Mark Zuckerberg, is gracing the podcast circuit to extol the virtues of open-source LLMs and Facebook’s noble quest to democratize AI. Color me impressed – they’ve actually released this highly valuable intellectual property into the wild. My initial reaction? A rush of excitement, a mad dash to spin up a machine and explore this cutting-edge model firsthand.
Reality, of course, soon set in. The hardware and infrastructure costs are significant. Undeterred, I pondered the possibility of capitalising on this AI goldmine as a new business venture. However, the path to monetisation is complex, and the potential rewards are uncertain. It would require someone with a high risk tolerance to leap into this level of the AI space – essentially total reliance on a third-party system that will be outdated quickly.
Let’s talk numbers, shall we? To deploy your very own private instance of Llama 3.1:405B, you’re looking at a cool $100K in hardware alone. Yes, the next model iteration will likely arrive sooner than later, but this rapid pace of innovation also presents opportunities for early adopters to stay ahead of the curve. It’s enough to make one’s head spin with possibilities.
So, the question remains: is this open-source AI model truly democratising access? The answer is nuanced. While it may not be immediately accessible to everyone, it lowers the barrier to entry for businesses and individuals with the resources to explore and innovate. The democratisation of AI is a journey, not a destination, and open-source models like this are an expensive step in the right direction.
Here is a back of the napkin breakdown of what the costs would look like (all costs are estimates from a quick google search):
Specifications
| Component | Details |
|---|---|
| GPUs | Type: NVIDIA A100 or equivalent Number: At least 8 GPUs VRAM: 80 GB per GPU |
| CPU | Type: AMD EPYC or Intel Xeon Cores: 64+ cores Clock Speed: 2.5 GHz or higher |
| Memory (RAM) | Size: 1 TB or more |
| Storage | Type: NVMe SSDs Capacity: 10 TB or more |
| Networking | Type: High-speed interconnect (InfiniBand) Speed: 100 Gbps or higher |
| Power Supply | Wattage: At least 5000W, preferably with redundant power supplies |
| Cooling | Type: Advanced liquid cooling or equivalent Capacity: Sufficient for high thermal output |
Cost Estimates
Hardware Costs
| Component | Cost Details | Total Cost |
|---|---|---|
| GPUs | Cost per GPU: $10,000 (NVIDIA A100) | $80,000 (8 GPUs) |
| CPU | Cost per CPU: $6,000 (AMD EPYC 7763) | $12,000 (2 CPUs) |
| Memory (RAM) | Cost per TB: $5,000 | $5,000 (1 TB) |
| Storage | Cost per TB: $300 | $3,000 (10 TB) |
| Networking | Cost: $10,000 | $10,000 |
| Power Supply | Cost: $2,000 | $2,000 |
| Cooling | Cost: $5,000 | $5,000 |
Additional Costs
| Component | Cost |
|---|---|
| Chassis/Case | $3,000 |
| Motherboard and Components | $5,000 |
| Cabling and Miscellaneous | $2,000 |
Total Estimated Cost
| Category | Estimated Cost |
|---|---|
| Hardware | $127,000 |
Running Costs
| Component | Cost Details | Monthly Cost | Annual Cost |
|---|---|---|---|
| Electricity | Power Consumption: 5000W Cost per kWh: $0.10 |
~$360 (24/7) | - |
| Maintenance | - | - | $10,000 |
| Software Licenses | Cost: Variable | - | - |
Summary
| Category | Cost |
|---|---|
| Initial Setup Cost | Approximately $127,000 |
| Monthly Running Cost | ~$360 (electricity only) |
| Annual Maintenance Cost | $10,000 |