NVIDIA

Origins & Founding (1993)

• Jensen Huang, Curtis Priem, and Chris Malachowsky founded NVIDIA over a Denny’s breakfast
  • All three had prior experience at Sun Microsystems and SGI
  • Huang had a singular vision: a chip that could accelerate graphics computation in software-programmable ways
• The name came from the Latin root for “envy” — they wanted the company to be envied
• First office: a rented space in Sunnyvale with $40,000 in seed funding from Sequoia Capital

The GPU Bet

• NVIDIA coined the term “GPU” (Graphics Processing Unit) with the launch of the GeForce 256 in 1999
  • The GeForce 256 could perform 10 million polygons per second
  • Transform and lighting (T&L) moved from CPU to GPU — a fundamental architectural shift
• The real breakthrough: parallel processing
  • Where a CPU has 8–64 cores, a GPU has thousands of smaller cores
  • Perfect for matrix math — the foundation of modern AI

Near-Death Experiences

• 2002: The dot-com bust nearly killed the company
  • Revenue collapsed; headcount cut from 800 to 550
  • Jensen chose not to lay off engineers — a bet on the coming ramp
• 2008: Financial crisis hit hard
  • Took a $196M charge on defective chips in HP laptops (MCP chip failure)
  • Required another painful restructuring
• 2018–2020: Crypto bust + Mellanox acquisition concerns
  • GPU demand collapsed after the crypto mining craze ended
  • The $7B Mellanox acquisition looked risky at the time — became transformative

CUDA — The Real Moat

• Launched in 2006, CUDA let developers write general-purpose code for NVIDIA GPUs
  • The key insight: researchers and scientists, not just game developers, could use GPU compute
  • CUDA is written in C — dramatically lowering the barrier to entry
• The flywheel effect
  • More developers → more CUDA libraries → deeper lock-in → more enterprise sales
  • Competitors (AMD ROCm, Intel oneAPI) have never caught up on the software side
• By 2023, most AI researchers had never written code for anything but CUDA

The AI Inflection Point (2012–present)

• 2012: AlexNet wins ImageNet using two NVIDIA GTX 580 GPUs
  • Hinton’s lab proved deep learning worked at scale on GPU hardware
  • Orders for research GPUs started flowing into NVIDIA
• 2016: Jensen shifts the entire company toward AI/data center
  • Re-architects the A100, H100, and GB200 specifically for transformer workloads
  • Data center revenue goes from $0 to $47B (FY2024)
• 2023: ChatGPT drives H100 demand beyond anything in semiconductor history
  • Lead times of 12+ months; grey-market H100s selling at 3× list price
  • Jensen calls the moment “the iPhone moment of AI”

Jensen Huang’s Leadership Philosophy

• Deliberately runs NVIDIA as a “no middle management” company
  • ~60 direct reports — unheard of in large tech firms
  • Forces high-bandwidth information flow; kills bureaucratic filtering
• The “one-on-one in public” culture
  • Critical feedback happens in group settings — everyone learns simultaneously
  • Uncomfortable for new hires; deeply effective for company alignment
• “Intellectual honesty over harmony”
  • Famous for killing projects that aren’t working, quickly and openly
  • The Omniverse pivot, the ARM deal cancellation — both handled with speed

Competitive Moat Analysis

• Hardware: Leading-edge process nodes (TSMC 4N and 3N), HBM memory integration, NVLink interconnects
• Software: CUDA ecosystem with 4M+ developers, cuDNN, TensorRT, Triton inference server
• System-level: DGX systems, NVLink Switch, Quantum InfiniBand — full-stack vertical integration
• The real moat: 15 years of CUDA developer lock-in is nearly impossible to replicate

Financials & Scale (as of late 2023)

• Market cap crossed $1 trillion in 2023 — faster than any company in history
• Gross margin: ~74%, among the highest for a hardware company ever
• Data center segment now >75% of total revenue
• The H100: ~$30,000 list price, estimated $10,000–$12,000 COGS → extraordinary economics