Marcus Vale: Ben, I've been stewing on something since Monday — you ever read a press release and feel like you're watching someone grade their own exam?
Ben Okonkwo: Hm, that's a very specific energy — what did you find?
Marcus Vale: Bristol Myers Squibb. July 20th. They just announced their second NVIDIA DGX SuperPOD — built on eight Vera Rubin NVL72 systems — and they called it, straight-faced, 'the most powerful and energy-efficient single-owned NVIDIA infrastructure in life sciences.' Their words, no third party, no benchmark, nothing.
Ben Okonkwo: Wait — no independent verification at all? Not even an internal audit they're citing?
Marcus Vale: Nothing. The proof is a hardware number — 10x performance per megawatt versus the predecessor system. Which is a real spec, I'm not dismissing it. But it's an energy efficiency metric for a data center, not evidence that a molecule moved faster through preclinical.
Ben Okonkwo: And that gap — between what NVIDIA can measure and what BMS actually needs to prove — that's the whole episode, isn't it.
Marcus Vale: Greg Meyers, their Chief Digital and Technology Officer, says BMS is 'beginning to see it pay off in our pipeline and operations.' Beginning. Three years after the first SuperPOD. Zero data attached. I mean — what does that sentence actually mean?
Ben Okonkwo: It means we're going to have to figure out what evidence would even settle this, because right now the claim is basically unfalsifiable.
Marcus Vale: But unfalsifiable is doing a lot of work there — because the hardware is genuinely real. Like, I don't want to wave that away. 72 Rubin GPUs per rack, 36 Vera CPUs, NVLink 6 interconnects tying the whole thing together. 3.6 exaFLOPS per rack. That's not nothing.
Ben Okonkwo: No, it's not nothing — and that's actually what makes this hard to call. The specs are real. The number is real. Here's what the number doesn't tell you.
Marcus Vale: Go.
Ben Okonkwo: Think about a hospital buying the fastest MRI scanner on the market. Genuinely fastest. Scans in half the time. And they announce it as a breakthrough in patient outcomes. The machine is impressive — but 'faster scan' and 'healthier patients' are two completely different claims. One is a hardware spec. The other is a clinical result. BMS is announcing the scanner.
Marcus Vale: And the patient outcomes are... nowhere.
Ben Okonkwo: Right — BMS deployed their first DGX SuperPOD around 2023. So we're now, what, three years into this. Three years. Not a single published benchmark on pipeline velocity. No earnings call footnote saying 'AI infrastructure cut our target-to-IND timeline by X months.' No peer-reviewed paper on target quality. I mean — in a regulated industry where FDA submissions require you to actually document your methodology, the absence of that data isn't caution. It's a choice someone made.
Marcus Vale: That's — wait, three full years and Greg Meyers gives us 'beginning to see it pay off.' That's the whole disclosure?
Ben Okonkwo: That's the whole disclosure. And look, I want to be fair — there are legitimate reasons proprietary drug discovery data stays internal. Competitive sensitivity is real. But you'd expect something. A conference abstract. An investor day slide with a leading indicator. Attrition rate, target quality score, anything. The silence isn't neutral.
Marcus Vale: So the bet question is: what would actually settle this? Because the hardware ships whether the outcome proof exists or not.
Ben Okonkwo: Actually — here's where I'll give the announcement real credit, because there is a kernel that's genuinely new. BMS is explicitly saying they're moving compute access away from a small group of computational specialists and out to scientists across all their global research sites. That's not a press release line. That's an organizational restructuring decision.
Marcus Vale: Wait — across all global sites? Not just Princeton?
Ben Okonkwo: All of them. And think about what that actually means on the ground. A medicinal chemist at the New Jersey site — previously she submits a docking job to a queue, waits, gets results back sometime Tuesday afternoon, maybe Wednesday. Now she's running molecular screening models herself, directly, no specialist intermediary. That's a real workflow change.
Marcus Vale: I actually — no, I think that's the right concession. That's not nothing. Removing the computational specialist bottleneck, that's real throughput.
Ben Okonkwo: It is. And then BMS layers on top of that: agentic research workflows. AI systems that autonomously execute hypothesis generation, molecule screening, multi-step tasks. Not a single query — an agent running a sequence. That's the use case they're building toward with this infrastructure.
Marcus Vale: Which is exactly when the governance question becomes — I mean, that's where I go cold on this. Fast.
Ben Okonkwo: Because that chemist running an autonomous screening agent at her terminal — which model version is she on? Is that run logged? If a molecule from that session eventually enters a regulatory submission, the FDA is going to want an audit trail of every computational step. The announcement says nothing. Zero mention of reproducibility infrastructure, audit logging, version control — nothing.
Marcus Vale: They announced the factory floor. Not the quality control system.
Ben Okonkwo: Exactly — and frankly, the hardware commoditizes in eighteen months anyway. What we haven't seen from BMS at all is whether they've solved data integration and organizational process, which is what would actually determine if any of this moves a molecule. That's the part I want to get to.
Marcus Vale: And that's the actual verdict, isn't it. The hardware commoditizes — NVIDIA ships a 10x gain on Vera Rubin NVL72 today, someone ships another 10x in twenty-four months. That clock is already running. So what's left? What does BMS actually own after the spec advantage evaporates?
Ben Okonkwo: Proprietary biological data. That's it. That's the only thing that doesn't commoditize on an eighteen-month GPU cycle. If BMS has genuinely integrated three years of internal screening data, target validation data, clinical attrition data into models trained on this infrastructure — that's a moat. A real one. But the announcement tells us... actually nothing about whether that's happened.
Marcus Vale: Nothing. Zero.
Ben Okonkwo: And here's what makes me genuinely uncertain — Lilly has been running their own AI-accelerated discovery platform for two years now. No supercomputer fanfare, no 'most powerful' press release. And their pipeline velocity looks roughly the same as BMS's. Now, I can't prove causation from that, but it's at least worth asking: is the DGX SuperPOD doing the work, or is it — I mean, is it the data integration and the scientists that BMS isn't announcing?
Marcus Vale: Wait — Lilly's moving at the same pipeline velocity without the hardware arms race?
Ben Okonkwo: Comparable velocity, publicly available pipeline data. Which doesn't indict BMS's strategy — but it does complicate the 'compute is the differentiator' claim pretty hard.
Marcus Vale: Okay, I'll take the other half of this though — the on-premise ownership. BMS chose to own this infrastructure, not rent it. That's a deliberate data-control posture. And NVIDIA's framing it as an 'AI factory' — their term — because they want BMS locked into training proprietary models on their stack. Once those models exist, once those workflows are certified, switching costs are enormous. That's NVIDIA's real win here, full stop, regardless of what BMS discovers.
Ben Okonkwo: Agreed — and that's actually the most defensible claim either of us can make today. NVIDIA is using BMS as the life sciences reference customer, the beachhead. BMS gets the hardware. NVIDIA gets the lock-in. What we genuinely don't know is whether BMS has the biological data quality and organizational process behind it to pull ahead of Roche or Lilly — and until Greg Meyers gives us a number that isn't a megawatt figure, we're just guessing.
Marcus Vale: Fine. I'll half-concede this. Maybe the hardware is everything Greg Meyers says it is. Maybe eight NVL72 racks running 3.6 exaFLOPS per rack actually does what BMS needs. And in three years, he'll have the pipeline data to prove it. That's the bet. I just want to be honest that it's a bet, not a result.
Ben Okonkwo: Yeah. That's — I mean, that's where I land too, actually. And look, 'beginning to see it pay off' is doing enormous work as a disclosure. It's not a lie. It's just not evidence. The announcement that will actually matter — for BMS, for Lilly, for Roche — isn't the next supercomputer. It's the first time any of them publishes a molecule that provably wouldn't exist without AI-assisted discovery. A real therapeutic advance, novel modality, reduced clinical attrition rate with AI on the causal path. Until that paper exists, every 'most powerful AI factory' announcement is a press release about a bet.
Marcus Vale: Exaflops and press releases. That's the genre we're in. Thanks for keeping me honest on the mechanism side — I kept wanting to call the winner.
Ben Okonkwo: And you kept me from hiding behind 'we can't know yet' forever. Good trade.
