The least interesting part of an AI accelerator is the accelerator. That’s a strange thing for a security researcher to say, but I’ve spent enough time looking at how these boards fail to believe it. The mainstream story about AI hardware is all about compute — teraflops, memory bandwidth, interconnect fabric. Meanwhile the components that decide whether your accelerator runs correctly, throttles unpredictably, or lies to your monitoring stack sit quietly on the power rail, and almost nobody in security is watching them.
Which is why Infineon’s announcement on 7 September 2026 caught my attention more than most model releases have this year.
What Infineon actually announced
Infineon Technologies introduced the TDA235E5 and TDA235E0, a dual-phase smart power stage family built for the power density demands of next-generation AI accelerators. The parts combine Infineon’s OptiMOS 6 MOSFETs with a dual-phase driver IC in a compact package. Infineon also raised its 2026 chip revenue target by 50% to 1.5 billion euros, citing demand in the AI power supply segment, after sales in that area nearly tripled year-on-year to more than 700 million.
One correction worth making up front, because it shapes how I read the coverage: I could not verify the claim that Infineon is sampling 120-amp power chips for AI accelerators. That specific number does not appear in any source I checked. The dual-phase smart power stage family is real and documented. The amperage figure circulating alongside it is not something I can confirm, and I’m not going to repeat it as fact.
That gap is the security story
I bring it up because the distance between “a real product exists” and “a specific number attached to it in a headline” is exactly the distance attackers and sloppy procurement both operate in. If a spec figure can drift into trade coverage without a source, it can drift into a bill of materials, a datacenter capacity plan, or a vendor risk assessment just as easily. Provenance discipline is not pedantry. It’s the same muscle you use when you refuse to trust an unsigned firmware image.
Smart power stages are software now
The word doing the heavy lifting in “dual-phase smart power stage” is smart. These are not passive components. A modern power stage reports current, temperature, and fault conditions upstream to a controller, which feeds a platform management layer, which feeds your telemetry dashboards and your autoscaling logic. That’s a data path. Data paths have trust boundaries, and trust boundaries are where I make my living.
A few things I think AI infrastructure teams should be asking as this class of part lands in production boards:
- Who can write to the power management bus? If a compromised BMC or a rogue container with unexpected sysfs access can adjust voltage regulation parameters, you have a physical denial-of-service primitive sitting inside a logical access control failure.
- Do you trust your own telemetry? Current and thermal readings from the power stage often drive throttling and health decisions. Falsified readings can either cook a board or convince your orchestrator to drain healthy capacity.
- What happens under a crafted load? Inference workloads have distinctive power signatures. Deliberately shaped workloads that push transient current behavior are a legitimate reliability concern, and a legitimate abuse vector on shared infrastructure.
- Where did the part come from? Power components are a favorite target for counterfeiting and relabeling precisely because they look boring. Supply chain verification should not stop at the silicon everyone brags about.
The side-channel angle nobody wants to hear
Higher power density means larger, faster swings in current draw for a given workload. For those of us who study side channels, power traces have always been a useful signal. I’m not claiming any specific attack against these parts, and I haven’t tested them. But the general direction of travel — denser power delivery feeding accelerators running high-value models — means power-based inference about what a machine is doing gets more interesting, not less, over time. If you run multi-tenant AI infrastructure, that belongs in your threat model as an open question rather than a settled one.
Follow the money, then follow the attack surface
Infineon’s 50% target increase tells you where the volume is going. When a component category grows that fast, deployment outpaces scrutiny almost every time. New parts land in new board designs, new board designs land in new datacenter builds, and the security review shows up eighteen months later asking why the management bus was wide open.
My suggestion is simple. When your team evaluates the next generation of accelerator hardware, spend a session on the power subsystem specifically. Ask the vendor for the management interface documentation, not just the efficiency curves. Verify claims against primary sources, including the ones in this article. The boring components are where the interesting failures live.
🕒 Published: