Intel's Foundry Fix Lags As Hyperscalers Hold Back Volume

Intel Foundry Services reported $5.8 billion in second-quarter 2026 revenue, with a mere $293 million from external customers, highlighting a 20-to-1 internal-to-external ratio. This disparity underscores the lag in merchant foundry readiness behind manufacturing recovery. The operating margin improved to negative 36.2 percent from negative 71.7 percent a year prior, and operating losses decreased to $2.1 billion from $3.2 billion. However, external revenue remains insignificant against the $17.8 billion total foundry business Intel recorded in full-year 2025.

CEO Lip-Bu Tan positions Intel's external strategy on custom ASIC partnerships rather than direct competition with TSMC. Fortinet's SP6 security processor, manufactured on Intel 4 node with Intel handling back-end design and manufacturing, represents a notable external win. Intel's 18A-P has entered risk production, showing progress in defect density and transistor performance, with 14A scheduled for high-volume manufacturing in 2028. Tan's claim that 14A will match TSMC's timeline underscores the urgency of securing external volume on 18A-P.

Intel's foundry operations are captive to its internal product demands, with CFO David Zinsner describing AI chip supply as "hand-to-mouth." Production lines have been redirected from consumer chips to Xeon server processors to meet internal AI priorities, leaving external customers with an unclear queue. Intel has not disclosed fab location, packaging technology, production timetable, committed volumes, or financial terms for the Fortinet engagement. The purpose-built silicon business is nearing a $2 billion annual run rate and could reach $4 billion, reflecting internal and semi-captive custom silicon rather than merchant foundry bookings. External foundry revenue was $174 million in Q1 2026 and $307 million for all of 2025.

The critical issue is the lack of volume commitments from AI hyperscalers on nodes crucial for inference silicon. Microsoft has confirmed a custom silicon collaboration without naming a node; AWS is developing custom Xeon and AI fabric chips; Apple has reportedly received an 18A-P design kit for evaluation. None indicate high-volume production commitments, and receiving a design kit is not the same as booking wafer starts. Under prior CEO Pat Gelsinger, 18A was pitched as a dual internal-external node; yield struggles led Tan to redirect external marketing to 14A, with recent improvements bringing 18A-P back into customer discussions. BNP Paribas analyst David O'Connor gives Intel 12 to 18 months to secure a major external 14A customer, which is crucial for Intel's long-term manufacturing presence. The Ohio fab is delayed until at least 2030.

For architects considering Intel Foundry against TSMC for custom AI inference chips, the Fortinet deal validates Intel 4 for complex ASIC execution. However, leading-edge deployments require 18A-P or 14A, with no hyperscaler committing production volumes on either. Architects should await concrete customer signals from Intel, promised for the second half of 2026, before integrating its nodes into supply-chain plans.