MUNICH, 26 AUG 2026 — Infineon is buying C2i Semiconductors, a fabless design house in Bengaluru that builds software-defined multiphase controllers and smart power stages for AI servers. Terms were not disclosed, and the deal is expected to close in the third quarter.
It is a small acquisition in a part of the machine that almost never gets written about, and it addresses a constraint that is becoming as real as the supply of accelerators themselves.
The layer nobody photographs
Every article about AI infrastructure carries a picture of a rack. Inside that rack, between the power supply and the accelerator, sits a set of components whose job is to convert a relatively high voltage into the very low voltage a processor actually runs on, and to hold it steady.
A modern accelerator draws enormous current at around a volt. Delivering that means splitting the job across many parallel converter stages, switching them out of phase with one another so their ripple cancels, and coordinating the whole arrangement with a controller. That is what a multiphase controller does, and what C2i makes.
It is unglamorous and it is not optional. No amount of accelerator supply helps if the power cannot be delivered cleanly to the die.
Why transient response is a real limit
The difficulty with an AI workload is not average power, but how violently the demand changes.
An accelerator running a training step can go from comparatively idle to drawing hundreds of amps in a matter of microseconds, and back again, thousands of times a second. Each of those swings tugs at the voltage rail — current rushing in makes it sag, and the sudden removal of load makes it overshoot.
The processor tolerates only a narrow band around its target voltage. Sag too far and it becomes unstable, so the safe response is to clock down and take the performance loss. Overshoot too far and the damage is cumulative.
Power delivery, then, is a performance question rather than only an efficiency one. A regulator that reacts faster lets the chip run closer to its limits without protective throttling, which is capability recovered from hardware already bought and already installed. In a market where accelerators are the scarce item, extracting more from the ones on hand is a straightforward financial win.
What Infineon was missing
The purchase makes more sense once you look at what Infineon already had. Its power portfolio is built on device technologies — silicon, silicon carbide and gallium nitride — which are the switches themselves, the components that turn current on and off.
The controller is a different discipline. It is mixed-signal and increasingly digital design: sensing what the load is doing, deciding how each phase should respond and doing it in a control loop measured in nanoseconds. Being good at building switches does not make you good at deciding when to operate them.
This is a move from selling components to selling a matched set. A customer designing a board can buy switches from one supplier and a controller from another and do the integration themselves, or buy a matched set characterised to work together. The second is worth more, and it is only available to a vendor holding both halves.
The smart power stage in C2i's description is the same logic in a package. Putting the driver, the switches and the current sensing into one component, physically close to the load, shortens the connections between them — and in a circuit switching this fast, the stray inductance of a few millimetres of copper is a real quantity that degrades exactly the transient response the whole design is trying to protect.
Software-defined power cuts both ways
The word doing the work in the description is software-defined. Traditionally this control loop was fixed in analogue circuitry, tuned at design time and unchangeable afterwards.
A programmable controller can be tuned per workload, adjusted as the characteristics of the load become better understood, and updated in the field. For a platform whose workloads change faster than its hardware, that flexibility is valuable.
It also puts firmware in the power path. Anything programmable can be misconfigured or carry defects. In principle it can also be reached from the management network. The consequences of a fault sit closer to the physical layer than most software failures do, and the industry has comparatively little operational experience with this surface.
None of that is an argument against the approach, which is clearly where the market is going. It is an argument for asking who signs the firmware and how it is updated, questions rarely asked of a voltage regulator.
India's semiconductor story is design, and it already works
Public discussion of India's ambitions is dominated by fabrication plants — enormous, slow, heavily subsidised projects. The part that already works, which is design, attracts far less attention.
C2i is fabless. Its value is entirely in engineering — circuit design, control algorithms and the verification work behind them — and that requires people rather than a multi-billion-dollar facility. Infineon already employs around 2,800 people in India and is establishing a centre of excellence for power technology there, which describes a company deepening a design presence it has spent years building.
This is the same shape as IBM's sovereign AI arrangement with Sarvam. In both, the asset being acquired or engaged is intellectual rather than industrial. A country can become significant in semiconductors through design decades before it becomes significant in manufacturing, and India is doing the former now.
What it means from here
The acquisition points to where the next constraint sits. Data centre capital expenditure is on course to pass a trillion dollars, and the discussion of what limits a buildout has moved from chips to power — grid connections, generation, transmission.
That conversation stops at the building. Inside it there is a second power problem, measured in microseconds rather than in years of permitting, and it decides how much useful work the installed hardware actually performs. An operator can secure a grid connection and still lose capability to protective throttling at the board.
Anyone specifying capacity should add a question to the list. Alongside how many megawatts a site has contracted, it is reasonable to ask how the power delivery on the boards is specified, because that is the difference between accelerators that run at their rated limits and accelerators that spend part of their life clocked down.
The price would have told us how strategic Infineon considers this. It has not been disclosed, so the size of the bet is unknown; only its direction is.
