The AI boom has a profit problem hiding in plain sight. OpenAI spent $34 billion in 2025, and both it and Anthropic now put more than half their revenue toward inference, the cost of actually running their models. Hyperscaler capital spending is heading toward $1 trillion next year. Data center operators are stuck choosing between systems that are fast but expensive and systems that are affordable but too slow for customers.
Marc Bolitho believes the way out is a 400-year-old idea. His company, Tensordyne, based in Sunnyvale, California, and Munich, uses logarithmic math to turn costly multiplication into simple addition. He says the approach delivers 13 times the throughput of Nvidia’s Blackwell systems on 75 percent less energy for agentic AI. With its first chip now in production at TSMC and a Series D on the horizon for early next year, Bolitho explains why the best technology doesn’t always win, and what it takes to sell something new to the most conservative buyers in tech.
Tell us more about Tensordyne’s business model.
Tensordyne builds AI inference systems—the hardware and software that run AI models. We sell these systems to hyperscalers, neoclouds and large enterprises.
Demand for AI is rising, but so are the cost and power required to run it. OpenAI’s spending hit $34 billion in 2025. OpenAI and Anthropic now spend more than half of their revenues on inference, and hyperscaler capex is heading toward $1 trillion next year.
Within the confines of the current hardware, data center operators are forced into a choice between systems that are fast but expensive, or those that are economical but too slow for customer expectations. Neither path scales into a sustainable AI future—the market is waking up to the fact that AI is only good when it can deliver on both speed and cost.
Multiplication in AI is expensive, so we use a logarithmic number system that turns them into simple, inexpensive additions. That frees up power and compute space on the chip, which we reinvest to right-size the chip for inference so that we can deliver 13x higher throughput than Nvidia’s Blackwell systems and consume 75 percent less energy for agentic AI. The aim is to make AI profitable as companies continue to build great products with it.
How have you grown?
Our first commercial chip has completed tape out and is now in production at TSMC. We’ve secured pre-orders and letters of intent from data center operators and are looking at our Series D early next year.
We’d proven the science worked in our earlier products before bringing it to generative AI inference. Napier is an industrial product that will ship and integrate into existing and new data centers.
Tensordyne’s AI system is built on your proprietary technology that changes the economics of how AI is run. What have you learned about bringing genuinely new technology to market?
The science is the part you can control—either it works, or it doesn’t. From years of research and building a successful first chip with it, we knew the math worked.
Logarithmic math is a 400-year-old idea that we brought into artificial intelligence. What we cracked was how to take that centuries-old concept and actually commercialize it at data-center scale, and we’ve patented the approach to protect that lead. No one else has done it.
But genuinely new technology doesn’t sell itself, and you shouldn’t expect it to. Data center operators, for one, are betting their uptime and margins on whatever you end up putting in their data center, so the parties that buy infrastructure are conservative for good reason.
They want both technological efficiency and operational efficiency—fast set-up times and little to no downtime. And in case there is a fault, ultra-fast and easy replacement of parts without disruption of operations.
The best technology doesn’t always win. The ones that are easy to use and integrate are the ones that have a high penetration rate. We deliberately built a system that looks familiar to operators of the incumbent’s hardware, so a customer can move things around without re-architecting everything they already own.
On top of that, we built a system that is fully air-cooled which makes our system ideal for 80 percent of data centers out there that are not fit for water-cooling.
How do you see the economics of AI compute reshaping decisions for business leaders over the next few years?
I think compute economics should already be a board-level discussion for any company growing in the age of AI. Right now, companies are meeting the real cost of AI as usage scales far faster than unit costs drop, and they are hitting hard ceilings on what they can spend. The leaders who win will be the ones who built with a profitable option for AI in mind.
A few things compound this. Models are getting larger as agentic work gets more complex. Video, as a modality beyond text, is far more demanding. All of that pushes cost up at exactly when everyone is desperate to deploy AI more widely.
The goal we’re working toward is to make high-quality inference behave less like a luxury and more like an affordable utility. Only when that happens can the question for AI shift from “can we afford this?” to “what can we build with it?”
