Nvidia, SK Hynix, Broadcom: Top AI Data Centre Buys as $300B+ CapEx Cycle Builds Momentum

The Quick Take

• Nvidia, SK Hynix, and Broadcom named top AI infrastructure buys as data centre spending set to surge
• Hyperscale CapEx commitments from Amazon, Microsoft, and Google underpin a multi-year GPU and HBM demand cycle
• Analysts cite data centre infrastructure as a 'secular, not cyclical' growth story with 30-40% YoY spend growth

With hyperscale cloud providers collectively committing over \$300 billion in data centre capital expenditure for 2026 and 2027, analysts have converged on three primary beneficiaries in the AI infrastructure stack: Nvidia (NVDA) for GPU compute, SK Hynix for high-bandwidth memory (HBM3E), and Broadcom (AVGO) for custom AI accelerators and networking silicon. Both articles in this cluster argue that the buildout cycle is secular rather than cyclical — driven by structural AI adoption across enterprise, government, and consumer verticals — and that the 30-40% YoY growth in data centre spending represents a multi-year investment thesis rather than a one-cycle surge.

Nvidia's dominance in training accelerators is well-documented, but the articles highlight that inference demand — which is longer-tail and more volume-sensitive than training — is the next leg of Nvidia's growth story. As AI models are deployed at scale in production environments, inference chips need to be provisioned per-request rather than per-training-run, creating a multiplicative demand dynamic. SK Hynix benefits because HBM (high-bandwidth memory) is a non-substitutable component of Nvidia's GPU architecture; with HBM3E supply tight and HBM4 qualification cycles starting, SK Hynix maintains pricing power through at least FY2027. Broadcom's angle is custom silicon: hyperscalers increasingly want proprietary AI accelerators (Google TPU, Amazon Trainium) built on Broadcom's ASIC architecture.

For investors, the risk in the data centre AI trade is timing the capex cycle's peak. Prior semiconductor supercycles (DRAM in 2017-2018, NAND in 2021-2022) saw brutal corrections when supply caught up with demand. The AI cycle's advocates argue it is different because demand is software-driven and open-ended — the more AI models are deployed, the more compute they generate demand for — rather than being tied to a single end-market like smartphones or PCs. The counter-argument is that hyperscaler CapEx is discretionary, and CFOs facing slowing cloud revenue growth could trim data centre budgets faster than the market currently prices.

Synthesized from 2 sources.