Marvell's $120B Google Win, Modular's Open Compiler Play, and the Fight Over NVIDIA's Ohio Scale-Back

Patrick Moorhead and Daniel Newman break down Marvell's $120 billion custom-silicon deal with Google, Qualcomm-owned Modular's open-source compiler push at ModCon 2026, and state-level pushback against AI data center construction in Texas and Pennsylvania. The hosts also flip roles to debate whether NVIDIA's reduced Ohio financing guarantee signals cracks in the AI infrastructure buildout.

The handpicked topics for this week are:

Marvell and Google Finalize $120 Billion Custom-Silicon Deal: Marvell secured a six-and-a-half-year agreement with Google worth close to $20 billion annually, alongside a warrant covering approximately 7% of the company. The deal extends Marvell's position across all three major hyperscalers through attach silicon: accelerators, optical and networking connectivity, memory interfaces, and near-memory compute built around the TPU ecosystem. Patrick Moorhead noted Marvell's CXL memory-pooling work signals where next-generation architecture is heading beyond current TPU designs.

(The Decode)

Modular Delivers on Open Compiler Promise at ModCon 2026: Modular open-sourced its Mojo compiler under Apache 2.0 and launched Modular Cloud, a platform letting enterprises arbitrate across both AI models and underlying hardware architectures. Moorhead reported the update removes device restrictions from the Max license and adds support for AWS Trainium and Qualcomm's own data center accelerators, with an alliance program set to launch by year-end. He argued the move gives Qualcomm indirect revenue exposure across every token served, extending its footprint from data center to edge, robotics, and industrial IoT. (The Decode)

OpenAI Balances a Safety Push With Codex-Driven Enterprise Wins: OpenAI paused reinforcement learning for its newest model and strengthened its data-retention policies, while high-profile Codex results—including Asana compressing an estimated five-year engineering effort into two weeks—showcased its enterprise impact. Daniel Newman saw the timing as an effort to reset the narrative around safety and transparency after months of competitive pressure from Anthropic. Patrick Moorhead also examined OpenAI's use of activation classifiers, which monitor internal model activity token by token and trigger deeper review only when risks are detected—a safeguard that adds roughly 20% to inference costs. (The Decode)

Texas and Pennsylvania Tighten Rules on AI Data Center Development: Governors in both states issued new restrictions within two weeks of each other, with Texas pausing new approvals pending an audit, and Pennsylvania adding fresh permitting requirements. Moorhead reported that at least 40% of Texas's planned data centers run on self-sustaining natural gas power and never touch the ERCOT grid. Newman connected the political friction to a communications gap, noting recent polling shows more support for a nuclear plant nearby than a new AI data center. (The Decode)

Micron Commits $10 Billion to Boise Research Labs on Top of $50 Billion Fab Buildout: Micron unveiled a decade-long research lab investment in Boise, Idaho, alongside its existing $50 billion fabrication expansion, part of a CapEx program Newman noted totals roughly a quarter trillion dollars in commitments. Both hosts framed the spending as reinforcing memory's strategic status alongside leading-edge logic, with Newman citing HBM's technical differentiation as harder for competitors to replicate than standard NAND or DRAM. Moorhead pointed to Micron CEO Sanjay Mehrotra's argument that domestic memory manufacturing will determine AI leadership. (The Decode)

The Flip: Is NVIDIA's Ohio Scale-Back a Capex Crack or Disciplined Structuring?: NVIDIA reduced its financing guarantee for OpenAI's Ohio data center from $250 billion to under $120 billion, a shift Moorhead argued as the first visible sign of balance-sheet discipline overtaking the demand story. Newman countered that the adjustment is one component of a financing platform exce