Overnight, the unwavering faith in AI stocks faced a reality check—from within their own ranks.

On July 1, Meta's shares surged nearly 9% in a single trading day, adding a staggering $100 billion to its market capitalization. Meanwhile, Micron and SanDisk tanked over 10%, while leading AI infrastructure players like CoreWeave and Nebius plummeted 15%.

The catalyst behind this stark market divergence? A single announcement: Meta would begin selling excess computing capacity to external parties. What initially seemed like a straightforward asset optimization move has ignited the most significant split in AI market sentiment this year:

Optimists argue this merely reflects normal "generation turnover" in hardware—retiring older GPUs won't impact procurement of cutting-edge chips. Pessimists, however, see Meta's move as conclusive evidence of industry-wide overcapacity and a potential inflection point for AI capital expenditures.

While this debate remains unresolved, the market has already begun rewarding cloud providers that demonstrate "strategic restraint."

Immediately following Meta's announcement, its stock price soared. This contrasted sharply with the previous month's performance, when the four major cloud service providers (CSPs) collectively underperformed the Nasdaq: Microsoft (-17.15%), Amazon (-11.93%), Meta (-10.86%), and even top performer Google (-6.18%).

What's driving this divergence? How should we interpret Meta's computing power strategy? Silicon-Based Gentleman examines the implications.

/ 01 / The Market Rewards Restraint

This market split reflects a fundamental shift in Wall Street's valuation framework for cloud infrastructure investments.

Through 2024, "compute scarcity" reigned as the market's ultimate currency.

Google's Q1 results showed capex surging 91% YoY to $12 billion (beating $10.3B estimates), triggering a 12% after-hours stock jump. Amazon's Q3 2024 capex hit $22.6 billion (+81% YoY), exceeding forecasts by $3 billion and lifting shares 5% after hours.

The prevailing logic was simple: aggressive spending equaled bullish signals. But the tide began turning in early 2025.

In July 2025, Google raised its full-year capex guidance to $85 billion (crushing $70B estimates). After a brief dip, shares ultimately climbed 2%. Three months later, Microsoft reported $34.9 billion in quarterly capex (+74.5% YoY), far exceeding expectations—yet its stock fell nearly 4% after hours.

By 2026, sentiment had completely reversed.

In late April, Meta delivered blockbuster results ($56.3B revenue, +61% net profit YoY). Yet merely raising its full-year capex midpoint by 8% sent shares plunging 7% after hours. Two months later, Amazon CEO Andy Jassy revealed 2026 capex would reach $200 billion (vs. $146.6B consensus), triggering an 11% after-hours collapse.

The pattern is clear: Two years ago, expansion drove rallies. Last year, cracks appeared. This year, even "better-than-expected" spending fails to impress—any hint of sustained outlays triggers valuation punishment.

This revaluation has created sector-wide weakness. Over the past month, the four major CSPs have significantly underperformed the Nasdaq (-2.81%): Microsoft (-17.15%), Amazon (-11.93%), Meta (-10.86%), Google (-6.18%).

The market's fear is palpable: It's not spending itself that worries investors, but the growing uncertainty about when—or if—these investments will pay off.

This explains the violent reaction to Meta's announcement. Simply monetizing idle capacity triggered a $100 billion single-day valuation surge.

Meta isn't alone. All major AI infrastructure players have quietly pivoted from "full-throttle chip stockpiling" to "internal utilization plus strategic sales."

Two unavoidable realities drive this shift:

1. Capex scales now threaten cash flow stability.

Amazon's $200 billion 2026 plan has Morgan Stanley projecting -$17 billion in full-year free cash flow. Even cash-rich Meta faces J.P. Morgan's -$4 billion FCF estimate for 2026. Google's $175-185 billion capex plan prompted Morgan Stanley to slash its per-share FCF forecast by 58%.

Only Microsoft, with its diversified business, maintains positive FCF growth—though its capex still consumes 63% of operating cash flow.

2. AI adoption lags behind inflated expectations, creating surplus capacity.

Meta's Llama models haven't produced commercial blockbusters despite their influence. Only OpenAI and Anthropic have achieved true scale. The trillion-dollar CSP investments haven't yet created proportional internal demand, making periodic idleness inevitable.

/ 02 / The AI Boom's First Major Schism

The Meta debate isn't mere market noise—it represents the first systemic challenge to AI's foundational narrative.

For two years, the consensus was clear: compute would remain scarce, and capex would only rise. Now, cracks are appearing.

First, industry leaders now explicitly consider "overbuilding" risks.

Mark Zuckerberg recently acknowledged, "If one day we feel we've overbuilt," signaling that top executives now factor in overcapacity. This expectation shift ripples upstream—when overbuilding becomes possible, locking in capacity at any cost becomes untenable, weakening order commitment.

Bloomberg-compiled analyst forecasts show major cloud capex growth slowing sharply: from ~100% YoY in 2026 to 22% in 2027, then single digits. Even optimistic J.P. Morgan only projects 50% growth for 2027, acknowledging the broader slowdown.

Second, compute now has a "secondary market." Pessimists argue simply: if idle capacity can circulate freely, "new demand = new procurement" no longer holds.

Recent moves validate this: xAI leased its 220,000-GPU Colossus cluster to Anthropic for $1.25 billion/month through 2029 ($40+ billion total). Now Meta joins this market.

In essence, the compute market now has a "stock supply" component. Future demand assessments must account for both new deployments and activated existing capacity.

Bulls counter that Meta's move doesn't signal an industry peak, citing two arguments:

1. Industry expansion shows no signs of slowing.

By late 2025, Meta's compute capacity equaled ~2.5 million H100s (~2GW). Its 2026 capex center is $135 billion, adding 2-3GW for a ~5GW total.

Future industry expansion is even more massive. Huayuan Securities estimates:

• Google: 25GW total by 2028 (5GW from Anthropic contract, assuming 25% of GCP demand)
• Amazon: 20GW total by 2027 (18GW from Anthropic/OpenAI needs, plus doubling from 6.5GW in 2025)
• Microsoft: 20GW total (from $250B OpenAI Azure contract, assuming 25% of demand)

Add OpenAI's 10GW Stargate projects with NVIDIA and Broadcom, and planned new capacity exceeds 85GW. Meta's potential 5GW lease is negligible by comparison.

2. Only aging capacity is being leased—new procurement remains unaffected.

Morgan Stanley confirms Meta only leases older H100/A100 inference cards externally, while retaining 100% of GB200/GB300 and next-gen Rubin training chips for Llama development.

The timing of Meta's June Crusoe contract (1.6GW) and July compute sale announcement proves this: selling existing capacity while buying new chips on parallel tracks.

While both sides make compelling cases, Meta's move alone won't resolve the cycle debate. Current evidence suggests the market narrative is shifting toward "strategic capex restraint" among AI cloud providers. A clearer picture may emerge after Q3 earnings in August.

By Yuan Yuan