A Bigger AI War Is Starting
Why Did the AI Giants Suddenly Become Obsessed with Life Sciences?
A few years ago, if you asked me where the biggest commercial battlefield for AI would be, I probably would have said search, advertising, office software, code, and customer service.
Those answers were reasonable. They still are. AI is already making money in those places. You ask it to write emails, clean up slides, generate code, summarize calls, or run a customer support bot, and companies understand the value immediately. Save time. Cut cost. Move faster. That looks like the natural path for AI commercialization.
But lately I keep coming back to a different thought.
The thing that really excites the AI giants may not be sitting on your laptop screen. It may be sitting under a microscope.
Look carefully at what has happened over the past two years. Almost every top AI company is pushing into life sciences, each from a different angle.
Google DeepMind built AlphaFold 3, putting proteins, DNA, RNA, small molecules, and antibodies into one prediction framework. Its sister company, Isomorphic Labs, is working with Novartis, Eli Lilly, and Johnson & Johnson to push AI models into real drug discovery.
OpenAI is working with Lilly on new antimicrobials and with Retro Biosciences on protein engineering for cell reprogramming. Microsoft is building BioEmu to model how proteins move. NVIDIA is turning biology into a market for GPUs, BioNeMo, NIM microservices, and AI infrastructure. Meta’s earlier ESM protein language model work helped seed EvolutionaryScale, whose ESM3 model is trying to make protein generation feel closer to image generation. Anthropic is plugging Claude into scientific workflows through Benchling, PubMed, 10x Genomics, BioRender, and related tools. Amazon Web Services is playing the cloud, model, and automation layer for companies like Bayer and Exscientia.
That signal deserves attention.
The AI giants did not all wake up one morning and develop a sentimental love for biology class. They see a larger opportunity. If AI can move from generating text to generating testable scientific hypotheses, the ceiling changes completely.
Writing an email may save a company a few minutes.
Designing a molecule that eventually enters the clinic can change the fate of an entire company.
That is the lure of life sciences.
After AlphaFold, Google Wants to Turn a Nobel Prize into Drugs
This story has to begin with DeepMind.
When AlphaFold 2 cracked the protein structure prediction problem, a lot of people realized something for the first time: AI could write essays, play Go, recognize images, and also solve scientific problems that had frustrated researchers for decades.
Then came AlphaFold 3, with a much larger ambition. When DeepMind and Isomorphic Labs announced the model in 2024, they said it could predict the structure and interactions of all life’s molecules, including proteins, DNA, RNA, ligands, antibodies, and more.
The phrase “predict structure” sounds abstract. The important part is more practical: drug discovery happens through interactions.
A drug works because it binds to a protein pocket, blocks a pathway, changes a signal, or alters a biological process. An antibody works because it recognizes an antigen. A DNA variant matters because it can change expression, splicing, or disease risk. For decades, scientists have attacked these questions with experiments, experience, luck, and a lot of expensive trial and error.
AlphaFold 3 tries to move part of that search into the computational world.
DeepMind’s own research blog says AlphaFold 3 improved predictions for interactions between proteins and other molecule types by at least 50% versus existing methods, with some categories doubling in accuracy. That does not mean drug-development success rates suddenly jump by 50%. Biology is not that kind of machine. But it explains why Isomorphic Labs has become such an important Alphabet asset.
The logic is simple.
DeepMind produces the scientific breakthrough. Isomorphic tries to turn the breakthrough into drug programs.
The company started a collaboration with Novartis in 2024, initially focused on three difficult small-molecule targets, then expanded the work in 2025. Its Lilly collaboration focuses on undisclosed small-molecule targets. Its Johnson & Johnson partnership goes broader, covering multiple targets and multiple modalities, including small molecules and biologics.
The message is direct: Google does not want AlphaFold to remain a citation machine or a scientific trophy. It wants to know whether AI can become the engine of a new kind of drug company.
That is why Alphabet’s line of attack is so interesting. Alphabet is doing something more ambitious than selling cloud compute to pharma or adding a few science plugins to a general chatbot. It has a chain that runs from frontier models to scientific teams, from scientific teams to a drug discovery company, and from that drug discovery company to large pharmaceutical partners.
If AI drug discovery eventually produces a group of major medicines, Alphabet may have one of the earliest full-stack templates.
OpenAI Wants to Prove Intelligence Can Invent
OpenAI’s entrance into life sciences feels more like OpenAI.
It has not yet built a full drug company around the effort. Instead, it is showing examples that support a much bigger claim: models can invent useful things.
In June 2024, Eli Lilly announced a collaboration with OpenAI to use generative AI to discover novel antimicrobials for drug-resistant pathogens. Antimicrobial resistance is one of those global risks that remains strangely underpriced in public attention. Traditional antibiotics can be commercially unattractive, while the science is difficult and the public-health need is enormous. Lilly’s partnership with OpenAI is basically asking: can large models help humans explore antimicrobial chemical space that has been neglected or poorly searched?
The more dramatic case is OpenAI’s work with Retro Biosciences.
Retro works on longevity science, including cell reprogramming. One of its areas of interest involves Yamanaka factors, a set of proteins that can push ordinary cells back toward a younger, stem-cell-like state. The problem is efficiency. Reprogramming takes time, and only a small share of cells complete the journey.
MIT Technology Review reported that OpenAI and Retro built a model called GPT-4b micro to redesign those factors. Preliminary results suggested that model-designed variants improved certain reprogramming markers by more than 50 times compared with the original factors.
We should be careful here. No drug has come out of this yet. No commercial product has come out of it either. It is more like a scientific demonstration.
But the demonstration matters because it hits the exact point OpenAI wants to make. AI can help scientists read papers, write code, and organize lab notes. It may also propose designs that human scientists did not think of, then produce effects that show up in the lab.
That matters to OpenAI far beyond longevity.
If a model can design proteins, it starts to approach the center of scientific discovery. If a model can propose an experimental hypothesis, watch the lab test it, absorb the result, and improve the next proposal, the system starts looking less like a chatbot and more like the early outline of a scientific engine.
From an investment lens, OpenAI’s life-sciences line has strong option value. It does not yet have the clean commercial drug-discovery path that Isomorphic is trying to build. But it has a larger story: general intelligence should eventually accelerate human invention.
That is why OpenAI has to be here. If the AGI story stays trapped in customer support, office work, and coding, enterprise-software valuations will define its imagination. Life sciences gives OpenAI a different story: superintelligent tools might change medicine, aging, disease, and human lifespan.
That story is too large for OpenAI to ignore.
Microsoft and NVIDIA Are Building the Scientific Engine Room
Compared with OpenAI and DeepMind, Microsoft’s life-sciences strategy looks less theatrical. It is also very serious.
Microsoft’s AI for Science team is building BioEmu around one key fact: proteins move.
A lot of people hear about AlphaFold and assume that once you know a protein’s structure, the drug discovery problem is mostly solved. Reality is messier. Proteins wiggle, fold, unfold, expose hidden pockets, shift domains, and occupy different states. A drug’s ability to bind often depends on those moving states.
BioEmu uses generative deep learning to model protein equilibrium ensembles. Microsoft’s research page says BioEmu can generate thousands of statistically independent protein structures per hour on a single GPU. Its training data integrates more than 200 milliseconds of molecular dynamics simulations, static structures, and experimental stability data. It can also predict free energies close to long-timescale molecular dynamics and experimental measurements.
That sounds technical. Underneath it is a business equation.
Traditional molecular dynamics simulation is expensive and slow. Experiments are even more expensive and even slower. If systems like BioEmu can amortize part of that cost into fast model generation, they can change the cost curve of scientific computation.
NVIDIA is looking at a different layer.
Once life sciences enters the model era, the field needs compute, software stacks, deployment tools, and inference services. That is NVIDIA’s home territory.
BioNeMo is NVIDIA’s toolkit for drug discovery and digital biology. It includes models for protein structure, generative chemistry, molecular docking, DNA sequence analysis, and single-cell RNA analysis. NVIDIA is also packaging these capabilities through NIM microservices so drug companies can integrate them more easily in the cloud or on premises.
This is the business NVIDIA understands better than anyone: turn a new computing paradigm into an infrastructure market.
If every large pharma company eventually trains its own molecular models, runs protein dynamics simulations, deploys AI lab platforms, and processes genomic, cellular, chemical, and imaging data at scale, GPU demand will not come only from chatbots. It will also come from animal models, cell images, genome sequences, protein spaces, and chemical universes.
NVIDIA does not need to know which AI-discovered drug wins. If the entire industry believes drug discovery needs more computation, NVIDIA gets to stand near the toll booth.
That is why life sciences may be an underappreciated long-term demand source for NVIDIA. Today, investors stare at data centers and cloud capital expenditure. Five years from now, biological computation could become another durable layer of AI demand.
The Seed Meta Left Behind Is Growing into Programmable Proteins
Meta looks quieter in this race, but its earlier ESM work matters.
ESMFold and the ESM Metagenomic Atlas predicted structures for hundreds of millions of metagenomic proteins, pulling unknown proteins from soil, oceans, microbes, and the human body into a structural map. The people and ideas behind that work later became central to EvolutionaryScale.
EvolutionaryScale’s ESM3 is an ambitious model. It works across protein sequence, structure, and function. The company says ESM3 has 98 billion parameters, was trained with more than 1 x 10^24 FLOPs, and generated a new green fluorescent protein. That protein had only 58% sequence similarity to the closest known fluorescent protein, which the company described as equivalent to simulating more than 500 million years of evolution.
That sounds like science fiction. But it captures the most fascinating part of protein engineering.
Nature has spent billions of years searching through protein space. Humans have mostly modified what nature already gave us. If a model can learn the language of proteins, it may be able to jump into new regions that nature never explored but physics still allows.
That is the imagination behind programmable biology.
Today we generate images with AI and barely think about it. You type a prompt, and the model searches image space for something that fits. One day, a scientist may type a functional constraint, and the model may search protein space for an enzyme, a binding protein, a delivery system, or a cell-therapy component.
Of course, the validation gap is huge. A human can judge an image in a second. A protein has to fold, function, avoid toxicity, and be manufacturable. The lab gets the final vote.
Still, the direction has changed.
In the past, we ordered from the menu nature had already written. Now AI companies want to help write new menus.
Anthropic and AWS Are Going After the Scientist’s Daily Workflow
Life-sciences AI does not have to begin with the design of a new molecule.
Some value will arrive in more ordinary places first: reading papers, drafting protocols, organizing data, searching internal records, generating compliance documents, planning clinical recruitment, and connecting lab notebooks with databases.
Anthropic’s Claude for Life Sciences is aimed at that layer.
It connects Claude to Benchling, PubMed, BioRender, Wiley Scholar Gateway, Synapse.org, 10x Genomics, and related platforms. In practical terms, scientists do not have to drag everything into a blank chat window. AI moves into the tools they already use.
That matters.
Scientific work contains a lot of reading, cleaning, formatting, protocol writing, database querying, first-pass analysis, and regulatory paperwork. The “eureka” moment is precious. The daily friction is relentless. If Claude can remove some of that friction, commercial value may arrive faster than outsiders expect.
Sanofi’s use of Claude points in the same direction. Pharma companies are not treating AI as a lab toy. They are putting it across the value chain, from R&D to internal knowledge work to commercialization.
AWS is doing something similar from the infrastructure side.
AWS worked with Bayer on a six-week project using generative AI to predict chemical reaction conditions. To outsiders, reaction conditions may sound unexciting. To chemists, they matter enormously. Making a molecule often means navigating solvents, catalysts, temperature, pressure, reagents, and sequence of steps. Better prediction can save real time.
AWS also supports Exscientia’s AI-powered drug discovery platform, which combines generative design with robotic lab automation. That loop, where the model proposes a design, robots run the experiment, and the data returns to the model, is one of the most important structures in future drug discovery.
Anthropic and AWS therefore represent a more pragmatic route. Put AI into the daily workflows of researchers and pharma teams. Start by saving time. Move gradually toward scientific discovery.
That path may be less glamorous than “AI designs a new drug,” but it may commercialize faster.
Why Now?
Saying “healthcare is a huge market” is too shallow.
Healthcare has always been huge. Pharma has always been profitable. So why are AI companies rushing in now?
Because several conditions matured at the same time.
Start with the deepest shift: life began to look like language.
Proteins are sequences. DNA is sequence. RNA is sequence. Chemical reactions can be tokenized. Lab protocols are text. Papers are text. Clinical records are full of text. Transformers are very good at learning patterns from large-scale sequences. They do not make biology simple, but they allow AI companies to see many biological systems as things with learnable grammar.
Then came data. Biology finally became large enough.
Sequencing, single-cell biology, spatial omics, cryo-EM, protein databases, high-throughput screening, and automated labs have made living systems more measurable. AlphaGenome can read one million DNA base pairs at a time and predict regulatory signals across tissues and cell types because the scientific world spent years measuring the underlying system.
Even more important, the experimental loop is forming.
In the past, AI models could make suggestions, but scientists still had to arrange experiments slowly. Now robotic labs, automated synthesis, and cloud data systems are shortening the prediction-experiment-feedback-prediction cycle. AI changes drug discovery through that shorter loop, rather than through one magical prediction.
Pharma also needs a new engine.
Large drug companies face patent cliffs. Older drugs lose revenue to generics. New drugs are more expensive to develop. Clinical failure is brutally costly. They need better targets, stronger candidates, smarter trial design, and fewer dead-end experiments. If AI improves even a small part of that process, the economic value can be enormous.
And AI companies need a bigger proving ground.
If AI only writes marketing copy and handles customer service, it is useful but limited. Life sciences gives AI companies a harder exam. Can you propose new hypotheses? Can you design new molecules? Can you improve experimental outcomes? Can you accelerate discovery?
That is why life sciences matters so much.
It is a place where AI can prove that it creates knowledge.
The Talent War Is Becoming Half Scientist, Half Engineer
This race eventually comes down to people.
The AI industry used to fight over machine learning researchers, GPU engineers, distributed-systems experts, and product engineers. Now it needs another kind of person: someone who understands proteins, chemistry, cell experiments, sequencing data, and also knows how to talk to a model team.
There are not many people like that.
A pure AI team may misunderstand experimental noise and mistake beautiful metrics for real biology. A traditional pharma team may understand biology but struggle to turn internal data into a continuously learning AI system. The scarce person is the bridge.
That is why Isomorphic Labs emphasizes multidisciplinary talent. Anthropic is also pulling more life-science expertise into its orbit. In 2026, John Jumper, the AlphaFold co-creator and Nobel laureate, left Google DeepMind for Anthropic. The symbolism is hard to miss.
A person who helped build AlphaFold chose another frontier AI company as his next stop, rather than a traditional pharma company.
That tells us something. Top scientific talent is starting to see AI companies as new scientific institutions. In the past, elite biologists wanted to go to universities, major pharma companies, Genentech, or the Broad Institute. In the future, some may decide the best experimental platform sits inside an AI company, where there is compute, model talent, engineering speed, and a sufficiently unreasonable goal.
For investors, talent flow matters.
Capital follows talent. Platforms form around talent. Partnerships follow talent. The companies that can put AI researchers, computational biologists, medicinal chemists, wet-lab scientists, and product builders into one fast operating rhythm have the best chance of building a true life-sciences AI flywheel.
For Investors, Start with Infrastructure, Then Look at Drug Optionality
If I put this into an investment framework, I would separate it into two layers.
The first layer is infrastructure.
NVIDIA is obviously there. If drug discovery becomes a large-scale modeling, simulation, inference, and data-processing problem, GPU demand grows. Microsoft, Google Cloud, and AWS are there too. Pharma companies need cloud systems to train models, manage omics data, run automated experimental platforms, and call biological foundation models. Lab data and R&D workflow platforms such as Benchling also become more important because AI needs clean access to experimental records to matter.
The advantage of this layer is timing. It does not need to wait for one AI-designed drug to succeed in the clinic. Once the industry adopts the tools, infrastructure gets paid first.
The second layer is platform and drug optionality.
Isomorphic Labs, EvolutionaryScale, Retro Biosciences, Exscientia, Recursion, Insilico, Iambic, and similar companies represent the higher-upside layer. If AI improves candidate quality, shortens discovery cycles, or opens new modalities, the returns can be enormous. The risk is also enormous, because clinical trials still decide the truth.
Biology does not flatter anyone for long.
A molecule can look beautiful in a model and fail in a cell. It can work in a mouse and fail in humans. It can show efficacy and then fail on toxicity. It can make mechanistic sense and still fail commercially. So I would not frame AI drug discovery as “software eating pharma.” The better framing is that AI gives pharma and techbio companies a new discovery lever.
Large pharma can be both customer and winner. Lilly, Novartis, J&J, Sanofi, Bayer, and others have clinical development, regulatory knowledge, manufacturing, and commercialization capacity. If they integrate AI into their R&D systems, AI may strengthen them rather than replace them.
The thing to watch is proprietary feedback loops.
Public data can train powerful foundation models. But long-term advantage in drug discovery often comes from private experimental data: failed compounds, real assay results, cell images, protein engineering logs, synthesis routes, clinical signals, and the tacit knowledge scientists build over years.
Whoever turns every experiment into training signal owns an asset that is hard to copy.
That is the life-sciences version of the AI flywheel.
The Real Bet Is Learning the Operating System of Life
I think the AI giants are after something much larger than another revenue vertical.
They are looking for AI’s next source of legitimacy.
Chatbots brought AI into everyday life. Coding assistants brought AI into software production. Life sciences may bring AI into humanity’s most difficult knowledge-production system: understanding disease, designing drugs, engineering proteins, interpreting genomes, and shortening experimental cycles.
In the short term, much of the value will look ordinary. Read papers faster. Draft protocols faster. Clean data faster. Generate regulatory documents faster. Predict reaction conditions faster. None of that sounds sexy, but pharma will pay for it.
In the medium term, AI should improve the search efficiency of drug discovery: better targets, better molecules, fewer wasted experiments, faster feedback loops.
Over the long term, the exciting possibilities arrive: programmable proteins, AI-designed molecules, genomic regulatory models, automated experimental loops, and semi-autonomous discovery systems.
The path will be uneven. AI will make mistakes. Models will hallucinate. Experiments will fail. Clinical trials will embarrass beautiful theories. Some projects will disappoint. But for investors with a five-year horizon or longer, the question to watch is whether biology is becoming a domain that computation can learn more effectively.
My answer is yes.
Life sciences is slowly moving from experience-driven discovery toward a system driven by data, models, and experimental feedback loops. This transition will not finish overnight, but the direction is clear.
So when AI giants bet on life sciences, they are chasing far more than a short-term theme.
They are fighting for a much larger doorway: whoever learns to read life may define the next generation of scientific productivity.
If the last decade taught AI human language, the next decade may teach it the language of life itself.
That is what makes this race so fascinating.
Disclaimer: This article is for research and educational purposes only. It does not recommend buying, selling, or holding any security, fund, drug asset, or private company. Life-sciences and AI drug-discovery investments carry high risk, including clinical failure, regulatory changes, valuation volatility, and slower-than-expected technical adoption. Make your own decisions based on your own risk tolerance.