AI's Real Cost Is in the Balance Sheet

AI infrastructure conversations still start with chips and megawatts. They should start with a balance sheet. The capex is booked as an asset. The cash is already gone. That gap is now the bottleneck, and it is larger than the power problem.

The setup is simple. A hyperscaler spends tens of billions on GPUs, buildings, and power contracts. Accounting treats most of that as an asset. The cash left the account the day the invoice cleared. Over the next three to five years, depreciation walks the asset off the books. Until then, the income statement looks healthier than the cash flow statement, and the capital raising conversation is priced on the prettier number.

The asset that is really an expense

A GPU is not a factory. A factory produces the same widget for twenty years. A GPU cluster is a wasting asset with a three to five year useful life, and the useful life is shrinking as models and chips turn over. Booking it as property, plant and equipment is legal. Treating it as a long-lived productive asset in a capital raising deck is a story.

Cash flow from operations can look fine while free cash flow is deeply negative. The difference is capex. Capex that has to be repeated every cycle is not growth capex. It is maintenance of the franchise. Call it what it is: an operating cost with a delayed receipt.

This is why the 401(k) flood I wrote about last week matters: https://brawlersguide.com/post/401k-flood-capital-raising-private-markets. Retirement capital looking for contracted infrastructure yield is not the same as equity looking for GPU residual value. One of those stories survives a fiduciary checklist. The other needs a depreciation schedule an auditor will defend.

Depreciation is the delayed confession

Three to five years. That is the window. Nvidia's data centre GPUs are being depreciated on accelerated schedules at several of the large buyers. The cash went out in year zero. The expense hits later. If utilisation, pricing, or model demand misses, you still have the depreciation. You just do not have the revenue that was supposed to cover it.

The power procurement bottleneck I covered earlier is the physical version of the same problem: https://brawlersguide.com/post/ai-infrastructure-power-procurement-bottleneck. You can own the chips and still not run them. An idle GPU depreciates at the same rate as a busy one. The balance sheet does not care about your utilisation story.

The capital stack has to change

Equity was the first answer because the returns looked venture-like. They are not. The cash conversion cycle looks more like a utility with a wasting plant. That is project finance, asset-backed lending, and contracted offtake, not a growth equity round dressed up as infrastructure.

The firms that raise well from here will sell a depreciation schedule, a contracted megawatt, and a residual value an auditor will sign. The firms that raise badly will sell a TAM slide. The 401(k) channel, the evergreen, and the CIT are built for the first story. They are poison for the second.

What the next five years actually look like

Capex does not slow. The depreciation wall arrives anyway. Some of the current build will earn its keep. Some of it will be a stranded asset with a nice logo on the door. The difference will show up in cash conversion, not in press releases.

Watch three numbers: useful life assumptions in the 10-K, cash capex versus depreciation, and contracted offtake as a share of nameplate. If useful life is stretching while chip cycles are shortening, the confession is being delayed. If cash capex stays multiples of depreciation, the asset base is still being built faster than it is being earned. If contracted offtake is thin, you are holding residual value risk and calling it infrastructure.

The honest list

Three things can still break this. First, a demand miss. Training clusters that do not fill inference seats still depreciate. Second, a rate or credit event that shuts the cheap capital that funded the build. Third, accounting catch-up: if useful lives get shortened in a single reporting cycle, the income statement takes a hit that the cash already took years ago.

The play is not to stop building. It is to stop pretending the build is an asset in the way a toll road is an asset. Structure the capital to match the life. Match the offtake to the depreciation. Tell the truth in the deck before the auditor tells it in the notes. The bottleneck moved. It is not in the substation anymore. It is on page two of the cash flow statement.