Exclusive-Anthropic weighs building it own AI chips, sources say

Anthropic's run-rate revenue accelerated to over $30 billion in 2026 from about $9 billion at end-2025. The company is exploring designing its own AI chips amid a sector-wide chip shortage but has not committed to a specific design or team; it currently uses Google TPUs and Amazon chips and signed a long-term deal with Google and Broadcom. Anthropic has pledged $50 billion to bolster U.S. computing infrastructure, while industry sources estimate developing an advanced AI chip costs roughly $500 million.

Analysis

A strategic move by an AI developer toward in-house silicon flips the competition map from pure software scale to vertically integrated hardware-software stacks; that tends to compress gross margins for third-party chip vendors but creates outsized consulting/IP-design revenue for boutique design partners. Expect a material timing mismatch: hiring and tape-out cycles mean tangible supply-side effects will play out over 12–36 months, while market pricing of winners could re-rate within 3–6 months on partnership announcements alone.

The biggest non-linear effect is bargaining power over foundry and packaging slots. A new entrant with custom designs amplifies demand for limited advanced-node capacity and test/assembly resource pools, increasing pricing power for foundries and tooling vendors and forcing incumbents to prioritize hyperscaler contracts. That can create temporary supply squeezes for competitors reliant on commodity GPUs and give upstream designers (and foundry-linked acquirers) pricing leverage.

Key tail risks: (1) technical failure or catastrophic yield problems on first silicon will turn anticipated upside into a multi-year write-off; (2) strategic pullback — deciding to re-source instead of designing — would quickly re-rate partners down; (3) policy/export constraints or a sudden large foundry contract with a rival could reverse the dynamic within months. Watch hiring patterns, IP filings, and initial tape-out/wafer agreements as high-signal catalysts.

Contrarian read: the market will likely overpay for “control of the stack” narrative in the near term while underweighting execution risk and capital intensity. That suggests a two-layer approach: selectively own design/IP beneficiaries on asymmetric option-like exposures while keeping core cloud/hyperscaler exposure tactical until first silicon proves out.

Key Decisions for Investors

• Long AVGO 9–12 month call spread (buy nearer-term call / sell higher strike call) sized 2–3% portfolio: asymmetrically capture design/IP upside and Broadcom-design fee tailwinds. Target 40–80% return if partnership share expands; max loss = premium paid. Trim on +30% implied-vol compression or on formal multi-year design win announcement.
• Long GOOGL equity or 6–12 month calls (size 2–4%): exposure to TPU/cloud volume growth and partnership monetization. Use a 8–10% stop-loss on equity; option holders should target 25–50% upside tied to cloud revenue/capacity announcements.
• Pair trade — long AVGO (as above) funded by short META 6–9 month call (small size, net zero-ish cost): expresses hardware-design beneficiary vs. consumer/advertising AI execution risk. Keep short call size <50% of long position; unwind if META issues dedicated developer/infra announcements or if implied vol spikes >40%.
• Event-driven monitor & alert: set real-time alerts for (a) first silicon tape-out, (b) foundry contract (TSMC/GlobalFoundries/Intel) mentions, (c) formal hiring of chip-architect leadership. Move from option exposure to delta-equity only after successful wafer yields are reported (12–36 month horizon).