The Friday Briefing: Two different bets on silicon carbide

Infineon brought a 300 mm silicon carbide wafer to Munich this week. It’s about the size of a dinner plate, though I wouldn’t want to be the person who leaves a thumbprint on it. The company had already made working power devices on it in Villach, using equipment it normally uses for silicon.
Meanwhile, onsemi and the Czech government dropped their plans to expand silicon carbide production at Rožnov, a project originally put at up to $2 billion. Put those stories together and you have a fair question for the next supplier meeting. Where can the industry make these parts more cheaply, and where does it still make sense to add capacity?
What Infineon gets from a bigger wafer
Going from 200 mm to 300 mm gives you 2.25 times the wafer area. There’s room for more devices in each run, and using existing equipment gives Infineon a way to try the process in a working fab. What I’d like to see next is the yield. A larger wafer gives you more places to put a die, but the cost per good device depends on how many work, what the wafer costs and how reliably you can repeat the run.
At Rožnov, onsemi will continue making silicon and SiC devices at the existing site. The company and the Czech ministry pointed to changes in the SiC market when they called off the expansion. They gave no detail on what had changed in the demand or cost assumptions behind the 2024 proposal.
Infineon’s 300 mm wafer · Onsemi’s Rožnov decision
Nvidia after Jensen Huang
Jensen Huang resigned as Nvidia’s chief executive, president and a board member on Thursday, effective immediately. Colette Kress is keeping her finance job while serving as interim CEO, and Harvey Jones has become chair. Nvidia shares fell more than 12% in early US trading before recovering much of that loss. Reuters later put them down about 4% during the session.
Vera Rubin is already ramping into production, according to Nvidia. The NVL72 puts 72 Rubin GPUs and 36 Vera CPUs in a liquid-cooled rack, and every one of those racks has a delivery date that depends on more than the chips. Memory and packaging suppliers have to deliver, server builders have to assemble and test the equipment and the site needs the power and cooling to run it. Kress inherits all of those schedules.
Nvidia has also put money into companies building infrastructure around its chips. I’ll be looking at how Kress talks about those commitments and whether customers and suppliers are still working to the same delivery schedules. That’s where a change of leadership could start to have a practical effect.
Our full report on Huang’s departure and Vera Rubin · Nvidia’s transition announcement · Vera Rubin production update · Reuters on the resignation and early share move · Reuters’ later market update
Fractile takes on the cost of moving data
Fractile told Reuters it raised $720 million, with the UK’s Sovereign AI fund and the British Business Bank among its backers. The London and Bristol company is developing AI inference chips that use custom memory. The underlying problem is familiar. A processor can only work as quickly as it can get the data it needs, and moving that data also uses power. The figures I’d want from Fractile once customers can try the hardware are throughput and power on the models they actually run.
UK government announcement · Reuters on the funding round · Fractile’s approach to memory
Sandisk keeps e.MMC in play
If you’re maintaining a gateway that writes logs all day, Sandisk’s latest e.MMC part has a feature worth looking at. The iNAND CL EM181 gives the host the option of setting aside an SLC partition for frequently rewritten data. You give up some general storage capacity in the process, so the choice comes down to what the device writes and how often it writes it. The 64 GB to 256 GB range keeps the same package format, which gives teams already using e.MMC a familiar starting point for qualification.
A few things from Munich
Imec is aiming to put a silicon interposer design kit in engineers’ hands in early 2027, following work on a manufacturing flow that supports microbump pitches down to 20 µm. Later in the week, Luminovo showed buyers how to search across their bills of materials for boards using an obsolete part. We’ve covered both in the daily reports, alongside the discussion on bidirectional GaN and Single Pair Ethernet.
Something you might have missed
When the only faulty chip you have is still sealed, cutting into it is a decision you’d rather make with some idea of where to look. QuantumDiamonds’ microscope uses defects in a diamond crystal to sense the magnetic field generated by current. A map of that field can help the failure analysis team find where the current is going before anyone reaches for the saw.
The diamond microscope in our day-two report
Have a good weekend. If you came back from Munich with a firm price and a delivery date in writing, you may have found the rarest thing at the show.



