Meta raises full-year AI capex guidance to $135–145 billion while posting 14% net profit decline and 8% operating profit decline in Q2
Meta has already lost the first half of the AI game, and now it has no choice but to tell a different story—but so far it hasn't gotten that story quite right." A securities analyst named Wang Hong from a Shenzhen brokerage told Leiphone.net this morning, shortly after Meta released its Q2 earnings.
Net profit plunged 14 percent, operating profit fell 8 percent, free cash flow shrank to just $784 million, and full-year AI capital expenditure was raised yet again to $135 billion to $145 billion—this constellation of figures sent Wall Street into shock.
Interestingly, in the month-plus before the earnings release, Meta moved with remarkable frequency: its internal incubator quietly developed a homegrown model router to rival OpenRouter, "downgrading" AI tasks to save money; starting in June, it imposed token-usage caps on employees and built budget-tracking platforms; in early July, it was revealed to be formally establishing a cloud business division to rent out computing power to external parties.
Contradictory? Not at all. This signals precisely that Meta's AI war has shifted from "paying any price to join the game" to "every dollar must earn its echo"—the second half of the contest.
"The large-model competition landscape is taking shape. The dominant theme of the second half is: AI deployment, profit realization, and telling a coherent ROI story," Wang Hong emphasized.
To understand why Meta suddenly grew eager to "extract money from AI," one must first see how precarious its current hand has become.
Starting from ChatGPT's eruption of large-model fever in early 2023, roughly three and a half years have passed. Among the same cohort of old-guard internet giants, Microsoft seized OpenAI's first-mover advantage and captured the first wave of AI dividends; Google leapfrogged to become firmly ensconced in the first tier with Gemini; Amazon, lacking a powerful model, nonetheless has AWS—the world's largest cloud market share—as a backstop.
From 2023 to 2024, the Llama series was the de facto synonym for global open-source models. Zuckerberg himself took the stage proclaiming, "Open source changes everything."
Back then, the progression from Llama 1 to Llama 3 was a wild ride—from small-parameter versions crushing GPT-3 to a 405B behemoth approaching closed-source peaks. Llama was the undisputed "faith" and industry benchmark of the open-source community.
In the second half of 2024, OpenAI released o1, its chain-of-thought model. In early 2025, China's DeepSeek emerged out of nowhere. Meta began falling behind on the technological front.
A former Meta AI engineer, Li Mao, revealed to Leiphone.net: "At the time, Meta's overarching strategy was to embed AI into existing products, so it shifted more resources toward multimodal and lightweight on-device models, while neglecting to continue pressing the advantage in general-purpose large models. It underinvested in core language-model capabilities like reasoning and chain-of-thought."
"Who expected o1 and DeepSeek to appear out of nowhere, throwing the upper levels into chaos, forcing them to pivot and fight fires. In the scramble, they rushed to ship—with backfired results."
In April 2025, Llama 4 was rushed to market, but the community quickly discovered that the model submitted to leaderboards was a never-before-public "special edition," engulfing it in controversy over "benchmark fraud."
Llama 4's debacle became an inflection point, dragging Meta directly down from its throne atop the open-source model world.
"In the past, Llama ruled the open-source space unchallenged, but now there are far too many better alternatives," multiple veteran AI developers told Leiphone.net.
Statistics show that in 2026, the top 15 on Hugging Face's download rankings are dominated entirely by Chinese models—Qwen, DeepSeek, GLM, Kimi, MiniMax; in April this year, Qwen's single-model downloads reached 942 million, roughly equivalent to two Llamas. On OpenRouter's March call-volume rankings this year, the top four were all Chinese models.
Since Llama 4, Meta has produced no new open-source flagship. Its frontier model Behemoth has been shelved. The Super Intelligence Lab's first model, Muse Spark, released in April this year, went straight to closed-source—yet its API has repeatedly slipped in schedule, never fully opening to developers.
Once, Meta prioritized products at the expense of model fundamentals. But the irony runs deeper: Meta has also stumbled in marrying products with AI.
A former Meta technical director told Leiphone.net that within Meta's core products, "large-model usage traffic accounts for only 1 to 2 percent of platform activity; most still rely on traditional recommendation methods."
In TikTok-style video-browsing scenarios, large models trail far behind traditional recommendation algorithms—the latter can record tens of thousands of user interactions daily, while large models' context windows struggle to digest such massive signals.
A veteran from the open-source large-model community told Leiphone.net candidly: "What remains of Llama is prior ecosystem inertia, not capability leadership. Frankly, Meta can no longer play the open-source card effectively."
Selling models? For now, Meta can't outsell OpenAI and Anthropic. Making chips? Its custom MTIA is still in early stages. Models, compute, chips—Meta needs to dominate at least one. By the math, converting its world-class GPU reserves into a cloud business looks like the most realistic path available.
Recently, research firm Epoch AI published a ranking current through July 24. By equivalent NVIDIA H100 computing power, the global top 60 AI data centers break down as: Google dominates with 16; Meta has 11. In other words, on raw compute, Meta is number two globally, behind only Google.
Jefferies' industry survey also shows Meta's internal infrastructure utilization at roughly 65 percent; the remaining 35 percent spare capacity is ready-made "product." You might say Meta selling compute was inevitable.
Consider this: in the first half of the AI race, Meta was actually relatively conservative on capital expenditure. But in the past year, the wind shifted sharply.
In 2025, Meta spent $72.2 billion on capex; in 2026, the forecast has swung from a range of $115 billion to $135 billion all the way up to $125 billion to $145 billion. The bill nearly doubled in a single year.
In this latest earnings call, Meta again raised its 2026 full-year capex floor by $10 billion, to $135 billion to $145 billion, ratcheting up expectations repeatedly over mere months.
Q2 earnings show Meta's single-quarter capex at $31.08 billion, surging 83 percent year-over-year; free cash flow down to $784 million, a staggering 91 percent plunge. Essentially, Meta's cash flow has been nearly consumed entirely by the frenzy of AI infrastructure buildout.
Meta's metaverse bet still bleeds: Reality Labs, the division responsible for that business, has accumulated losses exceeding $80 billion since its 2020 founding; in 2025 alone, it lost $19.193 billion.
In many observers' view, the metaverse gamble's catastrophic failure has weighed heavily on Meta's management psyche—making AI a battle that simply must be won.
The earnings figures betray this deep anxiety: Meta's enormous AI research and compute capital spending remains stubbornly high, propped up almost entirely by advertising profits from its apps family.
But troubles compound. The cash cow that once seemed safest is beginning to slip. Waves of new AI products and technologies are eroding the advertising conversion efficiency of traditional feeds and recommendation algorithms. TikTok, meanwhile, continues displacing Instagram and Facebook in the young-user market.
As multiple secondary-market investors told Leiphone.net, in Wall Street's eyes, Microsoft's Azure and Google's GCP can fold AI spending directly into cloud-revenue growth curves. Meta, dominated by consumer-facing business, can monetize AI almost exclusively through advertising.
Therein lies Meta's anxiety: it needs a revenue channel beyond advertising—a way to convert massive AI investment into Wall Street-legible income.
In the eyes of many former Meta employees, over the past year-plus, the organizational and personnel turmoil has been unmatched among Silicon Valley giants.
In June 2025, Zuckerberg spent $14.3 billion to acquire roughly 49 percent of non-voting equity in data-labeling company Scale AI, poaching its 28-year-old founder Alexandr Wang to become Meta's first-ever Chief AI Officer.
On June 30, Meta's Super Intelligence Lab (MSL) was established, with Alexandr Wang and former GitHub CEO Nat Friedman co-leading.
Then came a rare Silicon Valley talent war: poaching multiple GPT-4o core researchers from OpenAI, with Zhao Sheng-jia appointed chief scientist; recruiting Pang Ruoming, Apple's head of large-model teams, with a compensation package exceeding $200 million. To the point that OpenAI Chief Research Officer Mark Chen publicly complained, "It feels like someone broke into our house and stole things."
In August 2025, MSL was reorganized into four pillars: Wang's TBD Lab, the venerable FAIR research institution, PAR (product-application research) under Friedman, and the infrastructure team—Meta's fourth AI-team reorganization in just half a year.
In May 2026, roughly 7,000 employees were reassigned to AI-related workflows, marking Meta's largest-ever internal talent shuffle.
Acquisitions didn't stop. Late 2025 saw reports of Meta acquiring agent-startup darling Manus, though Chinese regulators later blocked it. The acquisition also exposed fierce infighting between Meta's old guard and new blood internally—see "Exclusive: The Manus Acquisition's Cancellation and Lingering Turmoil: Meta's AI Power Struggles Reach a Turning Point" for details.
In October 2025, the AI division laid off 600 people; FAIR took a major hit. In May 2026, Meta cut roughly 8,000 more—10 percent of its workforce—and canceled 6,000 open headcount requisitions. An internal memo characterized it as "the human cost of AI transformation."
Last October, Tian Yuandong, a Chinese scientist who had worked at Meta for a decade with over 18,000 paper citations, departed with his team as part of an optimization. In November, LeCun, Meta's chief AI scientist for 12 years, left to start his own venture.
A Meta technical director told Leiphone.net that the company is pivoting from its historical "bottom-up" culture of small-team autonomy toward a "top-down" model where leadership sets direction and teams follow. Once the upper echelon's thinking becomes muddled, teams easily slip into redundant competition and rank-and-file guesswork.
A recruiter close to Meta told Leiphone.net: "Meta can afford Silicon Valley's costliest compensation packages, but a year into the Super Intelligence Lab, product leads and star researchers come and go. Money can buy people; it can't buy stability and consensus."
Today, the AI industry and capital markets are undergoing a profound transformation: from the first half's focus on technological breakthroughs and compute buildup to full pivot into the second half—a battle over free cash flow, ROI, and commercial delivery.
Yet the reality is stark: tech giants stack capex higher and higher; free cash flow shrinks sharply; the vast sums poured into AI show no returns in sight.
Markets have given a cold response to this "money-swallowing" state of affairs. Take Google, which reported earnings last week: despite cloud revenue exploding 82 percent, capex doubled and free cash flow turned negative—yet the stock still fell. Microsoft, Amazon, and Meta face similar scrutiny.
The AI industry has shifted from "competing on spending" to "competing on returns." Model capability gaps are narrowing, but commercial-loop gaps are just beginning to widen.
Leiphone.net will continue tracking Meta's earnings and the commercialization progress of global AI giants. Readers are welcome to add author xf123a on WeChat for discussion and exchange.