Concentration of Computing Power
Two AI companies, OpenAI and Anthropic, now command one-third of all newly installed global computing power. This trend shows no sign of slowing. By next year, the share held by these two entities is projected to reach fifty percent. Projections from SemiAnalysis founder Dylan Patel suggest that by the end of 2028, these two firms could hold the vast majority of effective available computing power on the planet.
Patel shared these findings on the Dwarkesh Podcast this August. His analysis indicates a rapid escalation in consumption. Anthropic and OpenAI will each exceed five gigawatts of power usage by year-end. This is a massive jump from their starting position at the beginning of the year, when each held less than two gigawatts. These companies are effectively consuming the total output of three large nuclear power plants annually to sustain their current operations.
The Economics of High-End Infrastructure
Profitability dictates which companies secure the most hardware. A year ago, the token business operated at a loss. That has changed. Anthropic and OpenAI now report positive margins on their models. Because they generate high revenue per megawatt, they can outbid competitors for access to data centers and hardware. This creates a positive feedback loop. High profits buy more chips, which build stronger models, which drive even higher profits.
Cloud providers act as landlords in this arrangement. They build the facilities while OpenAI and Anthropic lease the space. Some players have entered the market to build facilities without waiting for client contracts. Companies like SpaceX, led by Elon Musk, possess the capital to build infrastructure first and lease it to the highest bidder. Reports indicate Anthropic currently leases significant data center capacity from SpaceX at a cost of 1.25 billion dollars every month.
Research Priorities and Financial Implications
Most of this power does not go toward running models for users. Roughly fifty percent of the allocated computing power is reserved for research. Only about forty percent is used for inference. Even with gigawatts of power available, engineers struggle to make all hardware work in unison. Data transmission speeds often limit the effectiveness of massive clusters. This leads to research teams focusing on new architectures rather than just training larger models.
Investors face a dilemma as these companies move toward public offerings. The labs prioritize AGI research over dividends. Dylan Patel notes that even as revenue growth flattens, the purchase of new computing power continues to climb. This capital expenditure is aggressive. Cumulative spending from 2024 to 2029 is estimated at 11 trillion dollars. About 5 trillion of that total will likely require debt financing, which could increase interest rates across the broader market.
Regulation and Future Obstacles
Governmental bodies have started to react to this expansion. Local authorities in New York, Texas, and Ohio have introduced measures to restrict or tax the growth of data centers. These physical limitations may force companies to search for energy-rich regions outside of current hotspots. Still, the demand for power remains persistent. The race for AGI continues to drive these companies to secure resources regardless of the cost or local regulatory pushback.