Starting Our Own OpenAI: A $5.6M Model, a $34B Yearly Company, and Training Is the Cheapest Line

The article outlines a plan to start an AI company by replicating DeepSeek-V3's training cost of $5.576 million, but quickly reveals that this figure excludes prior research, ablations, data preparation, salaries, and hardware. It cites OpenAI's audited 2025 financials: $13.07 billion revenue, $34 billion total expenses, and a $20.92 billion operating loss, with $17.2 billion paid to Microsoft for cloud services. The author calculates that a year-long cluster for 2,048 H800 GPUs costs $42.2 million, while the final run uses only 15.5% of the annual GPU hours. Larger runs like Grok 4 are estimated at $490 million by Epoch AI, but using Lambda's public prices for interconnects, the cost exceeds $1.3 billion. On-demand capacity is sold out, and long-term contracts are take-or-pay. Purchasing hardware is also expensive, with GB300 NVL72 racks costing $6-6.5 million, consuming up to 155 kW, and requiring liquid cooling, which standard data centers cannot provide.