Meta ships its first proprietary AI model after a bruising delay, but soaring costs and a Q2 earnings miss have investors questioning the $145 billion bet
July 29th, 2026: Meta misses Q2 profit estimates as costs rise 55% year over yearNew here? Follow stories to track developments over time. Create a free account to get updates when stories you care about change.
Meta shipped Muse Spark, code-named Avocado, in April 2026—its first proprietary foundation model and a ground-up departure from the Llama architecture. The model ranked fourth on Artificial Analysis's Intelligence Index at launch, behind rivals at Google, OpenAI, and Anthropic, but Meta's stock rose about 10% in the five days after release.
The financial picture has since darkened. Meta raised its 2026 spending forecast to $125–145 billion at Q1 earnings, then missed Q2 profit estimates—costs rose 55% and free cash flow fell from $8.55 billion to $784 million.
Why it matters
The AI model behind Facebook, Instagram, and WhatsApp shapes what three billion people see, read, and believe every day.
No questions yet — be the first to ask.
Mark Zuckerberg, Founder, Chairman and Chief Executive ...
Amortized hardware and energy cost to train frontier AI models over time
Training compute of frontier AI models grows by 4-5x per ...
Inside the Frontier AI Model Race: Releases, Regulation, and ...
Joyce Lin
America and China Are Racing to Different AI Futures
Center for Humane Technology
Meta reportedly delays flagship 'Behemoth' AI model
CNBC Television
Images from Openverse under Creative Commons licenses. Videos from YouTube.
Curated perspectives — historical figures and your fellow readers.
Fictional AI pastiche — not real quote.
"They spent six hundred billion dollars to discover what every writer has always known: you cannot buy the muse, only the desk she refuses to sit at."
Fictional AI pastiche — not real quote.
"Aye, Mr. Zuckerberg has learned what every steelman knows: you can buy a furnace, but you cannot buy the knowledge of how to run it — and six hundred billion dollars' worth of brick and iron is worthless without the men of genius to light the fire within."
Exploring all sides of a story is often best achieved with Play.
Can you match the quotes to the right people?
— Who said this?
Tip: press 1– to answer.
points — sign up to put that on the leaderboard.
A number from this story, against one from elsewhere in the news — guess which is bigger, then keep the chain going. 5 rounds, 3 strikes; a miss costs a strike and resets your streak.
Keyboard: ↓/L lower · ↑/H higher
0 points — sign up to put that on the leaderboard.
Order five events from this story, oldest at top. Each in the right slot scores 1 — neighbours within one slot count too. Your previous result — green ✓ for exact slots, yellow ~ for off by one. Cards now in true chronological order.
Sign up to save your score and track a streak across stories.
Sixteen names from the news. Find the four hidden groups of four. Four mistakes max.
Sign up to keep a daily streak — a new puzzle lands every day.
Your progress in this debate will be lost.
Choose one persona for each side of the debate
Choose personas with different perspectives for a more dynamic debate.
Select debater for this side:
No debate personas available right now.
Select debater for this side:
No debate personas available right now.
Make your prediction before the referee scores
The referee scores both sides on
Pick the question both personas must answer in the final round
Debate Oracle! You called every round!
Sharp Instincts! You know your debaters!
The Coin Flip Strategist! Perfectly balanced!
The Contrarian! Bold predictions!
Inverse Genius! Try betting the opposite next time!
Parent company of Facebook, Instagram, and WhatsApp, now spending more on AI development than any other corporation.
Meta's consolidated AI division, formed by merging its generative AI, Llama, and product teams under a single mandate to build superintelligence.
April 2025 July 2026
Meta reported Q2 revenue of $60.8 billion, up 28%, but missed EPS estimates by nearly $1 per share. Total costs rose 55% to $42 billion, including $2.4 billion in legal charges and $1.2 billion in severance. Free cash flow fell from $8.55 billion to $784 million, and Meta stock fell nearly 10% in after-hours trading.
Meta launched Muse Spark 1.1 three months after the original model, with improvements to coding and multi-step agent performance. Independent testing by Vals AI measured a score of 69 on a standard coding benchmark; Meta's own evaluation claimed 80. It is priced at roughly one-third the cost of top-tier rival models.
Google restricted the volume of Gemini API capacity it would sell to Meta, citing its own supply constraints. Meta had been using Gemini for content moderation, scam detection, and internal coding tools; the cutback pushed Meta to rely more on Muse Spark for those workloads.
Meta reported Q1 revenue of $56.3 billion, up 33%, with net income of $26.8 billion. The company raised its 2026 capital expenditure guidance to $125–145 billion from a prior $115–135 billion, citing higher component costs. Stock fell more than 6% in after-hours trading.
Meta released Muse Spark, the model code-named Avocado, as the first output from Meta Superintelligence Labs—a ground-up rebuild that ranked fourth on Artificial Analysis's Intelligence Index, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Meta's stock gained roughly 10% over the five trading days that followed.
Internal testing showed Avocado trailing Google, OpenAI, and Anthropic models on reasoning, coding, and writing. Meta pushed the launch from March to at least May. Leadership discussed temporarily licensing Google's Gemini to power Meta AI products. Meta shares fell 4.5%, the steepest drop in over four months.
Google released Gemini 3.1 Pro with a 77.1% score on the ARC-AGI-2 benchmark—more than doubling its predecessor. Combined with strong releases from OpenAI and Anthropic in February, the competitive bar rose sharply just as Meta was finalizing Avocado.
Meta announced 2026 capital expenditure guidance of $115–135 billion, nearly double the prior year. Chief financial officer Susan Li said the company remained "capacity constrained" and that compute demand was growing faster than supply.
Reports confirmed that Meta's next flagship AI model, code-named Avocado, would be closed-source and proprietary—abandoning the open-weight philosophy that had defined the Llama series and attracted a large developer community.
LeCun, a Turing Award winner who founded Meta's AI research lab 12 years earlier, confirmed he would leave to build Advanced Machine Intelligence Labs, a startup pursuing a fundamentally different approach to AI.
After years of championing open-source AI models, Zuckerberg said Meta would likely not open-source all of its superintelligence models, citing competitive and security concerns—partly driven by Chinese firm DeepSeek's use of Llama's architecture.
Zuckerberg announced the creation of Meta Superintelligence Labs in an internal memo, consolidating the company's AI research, model development, and product teams under a single division led by Wang.
Meta acquired a 49% nonvoting stake in Scale AI and recruited its 28-year-old founder Alexandr Wang as Meta's first chief AI officer, making him one of the highest-paid employees in the technology industry.
Chief product officer Chris Cox restructured Meta's AI teams into an AI Products group (led by Connor Hayes) and an AGI Foundations unit (co-led by Ahmad Al-Dahle and Amir Frenkel) to speed product development.
Meta's largest planned model, Llama 4 Behemoth, was pushed from early summer to fall 2025 or later. Engineers grappled with whether its improvements over earlier versions justified a public release. The model was eventually shelved.
Meta released Llama 4 Scout and Maverick, but the launch was overshadowed by accusations that a specially crafted, unreleased variant had been submitted to the LM Arena benchmark to inflate scores. Meta denied the claims.
Joelle Pineau, vice president of AI research who led Meta's Fundamental AI Research lab since 2023, announced she would leave the company. She later joined Cohere as chief AI officer.
3 moments from history that rhyme with this story — and how they unfolded.
In June 2000, Yahoo replaced its in-house search technology with Google's engine, making Google the system powering Yahoo's search results. The deal gave Google a "Powered by Google" credit on one of the internet's most-visited pages, exposing millions of users to the Google brand for the first time.
Yahoo got a better search product and could focus on media and content. Google gained massive visibility and traffic.
Google used the exposure to build its own destination site, launched AdWords, and became the dominant search engine. When Yahoo ended the deal in 2004 and built its own search, it was too late to reclaim the market.
If Meta licenses Gemini, it risks a similar dynamic: Google gains public validation that its AI model is superior enough for a competitor to pay for, while Meta's own brand becomes associated with someone else's technology. The longer the licensing lasts, the harder it becomes to justify switching back.
Apple had built its Mac identity around PowerPC chips developed with partners IBM and Motorola. When those partners failed to deliver processors competitive with Intel's x86 chips—particularly on laptop power efficiency—Steve Jobs announced at the 2005 Worldwide Developers Conference that all Macs would transition to Intel hardware, a move many considered unthinkable.
Mac performance improved significantly, and the transition was completed faster than expected. Developers adapted within two years.
Apple treated Intel as a bridge, not a destination. It quietly developed its own ARM-based chips and in 2020 launched Apple Silicon, leaving Intel behind entirely. The interim dependency funded the time needed to build a superior alternative.
Meta's potential Gemini licensing could serve a similar bridge function—buying time to improve Avocado without shipping an inferior product. The question is whether Meta, like Apple, treats the dependency as temporary and invests in catching up, or whether it becomes a crutch.
IBM, the dominant computer company, needed an operating system for its new Personal Computer and licensed DOS from a small company called Microsoft rather than building its own. IBM believed hardware was the real competitive advantage and that the operating system was a commodity component. The deal gave Microsoft the right to license DOS to other manufacturers.
IBM shipped the PC on schedule and dominated the early personal computer market.
Clone manufacturers used the same Intel chips and Microsoft software to build cheaper alternatives. IBM's hardware advantage eroded, and Microsoft's Windows became the industry's true platform. Microsoft eventually became the world's most valuable company.
Meta's $600 billion infrastructure bet assumes data centers and compute are the durable competitive advantage. But if the AI model—the software layer—is where differentiation actually lives, Meta risks building enormously expensive infrastructure that any capable model provider could theoretically run on. The IBM precedent suggests that the company controlling the software platform, not the hardware, often captures the most value.