The Transformer paper (2017) sparks GPT-1, GPT-2, GPT-3. Hyperscalers race to build the largest AI supercomputers in history.
Hyperscalers poured hundreds of billions into data centers, custom silicon, and high-speed networking. Infrastructure became the dominant AI investment theme — before any product shipped.
| Ticker | Company | Thesis |
|---|---|---|
| NVDA | NVIDIA Corp | H100 era begins; $40K GPU demand |
| MSFT | Microsoft | Azure AI + $10B OpenAI investment |
| GOOGL | Alphabet | TPU custom silicon + GCP + DeepMind |
| AMZN | Amazon | AWS dominance + Trainium/Inferentia chips |
| TSM | TSMC | 100% of leading-edge AI chips manufactured here |
| ASML | ASML Holding | No EUV = no advanced AI chips. Period. |
| AVGO | Broadcom | Custom AI ASICs for hyperscalers; path to $100B AI rev |
| MRVL | Marvell Technology | AI networking + custom silicon for hyperscalers |
| DELL | Dell Technologies | AI server supercycle; enterprise deployment |
| MU | Micron Technology | HBM memory — you can't train LLMs without it |
| ARM | Arm Holdings | CPU architecture powering every AI edge device, Apple Silicon, Qualcomm, NVIDIA Grace; royalty model scales with AI silicon shipments |
| TSEM | Tower Semiconductor | Specialty analog/mixed-signal foundry; silicon photonics production node; acquired by Intel attempt blocked, now independent |
| GFS | GlobalFoundries | Mature node foundry for secure/defense chips; U.S.-based manufacturing beneficiary of CHIPS Act; less AI-direct but strategic |
Educational framework, not financial advice. Tickers and theses are illustrative of each phase's investment thesis, not recommendations.