Everyone wants to
build AI chips. Almost
everyone still buys
Nvidia's.
The chip is fast. The real moat is something you can't see on the box.
↓ scroll to read
✦ The bottom line
Rivals match Nvidia's chips on paper. They still struggle to take its customers. The moat is software, not silicon.
↓ the brief below
✦ Teach me
Cents kept per dollar of sales
Of every $1 paid for an Nvidia chip, the cents Nvidia keeps before paying for offices, R&D, marketing. Higher = customers pay up because they have nowhere better to go.
Wall Street calls this
Gross margin
75¢ on the dollar is *historically unusual* for a chipmaker. It tells you the moat is real, in number form.
The moat, measured · latest quarter
~75
¢
Of every dollar Nvidia sells, they keep about 75¢ before overhead. A typical chipmaker keeps closer to 40-50¢.
Source · 10-Q · Income Statement · Q1 FY27 · Filed May 20, 2026
How the moat got built
1990s
Nvidia builds gaming GPUs. Niche but lucrative.
2006
Launches CUDA — software that lets researchers program GPUs for anything.
2010s
Every AI researcher learns CUDA. The software lock-in builds, quietly.
Now
Switching to a rival chip means rewriting your AI. Most won't.
But moats *can* be attacked
Nvidia's biggest customers — Microsoft, Google, Amazon, Meta — are designing their own chips to escape paying Nvidia. Meanwhile Nvidia ships a new generation faster than rivals can catch the last one.
↳ A moving target moat. Each new chip generation extends the lead. The question is whether speed of iteration outpaces customer chip-design efforts.
Source · earnings-call · publicly disclosed customer chip-design efforts (Microsoft Maia, Google TPU, AWS Trainium, Meta MTIA) · Updated May 26, 2026
✓
Strong
75¢ per dollar. Software lock-in keeps the moat from leaking.