AI's Money-Burning Truth: Compute Eats 70% of Costs, Talent No Longer the Priciest

## Compute Is the 'Hard Currency' of the AI Arms Race  Epoch AI's latest data pulls back the curtain on AI companies' cash burn: at three leading firms—Anthropic, Minimax, and Zhipu AI—compute spending eats up 57% to 70% of total costs, dwarfing talent expenses. Anthropic's 2025 total spending hit $9.7 billion, with compute alone accounting for $6.8 billion—and that's just training and inference bills. On the surface, it looks like AI companies are hemorrhaging cash. But what really matters is this: **the core of AI competition has shifted from poaching talent to securing compute power.** Chips and computing infrastructure now carry more strategic weight on the balance sheet than engineers and scientists. ## The Compute Black Hole: Burn Rate 2-3x Revenue These three firms spend about 2 to 3 times their revenue, meaning the industry is still in deep loss mode. Zhipu AI's compute spending hit 58% of total costs, a classic R&D-heavy compute structure. Even with top AI engineers commanding astronomical salaries, talent costs haven't breached half of total spending. What does this mean? **Compute has become the heaviest financial burden for AI companies—and the key variable determining survival.** If you can't secure enough GPUs, you can't build the next model. If you can't lower compute costs, you'll never turn profitable. ## China vs. US Divergence: Open Source as a Cost-Cutting Play? Notably, Minimax and Zhipu open-source many model weights, while Anthropic stays closed-source. Open-source strategies lower the industry's entry barrier on the surface, but behind it may be a move born from compute cost pressure—better to trade models for ecosystem and influence than let them rot. For investors, **the open vs. closed source divide will directly shape compute demand distribution.** Open-source models boost inference compute demand, benefiting infrastructure providers; closed-source models lean on training compute, demanding higher chip performance. ## Three Signals Investors Must Watch First, **when will the compute cost inflection point arrive?** If a company's compute spending share starts declining while revenue growth holds, scale effects are kicking in. Conversely, if compute costs keep ballooning while revenue lags, that's a red flag. Second, **chip supply chain stability.** Compute is AI's oil—whoever gets choked loses. Watch Nvidia's capacity allocation, progress on domestic chip alternatives, and cloud providers' compute rental pricing trends. Third, **profitability model validation.** Every AI company is burning cash for scale, but the market will eventually reward those who can profit. When compute costs stop growing indefinitely and revenue starts covering expenses, that's the real investment opportunity. ## Conclusion: Compute Is Power, But Power Comes at a Price The AI arms race has entered a stage of deep pockets. Compute isn't everything, but without it, you have nothing. For crypto readers, this logic is familiar—just as Bitcoin's hashrate determines network security, AI's compute determines model capability. The story ahead won't be pretty: either compute costs come down, or companies go under. Investors need to track compute bills closely and place bets when the inflection point arrives.