Two hundred and seventy-eight billion

A number leaked this week that is large enough to be quoted without context, which is how it will mostly be quoted. OpenAI expects to burn $278bn of cash between now and the end of 2030.

The figure comes from an internal presentation prepared in July for a computing deal, first reported by the Financial Times and confirmed by Bloomberg, as summarised by Tech Times. The same document projects roughly $856bn of infrastructure spending over the period against about $840bn of cumulative revenue, with annual revenue rising from some $36bn this year to $350bn in 2030.

Put the lines together and the shape is ordinary, even if the scale is not. A company expecting to spend slightly more than it earns while building the thing it earns from is a utility or a railway, not an anomaly. The burn is the gap between when the concrete is poured and when the meter runs.

What is not ordinary is the dependence on a revenue forecast that implies roughly a tenfold increase in four years. At $40bn of annualised revenue today, OpenAI is already one of the fastest-growing companies in history. The projection requires it to keep being that for the rest of the decade.

The financing consequence is the part investors will focus on. The $122bn raised in March, the largest private round ever assembled, is projected to be spent by around 2028 on these numbers. That is the arithmetic behind the talks reported last week at a valuation above $1.2tn: the round is not opportunism, it is the schedule.

It also explains the change in how the company buys capacity. The hunt for 20 to 30 megawatt sites alongside the gigawatt campuses is what a company does when its spending plan is fixed years ahead and its revenue depends on serving users this quarter. Small, already-powered buildings convert capital into revenue faster than a campus that energises in 2029.

Caveats belong on a document like this. It is a projection in a presentation prepared to support a deal, which is a genre with a direction. The costs in it are contractual and the revenues are hoped for, and the honest reading is that the spending is far more certain than the income.

The systemic question is who absorbs the gap if the revenue line disappoints. Much of the committed spending sits with counterparties — chipmakers, cloud providers, the developers of those campuses — who have built their own plans on those commitments, which is how a single company's forecast becomes an industry's exposure.

For now the market is treating the burn as the price of the position. The counter-argument is not that the number is small; it is that a firm with $40bn of annualised revenue growing at this rate has more ways to fund itself than a firm with a plan and no product, and OpenAI has a product that a billion people use.

The figure worth watching is not the $278bn. It is the $36bn, and whether it becomes the $350bn. Everything else in the document follows from that one line.