Cathie Wood Just Went Bargain Hunting. Here Are the 3 AI Stocks She Bought.

Ark Invest added approximately $55 million of Nvidia, $26 million of Cerebras Systems, and $41 million of Broadcom in late August, reinforcing Cathie Wood's AI-infrastructure investment thesis. Ark forecasts AI spending will triple from $500 billion in 2025 to roughly $1.5 trillion by 2030, with Nvidia expected to retain a majority of the AI-server market despite increasing adoption of custom chips. Nvidia reported $96 billion in revenue, up 106% year over year, while Cerebras revenue rose 74% to $180 million and Broadcom AI semiconductor revenue increased more than 200% to $16 billion. The core risk is a slowdown in hyperscaler AI-capex spending, which would pressure all three companies.

Analysis

This is not a new fundamental catalyst; ARK flow is too small relative to NVDA and AVGO liquidity to alter institutional positioning. The actionable issue is the AI-compute mix: custom silicon displaces Nvidia primarily in mature, predictable internal workloads, while frontier training and heterogeneous enterprise demand retain a premium for Nvidia's software ecosystem and rapid product cadence. That makes a wholesale GPU-share-loss narrative premature, but it does cap NVDA's multiple if hyperscaler capex migrates from externally purchased GPUs toward internally designed accelerators.

AVGO is the cleaner expression of rising custom-ASIC penetration because it monetizes customer-specific designs without taking end-model risk, and its diversified infrastructure software base cushions a pause in accelerator orders. GOOG and AMZN receive a second-order benefit from lower inference unit costs and reduced dependence on a single merchant supplier; the value capture is more likely to appear in cloud gross-margin resilience than in near-term revenue acceleration. The key 1-3 month catalyst is hyperscaler capex commentary and any disclosed ASIC volume ramps; the 6-18 month determinant is whether custom-chip deployments expand beyond captive workloads without delaying model-development cycles.

CBRS is a high-beta technology option rather than a credible near-term share-taker. Its valuation requires sustained execution, customer diversification, and manufacturing/service capacity that are not established by benchmark claims; a single large-customer concentration, deployment delay, or gross-margin miss could drive material multiple compression. Consensus may be underestimating that inference economics can support multiple architectures, but is overestimating the speed at which enterprises can redesign software, procurement, and operations around a non-CUDA stack.

Key Decisions for Investors

• Prefer long AVGO over NVDA on a 6-12 month horizon: AVGO offers direct custom-silicon upside plus software diversification, while NVDA bears the greater risk of a mix-driven multiple reset. Size as a relative-value pair, long AVGO/short NVDA, only if subsequent earnings show ASIC AI revenue growth accelerating while NVDA data-center guidance decelerates.
• Maintain NVDA core exposure but do not chase ARK-related flow. Add only on post-results weakness if management preserves data-center growth guidance and networking/software attach remains resilient; reduce if two consecutive quarters show hyperscaler demand moderation or gross-margin pressure from product transitions.
• Use GOOG and AMZN as indirect beneficiaries of custom silicon rather than as pure chip substitutes. Accumulate on cloud-margin or capex-related drawdowns if management demonstrates inference-cost reduction without a corresponding slowdown in cloud demand; the thesis is falsified by rising depreciation materially outpacing cloud operating-profit growth.
• Treat CBRS as watchlist-only pending verifiable evidence of customer concentration, backlog conversion, gross margin, and funded manufacturing capacity. A small tactical long is appropriate only after earnings confirm durable growth and diversified customers; avoid shorting solely on valuation because low-float AI names can remain dislocated.