Key Takeaways
- NVIDIA's Data Center revenue reached $75.2 billion in Q1 FY2027, up 92% year on year, after $193.7 billion across FY2026. Q2 guidance is $91.0 billion of total revenue.
- The four largest US hyperscalers have guided to roughly $700 billion of 2026 capex: Amazon about $200bn, Microsoft about $190bn, Alphabet $180-190bn, Meta $125-145bn.
- Microsoft's quarterly depreciation rose to $9.0 billion from $5.8 billion a year earlier, while quarterly capex was $31.9 billion including finance leases — depreciation is 28% of capital spending.
- Amazon shortened some server useful lives from six years to five, effective January 2025, adding $1.4 billion of depreciation and cutting net income by $1.0 billion, or $0.10 per share.
The run-rate argument is obsolete, and by a factor of three
An earlier version of this piece argued toward a $100 billion annual run-rate for AI data-centre revenue. NVIDIA reported $193.7 billion of Data Center revenue for FY2026 alone, and then did $75.2 billion in the first quarter of FY2027 — a single quarter running at three times the annual figure once treated as the destination. The chart above shows five consecutive quarters: $39.1bn, $41.1bn, $51.2bn, $62.3bn, $75.2bn. Guidance for the current quarter is $91.0 billion of total revenue, and NVIDIA states it is "not assuming any Data Center compute revenue from China."
Restating the number is not the interesting part. What matters is what that revenue becomes on someone else's balance sheet.
One company's revenue is four companies' fixed assets
NVIDIA's Data Center revenue is, to a first approximation, the capital expenditure of a handful of buyers. Those buyers have told shareholders what they intend to spend in 2026. Amazon: "we expect to invest about $200 billion in capital expenditures across Amazon in 2026." Microsoft's Amy Hood: "For calendar year 2026, we expect to invest roughly $190 billion in capital expenditures which includes approximately $25 billion from the impact of higher component pricing." Alphabet raised its range to $180-190 billion. Meta raised its to $125-145 billion, from $115-135 billion, citing the same component pricing.
That is roughly $700 billion of capital spending in one calendar year from four companies. It is recorded as an asset. It becomes an expense later — and it is the same spending that shows up downstream as a cost of goods sold in software.
The depreciation gap is where the earnings are hiding
When a hyperscaler buys a GPU, free cash flow takes the hit immediately — cash out, no expense. Earnings take it gradually, as depreciation, over the asset's useful life. When capex rises fast, depreciation lags, and reported operating margin is flattered by the lag.
Microsoft's own numbers show the size of it. Depreciation expense was $9.0 billion for the quarter ended 31 March 2026, against $5.8 billion a year earlier — a 55% increase, per the 10-Q. Capital expenditure in that quarter was $31.9 billion including finance leases, the figure Amy Hood gave on the earnings call; the cash-flow statement shows $30.9 billion of cash paid for property and equipment. Depreciation is 28% of capex on the first basis, 29% on the second.
For a business with a stable asset base, depreciation and capex converge. The distance between $9.0 billion and $31.9 billion measures how far Microsoft is from that steady state, and every quarter of elevated spending pushes more expense into future income statements. Hood was explicit about the composition: "roughly two thirds of our capex was for short-lived assets, primarily GPUs and CPUs." Short-lived assets depreciate fast.
The same pattern is visible elsewhere. Alphabet's depreciation rose to $6,482M in Q1 2026 from $4,487M. Meta's servers-and-network depreciation reached $4.38 billion against $2.63 billion. These are 44% and 67% increases, and they are the early part of a curve, not the end of one — a dynamic that also explains how cloud margins expanded during peak capital spending.
Amazon already reversed its own assumption
The clearest evidence that useful-life assumptions are load-bearing comes from Amazon, which has changed them twice, in opposite directions.
Effective January 2024, Amazon extended server useful lives from five years to six, which reduces annual depreciation and raises reported earnings. Effective January 2025 it reversed course for a subset of servers and networking equipment, back to five years, and disclosed why: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning."
The cost was quantified: "an increase in depreciation and amortization expense of $1.4 billion and a reduction in net income of $1.0 billion, or $0.10 per basic share and $0.10 per diluted share, which primarily impacted our AWS segment." One year of useful life on a subset of one company's servers was worth ten cents of EPS. Applied across a $700 billion annual spend, the sensitivity of reported earnings to an accounting estimate is not a rounding error.
What this framework cannot establish
Depreciation lagging capex is not evidence of anything improper, and it is not a prediction that earnings will fall. If revenue grows faster than the depreciation catch-up, margins hold — which is what Alphabet's cloud segment has done so far. The lag is a headwind, not a verdict.
The published data also cannot tell you the economic life of a GPU, only its accounting life. Those are different numbers, and nobody outside the operators knows the first. If AI accelerators stay productive for eight years, current schedules are too aggressive and earnings understated. If they are obsolete in three, they are too generous. Amazon's reversal is one data point suggesting the second risk is live, but it is an estimate, not a fact about the hardware.
The argument would be falsified if depreciation growth stalled while capex kept climbing — the operators would be lengthening lives, making the gap a policy choice rather than a timing effect. It would equally be undercut if hyperscaler free cash flow kept rising through the spending peak.
The disclosure to read first
The number that will matter in the next four quarters is not capex guidance, nor NVIDIA's revenue — the same figure seen from the other side. It is the useful-life footnote in the property and equipment note of each 10-K. That single sentence determines how much of $700 billion hits earnings each year, it is set by management, and Amazon has just demonstrated it can move by a year in either direction without any change in the business.
This content is for informational purposes only and does not constitute financial advice. Always do your own research or consult a qualified financial adviser before making investment decisions.