John Barrus · VP of Product, Niobium
In January 2025, hackers dumped over 30 million smartphone location data points from Gravy Analytics onto a Russian cybercrime forum, exposing visits to reproductive health clinics, addiction treatment centers, and political rallies. But the hack wasn't what caught regulators' attention. Weeks earlier, the FTC had filed a complaint against Gravy Analytics, not because of any security failure, but because the normal operation of the business was itself a privacy violation. The company geofenced reproductive health clinics, birthing centers, and support groups, then sold audience segments like 'New Parents/Expecting' and 'Women's Health' to advertisers. The processing was the problem.
Privacy is power, and when computation requires plaintext, the power shifts to whoever runs the pipeline. That's the structural issue across industries, from ad targeting to healthcare claims to AI-powered analytics: someone in the middle always sees the data in the clear.
Fully homomorphic encryption changes that equation, allowing computation on data that never leaves its encrypted state. The service provider learns nothing. The power stays with the individual. The gap between that promise and real-world deployment is speed, and I came to Niobium to close that gap.
Through a combination of serendipitous timing and luck, Niobium is my third opportunity to take a new chip to market. The first chip was the Cloud TPU at Google. The second was Groq’s LPU. In each of these three cases, a fundamental compute bottleneck existed that the available hardware couldn’t solve, and a team had come together with the right architecture and the focus to do something about it.
At Niobium, that bottleneck is privacy.
A Recurring Theme
When the Cloud TPU was taking shape, the gap was obvious: machine learning workloads were growing fast, but the performance Google needed for its specific workloads wasn't available at the right cost or power envelope. Purpose-built silicon closed that gap, delivering matrix multiply performance that could scale to hundreds of chips at a fraction of the cost per operation.
At Groq, the thesis was about inference. AI models were becoming powerful, but serving them in real time was painfully slow. It was clear that the future of AI wasn’t just about training bigger models, it was about delivering results fast enough to be interactive. The team’s willingness to rethink the processor from scratch is what made Groq what it became.
Specialized hardware, built around a real workload with a clear bottleneck, tends to win.
The Team
The decision to join Niobium came down to three things: the desire to reduce exposure of private data, the chance to take another chip to market, and the team itself. From the first meeting, we clicked. The company has its roots in Ohio, and there's a kind of midwestern directness and friendliness to the culture that I found refreshing. Silicon Valley has an extraordinary concentration of talent and energy. That's why so many of us are here, but it can also be intense. At Niobium, the team is every bit as smart and hardworking, but there's an ease to the way people collaborate. Low ego. Best idea wins. It reminded me of the culture Jonathan Ross cultivated at Groq, and that kind of environment is where hard technical problems get solved.
The World Is Asking for This
In the past several decades, compute has scaled, storage has become abundant, and AI has crossed a threshold of competency. One constraint has stubbornly remained: meaningful computation still requires access to unencrypted data. And the data that could unlock the most value, including medical records, financial transactions, and proprietary intelligence is exactly the data organizations can’t or won’t expose.
In conversations with companies and institutions around the world, the message is consistent: people are actively looking for a solution to this problem. AI is consuming every piece of knowledge it can access, regardless of how sensitive that data may be, and organizations are increasingly concerned about what happens to it. The demand for privacy-preserving computation isn’t theoretical—it’s urgent.
Fully homomorphic encryption offers a fundamentally different model: computation on data while it stays encrypted. Some people point to trusted execution environments as an alternative, but TEEs still require you to hand your keys—and trust—to both the cloud vendor and the hardware vendor, and even then there are known risks of exposure. FHE eliminates those risks entirely. The math has been understood for years. What’s been missing is hardware fast enough to make it practical. That’s the problem Niobium is solving.
More Than a Chip
One lesson that carries across the Cloud TPU, Groq's LPU, and Niobium's mistic ASIC is that a new chip is never enough on its own. When the Cloud TPU arrived, perhaps a dozen people at Google could write programs for it. TensorFlow, JAX, and XLA changed that. Groq’s engineers built a software stack to run LLMs on racks of LPUs and offered developers a token processing API through GroqCloud. The hardware was the foundation, but the developer ecosystem is what created adoption.
Niobium is taking the same approach—providing the full stack so that any of the roughly 45 million developers in the world can build privacy-preserving applications, not just the hundred or so who understand FHE cryptography today. This is all being built into Niobium’s cloud, which we call the Fog.
This full-stack thinking is also what drew me from engineering into product. A Ph.D. in engineering gave me a deep appreciation for what great technology can do, but technology doesn’t expand on its own just because it’s impressive. It needs to be accessible and valuable for the people who will use it.
Why Now
Niobium has assembled a team that has built and launched 30 chips alongside some of the world's best cryptographers with years of FHE experience. The bottleneck is clear. The team knows how to close it. I've been lucky enough to see this pattern twice before, and I wasn't going to watch it happen a third time from the sidelines.
…
John Barrus is VP of Product at Niobium, where he focuses on making privacy-preserving computation accessible to every developer. Previously he served as VP of Product Management at Groq and Product Manager for Cloud TPUs at Google. He holds a Ph.D. in Mechanical Engineering from MIT. Niobium, Inc. is a brand of Niobium Microsystems, Inc.