Semiconductors

AI compute demand is now contracted in gigawatts, not units.

AMD's 6-gigawatt GPU deployment agreement with OpenAI and Broadcom's supply-backed guidance of $115 billion in FY2027 AI revenue mark the moment AI chip procurement shifted from discrete purchase orders to long-duration, volume-committed supply relationships. Broadcom's Anthropic trajectory crossed from 1 gigawatt of Ironwood in 2026 to an explicitly named 5 gigawatt TPU v8i deployment in 2027, with 10 gigawatts in line of sight for 2028. This is structural: hyperscalers are now underwriting multi-year compute buildouts measured in power capacity, not chip generations.

Physical infrastructure, not silicon supply, caps AI revenue realization.

Broadcom explicitly conditioned its FY2027-2028 AI revenue outlook on data center site readiness, stating land, power, and shell availability is the largest variable in deployment timing. Lam Research flagged Strait of Hormuz disruptions from the February 2026 U.S.-Israel-Iran conflict as a new input-supply risk affecting aluminum, helium, bromine, and sulfur, while Chinese rare-earth export controls create a hard November 2026 supply deadline. The binding constraint on semiconductor revenue is no longer chip design or wafer capacity; it is the physical stack below the silicon.

GPU vendors are acquiring software and systems layers to lock in AI stack.

NVIDIA's $11.9 billion acquisition of Hugging Face extends its competitive position from GPU supply into ownership of the open-source model distribution layer, directly mirroring hyperscaler vertical integration logic. AMD's acquisition of ZT Systems' design operations and preview of the Helios rack-scale platform signal an identical intent: moving from discrete chip sales to full-stack AI system competition. Marvell's acquisitions of Celestial AI and XConn Technologies extend its connectivity moat into photonic interconnect and PCIe/CXL switching, while its NVIDIA preferred equity partnership formalizes hyperscaler alignment as structural.

Export controls are permanently bifurcating semiconductor revenue geography.

Applied Materials' China revenue fell from $10.12 billion in fiscal 2024 to $8.53 billion in fiscal 2025, while Taiwan surged from $4.01 billion to $6.86 billion and Korea from $4.49 billion to $5.61 billion, a geographic rerouting driven entirely by U.S. export control policy. Lam Research's China share of revenue has declined from 42% in FY2024 to 34% in Q3 FY2026, while AMD absorbed a net approximately $440 million charge from MI308 restrictions. Texas Instruments discloses that products shipped into China represent approximately 50% of total 2025 revenue, a concentration that no amount of customer-HQ-based reporting obscures.