Marc Bolitho likes problems that hide inside other problems. In cars, the visible question was autonomy: could a vehicle understand the road quickly enough to make a good decision? Underneath sat a less cinematic constraint. All that sensing and computation consumed power, and in an electric vehicle every hungry processor quietly borrowed miles from the battery. Now, at Tensordyne, the visible question is artificial intelligence. Underneath sits a familiar nuisance: every answer has an electricity bill.
The continuity is almost too neat. Bolitho began his career designing sensors for electronic braking and steering. Three decades later he runs a company designing an AI inference system from mathematics through the rack. The objects have grown larger, but the engineer's instinct has survived: look past the feature, find the bottleneck, and ask what the system is really spending.
His route to that rack was unusually long for a Silicon Valley chief executive. Bolitho spent 28 years at TRW and ZF, advancing from product engineering through electronics, sensors, passive safety and advanced driver assistance. By his final chapter there, he was responsible for ZF's Electronics and ADAS business, reported at $2 billion in annual revenue, and a global team of 5,000. He left in 2022 with a grateful note to colleagues. It read less like an escape than the closing of a very large, very formative book.
A second act, timed by accident
Recogni appointed Bolitho CEO and a board member in July 2022. The company built visual-perception processors for autonomous driving, a field that fit his experience almost suspiciously well. He had spent years thinking about sensors, safety, latency and the strange arithmetic of vehicle power budgets. In an interview that year, he remembered wanting from an early age to understand how things worked and admiring people who solved problems. Recogni had provided a beautifully difficult one.
Then ChatGPT arrived at the end of 2022 and rearranged the AI market. Recogni had just finished an engine for autonomous driving. The team could sell the proven device, divide its effort between two markets, or make the next chip for data centers. Bolitho later described the choice in unusually practical terms: a startup has limited resources and capital, so the company studied product cycles, customer needs, market size and execution risk. It chose data-center inference.
A pivot is a clean word for a messy act. It asks a team organized around one customer and product to carry its best work toward another. It asks investors to believe that discipline, rather than panic, is driving the change. It also makes yesterday's proof oddly inconvenient. A working automotive chip was both evidence of capability and a reason to stay put.
Startup success is a team sport.Marc Bolitho
Bolitho's answer was organizational as much as technical. He says that on his first day as CEO he wrote a startup thesis with four rules. The public version begins with emotional steadiness: prepare for peaks and valleys, avoid getting too high or too low, and return to the mission the next morning. Another rule makes the team, rather than the function, the unit of success. If one group falls behind, others lend time or resources. Individual metrics cannot become little private kingdoms.
Prepare for peaks and valleys
Keep emotional amplitude smaller than operational reality. Resume the mission tomorrow.
Work as one team
Move people and time toward the constraint, even when another department owns it.
Make differences visible
The leadership team shares DISC profiles to understand how colleagues process and respond.
There is a pleasantly mechanical quality to this view of management. People are not cogs, but interfaces matter. Bolitho says Tensordyne's leadership team completes DISC surveys and sees one another's results. The purpose is not to reduce a colleague to four letters. It is to expose how each person takes in information and behaves with others, particularly when the organization is climbing or descending one of those inevitable hills.
The arithmetic beneath the spectacle
The new mission led below the chip. AI inference consists of repeated linear-algebra operations, and multiplication is costly in silicon area and energy compared with addition. Logarithms offer an old trick: a multiplication can become an addition in the log domain. Tensordyne's premise is to make that trick useful at modern model scale, with software handling conversion and deployment so the customer does not have to become a logarithm enthusiast.
The elegance is only the beginning. A fast chip can still wait on memory. A group of chips can still wait on networking. A rack assembled from mismatched parts can waste density, capital and power. Bolitho says the team concluded that being only a chip supplier would not create the winning position. The product therefore expanded into math, compute, memory, interconnect, software and system integration.
This is a useful paradox of focus. Tensordyne narrowed its target from a variety of automotive use cases to data-center inference, then widened the number of technical layers it had to control. A precise promise can demand a broad product. The point is not to own everything. It is to own every dependency that can ruin the promise.
In 2024, Recogni raised a $102 million Series C. It announced its Pareto logarithmic AI math and a collaboration with Juniper Networks on inference systems. In 2025 came a partnership with DataVolt and a new identity. Recogni became Tensordyne, combining the tensor at the heart of modern AI with the dyne, a unit of force. Bolitho wrote that names should make a promise. The old name described recognition; the new one tried to describe a whole machine in motion.
Names should make a promise.Marc Bolitho, on the Tensordyne identity
The promise acquired a product name with a sly historical wink. Napier honors John Napier, the mathematician associated with logarithms. In June 2026, Tensordyne announced that its 3-nanometer Napier accelerator had taped out, developed with Broadcom and HPE Juniper Networks. Tapeout is a milestone, not the end of the story: it marks the point at which a chip design is sent for fabrication. The claims now have to meet silicon, software, customers and time.
The operator inside the engineer
Bolitho's public manner is more systems review than founder mythology. He talks about market ratios, architecture, organizational balance and the conditions under which a team can perform. Even his enthusiasm has a budget. In an automotive interview, he made the case for purpose-built processors by translating efficiency into something a driver could feel: more vehicle range. At Tensordyne he translates it into the economics an AI provider can feel: tokens per watt and tokens per capital dollar.
His long first career matters here. Automotive electronics punishes casual thinking. A component lives inside a product with hard limits on heat, power, cost, timing and safety. Suppliers coordinate across continents and product cycles measured in years. Bolitho learned leadership in a world where interfaces are contractual and consequences arrive in metal. The AI market moves faster, but its infrastructure is becoming stubbornly physical. Data centers must still be powered, cooled, financed and assembled.
That makes his career less a leap between industries than a migration of constraints. Braking sensors became perception processors. Vehicle power became rack power. A global automotive team became a dual-hub AI company, with chip, hardware, systems and software work concentrated in the United States and mathematical AI talent in Munich. According to Bolitho, code, models and requirements move continuously between them. Two sites, one team, and many opportunities for a handoff to go wrong.
The bet is now legible. Training creates a model; inference pays the recurring cost each time the model answers, reasons, acts or generates video. Bolitho expects that recurring workload to dominate AI compute. Tensordyne wants to make each answer fast enough for the user and inexpensive enough for the provider. It is a business thesis expressed as an engineering constraint.
There is no tidy ending available yet, which is appropriate. The design has reached tapeout. The system still has benchmarks to produce and customers to satisfy. Bolitho's own rule is better suited to the moment than a victory lap: do not get too high, do not get too low. A career that began by asking how things work has arrived at a more expensive version of the same question. The answer, if it works, will be counted one token at a time.