Luis Ceze and the computer that doesn't look like one

In 2016, a team at the University of Washington and Microsoft put an OK Go music video into DNA. They included War and Peace and the Universal Declaration of Human Rights in more than 100 languages. The collection occupied less space than the tip of a pencil. Luis Ceze, a computer architect, was among the researchers behind it. A literature shelf, a diplomatic archive and a pop performance had acquired a rather unexpected landlord.

The total was 200 megabytes, a record for DNA data storage at the time. By everyday computing standards, that was a modest collection. By the standards of a molecule, it was a useful demonstration: digital information could leave its familiar electronic surroundings, survive translation into a different alphabet and come back intact. The experiment made a question tangible. What else might count as a computer?

Ceze has spent much of his career investigating the awkward space between an appealing idea and a working system. His projects have included approximate computing, molecular storage and software that helps machine learning models run on different chips. He also co-founded OctoAI, the Seattle company NVIDIA acquired in September 2024. His present title, VP of AI Systems Software at NVIDIA, sounds thoroughly corporate. The route there was considerably less tidy.

An engineer learns to leave something out

Ceze was born in São Paulo and earned electrical engineering degrees at the University of São Paulo. His doctorate took him to the University of Illinois Urbana-Champaign, where he completed a computer science PhD in 2007. That September, he joined the University of Washington faculty. His early professional experience included work at IBM Research on projects including Blue Gene, Cyclops and PERCS. These were questions about the machinery beneath the application, the parts users usually have the good fortune to ignore.

One research direction asked whether computers were spending energy on precision that an application did not need. A rendered image, for example, can tolerate certain small errors without becoming useless. The tricky part is identifying where that tolerance ends. Carelessness is cheap; a programming system that permits approximation while keeping it contained is an engineering problem.

EnerJ, a Java extension developed by Adrian Sampson, Ceze, Dan Grossman and colleagues, gave programmers a way to identify data that could tolerate approximation. Its type system separated that data from computations requiring precision. The project explored a deliberate trade between output quality and energy use. The researchers also published a correction to their energy calculations. Even a project about acceptable imprecision needed exact accounting.

That distinction helps explain the later work. Ceze's interest is in the relationship between what an application needs and what a machine actually does. Sometimes the opportunity lies in doing less work. Sometimes it lies in changing the material that holds the information. Either way, the specification deserves another look before anybody orders more hardware.

Four letters, several disciplines

DNA storage translates digital data into sequences of four bases, conventionally written A, C, T and G. Manufactured strands hold the encoded information; sequencing reads it back. These are synthetic molecules made for storage. The practical undertaking reaches beyond encoding: a useful archive also needs ways to preserve files, identify them, retrieve them and correct errors.

Ceze and Microsoft researcher Karin Strauss have co-directed the Molecular Information Systems Laboratory since 2015. Their work brought computer systems thinking to a medium with a very different set of operating instructions. Researchers demonstrated random access, retrieving selected files rather than decoding an entire collection. A library becomes considerably more attractive when borrowing one book does not require reading every other book first.

In 2019, the group demonstrated an automated journey from digital information into DNA and back again. The message was the word “hello.” Five letters made a small payload, but automation addressed a substantial obstacle: a storage service cannot depend on someone manually tending every intermediate laboratory step. The demonstration established that those steps could be connected under machine control.

The work remained an engineering research effort, with costs and speed still obstacles to commercial use. That is part of its interest. A successful experiment can establish an important capability without delivering a finished product. The 200-megabyte archive and the automated greeting answered different questions; each made the next question more specific.

The gap between a model and a machine

Meanwhile, machine learning was creating another translation problem. A trained model describes computation, but the hardware must execute it. Different processors have different instructions, memory arrangements and performance characteristics. Getting a model to run well can require substantial manual effort. A successful training run is therefore followed by another assignment, usually with less applause.

TVM, developed by a team including Tianqi Chen, Thierry Moreau, Ceze and other collaborators, addressed that gap. The 2018 research paper described a compiler that optimizes both the overall computation graph and the smaller operations within it. It could combine operations, adapt them to hardware and use a learned cost model to explore possible implementations. Machine learning was helping choose how to run machine learning.

The ambition extended across server GPUs, mobile GPUs, CPUs and new accelerators. Hardware diversity was the reason for the project: developers needed an approach that could travel. An optimization that works beautifully on one device has a rather limited social life if it cannot leave that device.

In conversations with customers, Ceze described a handoff problem. Moving from a data scientist's model to deployable software involved people, specialized tools and repeated decisions about where and how to run it. His commercial ambition was to make that process accessible through familiar software development infrastructure. The compiler could solve difficult technical work underneath; the customer should encounter a simpler task above it.

“We’ve always had a taste for unusual topics.”

Luis Ceze · 2021

When two hundred people turn up

At a TVM conference in late 2018, more than 200 people attended. Ceze later recalled that gathering as a signal that something was happening beyond the research group. Several key contributors were approaching the end of their doctoral work. Jason Knight, then at Intel and already a supporter of TVM, was also considering his next step. The ingredients for a company were becoming available at the same time.

OctoML formed in July 2019. Its five co-founders were Ceze, Tianqi Chen, Jared Roesch, Jason Knight and Thierry Moreau. The commercial proposition built on the open-source compiler while adding automation for customers. Naming the company also became a technical discussion: the founders wanted something short, pronounceable and memorable, with an octopus supplying a helpful collection of associations. Infrastructure companies are allowed a little marine life.

Ceze already knew the route from university research to a business. He had co-founded Corensic in 2008; F5 acquired the debugging startup in 2012. Madrona, an investor in Corensic, brought him aboard as a venture partner in 2018. By the time OctoML arrived, his Seattle connections ran through research groups, startups and the people who financed them.

The newer business evolved with its market. The OctoAI compute service launched in 2023, and the company adopted the OctoAI name by January 2024. Its work increasingly centered on generative AI applications and inference, the process of running a model to produce an answer. Customers wanted model choice and control over where their applications ran. They also wanted the bill to remain tolerable.

A different desk, a familiar problem

NVIDIA acquired OctoAI in September 2024. The commercial services were scheduled to wind down on October 31. Ceze joined the buyer, describing a path from university compiler research through Apache TVM adoption to products used by customers. He singled out the team and thanked investors and mentors, including Madrona's Matt McIlwain and Amplify Partners' Mike Dauber. The outcome involved a company, its people and its customer obligations, as well as its technology.

His subsequent work has kept the underlying concerns visible. In June 2025, he co-authored an NVIDIA technical article about FlashInfer, an open-source library for language model inference. The project began through collaboration among the University of Washington, Carnegie Mellon University and OctoAI. NVIDIA was releasing inference kernels into it for use in different serving engines. The paper received an Outstanding Paper Award at MLSys 2025.

The article identifies two needs: fast execution and the ability for developers to adopt improvements quickly. That second requirement matters when models and algorithms keep changing. A highly optimized component loses some of its usefulness if incorporating it takes too long. Performance belongs to the whole path from a new technique to a running application.

Ceze's October 2025 MICRO keynote carried the title “AI compilers and inference at scale: efficiency and velocity.” Its agenda included AI-generated code and implications for hardware and software design. In 2026, he appears among the organizers of the MLSys FlashInfer AI Kernel Generation Contest, which asks people and AI agents to produce efficient GPU kernels. The machinery doing the optimization is becoming part of the experiment.

The cook at the systems table

Alongside the companies are the people trained through the research. When his ACM Fellow recognition was announced in January 2023, the university counted 23 completed PhDs under his guidance and 29 patents based on his work. Those are dated figures, a snapshot of a career that includes teaching and mentorship as well as papers. The 2020 Maurice Wilkes Award, shared with Strauss for DNA storage and retrieval, likewise recognized work built through a wider team.

His interests away from the desk fit the same appetite for understanding how things come together. He enjoys cooking and eating, and describes learning about food from its cultural context down to its chemistry. “I love to cook it,” he said in a 2024 interview. A meal supplies its own form of peer review, usually with fewer footnotes and a firmer deadline.

There is a useful restraint in his account of research: promising bets can fail, and the failures disappear from the neat version of a career. That makes his public sequence of projects more interesting than a procession of prizes. A molecule offered a storage medium. A compiler offered a route between models and chips. A company offered a way to put the tools into other people's hands. Each required a different group of collaborators and a different test of whether the idea worked.

His current work concerns systems that make AI efficient and easier to deploy. The old molecular experiments concern a different medium and a different timescale. Set them beside each other and the recurring question becomes clear: how much useful computing are we leaving on the table because we have grown accustomed to the way the table is built? Ceze has made a career of pulling up a chair and examining the joints.