Everything About AI
NextFin News - Lori Calvasina, head of U.S. equity strategy at RBC Capital Markets, has drawn a sharp line for anyone betting against the S&P 500: worrying about valuation is not enough. To make a bearish case stick, you have to attack the AI earnings story itself. In recent commentary, Calvasina wrapped the market's "AI froth" directly into the earnings risk that she says is the only thing capable of breaking the rally - and then argued that even a 5% cut to consensus earnings leaves her price target largely intact. The message inverts the usual Wall Street bear setup, and it puts the entire debate over whether this bull market survives on one question: can the hyperscalers' record capital-spending boom pay for itself?
Layer 1: The Situation
Calvasina's framing matters because it names the actual fault line. For more than a year, the most common argument against the S&P 500 has been valuation: the index trades at a trailing price-to-earnings ratio of roughly 24 times, well above its long-run average, and concentration in the largest technology names has pushed market-cap weighting to extremes not seen since the dot-com era. Calvasina is effectively telling both sides that they are arguing about the wrong thing.
"We've done stress tests for people. And what we show them is that if you really want to get bearish on the market, you have to get bearish on earnings. It's not just enough to be worried about valuation pressure. And you have to attack that AI earnings story. We take the consensus down by 5%. And we're still getting to 8162 on that model." - Lori Calvasina, Head of U.S. Equity Strategy, RBC Capital Markets
That 5% stress-test figure is the hinge of her argument. RBC's rolling 12-month S&P 500 target stands at 8,150, set in late June and frozen at that point because several of the firm's models are built on forward rates of return. At the time the target was set, the index sat roughly 10.8% below that level. Even after shaving 5% off consensus earnings per share, the model still prints 8,162 - essentially unchanged.
The implication is uncomfortable for bears and bulls alike. For bears, it means a valuation-only thesis cannot move the index meaningfully lower unless it travels through earnings. For bulls, it means the entire 8,150 target is underwritten by one assumption: that the AI-driven earnings engine does not stall.
Calvasina has used blunter language elsewhere. In a recent post, she wrote that "the AI-driven investor froth has cooled, but not cleared" - a phrase that captures the current market mood almost exactly. Sentiment has cooled from its early-2026 extremes, volatility has compressed back toward normal levels, and yet the concentration of earnings growth in AI-related names has not.
The earnings backdrop she is defending is genuinely strong. RBC's own work shows the S&P 500 tracking roughly 47% year-over-year profit growth in the current quarter; even stripping out the outsized AI contributors, the index would still be growing at nearly 30%. By comparison, earnings grow about 7.6% in a typical year. That is not a market that looks expensive on a static multiple if you believe the earnings number keeps compounding.
But the composition of that growth is the problem. The same RBC research that highlights AI as an "extremely bright spot in the earnings backdrop" also shows how narrow the leadership has become. Tech, Energy, and Materials have been the sectors lifting consensus EPS growth forecasts for 2026, while the rest of the index has lagged. When one theme carries this much of the earnings story, the market is no longer diversified against that theme failing.
RBC is not predicting a crash. The firm has said it expects any setback to resemble "a tier 1 garden-variety pullback in the 5-10% range," and that a steeper 14% to 20% drawdown would probably require renewed recession fears to emerge. The firm's separate 12-month target of 7,900, set earlier in the year, was built on a framework it described as "AI in the fast lane, Middle East in the slow lane" - a two-speed backdrop in which AI-related sectors continue to outperform while geopolitical tensions weigh on broader activity. That model assumed 28% earnings growth for AI-linked companies, 6% growth for the remainder of the index, consumer inflation of 3.3%, no additional Federal Reserve rate changes, and a 10-year Treasury yield of 4.5%.
Layer 2: The Analysis
Why Valuation Pressure Alone Cannot Break This Market
The first layer of Calvasina's argument is mechanical. A stock price is earnings multiplied by a multiple. If the multiple compresses but earnings rise fast enough, the price can still go up. This is precisely what has happened through much of the AI rally: forward earnings estimates have been revised higher almost as quickly as multiples have expanded, leaving the index's forward P/E less stretched than the trailing number suggests.
RBC's own valuation work bakes in conservative macro inputs to defend against multiple compression: roughly 3% inflation, one Federal Reserve rate hike, and 10-year Treasury yields around 4.5%. Under those assumptions, the firm arrives at a trailing P/E of about 24.25 and still finds the index fairly valued at its target. The logic is that if rates and inflation behave, the multiple does not need to collapse - and if earnings keep growing at 20% to 30%, the multiple can even stay where it is.
This is why a bearish call built only on "the market is expensive" has failed repeatedly since 2023. Every time the S&P 500 has looked overbought, the next earnings season has delivered another round of upward revisions from the AI complex, and the index has absorbed the valuation concern. The market has been telling investors what it believes: the earnings are real, and they are growing.
The counter-argument - that trailing earnings are backward-looking and that the multiple is pricing in a perfection that cannot last - is valid but incomplete. It explains why the market is vulnerable; it does not explain what would actually push it down. Calvasina's point is that vulnerability without a trigger is just a condition, not a forecast.
The Earnings Case Is Where the Fragility Actually Sits
So what would a real earnings attack look like? It would start with the hyperscalers. Alphabet, Amazon, Meta, Microsoft, and Oracle have been spending at a pace that has no modern precedent. Consensus estimates for 2026 capital expenditure across the group have risen to roughly $770 billion to $772 billion - up about 15% since March alone - with 2027 spending approaching $1 trillion. Goldman Sachs notes that this equals approximately 100% of the group's cash flow from operations, and the group's net debt has risen by about $170 billion since the start of 2025 as spending has outpaced internally generated cash.
Here is the mechanism that Calvasina's "earnings case" must confront, and it is not a simple one. When a hyperscaler spends a dollar on a graphics processor, that dollar does not hit the income statement immediately. Only a portion is expensed in the year it is spent; the rest sits on the balance sheet as property and equipment, and then flows into earnings gradually as depreciation. Industry analysis suggests that across the largest spenders, reported depreciation runs at roughly a third of the annual capex bill - meaning the majority of today's spending is an earnings cost that has not yet arrived.
That timing mismatch is the hidden leverage in the AI earnings story. Today, the capex shows up as a drag on free cash flow but not fully on reported earnings. Tomorrow, the depreciation expense arrives regardless of whether the AI revenue has materialized. Goldman Sachs estimates that depreciation and amortization for the hyperscalers will climb from 7% of revenues in 2022 to 12% by 2027. That is a five-percentage-point margin headwind building in the background, invisible to investors who are watching this quarter's EPS beat.
This is why Calvasina's 5% stress test is both reassuring and revealing. A 5% cut to consensus is survivable - her model barely flinches, moving from 8,150 to 8,162. But a 5% cut is not what a genuine AI earnings disappointment would look like. If the hyperscalers' AI revenue fails to cover the depreciation wave now being stacked onto their balance sheets, the earnings revision would not be 5%. It would be a wholesale repricing of the index's largest constituents, and it would flow directly into the one channel Calvasina says is the only one that matters.
Cyclical Super-Cycle on Top of a Structural Shift
The critical judgment for investors is whether the AI capex boom is cyclical or structural. The answer is both, and they point in opposite directions across different time horizons.
The structural leg is real and durable. Artificial intelligence is a general-purpose technology, and the shift of enterprise computing workloads onto AI-accelerated infrastructure is a regime change that will not reverse. Demand for compute, for data-center power, for the software layer that runs on top of the chips - none of this is a bubble in the sense that it will disappear. The technology is being adopted, and the productivity gains, where they materialize, are permanent.
The cyclical leg is the capital-spending wave that finances the structural shift, and it behaves like every capex super-cycle in history: it overshoots. Semiconductor cycles, telecom fiber build-outs, the early-2000s broadband boom - in each case, the underlying technology was real, and in each case the spending wave ran ahead of the revenue it would eventually produce. The overshoot is not a sign that the technology is fake. It is a sign that capacity gets built before the demand curve catches up, and then margins absorb the difference.
Calvasina's own language tracks this distinction. The froth has "cooled, but not cleared." The structural adoption is intact; the cyclical excess in investor positioning and in capital allocation has not fully worked off. Her stress test is essentially a bet that the cyclical leg can be trimmed by 5% without breaking the structural thesis.
The risk is that the two legs get conflated in the market's pricing. If investors begin to treat a cyclical capex slowdown as evidence that the structural AI story is broken, the multiple and the earnings estimate would contract simultaneously - the exact scenario Calvasina says is required for a bear case to work. That is the double-leverage downside she is implicitly warning about, even as she maintains her bullish target.
Second-Order Effect: The Depreciation Wave No One Is Modeling Into the Multiple
The first-order effect of the AI capex boom is well understood: semiconductor earnings surge, equipment vendors rally, and the index's headline EPS grows at an unusual rate. The second-order effect is what the market has not priced, and it is where Calvasina's warning lands.
The capex boom is boosting semiconductor earnings today at the direct expense of hyperscaler free cash flow. As the spending converts into installed capacity, the depreciation expense migrates from the balance sheet to the income statement. The group that is funding the boom - the hyperscalers - will see their margins compressed just as the group that benefited from the build-out - the chipmakers - faces demand that is now growing off a much larger base.
This is the classic capex-cycle inflection. During the build-out phase, the suppliers capture the upside while the spenders defer the cost. After the build-out peaks, the spenders absorb the depreciation regardless of utilization, and the suppliers face comparables that are nearly impossible to repeat. The earnings baton passes from one group to the other, and the index-level number can stall even when the underlying technology continues to advance.
Calvasina's stress test does not appear to model this second-order rotation explicitly. A 5% cut to consensus is a uniform haircut; it does not capture the asymmetry of a margin squeeze concentrated in the largest index weights while the smaller beneficiaries roll over. That asymmetry is exactly why an earnings-driven bear case, once it starts, can move faster than a valuation-driven one. Multiple compression is gradual. Earnings revisions can happen in a single quarter.
Layer 3: Conclusion and What to Watch
The Strongest Counter-Thesis
The strongest argument against Calvasina's framing is that it assumes the hyperscalers will successfully monetize the capacity they are building. If AI revenue scales with the infrastructure - if enterprises continue to adopt agentic AI, if inference demand grows with the installed base, if the software margins materialize as expected - then the depreciation wave is covered and the earnings story compounds. In that scenario, Calvasina is right that valuation concerns are a distraction, and the 8,150 target is conservative rather than aggressive.
This counter-thesis has substantial backing. The same institutions warning about capex intensity also document that technology earnings estimates have been revised higher across all six industries since March, driven by "continued strong momentum in the AI infrastructure build-out and the emergence of agentic AI." The demand side of the equation is not empty.
But the counter-thesis has a specific vulnerability: timing. Even if AI revenue eventually covers the capex, the depreciation expense arrives on a fixed accounting schedule. If the revenue ramps more slowly than the capacity comes online, there is a window - likely measured in quarters, possibly stretching into 2027 - where reported earnings absorb the cost before the revenue catches up. That window is where Calvasina's "earnings case" risk lives.
The Falsifying Signal
Investors should watch one metric above all: hyperscaler AI revenue growth versus the depreciation trajectory. Specifically, if the group's AI-related revenue fails to grow at a rate that covers the incremental depreciation expense for two consecutive quarters - and depreciation is expected to rise from 7% to 12% of revenues by 2027 - then the structural earnings thesis is under genuine pressure, and Calvasina's 5% stress test would look too mild. A second confirming signal would be a sustained downward revision in consensus 2027 EPS for the S&P 500 exceeding 5%, which is the exact threshold at which her model begins to break.
Scenarios
- Base case: AI revenue scales broadly in line with capacity additions, depreciation is absorbed, and the S&P 500 grinds toward RBC's 8,150 target with mid-to-high single-digit returns driven by earnings rather than multiple expansion. This is Calvasina's world, and it remains the highest-probability outcome.
- Upside case: agentic AI adoption accelerates faster than expected, hyperscaler margins hold, and earnings revisions continue to run ahead of estimates. In this scenario the index could exceed 8,150, though multiple expansion would be required - something RBC explicitly does not assume.
- Downside case: AI revenue disappoints relative to the installed base, the depreciation wave hits reported earnings, and consensus 2027 EPS is cut by more than 5%. This is the earnings case Calvasina describes, and it would likely produce the "tier 1 garden-variety pullback in the 5-10% range" she has flagged - with a deeper 14% to 20% drawdown only if recession fears re-emerge alongside it.
The Takeaway
Calvasina has done investors a service by naming the actual battlefield. The debate over whether the S&P 500 is expensive is a sideshow. The real question is whether the largest companies in the index can turn a historically unprecedented capital-spending wave into earnings that cover its cost. Her answer is yes, with a 5% margin for error. The market's next move depends on whether that margin is wide enough for the depreciation wave now building just off the income statement.
Explore more exclusive insights at nextfin.ai.
