How to Raise $500B
Stargate, Export Controls & DeepSeek R1
In this week’s Trelis News:
OpenAI's $500B Stargate Data Center Plan
New GPU export controls limit H100s to 18 approved countries.
DeepSeek's R1 model
Cheers, Ronan
OpenAI's $500B Data Center Plan: Breaking Down the Stargate Initiative
OpenAI, Oracle, SoftBank, and Abu Dhabi's MGX announced the formation of Stargate, a new venture to build data centers supporting OpenAI's computing needs. The initiative represents a significant shift in OpenAI's infrastructure strategy and relationship with Microsoft.
Investment Structure and Timeline
- Initial deployment: $100 billion
- Total planned investment: $500 billion over 4 years
- Current OpenAI revenue context: Approximately $6 billion to date
- First site: Texas location (previously announced Oracle-OpenAI partnership)
Funding Commitments (est.)
- OpenAI: $19 billion
- SoftBank: $19 billion
- Oracle: Reported $7 billion potential commitment
- Oracle's current financial position:
- Cash on hand: $10 billion
- Total debt: $86 billion
Microsoft Relationship Details (est.)
- Existing Microsoft investment: ~$14 billion
- Current arrangement includes:
- Microsoft equity stake in OpenAI
- 20% revenue share of OpenAI profits
- API exclusivity agreement through 2030
Export Control Framework
Recent January 15th update establishes three tiers of countries for GPU exports:
- Tier 1 (18 countries): Permitted with exemptions
- Notable exclusions: Switzerland, Israel, Malaysia, Singapore
- Restricted nations: Include Russia and China
Restrictions apply to:
1. High-performance GPUs above specific compute thresholds
2. Large model as measured by FLOPS (except for open-source models)
DeepSeek Reasoning Model Performance
- Performance metrics:
- Comparable to o1-preview model on reasoning tasks
- ARC benchmark scores similar to lower-end o1 models
- Cost efficiency:
- V3 model: Less than $0.01 per task
- R1 model: Approximately $0.06 per task
- Training approach:
- Uses recursive improvement process
- Iterative optimization based on correct/incorrect answer scoring
- Results in increasingly detailed reasoning steps
Key question is what data was used and how that was generated.