Depreciation schedules do not often assume strategic importance, but for neocloud suppliers of artificial intelligence “compute as a service,” depreciation of graphics processing units does seem to matter.
Assume a $12 billion investment in GPUs. If one assumes a three-year depreciation cycle, that produces a $4 billion per year hit to earnings.
On the other hand, if one assumes a longer six-year cycle, annual depreciation is just $2 billion per year.
The danger is the balance sheet hit if real-world useful life turns out to be less than six years.
Microsoft, Google, and Amazon arguably can absorb a bad depreciation call because they have robust other sources of revenue.
Neocloud providers must rely almost exclusively on the revenue from their GPU rental businesses.
So depreciation policy is an existential, not cosmetic issue. A useful-life error doesn't dent one segment's margin, it distorts the entire income statement, since:
Debt is often GPU-collateralized. Many neocloud financings are underwritten against assumed residual values. Most GPU financing deals assume a uniform, one-size-fits-all depreciation curve. If the real curve is steeper, the collateral coverage on that debt erodes faster than the loan amortizes.
Contract duration and useful life need to line up. If repayment schedules are structured around a six-year life but revenues fall after three years, debt servicing can become strained.
So schedule length maps to equity valuation. Longer schedules:
Lower annual depreciation, leading to higher reported net income and earnings per share.
Improve profitability optics today but risks painful impairments: if hardware is retired or written down early, the deferred expense hits all at once (an earnings "cliff") rather than being smoothed.
Skeptics argue that extended depreciation distorts actual operating metrics.
Shorter schedules also affect valuations:
Cause lower near-term margins and earnings per share, which can produce a valuation discount on trailing/forward P/E versus a peer using a longer schedule, even if the underlying cash economics are identical.
Lower restatement/impairment risk, and probably a lower cost of capital over time if investors reward accounting conservatism with a quality premium once the market re-prices this issue.
How much does it actually matter?
Some argue that since depreciation is a non-cash item, "the market sees through it."
Free cash flow is identical whether the schedule is three years or six, since the cash left the building at purchase.
A DCF-based or FCF-multiple-based valuation should be unaffected by the choice.
Others argue that generally accepted accounting practices still move equity prices. In this view, GAAP figures can be misleading because they're susceptible to noncash charges like depreciation, and metrics like EPS don't fully reflect a company's profitability.
So multiple-driven valuation can be distorted by the schedule choice, and equity-linked debt covenants (leverage ratios, EBITDA-based tests) can be gamed as well.
As always, the assumptions matter. Some argue GPU useful life really is in the three-year range.
Others argue the GPUs still can be used for other operations, and therefore warrants a longer useful life.
In other words, even when a GPU cannot be used for training, it still has value for inference, other batch work or non-AI operations, generating revenue all the while.
Physical failure and retirement data for hyperscaler fleets do suggest older GPU generations last seven to nine years in production before physical retirement.
Either way, the industry seems to be settling on a six-year depreciation cycle for GPU hardware.
No comments:
Post a Comment