Not all of AI spending is “circular”, and Bristol Meyers is seeing productivity gains.
The drugmaker said it will be the first life sciences company to buy an Nvidia DGX SuperPOD based on its Vera Rubin systems. The chipmaker unveiled its Vera Rubin architecture earlier this year as the successor to its current generation of AI computing systems…
Robert Plenge, chief research officer at Bristol Myers, said the new capabilities would allow the company to cycle through many more potential drug candidates early in the drug development cycle. “Maybe before we could do 10 and now we can do dozens,” he said…Plenge also said that the company is already using AI tools to cut the time to make medicines to test in trials by 20% to 30%. That could even reach 50% in coming years, he said…
Greg Meyers, the company’s chief digital and technology officer, said the investment was driven in part by rapidly growing computing demands as Bristol deploys larger AI models across its research organization. It uses AI in all of its small-molecule and most of its large-molecule programs. He also said the new system will be more energy efficient.
@DrBob2 Bristol Myers is using a dedicated AI system which is specifically trained to do drug discovery based on chemical interactions. Dedicated AI applications are vastly less expensive to run and maintain than generalized frontier Generative AI models.
Dedicated AI and Generative AI are qualitatively different as well as quantitatively different. The cost of building, training and running a dedicated AI is orders of magnitude lower than Generative AI.
If you want, I can copy Gemini’s details about this.
A most interesting perspective I had not thought about!
Is Tesla FSD AI less expensive than Generative AI?
GoogleAI:
Yes, Tesla Full Self-Driving (FSD) AI is generally less expensive than frontier Generative AI at the enterprise infrastructure and training level, but it depends heavily on whether you are measuring consumer costs or corporate development budgets. [1, 2, 3, 4]
Comparing individual consumer costs
For an everyday end-user, accessing Tesla FSD requires a higher baseline financial commitment compared to standard Generative AI software, though both operate on a recurring monthly fee structure. [1, 2]
Tesla FSD Cost: A flat subscription of $99 per month. However, this carries a massive hidden premium: you must first buy or lease a Tesla vehicle to use it. [1, 2, 3, 4]
Generative AI Cost: Premium consumer tiers for foundational chatbots like ChatGPT Plus or Claude Pro typically cost $20 per month. They run on any existing smartphone or laptop, eliminating specialized hardware costs. [1, 2, 3]
Comparing enterprise development costs
At the corporate level, training and running a real-world autonomous driving system vs. a massive, multi-modal frontier Generative AI model reveal distinct architectural and financial differences.
Model training expenses
Tesla FSD: Tesla’s training pipeline is narrowly targeted toward processing spatial real-world video and telemetry. While Tesla operates massive computing clusters powered by tens of thousands of Nvidia H100 chips and its custom Dojo supercomputer, its core models are highly optimized to process a single domain (vision and driving physics). Training individual version updates costs tens of millions of dollars annually. [1, 2, 3, 4, 5]
Generative AI: Frontier foundational Generative AI models are exponentially more resource-intensive. Training a state-of-the-art Large Language Model (LLM) or multi-modal foundation model requires hundreds of billions of general text, code, image, and audio parameters. The initial training of a single frontier LLM version easily surpasses hundreds of millions of dollars, with projections for future iterations scaling past $1 billion. [1, 2, 3]
Computing and inference architecture
Tesla FSD: Tesla offloads the vast majority of its inference costs (the computing power required to execute the AI in real time) onto the user. Every Tesla features an on-board computer chip that processes the neural networks locally inside the vehicle using the car’s own battery power. Tesla pays almost $0 in cloud computing costs when a customer drives using FSD.
Generative AI: Generative AI providers bear massive, ongoing centralized inference costs. Every time a user generates text, an image, or a video, cloud data centers must spin up high-end server GPUs to process the request. Inference accounts for up to 90% of the lifetime cost of running a Generative AI platform, making scaling to millions of active users highly expensive for the provider. [1, 2, 3, 4, 5]
Drug design is a very specialiazed subject. Training AI for this purpose makes a lot of sense. I would be surprised if a general trained AI system could do this.
Its about fitting active groups into an enzyme site usually on a folded protein. You can optimize the distance and rigidity of the test molecule based on what is known to be active. You can also use electropositive or negative modifications to optimize performance. AI generates best guess improvements. Then you make them and test them. AI should do this faster and more precisely and has potential to learn as it gets more data.
