On July 1, 2026, Bloomberg dropped a "deepwater bomb" on the US stock market: Meta is establishing a cloud computing division called "Meta Compute," planning to sell its surplus AI computing power and model access to external clients. The news triggered an extremely polarized market reaction.
Meta’s own stock price soared 8.81% that day, closing at $612.91—the company’s best single-day performance in six months. Meanwhile, the AI hardware sector suffered its most brutal sell-off of the year. Leading memory chipmaker Micron Technology plunged 9.7%, SanDisk fell 10.82%; in optical communications, Corning tumbled over 13%, Marvell Technology dropped more than 7%. The Philadelphia Semiconductor Index fell 6.27% in a single day. Compute leasing companies took the hardest hit: Nebius plummeted over 14%, while CoreWeave dropped more than 13%.
At the heart of this seemingly contradictory market phenomenon is a single expectation: "AI capital expenditure has peaked." If Meta already has surplus computing power and needs to lease it out, the market naturally infers that Meta’s appetite for new upstream hardware purchases will slow dramatically. Storage chips, optical communications, GPUs—the "pick-and-shovel" sellers that were once in a buying frenzy during the AI compute arms race—now suddenly face the risk of reversed demand expectations.
The market split triggered by Meta’s announcement continued to spread in subsequent trading sessions. On July 2, the South Korean stock market opened down 5%, immediately triggering a circuit breaker. Samsung Electronics fell over 7%, SK Hynix dropped more than 8%. The Nikkei 225’s intraday losses widened to 2%. Notably, as the AI frenzy cooled, Bitcoin appeared to benefit from some capital rotation, rebounding above $61,000 at its peak.
But does this panic selling truly signal the end of "peak AI capex"? What are Meta’s real intentions behind its cloud initiative? To answer these questions, we need to look beyond short-term market sentiment and return to the core logic of Meta’s AI strategy.
From "Cost Center" to "Revenue Engine": The AI Efficiency Revolution in Meta’s Ad Business
Before Meta announced its cloud initiative, the market’s core focus for this social media giant was: What kind of returns can $125–145 billion in annual AI capital expenditure actually deliver?
The Q1 2026 earnings report offered a partial answer. Meta’s total revenue grew 33% year-over-year to $56.3 billion, with ad revenue reaching $55 billion—also up 33%. Even more noteworthy is the structure of this growth: ad impressions rose 19% year-over-year, while the average price per ad climbed 12%. This "volume and price both up" combination directly reflects the efficiency gains from deeply embedding AI into Meta’s ad system. Higher impressions signal that AI recommendation algorithms are boosting user engagement, while higher prices indicate advertisers are willing to pay a premium for more precise targeting.
These efficiency gains stem from Meta’s massive overhaul of its recommendation systems. On the Q4 2025 earnings call, Mark Zuckerberg noted that while Meta’s current recommendation engine had already driven over 30% year-over-year growth in Instagram Reels watch time in the US, the system was still "primitive" compared to what’s coming. The company is rebuilding its entire recommendation system into a scalable engineering framework similar to large language models.
On the deployment front, Meta now runs multiple AI models in its ad system, including GEM (Generative Experience Model) for ad ranking, Andromeda for ad retrieval, and Lattice for cross-system ad performance prediction. For example, in Q4 2025, Meta doubled the number of GPUs used to train the GEM model and adopted a new sequence learning architecture. The results were immediate: Facebook ad clicks rose 3.5%, and Instagram conversion rates increased by more than 1%.
Another case in point: the Adaptive Ranking Model, launched on Instagram in Q4 2025. According to Meta, after rollout, ad conversion rates for target users rose 3%, and click-through rates climbed 5%. With Meta’s platforms reaching over 3.5 billion users daily, even a one-point efficiency gain translates into billions in incremental revenue.
From the advertiser’s perspective, AI is fundamentally changing the logic of campaign decision-making. In Q1 2026, more than 8 million advertisers used at least one generative AI ad creative tool. Video generation tools helped advertisers boost conversion rates by over 3%. Meta’s Advantage+ automation suite lets advertisers create assets from scratch using AI for real-time personalized delivery. Ad buying is shifting from "experience-driven" to "algorithm-optimized."
According to WARC Media, Meta’s ad business grew 22% to $196 billion in 2025 and is projected to rise another 22.3% to $240 billion in 2026—far outpacing overall global social media ad market growth. Morgan Stanley’s January 2026 report even predicts Meta’s quarterly ad revenue will surpass Google Search for the first time in Q2 2026—a historic power shift in digital advertising if it materializes.
Monetizing Compute: Meta’s "Second Growth Curve"
If AI-driven efficiency gains in advertising are the "first leg" of Meta’s AI strategy, the planned cloud business is the "second leg"—transforming AI infrastructure from a cost center into a revenue center.
To understand why this is necessary, consider the numbers. Meta’s 2026 AI-related capex guidance is $125–145 billion. Among North America’s Big Four tech giants, Amazon has AWS, Microsoft has Azure, and Google has Google Cloud—all mature cloud businesses that can absorb AI infrastructure investment by selling compute directly to customers. Meta is the outlier. Its data centers and GPU clusters, in theory, only serve its own social platforms, ad systems, and AI R&D.
This creates a huge risk: if Meta’s internal demand for AI compute doesn’t meet expectations, that $145 billion becomes sunk cost. Data centers are built, GPUs bought, long-term power contracts signed—these are fixed investments, not as easily trimmed as a marketing budget.
At the May 27 shareholder meeting, Zuckerberg addressed this concern directly. He said cloud computing is "definitely on the table," revealing that "almost every week, outside companies approach us—either asking for API access or offering to pay a premium for our compute."
Media reports suggest Meta’s cloud business will follow a dual-track model. First, "Model-as-a-Service," letting external developers pay to access AI models hosted on Meta’s infrastructure—including Meta’s in-house Muse Spark, positioned as a rival to AWS Bedrock. Second, a more aggressive approach: leasing out raw GPU compute power—exactly what CoreWeave and Nebius do. No wonder their stocks crashed when the biggest customer announced plans to compete head-on.
On a deeper level, Meta is betting on a long-term scarcity of compute resources as AI workloads shift from training-centric to inference-centric. McKinsey estimates that by 2030, global data centers will require about $6.7 trillion in investment to meet compute demand, with $5.2 trillion of that for inference-side data center capex. The International Energy Agency projects global data center electricity usage will double to about 945 TWh by 2030. Goldman Sachs expects US data center power demand to rise from 31 GW in 2025 to 66 GW in 2027.
If this long-term trend holds, Meta’s cloud business isn’t just about "monetizing surplus capacity"—it’s about securing a strategic position for the coming era of AI inference demand.
Market Divide: Mispricing or Paradigm Shift?
Yet consensus on Meta’s cloud ambitions remains elusive. The expectation reset around "peak AI capex" is rapidly rippling through the global tech investment landscape.
Bears have a direct, clear logic: Meta renting out excess compute means supply now exceeds demand. This suggests Meta will sharply cut new orders for memory chips, HBM, and other hardware. If existing compute is already in surplus, Meta’s appetite for upstream hardware will slow dramatically. This logic punctures the previous market assumption of "infinite AI hardware demand."
Bulls have their own rationale. Some brokerages interpret Meta’s move as a challenge to AWS/Azure/GCP, but fundamentally see it as a way to commercialize massive AI capex. While short-term sentiment may hit cloud providers’ AI premium pricing, it shouldn’t be read as a demand negative for AI hardware. In fact, if Meta can externalize its self-built compute, it actually increases the sustainability of continued investment in GPUs, networking, optical modules, power, cooling, and data centers.
Other analysts point out that while Meta has stockpiled large-scale compute resources, it lacks AI foundation models with true industry competitiveness. Internal demand is insufficient to absorb all its compute, so it chooses to lease surplus capacity to leading third-party AI firms. If this business model proves out, Meta may not cut AI hardware purchases at all—instead, it could accelerate data center expansion to capture cloud market share.
D.A. Davidson analyst Gil Luria raised a sharper critique: Meta’s cloud ambitions signal the company is "giving up on frontier AI" in favor of selling compute. Since launching its superintelligence lab last year, Meta has released new Muse Spark models, but still trails Anthropic and OpenAI.
Meta’s Dual Logic: A Paradigm Shift in AI Commercialization
Viewed together, Meta’s ad efficiency revolution and compute monetization strategy paint a more complete picture.
On the ad side, AI has evolved from an "assistive tool" to "core productivity"—rebuilding the ad system’s operating logic from the ground up and directly converting technical investment into revenue growth. On the compute side, AI infrastructure is shifting from "cost center" to "revenue center"—with cloud services externalizing surplus compute and creating a path to recoup massive capex.
These two logics are not independent—they form a mutually reinforcing loop. Sustained ad growth funds continued AI infrastructure investment, while the cloud business provides "downside protection"—even if internal demand falls short, compute assets can still generate revenue through external leasing. This "dual-engine" structure is central to Meta’s effort to shift the AI narrative from "burning cash" to "making money."
On July 2, 2026, the S&P 500 closed at 7,485.02, down 0.19%. The Dow was flat. The Nasdaq Composite fell 0.66%, while the tech- and momentum-heavy Nasdaq 100 dropped 1.5% to 29,809.13. This market divergence reflects the ongoing revaluation across different links in the AI value chain. Meta soared, hardware crashed—not just a short-term blip, but a sign the market is repricing the entire AI compute value chain.
The market split triggered by Meta’s "compute sales" may be just the first signal of a new phase in AI commercialization. When AI is no longer just a "capex story," but offers dual returns—"efficiency gains" and "asset monetization"—the entire industry’s valuation logic is up for a reset. For investors, understanding the depth and breadth of this paradigm shift is likely far more valuable than debating any single day’s price swings.
FAQ
Q1: Why did Meta’s plan to sell compute through its cloud business trigger a crash in AI hardware stocks?
The market interpreted this as a sign that "AI capex has peaked." If Meta already has surplus compute and needs to lease it out, investors expect Meta’s future purchases of upstream hardware like memory chips and optical components to slow sharply. Previous AI hardware valuations were built on the assumption that "compute is never enough"—this news undermined that foundation.
Q2: How exactly will Meta’s cloud business operate?
Meta has created a "Meta Compute" division to lead this initiative. The business will follow a dual-track model: first, "Model-as-a-Service," allowing external developers to pay for access to Meta-hosted AI models (like Muse Spark), similar to AWS Bedrock; second, direct leasing of raw GPU compute. Both approaches directly challenge Amazon AWS, Microsoft Azure, and Google Cloud.
Q3: Where does Wall Street disagree on Meta’s cloud strategy?
Bears argue this shows Meta’s internal AI compute demand is weak and that the company is "abandoning frontier AI." Bulls see it as a way to commercialize massive AI capex; if the model works, it could actually accelerate data center expansion and drive upstream hardware demand.
Q4: How large is Meta’s 2026 AI capital expenditure?
Meta’s 2026 AI-related capex guidance is $125–145 billion. Previous guidance was $115–135 billion, later raised. This is nearly double the 2025 level, mainly allocated to AI data center construction, compute chip procurement, and model development.
Q5: What specific data demonstrates AI-driven efficiency gains in Meta’s ad business?
In Q1 2026, Meta’s ad revenue grew 33% year-over-year to $55 billion. Ad impressions rose 19%, and average price per ad increased 12%, achieving "volume and price growth." After GEM model improvements, Facebook ad clicks rose 3.5%, and Instagram conversion rates increased by over 1%. Following the rollout of the Adaptive Ranking Model, conversion rates rose 3% and click-through rates 5%.
