Goldman Sachs Predicts AI Infrastructure Spending Could Hit More Than $1 Trillion in 2027. 3 AI Stocks to Buy.

AI infrastructure capex is projected to exceed $700 billion this year and could rise to $920 billion-$1.4 trillion next year, underscoring continued demand across the AI supply chain. The article highlights Nvidia, AMD, and Micron as beneficiaries: Nvidia remains dominant in training and expanding into inference/agentic AI, AMD is positioned for inference and data center CPU growth, and Micron is seeing record DRAM-driven results amid undersupplied memory markets. The piece is broadly constructive on AI hardware names, though it is primarily an analyst commentary rather than new company-specific news.

Analysis

The market is still pricing AI capex as a single-wave GPU trade, but the next leg is a mix-shift story: training remains compute-heavy and winner-take-most, while inference and agentic workloads pull spend toward memory density, CPU orchestration, and software/compiler stack efficiency. That broadens the beneficiary set, but it also compresses moats in the most crowded layer because customers will increasingly dual-source around deployment economics rather than performance bragging rights.

NVDA still has the cleanest ecosystem lock-in, but the incremental upside is less about headline GPU share and more about monetizing the full rack: networking, CPUs, low-latency inference silicon, and software attach. The second-order risk is that as workloads mature, buyers push harder on total cost per token, which should force faster pricing discipline across the supply chain and could narrow the gross-margin spread between the incumbent and fast followers over the next 12-24 months.

AMD is the more interesting beta to the spend cycle because its upside is tied to broader deployment breadth rather than just frontier model training. If inference and agentic AI scale as expected, the key variable becomes memory capacity per package and CPU content per cluster, which favors AMD’s architecture but also gives system integrators and memory vendors bargaining leverage; that makes the trade less about one vendor winning and more about a multi-node capex expansion. MICRON is the cleanest way to express the scarcity angle, but the risk is that memory is the first place buyers ration if capex budgets get questioned, so the stock should be treated as a late-cycle beneficiary even if fundamentals remain strong.

The contrarian miss is that enthusiasm may be underestimating how quickly supply comes online in HBM and DRAM once pricing stays elevated for several quarters. If utilization normalizes into 2026, memory margins can mean-revert faster than most expect, while GPU/CPU demand may prove stickier because software switching costs are higher. That argues for owning the platform names on weakness and using memory strength as a tactical trade rather than a permanent core position.

Key Decisions for Investors

• Maintain core long NVDA into any 5-8% pullback over the next 1-2 months; favor the stock as the lowest-risk expression of AI capex, but trim if the forward multiple re-rates above the low-20s on 2028 estimates.
• Initiate a 3-6 month long AMD / short NVDA pair trade for investors seeking relative upside from inference and agentic AI adoption; AMD has more room for estimate revision, while NVDA is closer to consensus saturation.
• Go long MU on 2-4 week weakness, with a 6-12 month horizon; use a tight stop if DRAM pricing signals roll over, since the stock is most exposed to any pause in hyperscaler memory ordering.
• Consider a basket long of NVDA + AMD + MU sized as a thematic trade, but cap MU at half the weight of NVDA/AMD because memory has the shortest duration risk if capex guidance disappoints.
• For options, buy 6-9 month AMD call spreads rather than outright calls to express upside from inference/agentic share gains while limiting valuation and execution risk.