Anthropic explored MatX acquisition for chip development

Reuters reports Anthropic discussed acquiring AI chip startup MatX for roughly $7B, though talks have shifted to potential partnership discussions. MatX, founded by former Google tensor processing unit engineers, is reportedly seeking new capital at a ~$4B valuation. The move highlights Anthropic’s push to secure chip resources and accelerate in-house custom hardware development to support its fast-growing AI business.

Analysis

This is not a near-term GPU displacement story; it is a signal that frontier AI buyers are trying to arbitrage inference cost and vendor concentration. That matters for NVDA more in the multiple than the next quarter’s revenue: if the market believes custom silicon becomes a standard procurement lever, the terminal share of wallet for GPUs can compress even while total AI capex keeps rising. The time horizon is key — tape-out, software porting, and yield learning mean any real substitution is a 6-18 month issue, not a days/weeks trade.

The cleaner second-order winner is GOOGL. Any credible custom-chip initiative by a major model lab validates the TPU-style thesis that differentiated silicon plus software can lower unit compute costs and defend margins; it also increases the value of scarce ASIC talent and toolchains around Google’s ecosystem. Conversely, the supply chain beneficiaries may be foundry/packaging/EDA names, because more custom silicon implies more wafers and advanced packaging even if the accelerator mix shifts away from Nvidia.

Consensus risk is overreading the bearish NVDA angle. A single private-company exploration does not equal volume migration, and the hardest workloads will likely remain GPU-heavy longer than headlines suggest. What would falsify the bullish custom-silicon thesis is any evidence that these efforts stay exploratory — no hiring ramp, no silicon tape-out, or no disclosed deployment within 2-3 quarters — while NVDA’s data-center backlog and gross margins keep re-accelerating.

Key Decisions for Investors

• Prefer GOOGL over NVDA for a 1-3 month relative-value expression: long GOOGL / short NVDA on strength if the market starts extrapolating custom-silicon adoption. Risk/reward is modestly favorable because GOOGL has optionality to TPU/AI infrastructure upside while NVDA is more exposed to multiple compression than near-term earnings damage.
• Do not short NVDA solely on this headline. Use it as a watch item only; the thesis needs confirmation via customer procurement shifts, not press-level chatter. If NVDA holds its post-earnings breakout and data-center guidance stays firm, cover any tactical shorts immediately.
• If you want to express the second-order theme, accumulate on weakness in the custom-silicon supply chain rather than the AI model names themselves. The cleaner beneficiaries are advanced foundry/packaging proxies over a 6-18 month horizon, because the capex spend persists even if accelerator share changes.
• Set an alert on any Anthropic or peer disclosure of owned silicon/tape-out timelines. That is the first real catalyst that would justify a larger underweight in NVDA and a deeper rotation into GOOGL and AI infrastructure names.