NVDA

Walk into a hyperscaler data center — a building the size of several football fields, humming with tens of thousands of servers. Open one of the server trays. Inside, you will see between four and eight large green circuit boards, each with a GPU chip roughly the size of a postage stamp mounted under a massive heatsink. Those are NVIDIA GPUs — the chips that do the actual mathematical work of training AI models.

Each GPU is connected to stacks of HBM memory (High Bandwidth Memory — fast memory chips stacked vertically and bonded directly onto the GPU package, so data can move between memory and processor without crossing a circuit board). A single NVIDIA Blackwell GPU package is actually two GPU dies and eight HBM memory stacks fused together on a silicon interposer — a precision silicon base plate the size of a dinner plate, manufactured by TSMC using their CoWoS advanced packaging process.

But the GPU is only half the story. Between servers, NVIDIA's NVLink cables connect GPUs directly to each other at 7,200 gigabits per second — roughly 18 times faster than the fastest standard Ethernet connection. In a Blackwell NVL72 rack, 72 GPUs are wired together so tightly that the AI software treats them as one enormous processor. That rack-scale integration is what makes NVIDIA different from a company that just sells chips — they sell the entire compute fabric.

The reason NVIDIA went from a $400 billion company to a $4.5 trillion company in three years comes down to one architectural shift: AI models got big enough that the only way to train them is to wire thousands of GPUs together and treat them as one machine. That is exactly what NVIDIA's full-stack approach — GPU silicon + NVLink fabric + CUDA software — is built to do. The Blackwell generation entered full production in late 2025 and is sold out through mid-2026. The next generation, Rubin (6 new chips unveiled at CES 2026, first samples shipped to customers in February 2026), enters production in H2 2026. Analyst expectations for 2026 capex across the top 5 cloud providers and hyperscalers now approach $700 billion — and NVIDIA is the default recipient of that spend.