The US market wore two faces in the first half of 2026. The S&P 500 closed at a record in early June, and Micron doubled its market value from $500 billion to $1 trillion in just under 50 trading days. Yet in the same month, the Philadelphia Semiconductor Index dropped 10% in a single session and bounced back, and on June 23 the Nasdaq fell 2.2% and Korea's Kospi crashed 10%, tripping a circuit breaker.
It is a bull market that makes you motion-sick. At the center of the contradiction sits AI capex, capital expenditure. The sheer scale of what big tech is shoveling into data centers, and the market's doubt about when that money returns as profit, is throwing stocks up and down on every earnings print. This piece dissects the amplitude with numbers and news, from both the bull and bear sides.
1First, the sheer size of the number
The four big-tech names will spend an estimated $725 billion on capex in 2026, up 77% from roughly $410 billion a year ago.
Microsoft, Alphabet (Google), Meta, and Amazon. The combined capital expenditure these four have guided for 2026 totals about $725 billion. That is the kind of sum that goes into GPU clusters and data centers, and it is almost entirely AI infrastructure. For scale, it exceeds the total annual revenue of all but a handful of countries.
2026 capex guidance, by company (USD billions)
The direction is what matters. Alphabet raised its 2026 guide once, to $180–190 billion, and explicitly said it would "significantly increase" in 2027. The curve shows no sign of bending yet. The catch is that the steeper the curve gets, the more sharply the market asks the same question: so when, exactly, do you earn it back?
2Why strong earnings still sink the stock — the asymmetric era
The Q1 2026 earnings season showed this question at its starkest.
Classically, a stock rises when earnings beat. But in the spring of 2026, the market moved the opposite way. The variable it watches shifted from "this quarter's profit" to "will capex eat future profit?" Spend more on data centers today, and tomorrow's depreciation grows and margins compress. Faced with identical beats, the market reacted in opposite directions depending on a single line of capex guidance.
| Company | Q1 result | Capex signal | Stock reaction |
|---|---|---|---|
| Meta | Revenue +33%, EPS +62% (beat) | 2026 capex raised to $125–145B | -8 to -10% |
| Microsoft | Beat estimates | Record $37.5B quarterly capex, OpenAI risk | fell |
| Alphabet | Cloud +63%, $462B backlog | Raised to $180–190B + monetization proof | +7% (after hours) |
| Nvidia | Revenue $81.6B (+85%), DC $75.2B (+92%) | Q2 guide $91B, +$80B buyback | sold on the news |
Meta delivered a "perfect" quarter, revenue up 33% and EPS up 62%, and the stock still fell 8 to 10%. The sole reason: it lifted the top of its capex range to $145 billion. Microsoft was tripped up by a record $37.5 billion quarterly capital outlay and its OpenAI dependence. Even Nvidia, with 85% revenue growth, was pushed down by "buy the rumor, sell the news" profit-taking. A 30x forward multiple demands not "good" but "overwhelmingly good," and the result was merely good.
3June's whipsaw — what the volatility actually looked like
Once doubt crosses a threshold, a bull market swings between crashes and rebounds within days.
June 2026 was a textbook on volatility. Early in the month the Philadelphia Semiconductor Index (SOXX) plunged 10% in a single day. Nvidia shed more than $300 billion in market cap, Broadcom fell 12.6%, and Marvell cratered 17%. Yet within days AMD erased all of its losses and printed a new high, a "V-shaped" rebound, the classic pattern of a correction inside a bull market.
June 23 brought broader selling. The Nasdaq fell 2.2%, the S&P 500 fell 1.4%, and Korea's Kospi crashed 10%, tripping a circuit breaker. The trigger was a mix: doubt that AI data-center demand might be slowing, reports that the four companies' combined capex had already topped $452 billion, and crowded one-sided institutional positioning. NPR ran "Is AI one big bubble?" as a headline that week.
The essence of this amplitude is not a simple "bubble or not." It is that, facing the same data (surging capex), the market reads it as "proof of demand" on some days and "waste that will never be recouped" on others. When the interpretation wobbles, the price wobbles.
4The bull case — this time the demand is real
The strongest evidence justifying the capex is Google, and power.
① Google proved the "return"
Alphabet's Cloud revenue grew 63% year over year to $20 billion, with a 32.9% operating margin and a backlog that nearly doubled in a single quarter to $462 billion. Paid Gemini Enterprise users grew 40% in the quarter, and Search's AI Overviews are monetizing at a rate similar to traditional Search.
The decisive blow: across 2025, Google cut the unit cost per Gemini query by about 78% through TPU optimization. When the cost falls, AI features add to margins instead of eroding them. That is why the market rewarded Google alone.
② The power bottleneck is itself proof the demand is real
US data-center power demand is projected to double from 80 GW in 2025 to 150 GW by 2028. Yet of 12 GW of announced 2026 projects, only 5 GW is actually under construction; the rest is stuck as high-voltage transformer and switchgear lead times stretch to five years.
Supply that cannot keep up with demand means that, at least near term, this is not a "build it and let it sit" glut but a regime of excess demand short on sites and electricity. Texas becoming the largest market at 40 GW by 2028 comes down to power and land.
Nvidia's quarterly revenue of $81.6 billion, of which Data Center was $75.2 billion (+92%), is proof this demand is printing as actual revenue. The bull thesis is simple: this is not a dot-com promise of clicks, but demand real enough to slam into the physical limits of electricity and silicon.
5The bear case — the numbers don't add up
The core of the skepticism is the "investment-to-revenue gap" and the structure of "paying each other."
① Spending $8–10 to earn $1
US AI capex is estimated at $500 billion a year in 2026–27, while current AI revenue runs around $12 billion in consumer spending, roughly $25 billion annualized at OpenAI and $19 billion at Anthropic. The math implies the industry is investing $8–10 for every $1 of revenue. Enterprise surveys consistently report that about 95% of corporate AI projects show "no measurable ROI."
② Depreciation, the time bomb
GPUs have an economic life of 3–7 years, but assigning a longer accounting life books less depreciation and inflates profit. Critics estimate big tech may be understating depreciation by roughly $176 billion across 2026–28. The chips age fast while the debt is repaid slowly: a refinancing treadmill.
③ The circle that manufactures each other's revenue
Nvidia committed $100 billion to OpenAI, about 39% of Nvidia's annual revenue, and OpenAI is one of Nvidia's largest customers. Nvidia also holds 7% of CoreWeave and agreed to buy $6.3 billion of CoreWeave's unsold data-center capacity, capacity stocked with Nvidia GPUs. CoreWeave carries $14.2–21.6 billion in total debt against $3.34 billion of equity, with a $4.2 billion principal repayment due in 2026.
6So what should you watch — a multi-angle synthesis
This is not a binary "bubble or not." It is a capital-intensive regime where winners and losers diverge.
The bull and bear cases are really two faces of one picture. The demand is real (power bottleneck, Nvidia revenue, Google monetization), and at the same time the pace of investment runs far ahead of current revenue (the 8–10x gap, understated depreciation, circular financing). Both can be true, which is why the market sways each quarter as it argues over which side to weight.
📡 Five checkpoints for reading the swings
- The "direction" of capex guidance — less the absolute figure than the raise/cut and "slower growth" language. Two or more cutting at once is a strong turn signal.
- Evidence of monetization — cloud revenue growth, the speed at which backlog converts to revenue, falling unit cost of AI features. Whether the Google model spreads.
- Depreciation schedules — whether big tech is lengthening GPU useful lives (inflating profit) or shortening them. Read the 10-Q footnotes.
- Power and transformer lead times — once the physical bottleneck eases, supply catches up and the risk of a late-cycle glut rises.
- Stress in the circular financing — whether repayment or equity-raise noise emerges from leveraged links like CoreWeave and OpenAI. The weakest link cries first.
Compress the conclusion to one sentence: AI capex is neither a bubble nor a revolution, but a capital-intensive regime where both overlap in the same window of time. Within it, the companies that prove the return (Google) and those that just keep raising spend without proof increasingly walk different paths. The volatility is the sound of the market repricing that fork in real time, and it will not vanish soon. The investor who looks first at "evidence of return," not the earnings number, gets less motion-sick on this rollercoaster.
iKey sources
- Big-tech capex $725B, +77% — Tom's Hardware, CNBC, Fortune, Futurum (Feb–Apr 2026)
- Market rally, Micron $1T — CNBC, CNN Business "It's still all about AI" (Jun 2026)
- Meta/MSFT capex shock — Fortune, Yahoo Finance, ainvest (Apr 2026)
- Alphabet Cloud +63%, $462B backlog, -78% unit cost — CNBC, Investing.com, Fortune (Apr 2026)
- Nvidia Q1 FY27 $81.6B, DC $75.2B — NVIDIA Newsroom, StockTitan (May 2026)
- June selloff, Kospi circuit breaker — CNN Business, NPR, Intellectia (Jun 2026)
- ROI gap, depreciation, circular financing — Bloomberg "AI Circular Deals," Wikipedia (AI bubble), Sourcery Intel, Morgan Stanley (2026)
- Data-center power 80→150 GW, transformer lead times — Data Center Frontier, Sightline Climate (2026)