OpenAI's Mounting Losses Are the Toll of Building the Frontier, Not Its Obituary

According to internal documents cited in this report, OpenAI lost roughly $5 billion in 2024 and is projected to lose around $38 billion in 2025, with the elevated cost of training AI models named as the primary driver. We present these as figures attributed to leaked internal material, not as audited results.

The context matters. Training frontier models is one of the most capital-intensive activities in modern technology: compute, energy, data and talent all compound, and each new generation tends to cost more than the last. Losses of this scale are less a sign of a broken business than of an industry in a heavy build-out phase, where companies are paying upfront for capability that they hope to monetize later across products, enterprise contracts and infrastructure.

The short-term risk is real and worth naming. Spending tens of billions ahead of revenue concentrates power in whoever can keep writing those checks, raises the stakes of every strategic bet, and makes the whole sector vulnerable to a shift in investor patience. If the gap between burn and income doesn't narrow, the pressure to recoup costs could distort priorities toward whatever monetizes fastest.

Our reading: these losses are the price of admission to the frontier, not yet evidence that the frontier doesn't pay. The decisive variable is the trajectory — whether training efficiency improves and revenue scales faster than costs. If it does, today's deficits become the down payment on tools that compress drug discovery, scientific research and productivity, the long-term abundance we keep pointing to. If it doesn't, the market will force a brutal reckoning. The honest position is to treat a number like $38 billion as a measure of ambition and risk simultaneously, and to watch the slope, not the headline.