Based out of Amsterdam, full-stack AI cloud builder Nebius built its reputation by assembling immense, GPU-dense datacenters optimized specifically for large-scale AI workloads.
With roughly 30,000 GPUs currently operational, mostly high-end NVIDIA accelerators like A100s and H100s, and planned expansions underway in Finland and Kansas City to add tens of thousands more, the company ranks among the most capable GPU cloud upstarts on the planet.
This kind of raw GPU horsepower makes Nebius uniquely suited to ambitious, computationally demanding applications like large-scale model training, multimodal data analysis, and complex generative AI simulations.
This, of course, makes it attractive for the healthcare and life sciences sectors, fields where data volumes are exploding, where workloads like transcriptomic mapping, molecular dynamics, and disease-state modeling demand sheer computational muscle and ultra-low-latency infrastructure.
It’s against this backdrop that Nebius recently launched its inaugural AI Discovery Awards, offering substantial GPU cloud resources, in this case, $100,000 worth of GPU credits each, to early-stage teams pushing hard technical boundaries.
While Nebius didn’t disclose how many GPUs each winning startup used, these credits translate into thousands of GPU-hours on the top-tier hardware essential for rigorous scientific computing and AI experimentation.
The four winning life sciences companies, Ataraxis AI, Aikium, Transcripta Bio, and MetaSight Diagnostics, are deploying these resources in noteworthy ways for cancer diagnostics, drug discovery, genomics, and population-scale disease screening.
Ataraxis AI emerged from the NYU orbit, the brainchild of researchers Jan Witowski and Krzysztof Geras, who had spent enough time poring over pathology slides to realize how much a deeply trained neural network might surpass human judgment.
Their Kestrel model training involved millions upon millions of digitized cancer biopsy images. The payback has been real, tangible, and validated across fifteen institutions. Instead of waiting weeks for genomic tests, oncologists can get predictions about tumor recurrence and treatment responsiveness in a single day.
The improved accuracy (roughly 30% better than existing methods) speaks to the sheer scale of Ataraxis's computational approach, but also underscores a certain understated ambition: to make cancer treatment less like guesswork and more like exact science.
Out in Berkeley, Aikium took a different tack altogether, blending hardcore synthetic biology with generative AI, though this wasn't just another protein-modeling play.
The company's Yotta-ML² platform actually synthesizes and screens trillions of protein variants using mRNA display technology.
CEO Eswar Iyer's team aims squarely at “undruggable” proteins, the molecular targets that pharma companies often write off because they're so complex. What's fascinating isn't just the volume of data they're handling, but the way they're bending biology and computation together into a loop, identifying protein binders in months rather than years.
Meanwhile, down the coast in Palo Alto, Transcripta Bio quietly built its Conductor AI platform, a massive atlas linking how genes respond to various small molecules.
The goal is to use transcriptomic data at a scale few have attempted to map drugs' molecular footprints. The payoff has been rapid repositioning of approved medications for new uses, demonstrating the real-world utility of their system..
Across the pond, MetaSight Diagnostics took the prize for HealthTech by aggregating an enormous dataset of blood molecular profiles from half a million Israeli participants.
By matching detailed multiomic blood analyses to electronic health records, they've carved out precise, molecular-level signatures for diseases like colorectal cancer and liver fibrosis.
Competitions like the ones Nebius put together show the world that GPU cloud providers are able to handle complex, large-scale workloads.
And certainly, it’s not as if on-prem clusters or big public clouds like AWS or Azure are incapable of powering workloads like these, they very much can and do. But startups often find themselves caught between two challenging realities.
Building in-house GPU infrastructure is a massive upfront expense, a sinkhole of cash and engineering effort exactly at the moment when speed and scientific focus matter most.
Public cloud providers, meanwhile, promise near-infinite scale, but often do so at a premium, alongside complexity and general infrastructure less finely tuned to some specific scientific computing needs.
What Nebius and similar specialized AI clouds like CoreWeave, Lambda, and others offer, by contrast, is an intentional middle path. They deliver raw GPU muscle similar to what you'd find with a hyperscaler, but wrapped in a software stack, storage architecture, and orchestration framework purpose-built to streamline data-intensive AI workflows, the same that underpin modern computational biology, genomics, and drug discovery.
And speaking of data-intensive life sciences workloads, Nebius relies on VAST Data under the hood, to ensurethat massive AI workloads like genomics analyses or drug discovery pipelines don't stall out at the storage layer.
VAST’s disaggregated, flash-optimized architecture lets Nebius efficiently handle huge volumes of both structured and unstructured data, critical for workloads that blend genomics, pathology, or multiomic data at scale.
Unlike conventional storage setups, VAST’s platform allows Nebius’s GPUs to access a shared, scalable namespace without running into the performance cliffs or latency bottlenecks typical of large-scale, traditional file systems.
Put simply, when a startup needs to feed petabytes of raw biological data into thousands of GPUs simultaneously (and do it without tripping over tangled I/O paths) Nebius's choice to deploy VAST means those researchers can focus on the science rather than the plumbing.
It's the freedom to experiment, iterate, and scale without unnecessary complexity or financial overextension, allowing companies like Ataraxis, Aikium, Transcripta Bio, and MetaSight Diagnostics to remain laser-focused on the science itself, not on managing infrastructure.