AI Earnings Boom Hides $549B Depreciation Bill Coming Due
The five biggest hyperscalers are on pace to spend roughly $760B on AI infrastructure in 2026 while recognizing only ~$211B in depreciation. The other $549B sits on balance sheets, waiting to hit income statements. That deferred cost is the accounting story behind record S&P 500 earnings.
The S&P 500 is posting record profits while the biggest cost of the AI buildout hasn't hit the income statement yet.
Key takeaways
- The five largest hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle) are on track to spend roughly $760 billion on AI infrastructure in 2026 while recognizing only an estimated ~$211 billion in depreciation (per Roberts's analysis), leaving approximately $549 billion in deferred costs that will hit future earnings.
- S&P 500 earnings are tracking 20%-plus growth for a second straight quarter, with the full-year consensus estimate revised up from ~14% in February to north of 23%, the opposite of the historical pattern where, per Roberts, analysts trim ~2% at this stage of the calendar.
- The ratio that tells the real story: combined free cash flow for the five hyperscalers is projected to fall 91% to roughly $16 billion in 2026, even as combined net income is projected to rise 25% to approximately $506 billion, per RIA Advisors' analysis.
Lance Roberts, Chief Portfolio Strategist at RIA Advisors, published an analysis this week arguing that the 2026 earnings boom is partly an artifact of accounting timing: hyperscalers book AI hardware as capital assets and spread the cost across years, so the cash is leaving now while the expense arrives later. The reported earnings surge is borrowed from the future.
Alphabet handed investors the live example last week. The headline read "earnings up 294%," referring to EPS growth confirmed in Alphabet's Q2 2026 earnings release filed with the SEC. Strip out a roughly $99 billion paper gain on its stakes in Anthropic and SpaceX, and core EPS came in around $2.85 against a $2.88 estimate, per Roberts's analysis, with the underlying business posting solid but ordinary growth.
In Q2 alone, Alphabet spent $44.9 billion on capital projects, more than double a year earlier, while free cash flow swung to negative $5.9 billion even as operating income hit $40.8 billion. The cash is already out the door. The reported profit hasn't flinched.
The Gap Between Spending and Expensing
The five hyperscalers spent somewhere in the range of $443 billion to $448 billion on capex in 2025, per data compiled from SEC filings by CreditSights and Epoch AI respectively. Roberts's 2026 estimate runs to roughly $760 billion, directionally consistent with independent third-party projections in the $600 billion to $750 billion range, including a ~$750 billion figure from CreditSights. The depreciation and amortization those companies are expected to recognize against all that spending in 2026: approximately $211 billion, per Roberts's analysis.
That leaves a gap of roughly $549 billion sitting on balance sheets. Most of it isn't being expensed yet because the data centers housing the equipment are still under construction. The depreciation clock doesn't start until the assets go into service.
David Zion of Zion Research Group flagged the modeling problem directly, warning (as cited by Roberts) that "consensus D&A estimates could be systematically understated." Morgan Stanley's Todd Castagno, also as cited by Roberts, calls the current stretch "a golden window where everybody looks good." Chipmakers book revenue immediately when a hyperscaler buys a GPU. The hyperscaler spreads the cost across years. Both sides of that trade look healthy at the same time, which is precisely what makes the cycle look more durable than it is.
This is the same dynamic Roberts identified earlier in the year: the AI capex and GDP story has an accounting architecture underneath it that the headline numbers don't surface.
What the Cash Flow Numbers Actually Show
The full-year projection for 2026 makes the gap concrete: combined net income for the five hyperscalers rises an estimated 25% to roughly $506 billion, while combined free cash flow is projected to fall 91% to approximately $16 billion. These are Roberts's analyst estimates, not reported actuals, and they carry the uncertainty inherent in any forward projection. But the Alphabet Q2 print is already a real-time data point, not a forecast.
The analyst consensus dynamic compounds the risk. Over the past five years, full-year S&P 500 earnings estimates have been trimmed by roughly 2% on average at this point in the calendar, per Roberts. In 2026, the revision went the other way by approximately nine percentage points, from roughly 14% growth in February to north of 23% now. Every quarterly beat this season lifts the bar for the next one. When the depreciation wave eventually flows through income statements, those elevated estimates face a structural headwind with little cushion built in.
The AI capex buildout is already straining physical infrastructure; the financial infrastructure beneath the earnings story is under comparable pressure.
For a Bitcoiner, the second-order effect is straightforward. This is what corrupted money does to capital allocation: cheap fiat credit funds an asset class at a scale the income statement cannot honestly represent in real time, the cost signal gets suppressed by accounting convention, and the market cannot price the risk correctly until it's too late to redirect the capital. The $549 billion gap between what hyperscalers are spending and what they are expensing this year represents one of the largest deferred-cost overhangs in recent corporate history, and it is being cheered as record earnings. When that depreciation wave arrives and earnings disappoint relative to a 23%-growth consensus, equity multiples built on the AI productivity narrative will compress.
Bitcoin's cost, the energy required to mine each sat, is recognized immediately, in full, every block. The depreciation ramp carries no deferred expense, no consensus estimate to revise upward, and no golden window.
The falsifiable version of this thesis: if AI infrastructure generates revenue growth fast enough through 2027 and 2028 to absorb the depreciation ramp without meaningful earnings misses, if hyperscaler net margins hold or expand as D&A scales, then the golden window is a legitimate investment cycle and the deferred-cost thesis fails. Watch for hyperscaler guidance on revenue per GPU utilization, whether D&A consensus revisions track upward or stay systematically low, and whether Alphabet's free cash flow turns positive in 2027 even as depreciation ramps.
What to Watch
The depreciation clock starts when data centers come online, and a significant portion of 2025 and 2026 capex is still under construction. The 2027 earnings season is when the accounting math becomes unavoidable. Between now and then, the tell will be whether analyst D&A estimates start tracking upward toward actual spending or remain anchored to numbers Zion Research Group already flagged as potentially understated.
Update, July 24, 2026
Ed Dowd of Phinance Technologies published a new note this week arguing the AI capex cycle is hitting a wall on multiple fronts simultaneously. The primary source is his Substack, amplified via ZeroHedge. The sharpest claim is on private credit: Dowd calls private credit "the marginal credit producer for the last two years," a lightly regulated corner of finance that grew an estimated 50% to 75% across 2024 and 2025, and says that growth has now stalled, with some funds limiting withdrawals as investors ask for their money back. That freeze matters because Morgan Stanley estimated private credit could fund up to 50% of the external financing needs for the AI data center buildout, and when that spigot slows or gets expensive, the capex math breaks.
On the demand side, Dowd points to a concrete enterprise pullback that goes beyond generic ROI hand-wringing. Companies that rushed AI tools into workers' hands are now reining them in because costs at scale are biting, and enterprises are stepping back to assess ROI as AI budgets come in above initial cost projections. That dynamic is reinforced by open-source commoditization: Chinese-based DeepSeek and the new frontier model Kimi K3 are rivaling OpenAI and Anthropic at a fraction of the price, compressing the economics for anyone charging premium rates for proprietary inference. Dowd also flags a specific stress signal in the credit market: Oracle's credit-default swaps are "exploding" and the stock has been getting hammered.
The Goldman Sachs credit desk is arriving at a similar read. Current creditors are reassessing their exposure, as the Goldman Sachs credit desk has recently highlighted.
A financing reversal could turn prohibitively costly, pausing or dramatically slowing the cycle that has supported the S&P 500, roughly 45% of whose market cap is AI or AI-adjacent. The depreciation overhang detailed in this article was always the income-statement problem. Dowd's note adds the credit-market mechanism by which the spending itself could slow before that deferred bill even fully arrives.
Sources
Frequently Asked Questions
When a hyperscaler buys AI hardware, it records the purchase as a capital asset and expenses the cost gradually over years through depreciation. The cash leaves immediately; the income statement cost arrives in slow motion. During a buildout of this scale, that timing gap inflates reported profits today by deferring hundreds of billions in costs to future periods. Earnings look strong not because the underlying economics are unusually good, but because the full cost hasn't been recognized yet.
Roberts estimates the five largest hyperscalers will spend roughly $760 billion on AI infrastructure in 2026, against an estimated $211 billion in depreciation recognized this year, per his analysis. The remaining ~$549 billion sits on balance sheets and flows through income statements as assets enter service, likely accelerating through 2027 and 2028 as data centers under construction come online.
If AI revenue growth is fast enough to absorb rising depreciation without margin compression, multiples can hold. If it isn't, analysts will be cutting estimates from an already elevated ~23% full-year growth baseline, a combination that historically drives meaningful multiple compression. The free cash flow divergence (projected 91% decline against a 25% net income increase, per Roberts's analysis) is the early signal that the gap between accounting profit and economic reality is already wide.
