Endpoints published an article on Nvidia NVDA 0.00%↑ and their biotech pitch for mainly AI.
There are many points in the article that have been covered before in this Substack, but worth commenting some of the points made and how they relate to what we’ve discussed before.
“Nvidia is absolutely key to ensuring that AI is ultimately successful in healthcare,” Absci CEO Sean McClain, whose company uses Nvidia’s chips and software, said in an interview. “Even if we have our models, if Nvidia does not exist, AI is not going to make the impact in healthcare at the end of the day.”
I’ve written about AbSci ABSI 0.00%↑ before: a company dedicated to create therapeutic antibodies (the likes of the Regeneron therapy that Donald Trump received when he caught COVID19). AbSci is not a big company, and could still be classed as a startup in biotech, applying clever AI modelling to antibody discovery with the idea that they’ll create or partner in creating new antibody therapies for human health that will be worth a lot of money. Indeed, therapeutic antibodies is where the money is at, and there are many examples of multi-billion dollar blockbusters in areas like cancer antibody therapies and other indications.
AbSci is a “modern breed” of using in-silico tools to help in antibody discovery, and they have already published either preprints or peer-reviewed papers outlining their in-silico approach. One of their approaches is to start with an existing antibody therapeutic, and try and create another, different antibody, that is novel compared to the first, but maybe for the same market, and can be turned into a blockbuster. This approach, generating an “in-silico me too” blockbuster, can be very attractive: if it’s different enough from an existing blockbuster, the owner of the original can’t sue for patent infringement, but the new antibody can leverage all the clinical validation work from the first, although one would need to repeat it independently.
“The AlphaFold moment convinced us this was possible,” Kimberly Powell, Nvidia’s vice president of healthcare, told Endpoints News in an interview in San Jose.
The chipmaker had already spent years in biology’s weeds, in areas like molecular dynamics, genetic sequencing, and cryo-electron microscopy. By turning amino acid sequences into highly accurate 3D structures, AlphaFold led Powell to imagine what else could be possible.
“Once you can represent something in a computer, you can start building models,” Powell said. “Once you have models, you can start learning more about their interactions. This, to me, is a new dawn of computer-aided drug discovery, moving into many more areas than just the molecular dynamics simulations.”
Indeed, probably for everybody using AI in biotech, the news of what one could achieve with Alphafold2, when it came out, help in validate many related business models on the use of AI for biotech. The argument would be: “if Alphafold has demonstrated the value of AI in something as difficult as 3D structure prediction, this tell you how valuable AI will be in XYZ approach”.
For a subset of those business models in biotech startups, Alphafold2 was also the starting point for them to build their own in-silico drug discovery: take Alphafold2, modify certain parts of it, and build a custom in-silico platform that does something that the vanilla Alphafold2 wasn’t doing. In many cases, this meant generating some new data to train some new AI model.
Hardware remains the heart of Nvidia’s business. It’s currently building supercomputers for the Novo Nordisk Foundation in Denmark and for Amgen in Iceland — tapping into the latter’s massive genetic trove through its deCODE genetics subsidiary. But its software, tailored to the drug industry to put that compute into practice, is becoming a huge focus.
I didn’t know of any of these two ongoing efforts, but I am not surprised: Novo has so much money kicking about, they really are very well placed in spend it in cutting edge projects. On the other hand, deCODE has always been cutting edge. I had the privilege to visit them a few years ago, and get the tour of their facilities, and I was impressed in how well they’ve been able to follow the “incubator within a large pharma” model, being part of Amgen, but yet independent enough to be able to pursue cutting edge Science themselves, including AI.
In its latest offering, Nvidia introduced what it calls microservices last month. These are AI models ready to use out of the virtual box, with Nvidia charging $4,500 per GPU per year or $1 per GPU per hour to use these services. Nvidia says a pharma company can start using these models within minutes, requiring no AI expertise of its own. It lets drug companies do what they know best, while Nvidia provides its core strength — compute power and engineering know-how.
I remain a bit skeptical about the success of these: I am sure there will be many biotech entities using them for prototyping, to see how far they can take things, but in many cases, what biotech is looking for in Nvidia is the hardware, i.e. the GPUs and the CUDA-compatible stack of software that academia is so good at creating and maintaining.
Andrew Gostine, CEO of a hospital automation startup called Artisight, said his company has tried other chipmakers, even as Nvidia has invested in Artisight, but found the effort to optimize models and fit them into the startup’s ecosystem a “huge pain.”
“The amount of money I would have to spend to buy AI talent to do that optimization, Nvidia gives it to you for free — if you buy their super-expensive GPUs,” he said while presenting at last month’s Nvidia developer conference.
This is indeed one of the biggest moats that Nvidia has in biotech AI: they were there first, the first supercomputers where built by their A100s and H100s, and the software stack is well understood, well benchmarked, ready to go for many model types. Although this is their biggest moats, it comes at a cost: demand for their GPUs has been so high, that it’s difficult to build new supercomputers with Nvidia products, which prompts people in biotech, sometimes with tight budgets, to ask the question: “we know we could do this with Nvidia GPUs, but could we do it with AMD GPUs or some other provider, at a lower cost?”. At the end, it comes to the usual balance: do you have more time than money, or more money than time?
Nvidia’s venture capital bets
The article shows a thorough list of investments that the venture arm of Nvidia has made so far in biotech. For a 2.2 trillion dollar company, it’s not a lot of money, but it does help cement the perception that Nvidia takes biotech seriously.
This is all venture capital investment, but Nvidia also announced that they had invested in Recursion Pharma, which trades in the Nasdaq ( RXRX 0.00%↑), and is well positioned to capitalise both on the investment but also the special relationship they now have with Nvidia: RXRX can no doubt consider themselves a “GPU rich” biotech company now with this investment and partnership with Nvidia.
Overall, I don’t think this is the last time we will read about Nvidia in the context of biotech, and how big a play they have and can play further in the future. I still remember fondly spending time trying to iron out the kinks of using Nvidia GPUs on Linux machine, 15 years or so ago, so that writing CUDA code on my Linux computers would be as seamless as it was back then for MS Windows. Little did I know how serious would biotech take Nvidia GPUs in 2024 back then. But I was onto something…