Samsung secured a $200 billion supply agreement with Broadcom running through 2030, covering 2-nanometer logic wafers, high-bandwidth memory, and advanced packaging for AI infrastructure. The deal averages $40 billion annually over seven years and represents Broadcom's largest single-vendor commitment in its history. Samsung did not disclose capacity reservation percentages, but the scale implies dedicated cleanroom allocation at the Hwaseong and Pyeongtaek fabs.
Broadcom's procurement strategy typically locks 18-24 months of forward capacity with multiple foundries. This structure extends that horizon to 84 months with a single counterparty, a departure consistent with hyperscale AI infrastructure buildouts where node transitions happen slower than mobile. The 2nm node refers to Samsung's gate-all-around transistor architecture, first sampled in Q4 2025 and entering volume production in Q2 2026. HBM inclusion signals Broadcom's vertical integration into memory-intensive AI accelerators, mirroring Nvidia's shift toward co-packaged optics and in-package memory. Advanced packaging here means 2.5D interposers and through-silicon vias, not chiplet disaggregation — Broadcom designs monolithic ASICs, not tile-based GPUs.
The agreement matters because it forces Samsung Foundry to prioritize Broadcom's 2nm allocation over yield-sensitive mobile customers, particularly Qualcomm and MediaTek, who historically received first wafer starts. Broadcom's AI revenue grew 74 percent year-over-year in fiscal Q2 2025, driven by custom accelerators for Google, Meta, and ByteDance. Locking Samsung capacity through 2030 suggests Broadcom projects hyperscaler demand will absorb $200 billion in silicon without price elasticity — a bet that AI infrastructure capex remains above $250 billion annually through the end of the decade. Samsung's margin profile changes: AI ASICs tolerate lower yields than smartphone SoCs because hyperscalers buy full wafer lots and bin chips in-house. This lets Samsung push 2nm into production faster, but at 12-18 percentage points lower gross margin than Apple's N3 volume at TSMC.
Allocators should watch Samsung's Q3 2026 earnings call for updated foundry margin guidance and any HBM4 qualification timelines tied to Broadcom's roadmap. Broadcom's next analyst day, typically held in December, will clarify whether this deal includes equity-like volume commitments or fixed-price tranches — the difference determines how much demand risk Samsung absorbs. TSMC's response matters: if Arizona N3 production ramps in late 2026 as planned, Broadcom could route iPhone modem volume there, freeing Samsung 2nm capacity for AI exclusively. The procurement structure also implies Broadcom negotiated $8-12 billion in upfront tooling credits, standard for deals above $100 billion, which would appear in Samsung's Q2 2026 cash flow statement under customer advances.
The seven-year horizon tells you Broadcom believes AI accelerator design cycles will slow from 18 months to 30-36 months, matching datacenter refresh economics rather than consumer upgrade cycles.