Anthropic weighs building its own AI chips- Reuters

Anthropic is considering designing its own AI chips amid a shortage of processors needed to run advanced models, though plans remain early-stage and the company may still choose to buy chips. The firm recently signed a long-term deal with Google and Broadcom for Google TPUs and currently uses multiple chip types to run Claude. Similar in-house chip discussions are underway at other major AI players, including Meta and OpenAI, highlighting sector-level supply pressure and strategic moves to secure compute capacity.

Analysis

Anthropic’s exploration of in-house silicon is a classic vertical-integration option: low probability near-term (execution and fab access) but high impact if pursued. If they move beyond design into production, it shifts spend from cloud TPU/GPU time toward upfront CapEx and bespoke server builds, increasing demand for systems integrators and packaging services while compressing per-unit revenue for cloud incumbents over a multi-year window.

Near-term (0–12 months) the market reaction will be driven by procurement behavior — more bespoke designs raise orders for board-level components and chassis from OEMs, and increase demand volatility for wafer starts at TSMC-style foundries; this benefits fast-turn server OEMs and specialty component suppliers more than general-purpose GPU makers. Over 12–36 months the second-order supply-chain winners are companies that capture integration risk (thermal, interposer, firmware) rather than pure-play fab or model licensors, because ML stack optimization becomes the gating factor for production-grade accelerators.

Key reversal risks: Anthropic concludes that in-house chips don’t materially beat external TPUs/GFLOP per dollar, or long-term contracts (and ecosystem lock-in) with cloud partners remain cheaper — both would sharply reduce the upside for suppliers of custom hardware. Regulatory/security concerns and fab capacity shocks are lower-probability but high-consequence tail events that could accelerate either consolidation or a pivot back to external procurement, shifting winners and losers in under 6–24 months.

Key Decisions for Investors

• Long SMCI (3–12 month horizon): overweight server OEM exposure to capture rising demand for custom rack solutions. Target +35–50% upside if anthopric-scale bespoke builds materialize; set tactical stop at -18% to limit execution/fabrication risk.
• Long AVGO (6–18 month horizon): buy stock or 12-month call spread to play networking/accelerator interface silicon demand. Expect 20–35% upside if OEMs standardize on Broadcom interconnects; downside limited to ~15% from multiple compression in a weak procurement cycle.
• Pair trade (12–24 months): long SMCI + AVGO vs short GOOGL (equal notional). Rationale: capture systems/integration upside while hedging cloud-provider revenue retention risk; target 2:1 upside/downside (30%/15%). Rebalance quarterly as contract disclosures or fab announcements arrive.
• Hedge with GOOGL put spread (12 months): buy protective put spread to guard against a realization event where Anthropic or competitors internalize chip production and reduce long-term TPUs demand. Cost is limited to premium; payoff begins if GOOGL falls >~12–15% within 12 months.