NVIDIA's AI factories: what Chartered Accountants can learn
2 Min Read
NVIDIA published a post this month making a case that its chips should be treated as infrastructure. Not as products a company buys and writes off, but as long-lived assets you can lend against, the way banks lend against toll roads and power stations.
To back that up, it announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, aimed at mobilising more than $500 billion of third-party capital over time.
NVIDIA says it will provide residual value support on up to 25% of the financing, project by project. If the chips turn out to be worth less than expected when the loan matures, NVIDIA covers a slice of the shortfall.
Which means NVIDIA has attached its own money to an estimate that appears in somebody else's fixed asset note, and that somebody else gets to choose.
🧠 Fun fact: "Asset" comes from the Anglo-French assetz, meaning "enough". It was a legal term first: a deceased estate had assetz if there was enough in it to settle what the deceased owed. So for a couple of centuries the definition of an asset was simply whether it covered the debt against it. Feels topical.
A Quick Word From Today’s Partner
If you’re reading this the chances are that you’re probably on your way to being a chartered account, and you’re either:
Stuck in the books
Putting a shift in at articles
Recently qualified
At this stage in your career it’s important that you think about protecting the career and assets that you’re going to be building for yourself.
So today’s partner is PSG Insure - Bedfordview. They work with 1 000+ South African professionals and business owners, so whether you’re looking to be a corporate weapon, or start your own gig they’ve got your back on insurance.
Having a good broker in your corner goes a long way in protecting your lifestyle, by making sure you get the right advice and consequently the best insurance cover, as well as someone who fights for you at claim time.
Don’t leave the protection of what you’re working so hard to build until it’s too late.
Reach them here → https://advisers.psg.co.za/branch-office/bedfordview
What an AI factory is
Worth clearing up the vocabulary first.
An "AI factory" is NVIDIA's term for a data centre built to do one job. A normal data centre stores your files and runs your email. Whereas, an AI factory takes in electricity and produces model outputs, either training new AI models or answering queries from ones already built. What this looks like is rows of servers, each holding a stack of NVIDIA chips, running flat out.
Those chips get rented by the GPU-hour: you rent one chip, for one hour, at one price. NVIDIA quotes one-year H100 contracts moving from about $1.70 per GPU-hour in October 2025 to roughly $2.35 by March 2026, with the newer B200 fetching between $5.30 and $7.05. The H100 and B200 are just successive generations, the way phone models are. The A100 came in 2020, the H100 after it, and Blackwell chips like the B200 are the current run.
The businesses buying them at scale get called neoclouds. They buy enormous quantities of GPUs, put them in a building, and rent the capacity out. Their entire balance sheet is essentially one asset class, bought with borrowed money, and that's who these financing platforms are aimed at.
Nobody actually knows how long a GPU lasts
Here's the question we are all hoping someone else answers.
When NVIDIA releases a chip that's twice as fast, does the previous one become worthless? The instinct says yes, because that's how consumer technology works (and impairment in IAS 36). For instance, your ten-year-old laptop is just a doorstop now.
NVIDIA's argument is that this instinct is wrong, for reasons specific to how AI compute gets used.
Firstly is that older chips get demoted rather than discarded. Training a new model from scratch needs the fastest hardware available. But, running an already-trained model to answer a user's question, which the industry calls inference, needs far less. So last generation's chips slide down to the cheaper job, and there's plenty of that job going around.
The second is software because NVIDIA's chips run on CUDA, its own software layer, and NVIDIA keeps improving it. That means the same physical chip does more work in year four than it did in year one, without anyone touching the hardware. It's an unusual claim to make about plant and equipment, and it's doing a lot of work in the argument.
The third is that the chips are fungible. They can be pulled out and redeployed for a different customer in a different building, which is what makes a secondhand market possible at all. NVIDIA's exhibit is the A100: launched in 2020, still commercially deployed six years later, with some customers apparently planning around a ten-year economic life.
That's NVIDIA's case, published by NVIDIA, in support of NVIDIA's financing structure, which makes it advocacy rather than evidence.
Which makes this an accounting question
Now bring it back to something you can actually mark and think about in your lecture tomorrow.
Say a neocloud puts R18 billion of GPUs on its balance sheet, roughly a billion dollars. Depreciate that over three years and you're running R6 billion a year through profit or loss. Over six years, R3 billion. Same chips, same electricity bill, and a R3 billion swing in operating profit produced by nothing except a view on obsolescence.
The two levers compound, too. IAS 16 depreciates the depreciable amount, which is cost less residual value. So a company that believes GPUs hold their value gets a longer life and a higher residual, and both push the annual charge down. Assume the hardware is worth 20% at the end rather than nothing, and you've shaved a fifth off the charge before you've argued about the useful life at all.
IAS 16 even names the thing being argued about. Among the factors you must consider in setting useful life is technical or commercial obsolescence arising from changes or improvements in production (IAS 16.56(c)). Every new chip generation is a change or improvement in production. The standard has been asking NVIDIA's question all along.
So look at what the guarantee actually is. NVIDIA's money is at risk precisely when its customers' useful life estimates turn out to have been too generous. Those estimates are set by the customers and signed off by their own auditors, and what they are forecasting is obsolescence in a product line whose release schedule NVIDIA controls. That's the loop, and it's more interesting than the usual complaint about vendors funding their own sales.
The last piece is how a wrong estimate gets fixed. Useful life and residual value must be reviewed at least at each financial year end, and any change is a change in accounting estimate (IAS 16.51), applied prospectively (IAS 8.36). A company that spent four years depreciating GPUs over six and then concedes they were three-year assets reopens nothing. It writes the stub off over what's left and moves on. No restatement, and no comparative to explain.
That's the correct treatment as per the standard. However, it does mean the most consequential number in this entire trade can be wrong for years and then get resized without anyone calling it an error.
Insights From The Pocket CA
We've partnered with Pocket CA, an AI tool built for accountants, like your friend if they actually listened in lecture.
Here’s the technical detail in this scenario from the number-crunching machine:
IAS 16 Property, Plant and Equipment
The depreciable amount is cost less residual value, allocated systematically over useful life (IAS 16.50 and 16.53).
Residual value and useful life must be reviewed at least at each financial year end, and any change is a change in accounting estimate under IAS 8 (IAS 16.51).
Factors in determining useful life expressly include technical or commercial obsolescence arising from changes or improvements in production (IAS 16.56(c)).
Useful life is the asset's expected utility to that entity, and may be shorter than its economic life (IAS 16.57). NVIDIA's decade argument is about economic life, which is not the number the operator depreciates over.
IAS 8 Basis of Preparation of Financial Statements
A change in estimate is recognised prospectively, in the period of change and in future periods where it affects both (IAS 8.36).
Prospective recognition means the change applies from the date of the change onwards (IAS 8.38). Comparatives stay untouched.
IAS 37 Provisions, Contingent Liabilities and Contingent Assets
A provision only once there's a present obligation from a past event, a probable outflow and a reliable estimate (IAS 37.14), with "probable" meaning more likely than not (IAS 37.23).
Until then it's a contingent liability, disclosed but not recognised, unless the possibility of outflow is remote (IAS 37.27 and 37.86).
Worth knowing: IFRS 9's definition of a financial guarantee contract turns on a debtor failing to pay (IFRS 9 Appendix A). A promise about what an asset will fetch doesn't obviously fit there, which is how a very large obligation can live in the notes rather than on the balance sheet.
Check out the → Pocket CA
The Bottom Line
Strip out the headline number and what's left is a company betting on an accounting estimate.
The interesting part is that NVIDIA has taken a position which costs it nothing so long as its customers' depreciation assumptions hold, in a framework where getting those assumptions wrong is corrected prospectively and never called a mistake. Everyone involved is applying the standards properly, which is exactly why it's worth looking at rather than complaining about.
If you're in articles anywhere near heavy technology capex, the version of this on your desk is simpler than it looks. Find the useful lives disclosure, then ask when it last changed and what evidence supported the change.
Until next week,
The Journal Entry Team
If you enjoy our content we’d love it if you showed your friends and peers!
Please copy and paste this link to others:
