The AI Supercycle Isn't Over -- It Just Moved. 5 Growth Stocks Riding the Next Wave.

Key event: Anthropic placed a $21 billion order for TPUs and Alphabet's TPU-based servers are estimated to deliver a ~44% lower total cost of ownership versus equivalent Nvidia GPU servers, strengthening Alphabet and Broadcom in inference ASICs. Broadcom is positioned as a leader in data-center networking (Ethernet switches, optical interconnects) and ASICs; Arista should benefit by packaging Broadcom components with its EOS platform; AMD and Arm are expected to gain from rising demand for data-center CPUs as agentic AI grows, with Arm forecasting a $100 billion data-center CPU market in five years and targeting a 15% share.

Analysis

The market rotation toward inference and agentic workloads re-prices the relevant supply chain: value shifts from raw accelerator TFLOPS to throughput-per-dollar across a multi-component stack (switches, optics, ASICs, CPUs, software). That creates longer, stickier purchase cycles for vendors that can sell integrated subsystems and recurring software/diagnostic licences, and creates an opportunity for component suppliers to extract margin via constrained optical and board-level supply.

A key second-order dynamic is hyperscaler procurement strategy: the more vertically integrated their stacks become, the greater the risk of displacing third-party incumbents downstream (component vendors that lack software lock-in) but the larger the prize for firms that supply both silicon and orchestration layers. Concurrently, CPU demand for agentic workloads shifts server bill-of-materials toward memory bandwidth, I/O, and board engineering, favoring vendors with footprint scale in those domains and penalizing low-margin, commodity suppliers.

Main risks are non-linear: rapid software advances (quantization, compiler optimization) or a meaningful reduction in inference per-transaction compute could compress addressable demand within 6–18 months, while export controls or optical-component shortages could accelerate pricing power for suppliers. Watch for discrete catalysts over the next 3–24 months—hyperscaler capex cadence, announced design wins, ASIC-to-GPU replacement ratios, and memory/I/O benchmarks—as these will drive durable revenue re-rating or a swift reversal.

Key Decisions for Investors

• Long AVGO (12–24 months): 3–5% portfolio weight as a core exposure to the higher-margin components and firmware stack; target +25–40% on successful hyperscaler refresh cycles, stop-loss 12–15% below entry. Rationale: asymmetric payoff if component ASPs re-rate with limited cyclic supply.
• Pair trade — Long ANET / Short NVDA (6–12 months): buy ANET equity or 12-month calls (size 2–3%) and finance with a short NVDA call spread (sell 1–2% notional 9–12 month near-ATM calls, buy further OTM calls as hedge). Risk/reward: protects against a GPU rebound while capturing networking share gains; max loss capped by the call spread premium, upside >30% in ANET if large-scale switches roll out.
• Long ARM + AMD (18–36 months): equal-weight positions totaling 4–6% portfolio to play CPU-led capacity growth for agentic AI. Use long-dated calls on ARM (24–36 months) to amplify upside; risk: semiconductor design-cycle slippage or licensing pushback could impair returns—limit size and use 20% trailing stop.
• Options hedge idea — Sell NVDA call premium (3–6 months) to monetize high gamma ahead of potential demand-normalization; keep position size small (≤1% portfolio) and buy protective puts 20–30% OTM to cap tail risk. Reward: collect premium if GPU replacement slows; risk: fast GPU re-acceleration could produce sharp losses without proper hedges.