not safe or not effective. At this point, he says, it is probably easier to send things to space, or put people on the moon, than to get a new medicine approved.

A lot of that cost comes from how artisanal the industry is. The physical realm has been systematized: Illumina put sequencers on every bench, and reagents and consumables followed. The digital realm—how people collaborate, how data is structured and shared, the workflows for collecting it—is bespoke, invented one-off by every company. Those companies are playing a one-time game: survive until year six or seven, show clinical success, and get bought. They are not building for scale or durability. Guo adds the structural point: if you were founding the next great pharma with a whole portfolio, you would invest in systems. After seven or eight years of good data, most people sell rather than roll the dice for another decade.

People get very focused on clinical trials because they are the biggest line item. Wickramasekara thinks that is a bit of a red herring. There are operational problems—badly designed studies, hard recruitment, a huge sticker price—but a lot of molecules are just not good. The industry needs better molecules and needs to move them to people faster. He keeps quoting Dario Amodei from *Machines of Loving Grace*: the returns to intelligence in scientific progress are very high. GLP-1s and Keytruda are the footnotes. The core GLP-1 science had been sitting on the shelf since the 1990s; obesity was an unfundable category five years ago because the trials are huge, long, expensive, and everything had failed. Insight and conviction turned it into what will probably be the best-selling drug of all time. Keytruda bounced through acquisitions, sat at the bottom of an out-license list, then a competitive threat appeared and Merck had the courage to go all-in. It is now the world's top-selling drug.

The people in the system are rational. The problem is that we know so little about biology, our ability to predict what will happen in the clinic is so poor, and the cost to get there is so high. Failure in software is a year or two and a couple of million dollars. Failure in biotech is four, five, six years, a big team, and hundreds of millions. The question, he says, is how you compress that so you get feedback faster.

Benchling AI has two major pieces. The first is simulation: taking open-source, proprietary, and companies' internal models and putting the right one at the right moment in the scientific workflow, already set up, so a wet-lab scientist without computational skills can use it, with results linked to everything else in Benchling. That ladders up to recommending the next best experiment from a scientist's own history plus the public literature—his way into the co-scientist problem. The second is agents. A deep-research agent, in the spirit of Anthropic's, runs over Benchling data with the Benchling data model. Questions that used to take weeks or months come back in a couple of hours.

The case he finds most persuasive: a customer getting ready to run mouse studies was evaluating 20 mouse models. Deep research looked at all the historical mouse work they had ever run and found that several of the models had already been done. The conclusions were trapped in a lab notebook from many years ago, from a company that had been bought, with the people long gone. The experiment would have taken eight months and a huge cost. They skipped it. So much of science, he says, lives in folklore and institutional knowledge and is lost over time. The job is to unlock organizational memory and make scientific data reusable—provided it was structured into Benchling in the first place, a premise he spent a decade selling.

Guo asks the harder product question: how do you make agents transparent to specialists outside the model's home domain? Engineers already rubber-stamp a lot of generated code; a wet-lab scientist looking at computational analysis may not even know where to be suspicious. Wickramasekara had just come back from a month on the road in Boston, London, and other scientific capitals outside San Francisco. Most people are not using much AI in R&D yet. They want to, they are primed, but concerns about accuracy, IP, security, and legal are everywhere, and they get larger the farther you go from SF. His line: we have GPT, but there is no chat. Chat, metaphorically, was the interface that made software AI take off, and bio has not found its equivalent. In a vertical, 90 percent of the work is translation. The AI that wins is the one people actually use.

Projected out a few years, a fully autonomous design-make-test-analyze loop with robots delivering drugs is something he would love, and he is more optimistic on a longer time scale. For the next one to two years he is more bullish on augmentation. He frames it as Waymo versus Tesla: the Waymo path to autonomy needs a lot of money and patience, and companies that have those ingredients should take it. The rest of science can take steps, pick off low-hanging fruit one experiment at a time, and see whether 7 to 10 years can come down to 2 to 3, with fewer specialized roles and a cheaper path to market. Radiology is Guo's parallel: ML people have been predicting the death of radiologists for 10 years, she thinks 40, and the model that worked is the copilot, because someone has to be accountable—and get sued. He buys that for clinical practice and is more hopeful that experimental decisions can be automated.

Large pharma's temperature on AI, from his customer base, is real excitement with AI officers in place, plus a pragmatic, methodical posture. Most of the large pharmas he works with have Copilot and a lot of pilots; he has not seen their R&D organizations transformed yet. Where they have leaned in, and where they have an advantage, is experimental data generation at a scale no biotech startup can match. They can and should train their own models. He expects unique predictive models to come out of pharma computational groups, similar to what is happening in open source.

The model ecosystem has grown fast. Open source was not really a thing in biology; in the last two years it feels here to stay. Structural-prediction models such as Boltz are going to change biology. Eli Lilly's TuneLab takes internal models and makes them available to the broader scientific ecosystem on a give-to-get, federated basis so the company never sees the actual scientific data. Structure prediction and antibody developability look fairly tractable. What is still missing are models that predict what happens in patients, and almost nobody is working on manufacturing-process optimization—even though faster discovery leaves even less time to invent a high-yield, cheap process.

Naysayers complain that years into the premise there is still no AI-discovered drug out the other end of the pipeline. Wickramasekara thinks they have a type-a-disease-get-a-molecule worldview. His mental model is thousands of cumbersome steps, a game of making each one a little better, and years after the molecule that can still be compressed. A better metric is the share of experiments already touched by some predictive capability, simulation, or AI, a share he bets is getting higher every day. Pure model companies collecting $100 million upfronts from five pharma customers are unlikely to remain that, because model-building is commoditizing too fast. More likely paths: morph into a biopharma with a pipeline, or distribute models the way software is sold, fee-for-service inside a network like Benchling, which already hosts Chai, AlphaFold, and other open-source models. Data transactions should exist in a field whose currency is data, and almost do not, because nobody trusts anyone else's format or provenance. Tooling and normalization might even let people sell negative data into a shared pool.

The old Wall Street line is that the only way to create value in pharma is assets, not tools. Wickramasekara has been making tools for 13 years. Thermo Fisher and Danaher are sneaky-big companies that systematized the physical realm—instruments, reagents, services around them. Something that rhymes should exist on the digital side; the technology just was not there. When Benchling started in 2012, cloud was the norm everywhere else, and most of life science was still paper, on-prem, and spreadsheets. The first couple of years were evangelism: bring science online. The next ten were convincing people that structured data mattered. Paper and Excel are easier. AI makes the benefit immediately obvious, which he thinks will be a tailwind with old skeptics. The Anthropic partnership, he says, is mission alignment: Dario is a scientist, the lab has said science is the next frontier after code, and Benchling customers care about trust. Benchling and Claude already interoperate natively, so scientists can generate reports and ask questions from a familiar interface.

Thirteen years of company-building, Guo asks, in two-minute takes. The most recent hard decision was going all-in on AI. Cofounder Ashu gave up all of his direct reports. Wickramasekara called a bunch of friends. The power of being a cofounder, he says, is moral authority: I do not need any of my Legos anymore, anyone else can have them, and if I look stupid that is okay. Ashu wrote a post that he was quitting his job to do this other thing; customers called to console Wickramasekara. It was a metaphorical resignation into full-time AI. The decision was controversial inside a company living through the biotech bust. The team wanted to focus on the basics; customers were laying people off and shutting down; was AI a distraction? What convinced them was Ashu hands-on-keyboard over Christmas, playing with a model and building for himself, plus a sentence that goes back to the product thesis: if we do not do this for our customers, who will? Translating Bay Area breakthroughs into useful vertical applications in complex, regulated domains is exactly Benchling's seat.

His most trusted algorithm is still go talk to customers. Every time the company, or his own mood, has been at its lowest, it is because he has gotten too far from them. Product-market fit is a moving target; a bunch of biotechs that were going to take over the world suddenly were not. He still spends 30, 40, 50 percent of his time with customers, and he has to role-model that so the whole company does it. A former employee, Malay, had told Guo that Saji holds the macro and the customer detail at once. Wickramasekara jokes that the compliment is like being told you have a large context window.

In a vertical, underserved customers first insist they are snowflakes that general-purpose software cannot handle. Talk to enough of them and they almost all want the same thing. Five to ten customers, not overweighting one modality or one company size, are a pretty representative model for what the entire industry needs. Everything successful they have built came from going very deep with a couple of customers until those customers were super happy, then watching it take off. Do things that don't scale, the YC line, works 10x better in a vertical. The AI-tools team is on daily and weekly calls with five or ten customers. The day before the taping, a customer texted about an FDA meeting: a report that used to take a day ran in the deep-research tool in about five minutes, and it was perfect.

Getting scientists to work inside a software company has left him with battle scars. If he had only hired software people who knew biology, he would have exhausted the pool ten years ago. Software people and science people have to sit together and learn. The sharpest conflict is incentives. In academia labor is almost free—grad students—and the currency is publishing papers to get more funding. A company has to sell software. The most impactful move has been repetition: the better the business does, the more innovation they can bring customers, and if Benchling does not do it, who will? He thinks they are the only independent scaled player building this kind of R&D software. Guo mentions a founder friend who asks research scientists how they feel about capitalism, as a sorting function. Wickramasekara laughs and says he will try it.

What can the two worlds learn from each other? Wickramasekara's answer is not what Guo expected. Biotech and pharma should learn from how tech communicates and tells stories, including the go-direct wave. Most people could not name five scientists or five pharma CEOs; everyone is on a first-name basis with Sam or Jensen. There is a hero's journey in tech. Talking about patients and scientists would change public perception of how hard it is to make a medicine. Moderna and Pfizer helped the world reopen during COVID, and Zoom still gets more credit. Gilead turned HIV from an 1980s death sentence into something close to a cure, and most people have no idea. Faceless companies are easy to hate and easy to underappreciate. They need to tell the story and go direct.

What tech can learn from bio is rigor, validity, and accuracy. Move fast and break things works in some domains. Once you need credibility with regulators to put things in patients, that stuff matters, and biopharma has figured out how to be safe. The US is still the gold standard for delivering medicines safely. Guo had asked a top lab scientist how to use models in quality-sensitive fields and was told to wait, or to focus on less rigorous use cases. That struck her as unambitious given the returns to intelligence. Wickramasekara agrees it is easy in Silicon Valley to oversimplify other people's jobs, and he is still optimistic the tech companies will figure it out, because everyone is excited about AI.

Outside Benchling, what he is excited about is personal again. He had not written a line of code in eight or nine years—he stopped around the time React showed up—and agentic coding tools have given him back the whimsy of building something quickly. His mother is a ChatGPT user, Google-search-plus-plus for now, and he no longer has to be IT support at Christmas. Guo's kids are using AI tutors; learning, she says, feels like play for the first time. The UX of natural language and voice is changing who gets to use the technology, including audiences that were never the lucrative 15-to-40 tech adopter. That, too, is a kind of scientific progress.