Read the bridge between AI operating loss and adjusted EBITDA

SpaceX's AI segment reported a $1.257 billion operating loss and $1.146 billion of adjusted EBITDA for the second quarter of 2026. Both figures appear in the same earnings release. They describe different measures, and the company supplies the bridge between them.

The reconciliation on page six adds back $1.885 billion of depreciation and amortisation, $516 million of share-based compensation and $2 million of restructuring costs. In millions, the arithmetic is −1,257 + 1,885 + 516 + 2 = 1,146.

That bridge is where I would start an AI infrastructure discussion. Picking whichever endpoint supports an optimistic or pessimistic story throws away information the filing has already provided.

Separate the questions before interpreting the answers

Operating income asks about a period's operating result under the reported accounting treatment. Adjusted EBITDA removes specified items. A reader can reasonably find both useful, provided the adjustments stay visible.

Depreciation and amortisation are the largest part of this particular bridge. The table does not identify that entire amount as GPU depreciation. It would be a mistake to give a broad accounting category a more specific hardware label simply because GPUs dominate the story around AI.

My interpretation is narrower: the difference is large enough that an evaluation of the business must explain how it treats the cost of assets and compensation. Calling the segment profitable without naming the metric leaves that explanation unfinished. Calling the adjustment worthless would be equally unhelpful.

As an illustration, I would ask an engineering and finance team to review a hypothetical cluster proposal in three adjacent views. One shows operating expense, including the relevant depreciation assumptions. Another shows the cash payments over the contract and construction schedule. The third shows the operating metric management proposes to monitor.

The useful discussion happens where those views diverge. A payment can occur before the asset contributes to service. An accounting expense can continue after the original cash outlay. Neither timing difference automatically tells us whether the workload earns an adequate return.

A segment result is not a funding trace

The release also reports $15.828 billion of AI capital expenditure out of $18.369 billion in total. Connectivity generated operating income, while Space reported an operating loss. AI was therefore not the only segment with an operating loss.

It is tempting to connect a profitable segment directly to another segment's capital spending and call that the funding model. A segment income table does not establish those cash movements. Financing, group cash and the timing of expenditure need their own evidence.

For a practical reading exercise, take any AI operator's results and annotate each headline number with its unit, period and accounting measure. Put capital expenditure in a separate row from operating expense. Then find the reconciliation for every adjusted measure used in the argument.

The remaining uncertainty becomes useful work. Which costs recur? What assumptions determine the asset's useful life? How does utilisation affect the investment case? Those questions can be investigated without pretending that one quarter settles the economics of an entire buildout.

I would rather leave a funding arrow undrawn than draw it from an income statement that does not support it. The missing evidence is a reason to inspect cash flows and financing disclosures, not a licence to fill the gap with a persuasive narrative.

For an AI buildout, read the accounting bridge before choosing the profitability headline.