Meta to put AI chip into production in September as it looks to double computing capacity, memo shows

Meta plans to start manufacturing its in-house AI chip “Iris” from September, following six weeks of testing, aiming to expand computing capacity to 7GW in 2026 (up from adding 1GW H1 and 2.5GW by year-end) and 14GW by 2027. The company expects to spend up to $145B on AI infrastructure this year and secured long-term supply agreements (Samsung memory, Sandisk flash, Sumitomo fiber) amid memory “chipflation” pressures. While shares initially fell on the chip manufacturing update, they later recovered and were up 4.6% as Meta announced developer access to an AI coding model competing with OpenAI and Anthropic.

Analysis

Meta’s real value capture here is not “AI ambition” but procurement optionality: custom silicon can turn a structurally rising compute bill into a slower-growing opex line, which matters far more at its scale than at peers’. The market should treat this as a margin-defense story first and an innovation story second; if the internal chip cadence is real, it compresses the economics of merchant GPUs over 6-18 months, but only after a long validation cycle.

The immediate winners are the silicon enablers and supply bottlenecks that remain sticky even when hyperscalers internalize design: AVGO and TSM have the cleanest exposure, while memory/interconnect suppliers benefit from the broader buildout as the constraint shifts away from pure compute. The loser set is NVDA/AMD, but the first-order revenue hit is likely overstated; custom chips usually take the lowest-ROI inference workloads first, so any bear case on merchant GPU demand is a 2026 story, not a next-quarter story.

The contrarian view is that consensus is too focused on “Meta beating Nvidia” and too little on the fact that AI infrastructure inflation is spreading across the stack. If power, memory, and networking stay tight, lower chip dependency simply reallocates spend rather than reducing it, which supports suppliers like SNDK/SSNLF and keeps the AI capex cycle intact. The thesis is falsified if Meta’s capex/power ramp slips or if Broadcom/TSMC commentary fails to show any custom-silicon volume inflection over the next two earnings cycles.

Key Decisions for Investors

• Overweight META on any post-rally consolidation: the setup is 3-6 months of margin leverage if custom silicon reduces compute intensity; cut if next earnings shows capex or deployment slippage.
• Long AVGO / long TSM as the cleaner 6-12 month expression of hyperscaler custom silicon adoption; better risk/reward than shorting a single GPU name because they monetize design/foundry volume regardless of who wins the accelerator share.
• Keep NVDA and AMD on a watchlist rather than a hard short: only consider a bearish options structure after evidence that Meta is migrating meaningful inference workloads off merchant GPUs, ideally confirmed in the next 1-2 quarterly updates.
• Add SNDK and SSNLF on pullbacks for a 6-18 month supply-constrained memory trade; the risk is demand normalization, but the supply agreements imply pricing power persists longer than the market usually discounts.