The Fab Boots With CUDA — The Week Nvidia Became Every Fab's Operating System

Three days, four layers. Between June 5 and June 8, Nvidia quietly slid software, simulation, and co-design into every manufacturing layer it doesn't own. Not the GPU layer — that fight is over. The fab layer. The mask, the tool, the test, the memory next to the die. By Sunday night, the headline that ran the loudest — Senator Warren calling Jensen Huang to a June 11 hearing on China sales — Huang has since declined and counter-offered a headquarters visit — was the wrong picture. The right picture was that Huang already had bigger problems to solve than China: he was rebooting the global fab on his own software stack.

Start with photolithography. On June 8, Nvidia released cuLitho computational-lithography libraries explicitly tuned to TSMC's process nodes. That's not marketing. Computational lithography (cuLitho) — using GPUs to solve the inverse problem of mask design at N3 and N2 — has been TSMC's quiet bottleneck for two nodes, and Nvidia just shipped the libraries that sit between the mask-design EDA flow and the foundry. Cadence and Synopsys still own the schematics; Nvidia now owns the math.

Then fab simulation. Micron and MetAI announced SimReady fab digital twins on Nvidia Omniverse — a physical-AI runtime for the fab floor itself, robotics-grade, with closed-loop simulation for automation. Not a small pilot; a platform-level co-announcement. The Omniverse SDK becomes the operating layer for how Micron commissions, debugs, and trains the robotic fleet that runs DRAM and HBM cleanrooms.

Then memory co-design. SK hynix and Nvidia disclosed a multiyear technology partnership to "advance memory for AI factories" — a phrase that lands as marketing but reads, in the fine print, as an HBM4/HBM4E joint roadmap with Nvidia inside the spec from cell architecture to controller. SK hynix is the leading HBM3E vendor; pulling it into a multiyear engineering pact locks the supply that Broadcom's customers cannot get and that AMD is still trying to qualify.

And then the floorplan. TSMC bought additional land adjacent to its Arizona complex specifically to meet AI demand from US customers — read: Nvidia. The Arizona buildout was originally a political artifact; it has now become an Nvidia-customer artifact. The capacity isn't being built for "the US"; it's being built for the customer that just shipped cuLitho to the litho team that uses it.

Stack it up: photolithography (cuLitho × TSMC), fab simulation (Omniverse × Micron), memory co-design (multiyear × SK hynix), capacity anchor (Arizona × Nvidia demand). Five years ago, Nvidia bought wafers. This week, Nvidia booted the fabs.

The market got the wrong headline. Broadcom missed, Korean memory sold off (Samsung Electronics and SK hynix both fell hard on Friday), Seeking Alpha downgraded Nvidia to Hold, and Senator Warren summoned Jensen Huang to testify on China-bound H20 and B20 chips — Huang declined, counter-proposing a headquarters visit. The political tape is real — the BIS probe into chip diversion is open, and there will be hearings — but it's a sideshow to the structural move. The fab-stack acquisition is the M&A story Nvidia never has to announce. No regulators, no exchange ratio, no integration risk. Just SDKs, partnerships, and a piece of Arizona dirt.