Peter H. Diamandis
August 28, 2026
Why Companies Lock in Older GPUs: The A100 Economics Through 2029
TL;DR
Companies are contracting for 6-year-old A100 GPUs through 2029 because their marginal costs drop to just electricity once hardware is paid off, while newer efficient models allow 30 smaller models to run on one chip instead of requiring 16.
Key Takeaways
- 1Core clients have signed contracts for A100 GPUs extending to 2029, locking in 6-year-old hardware that Nvidia introduced in 2020
- 2Once A100 hardware costs are fully paid, the only marginal cost becomes electricity, making older paid-off chips economical for constant workloads
- 3A100 efficiency has dramatically improved: GPT-4 required 16 A100s in 2022, but now 30 smaller 5-10 billion parameter models can fit on a single A100 and outperform it
- 4All new models run on CUDA, making A100s compatible with both current and future software, keeping older hardware viable longer
- 5Companies are financing GPU capacity before the next generation chip breakthroughs, similar to timing an IPO before new competition emerges
Notable Quotes
“Once you've paid off the initial bulk order of it, it's about electricity turning into intelligence. That's your marginal cost.”
“Now you can have literally a 10 billion parameter model or a 5 billion parameter model that outperforms that in terms of you can fit like 30 of those on one chip rather than needing 16 chips.”
“It's before the next generation chips. It's before the chip breakthroughs just like it's now a great time for Anthropic to come to IPO before Grock comes and takes their lunch.”
Chapters
1. The A100 Contract Surprise
Core announced that some clients have contracted for A100 GPUs through 2029, which is remarkable because the A100 was introduced in 2020, making it a 6-year-old chip with 40-80 GB of RAM configurations.
2. The Paid Hardware Economics
Once A100 hardware costs are fully paid off through bulk orders, the remaining marginal cost becomes only electricity, making the chips economical for consistent workloads spanning multiple years.
3. Improving Model Efficiency
Model efficiency has improved dramatically since 2022 when GPT-4 required 16 A100s to train; now 30 smaller 5-10 billion parameter models can fit on a single A100 and outperform the older generation.
4. CUDA Compatibility and Cycle Timing
All new models run on CUDA architecture, ensuring A100s work with current and future software; companies are also timing GPU financing before the next generation chip breakthroughs, similar to IPO timing strategies.
Key People & Entities
- Nvidia
- GPU manufacturer; introduced the A100 in 2020 and created CUDA architecture
- Jensen
- Nvidia executive mentioned in connection with GPU architecture decisions
- Core
- Company that announced client contracts for A100 GPUs through 2029
- Anthropic
- AI company used as example of timing IPO before competitive breakthroughs
Glossary
- A100
- An Nvidia GPU introduced in 2020 with 40-80 GB of RAM per chip, designed for large-scale AI model training and inference
- CUDA
- Nvidia's parallel computing architecture that allows software to run across different generations of Nvidia GPUs
- Parameter model
- An AI model defined by the number of learnable weights; a 10 billion parameter model has 10 billion adjustable values
- Marginal cost
- The additional cost of producing one more unit after fixed costs have already been paid; in this case, electricity for running paid-off GPUs