Jill Mesirov, PhD
Jill Mesirov, PhD
Featured
Deck LATEST

The future of cancer research runs on supercomputers

Sanford Burnham Prebys recruits one of the world’s leading computational biologists to accelerate AI-driven biomedical discovery through advanced computing

For decades, supercomputers have reshaped our understanding of the universe through feats like simulating stellar explosions, modeling global climate patterns, and engineering next-generation aircraft. Today, these immense computational capabilities are being directed inward to address one of science’s greatest challenges: deciphering the complex language of human biology.
 
This shift was underscored this week by the appointment of Dr. Jill P. Mesirov, a pioneer in computational biology and cancer genomics, as Distinguished Professor and Senior Vice President for Computational Science at Sanford Burnham Prebys. Her recruitment marks a strategic effort to integrate advanced computing, artificial intelligence, and data science into the core of the institute’s biomedical research. For the high-performance computing (HPC) community, this move is more than just a key hire; it signals the definitive convergence of supercomputing, AI, and modern medicine.

Biology has become an HPC problem

Biological research has undergone a remarkable transformation during the past twenty years. Sequencing a human genome once required years of effort and billions of dollars. Today, thousands of genomes can be sequenced in days. Single-cell sequencing technologies now generate millions of individual cellular measurements from a single experiment, while spatial transcriptomics and advanced imaging systems produce multidimensional datasets measured in petabytes.
 
Extracting meaningful biological insight from these data is no longer primarily an experimental challenge.
 
It is a computational one.
 
Modern cancer research depends upon algorithms capable of integrating genomic, transcriptomic, proteomic, metabolomic, and clinical datasets simultaneously. These analyses involve billions of variables and demand computational infrastructures that resemble those found in national supercomputing centers.
 
Mesirov has spent her career developing precisely these kinds of computational approaches, helping establish data science as a central pillar of biomedical research. Her work has contributed to widely used computational tools and analytical frameworks that enable researchers to interpret complex genomic information and identify molecular mechanisms underlying disease.

Beyond bioinformatics

The title “computational biology” scarcely captures the breadth of modern biomedical computing.
 
Today’s computational scientists build machine-learning models capable of identifying previously unknown cancer subtypes, predicting patient responses to therapy, reconstructing cellular signaling networks, and discovering molecular biomarkers hidden within enormous genomic datasets.
 
Each of these workflows requires sophisticated numerical methods executed across large-scale computing systems.
 
Genome-wide association studies routinely analyze millions of genetic variants across hundreds of thousands of individuals.
 
Single-cell RNA sequencing experiments may profile millions of cells simultaneously.
 
Deep-learning pathology systems process gigapixel microscope images using thousands of GPU cores.
 
Drug discovery platforms evaluate billions of molecular interactions through simulation and AI-guided optimization.
 
Collectively, these workloads have made biomedical research one of the fastest-growing consumers of high-performance computing resources worldwide.

Supercomputers as biomedical instruments

Traditional scientific instruments observe nature.
 
Supercomputers increasingly function as instruments themselves.
 
Rather than collecting photons or particles, they construct mathematical representations of biological systems, allowing scientists to investigate processes that cannot be observed directly.
 
Large GPU clusters now train foundation models on genomic sequences, protein structures, electronic health records, and multimodal imaging data. These models can identify relationships that would be impossible for human investigators to recognize manually.
 
Increasingly, biological discovery begins not in the laboratory but inside large-scale computational infrastructure.
 
This shift explains why research institutions are investing heavily in computational leadership alongside experimental expertise.
 
By recruiting Mesirov, Sanford Burnham Prebys is reinforcing the idea that future biomedical breakthroughs will emerge from close integration between laboratory science and advanced computing. The institute has emphasized expanding capabilities in data science and AI as part of its broader research strategy.

AI changes the scale of discovery

Artificial intelligence is rapidly changing every stage of biomedical research.
 
Deep neural networks now predict protein structures with remarkable accuracy.
 
Generative AI models assist researchers in designing new therapeutic molecules.
 
Machine learning accelerates image segmentation, genomic classification, biomarker discovery, and clinical decision support.
 
Yet AI itself depends upon extraordinary computational infrastructure.
 
Training state-of-the-art biomedical foundation models requires clusters containing thousands of GPUs connected through high-bandwidth interconnects, supported by distributed storage systems capable of delivering terabytes of data every second.
 
The resulting computational demands rival those of traditional scientific supercomputing applications.
 
As AI becomes embedded within biomedical research, institutions capable of combining biological expertise with leadership-class computing infrastructure will possess a growing competitive advantage.

The rise of computational medicine

Medicine is steadily becoming a predictive science.
 
Instead of reacting after disease develops, researchers increasingly seek to model disease progression before symptoms appear.
 
Digital representations of tumors can simulate therapeutic response.
 
Network models identify previously unknown disease pathways.
 
Multiomic analyses reveal subtle molecular signatures long before conventional diagnostics detect abnormalities.
 
These capabilities depend upon sophisticated computational pipelines integrating simulation, statistical inference, machine learning, uncertainty quantification, and large-scale data management.
 
Each represents a mature discipline within high-performance computing.
 
Rather than replacing laboratory experiments, supercomputers now guide them.
 
Scientists can prioritize promising therapeutic targets computationally before committing years of experimental effort.
 
This dramatically shortens the path from hypothesis to discovery.

Converging scientific disciplines

The significance of Mesirov’s appointment extends beyond cancer research.
 
It reflects a broader transformation occurring across scientific computing.
 
Historically, computational biology evolved separately from traditional HPC disciplines such as computational fluid dynamics, astrophysics, and climate modeling.
 
Today, those boundaries are dissolving.
 
Shared technologies, including GPU acceleration, distributed computing, cloud-native workflows, AI frameworks, high-performance storage, and advanced visualization, are becoming universal scientific tools.
 
The same accelerator architectures used to simulate galaxy formation now train genomic foundation models.
 
Parallel computing techniques originally developed for physics increasingly drive precision medicine.
 
The future of supercomputing is no longer defined by scientific discipline.
 
It is defined by computational capability.

Building the next generation of discovery

Perhaps the most inspiring aspect of this appointment is what it represents for the future of biomedical research.
 
Scientific progress has always depended upon better instruments.
 
Microscopes revealed cells.
 
DNA sequencers revealed genomes.
 
Today, supercomputers reveal patterns hidden within biological complexity.
 
Every additional GPU, every faster interconnect, every more efficient algorithm expands researchers’ ability to understand disease at unprecedented resolution.
 
By bringing one of computational biology’s most influential leaders to Sanford Burnham Prebys, the institute is making a clear statement about where biomedical science is headed.
 
The laboratories of the future will still contain microscopes, sequencers, and imaging systems.
 
But they will also rely upon leadership-class computing clusters, artificial intelligence, and computational scientists capable of translating massive datasets into actionable biological knowledge.
 
For the supercomputing community, that evolution represents one of the most exciting frontiers in computational science.
 
The next life-saving medical breakthrough may not emerge solely from a laboratory bench.
 
It may first appear within the processors of a supercomputer, where mathematics, biology, and artificial intelligence converge to reveal discoveries that would otherwise remain invisible.
Featured

Meta’s next frontier may not be social media; it may be supercomputing

Tyler O'Neal, Staff Editor LATEST

Reported plans to commercialize AI infrastructure could transform Meta from one of the world’s largest consumers of supercomputing into one of its largest providers.

This transition signifies a fundamental shift in the global technological landscape, where the primary barrier to entry for AI innovation is no longer just talent or algorithms, but the sheer availability of high-performance hardware. As Meta pivots toward potentially offering commercial cloud services, it underscores the realization that compute power has surpassed traditional software assets in strategic importance.
 
Key Implications of Meta’s Potential Infrastructure Commercialization:
  • Commoditization of Compute: Access to high-density GPU clusters is evolving into a utility-like resource, placing it alongside electricity and bandwidth as a foundational requirement for enterprise growth.
  • Capital Efficiency: By monetizing currently idle capacity, Meta can offset the staggering costs of its multi-billion dollar data center investments, turning depreciating capital expenditures into robust revenue streams.
  • Expansion of the AI Ecosystem: Lowering the barrier to accessing world-class training environments democratizes the ability for smaller enterprises and research institutions to develop frontier-level AI models without the impossible cost of building their own physical infrastructure.
  • Structural Market Shift: This move threatens to disrupt the existing cloud hierarchy, challenging incumbents like AWS and Azure while also putting pressure on specialized AI cloud startups that previously occupied this niche.
  • Convergence with Scientific Computing: The blurring lines between large-scale AI training and traditional HPC workloads suggest that the future of scientific discovery, from medicine to climate science, will increasingly rely on the same infrastructure originally engineered for social media recommendation engines and LLMs.
The move marks a departure from the “walled garden” approach of previous tech eras. By inviting outside developers to run workloads on its proprietary systems, Meta is signaling that the competitive advantage in the next decade will belong to those who provide the foundational machinery upon which the rest of the industry is built. If this infrastructure becomes a public or semi-public utility, the company will have effectively positioned itself as the underlying engine of the broader artificial intelligence economy.

A different kind of supercomputer

Traditional supercomputers have historically been constructed for a single organization.
 
National laboratories build machines for scientific discovery.
 
Universities construct clusters for research.
 
Enterprises deploy HPC systems to solve engineering and manufacturing problems.
 
Meta’s infrastructure follows a different philosophy.
 
Rather than running tightly coupled scientific workloads using MPI-based parallelism, Meta’s AI clusters are optimized for enormous distributed training jobs involving trillions of model parameters. Tens of thousands of GPUs communicate simultaneously using ultra-high-bandwidth fabrics, while advanced storage systems stream petabytes of training data to accelerator nodes with minimal latency.
 
Although these systems differ architecturally from traditional capability-class supercomputers, they represent some of the largest computational installations ever assembled.
 
Their purpose is not climate modeling or astrophysics.
 
Their purpose is intelligence.
 
If Bloomberg’s reporting proves accurate, Meta may soon begin exposing that infrastructure to external users, allowing organizations to rent access to the same GPU clusters powering the company’s AI ambitions.

The economics of AI infrastructure

Modern AI data centers require investments measured not in millions but in tens of billions of dollars.
 
Each facility demands thousands of GPUs, advanced networking equipment, liquid cooling systems, substations, backup power generation, and increasingly dedicated energy sources capable of supporting hundreds of megawatts of continuous operation.
 
Meta has guided toward AI infrastructure spending as high as $145 billion in 2026, reflecting one of the largest capital investment programs in computing history. Industry-wide, major technology companies are expected to spend well over $700 billion on AI infrastructure this year.
 
Such investments fundamentally change the economics of computing.
 
Historically, cloud providers built infrastructure after customer demand materialized.
 
Today’s AI race reverses that model.
 
Companies are constructing enormous GPU capacity first, anticipating future demand for model training, inference, and agentic AI workloads.
 
The consequence is inevitable: At certain times, portions of these massive AI supercomputers will sit idle.
 
Commercializing unused capacity transforms what would otherwise be depreciating capital assets into revenue-generating infrastructure.

Compute becomes the product

The reported initiative reflects a broader industry trend.
 
Increasingly, the most valuable product is not necessarily the AI model itself.
 
It is the computational platform that can train and serve those models.

This distinction matters.

Training frontier AI models requires extraordinary computational density, often involving synchronized execution across thousands of accelerators connected through high-speed interconnects such as InfiniBand or custom Ethernet fabrics. These systems incorporate distributed storage, sophisticated scheduling software, fault-tolerant checkpointing, and optimized collective communication libraries that maximize GPU utilization.
 
Building such environments requires years of engineering experience.
 
For many enterprises, renting access to an existing AI supercomputer is significantly more practical than constructing one internally.
 
Should Meta commercialize its infrastructure, it would effectively be selling access not merely to GPUs, but to one of the world’s most sophisticated AI computing environments.

Challenging the AI cloud landscape

The reported strategy would position Meta alongside established hyperscale cloud providers while simultaneously challenging specialized AI infrastructure companies.
 
Unlike traditional cloud platforms that evolved from general-purpose virtual machines, AI infrastructure providers focus on delivering accelerator-rich environments optimized specifically for machine learning workloads.
 
This emerging “AI cloud” market emphasizes:
  • Massive GPU clusters
  • High-bandwidth networking
  • Distributed AI training
  • Inference optimization
  • Foundation model hosting
  • Large-scale storage architectures
  • Advanced orchestration software
Bloomberg reported that Meta is evaluating both raw compute rentals and hosted AI model services, similar to existing offerings that allow developers to access foundation models without managing underlying infrastructure.
 
That combination would allow customers to choose between renting hardware directly or consuming AI models as managed services.

From internal infrastructure to public utility

Perhaps the most remarkable aspect of the reported strategy is philosophical rather than technical.
 
For much of its history, Meta’s infrastructure existed solely to support Facebook, Instagram, WhatsApp, and the company’s internal AI research.
 
Opening those systems to outside developers would fundamentally change their role.
 
Instead of operating as private computational assets, they would become shared digital infrastructure supporting thousands of organizations.
 
This transition mirrors an earlier evolution in computing.
 
Amazon Web Services originated from infrastructure Amazon built for its own retail operations before becoming the world’s largest cloud platform.
 
Many observers now wonder whether AI infrastructure is entering a similar phase.
 
The difference is scale.
 
Modern AI clusters rival traditional leadership-class supercomputers in computational capability while serving entirely different workloads.

Implications for scientific computing

Although the reported initiative targets enterprise AI, its implications extend into scientific computing.
 
Many HPC applications increasingly incorporate machine learning alongside traditional numerical simulation.
 
Drug discovery combines molecular dynamics with foundation models.
 
Climate science augments numerical weather prediction using neural networks.
 
Materials science integrates density functional theory with AI-guided search.
 
Access to large GPU clusters is becoming essential across nearly every computational discipline.
 
If additional commercial AI infrastructure becomes available, research institutions may benefit from expanded computational capacity without bearing the enormous capital costs associated with constructing comparable systems.
 
The distinction between AI infrastructure and scientific supercomputing continues to blur.
 
Increasingly, they are converging into a single computational ecosystem.

Infrastructure becomes the competitive advantage

Perhaps the most important lesson is that the AI race is evolving.
 
The first phase centered on developing larger language models.
 
The second emphasized acquiring the world’s best AI researchers.
 
The emerging third phase focuses on infrastructure itself.
 
Owning vast computational resources is becoming a strategic advantage comparable to owning intellectual property.
 
The companies capable of deploying gigawatts of power, networking hundreds of thousands of accelerators, and operating hyperscale AI clusters may ultimately possess the strongest competitive position, not simply because they build better models, but because they own the machines on which future models will be trained.
 
If Meta ultimately launches a commercial compute business, it would underscore a profound shift in the economics of artificial intelligence.
 
The world’s largest social networking company would also become one of the world’s largest supercomputing providers.
 
For the HPC community, that possibility reinforces an increasingly clear reality.
 
The future of artificial intelligence will not be determined solely by algorithms.
 
It will be determined by who owns, builds, and operates the supercomputers capable of bringing those algorithms to life.
Featured

IBM’s sub-1 nanometer chip breakthrough: A genuine revolution, or another semiconductor science project?

Tyler O'Neal, Staff Editor LATEST
The semiconductor industry has long prioritized smaller transistors, higher chip density, and faster computing. On June 25, IBM unveiled what it describes as the world’s first sub-1-nanometer technology: a 0.7-nanometer (7-angstrom) transistor architecture dubbed "Nanostack." IBM claims this breakthrough offers up to 50 percent greater performance or 70 percent better energy efficiency than its 2021-era 2-nanometer technology, potentially cramming nearly 100 billion transistors onto a fingernail-sized chip. While the announcement has generated significant excitement regarding this long-awaited milestone in silicon scaling, a critical question remains: Does this represent the future of commercial computing, or is it merely an impressive laboratory demonstration destined to remain out of reach for mass production?

The end of traditional scaling

The significance of IBM’s announcement lies not in the transistor dimensions themselves but in how the company achieved them. For years, the semiconductor industry has relied on shrinking transistor features to improve performance and efficiency. However, as dimensions approach atomic scales, traditional methods become increasingly difficult. Leakage currents, quantum effects, manufacturing tolerances, and escalating fabrication costs have made each successive node dramatically harder to commercialize.
 
IBM’s answer is a three-dimensional architecture called Nanostack. Rather than continuing to flatten more transistors onto a two-dimensional surface, the company vertically stacks and staggers transistor structures, effectively extending Moore’s Law into the third dimension. IBM describes the technology as the industry’s first known three-dimensional nanosheet-based transistor design. From a research perspective, this is a meaningful accomplishment. From a manufacturing perspective, it raises entirely new challenges.

The supercomputing angle

If IBM’s projections prove accurate, the implications for high-performance computing could be substantial. Modern AI training systems consume enormous amounts of electricity and require massive accelerator clusters. Every percentage gain in energy efficiency translates directly into lower operating costs and higher computational throughput. IBM researchers estimate that AI accelerators built with 7-angstrom technology could achieve approximately six times the computational throughput of today’s leading accelerators, potentially reducing large language model training times from months to weeks.
 
For exaflops supercomputers and hyperscale AI data centers, those gains would be transformative. The challenge is that projected performance gains inside a laboratory environment rarely translate directly into production systems. Memory bottlenecks, packaging constraints, thermal limitations, interconnect overhead, and software inefficiencies frequently erode theoretical advantages. The history of computing is filled with architectures that looked revolutionary on paper but delivered far less dramatic gains in deployed systems.

IBM’s manufacturing problem

Perhaps the biggest reason for skepticism is that IBM no longer manufactures leading-edge semiconductors. The company sold its microelectronics manufacturing business more than a decade ago and now operates primarily as a semiconductor research organization. Its business model depends on licensing technology to foundries rather than producing chips itself. That distinction matters. Creating a laboratory prototype is difficult. Producing millions of chips with acceptable yields, manageable defect rates, and commercially viable costs is exponentially harder.
 
IBM projects commercial deployment could occur within five years. The semiconductor industry has heard similar timelines before. The company’s 2-nanometer technology, unveiled in 2021, demonstrated IBM’s research capabilities, but it required years of ecosystem development before broader industry adoption became realistic. The same pattern may repeat with Nanostack. The technology’s success ultimately depends on whether foundries such as Samsung, Rapidus, Intel, or others determine that manufacturing complexity and economics justify deployment.

The AI gold rush factor

There is another reason to view the announcement cautiously. The semiconductor industry is currently experiencing unprecedented demand driven by artificial intelligence. Every major chipmaker is under pressure to demonstrate a roadmap that extends beyond current process technologies. As a result, announcements increasingly emphasize future potential rather than near-term products. IBM’s presentation heavily highlights AI, cloud infrastructure, and next-generation computing as beneficiaries of the technology. While those applications are plausible, they remain projections rather than demonstrated commercial outcomes. Investors and technology buyers should remember that the chip unveiled this week is a research platform, not a product roadmap.

Why this still matters

Skepticism should not be confused with dismissal. IBM deserves credit for advancing semiconductor science at a moment when many experts have questioned how much further silicon scaling can realistically proceed. The Nanostack architecture addresses one of the industry’s most pressing challenges: how to continue increasing transistor density when conventional scaling approaches are nearing their physical limits. Whether or not the exact 0.7 nm implementation reaches production, the architectural concepts behind it could influence future generations of processors, AI accelerators, and supercomputing hardware. In that sense, the announcement may prove more important as a blueprint for future semiconductor design than as a specific manufacturing node.

The bottom line

IBM’s sub-1 nanometer chip technology represents a significant research achievement and offers an intriguing glimpse into the future of semiconductor design. The company’s Nanostack architecture demonstrates that innovation in transistor structures continues even as traditional scaling approaches their limits.
 
But history urges caution. The semiconductor graveyard is littered with breakthrough prototypes that never became commercially viable products. Until major foundries demonstrate manufacturable processes, competitive yields, and sustainable economics, IBM’s 0.7 nm technology remains exactly what it is today: A fascinating laboratory success. Whether it becomes the foundation of the next decade of supercomputing, or merely a milestone on the road toward some entirely different architecture, remains an open question.