The Compression Company develops neural compression codecs that reduce the size of Earth observation satellite data before transmission, enabling operators to downlink more imagery without changing their hardware. Pulse 2.0 interviewed The Compression Company Co-Founder and CTO Joe Griffith to learn more.
Joe Griffith’s Background
When asked about his background and the experiences that led him to establish The Compression Company, Griffith shared:
I studied computer science at university, but I learned most of what I know from YouTube lectures and coding in my bedroom. I got into programming because I wanted to make video games, then got completely derailed by a YouTube video of someone using evolutionary algorithms to play Mario.
That sent me down a rabbit hole into neural networks that I never came back from. I completed a master’s degree and then started a Ph.D. in neuroscience-inspired AI at Imperial College London, where I met my co-founder, Michael Stanway. I dropped out when we found a problem worth building a company around, which became The Compression Company.
How The Compression Company Started
When discussing how the idea for The Compression Company came together, Griffith explained:
Michael and I were both conducting research at Imperial College London. We joined Entrepreneur First, an accelerator that helps technical people find co-founders and build companies. Its program pushes you to talk to as many industries as possible and find a problem worth solving before you write a line of code.
My Ph.D. research was in how neural networks learn to represent data efficiently, which is essentially what compression is: finding a smaller, smarter way to encode information. So, as we spoke to people across different industries, we kept asking the same question: Where is data volume a painful bottleneck?
The answer that kept coming back loudest was satellites. Earth observation satellites today have incredible sensors, but the radio link back to Earth is narrow, and you can’t physically widen it from orbit. Operators are forced to throw away data, downsample it, or wait for more ground station passes. It’s a physics constraint, not an engineering oversight.
We realized that neural compression, using neural networks to represent satellite imagery far more efficiently than traditional codecs, could change that equation entirely. Operators could get the same data down using a fraction of the bandwidth or get far more data within the same window. That was the founding insight.
Launching Code Into Space
When asked about his favorite memory from working at The Compression Company, Griffith recalled:
December 2025. Tilebox, a company that builds satellite tasking infrastructure, told us it had an upcoming orbital test and asked whether we wanted our compression codecs included. It would be the first time our code ran in space, so of course we said yes.
The catch was that we had two weeks to prepare codecs we had expected to spend three months developing. The deadline was December 27. I spent Christmas debugging CUDA kernels.
We made it, and the code launched on a SpaceX Transporter mission in early 2026. Afterward, I completely collapsed, but our code was on its way to orbit.
Neural Compression Codecs
When describing The Compression Company’s core products and features, Griffith detailed:
We build neural compression codecs for Earth observation satellite data.
When a satellite captures an image, it needs to transmit that image back to a ground station by radio. Modern sensors capture far more data than the link can handle, so operators constantly trade off image quality, coverage, and delivery speed. Our software makes satellite images 10 to 50 times smaller with configurable quality bounds while preserving what analysts and automated pipelines need downstream.
The codec runs onboard the satellite, compressing data before transmission. Operators can downlink more data per pass, increase revisit rates, or add new data products, all without hardware changes.
The technical advantage is that our codecs learn the statistical structure of specific sensor types, so a codec trained on multispectral imagery understands that data far better than any general-purpose standard.
We also configure where quality matters: lossless compression for high-value targets and aggressive compression for clouds or open ocean, all tunable for each mission.
Building Trust In The Space Industry
When asked about challenges in the space sector and how The Compression Company has addressed them, Griffith noted:
The space industry is conservative for good reason. When your hardware is hundreds of kilometers up, you can’t fix bugs in person. Operators need deep trust in any new software before it touches flight systems, which is a high bar for a startup asking them to run novel AI onboard.
We’ve addressed this by letting the results do the talking. Our orbital deployment with Tilebox and DPhi Space gives operators real, in-space validation.
We’ve also built benchmarking tools where operators can see exactly how our codecs perform on their specific sensor data, with transparent comparisons against JPEG2000 and other standards, before committing to anything.
Don’t ask anyone to trust us. Just show them the numbers.
Technology Evolution
When discussing how The Compression Company’s technology has evolved since launching, Griffith described:
Early on, we were working with standard neural compression architectures on publicly available imagery. Since then, we’ve built codecs optimized for specific satellite sensor types, including hyperspectral instruments that capture data across dozens of wavelengths beyond visible light.
We’ve also built an automated experimentation pipeline that trains, evaluates, and iterates on codec designs with minimal manual intervention.
AI coding tools have been a big part of that acceleration. I use Claude Code daily, and workflows that used to take a week now take hours. That speed feeds directly into how quickly our compression technology improves.
Major Company Milestones
When asked about The Compression Company’s most significant milestones, Griffith said:
Raising our $3.4 million pre-seed round led by Long Journey, an early backer of SpaceX, gave us the runway and validation to execute.
Getting into Entrepreneur First was important early on because it gave us the structure to turn a shared research interest into a real company.
Our orbital deployment on a SpaceX Transporter mission in early 2026 will be our first time running neural compression in space on real satellite data, and it is the milestone I’m most proud of.
Market Opportunity
When discussing the total addressable market The Compression Company is pursuing, Griffith stated:
Every satellite operator generating high-volume imagery faces the same downlink bottleneck. Our addressable market is any operator where compression meaningfully affects economics, and that number grows with every new constellation and every improvement in sensor resolution.
Beyond satellites, neural compression applies anywhere large volumes of sensor data create bottlenecks, including autonomous vehicles, LiDAR, hyperspectral imaging, medical imaging, and fiber-optic sensing.
We’re starting with satellites because the physics make compression uniquely valuable. You cannot widen the radio link from orbit. But the technology generalizes.
The Compression Company’s Differentiators
When asked what differentiates The Compression Company from its competition, Griffith emphasized:
Traditional codecs like JPEG2000 and CCSDS were designed decades ago as general-purpose tools. They don’t understand the data they’re compressing.
Our neural codecs learn the statistical structure of specific sensor types, achieving 10 to 50 times compression compared with 2 to 3 times lossless compression for traditional approaches.
We’re also building this as a product, not a research project. There are plenty of academic papers on neural compression, but turning a prototype into something that runs on satellite hardware, integrates with ground station workflows, and handles the range of sensor data operators actually produce is a different challenge entirely.
Future Goals
When discussing The Compression Company’s future goals, Griffith shared:
In the near term, we plan to validate the technology through our orbital deployment and move into commercial pilots.
In the medium term, we aim to become the default compression layer for Earth observation.
Longer term, we plan to apply what we’ve learned to other domains where data volume is a bottleneck. The underlying technology, learned compression tuned to specific data types, extends well beyond space.
Using AI To Build AI
When invited to discuss additional topics related to The Compression Company, Griffith concluded:
AI is not just what we sell. It’s how we build.
We’re shipping at a pace that would have required a much larger team two years ago. AI coding tools have changed what small teams can accomplish.
If you’re a founder and you’re not deeply integrated with these tools, you’re leaving enormous leverage on the table.
