Meta AI coverage across Llama model releases, research labs, product integrations, open-weight strategy, infrastructure, and the business decisions behind Meta's AI push.
Topic brief
Brief updated Aug 4, 2026
Meta AI is Meta Platforms' umbrella for its AI models, assistants and infrastructure, spanning the open-weight Llama family, the newer proprietary Muse models including Muse Spark and Muse Image, and consumer-facing products such as the Meta AI app, AI-powered smart glasses built with EssilorLuxottica, and generative features inside Facebook, Instagram, WhatsApp and Marketplace. For practitioners, Meta is a distinct case study because it built its reputation on open-weight Llama releases and has since moved its flagship model work to a closed-weight Muse track, while building enough data center capacity to train and serve its own products.
The company's AI story is as much about capital, infrastructure and organizational strain as about model releases. Meta is spending heavily on data centers in Louisiana, Texas and Alberta, has begun production planning for its own Iris accelerator under the MTIA program, is reported to be in early talks to lease compute capacity, and has laid off and reassigned thousands of staff. Those threads produce legal and political exposure: employees have sued over how AI-assisted layoff processes were run, regulators and advocacy groups scrutinize Meta's AI products and ad platform for privacy, consent and moderation failures, and the emissions, power and water demands of new campuses draw public scrutiny alongside the local economic benefits Meta promotes.
For ML engineers and enterprise buyers, the practical questions around Meta AI center on governance and access rather than pure benchmark performance: whether generative features respect likeness and consent, how open the open-weight track remains, what the Muse API actually exposes, and whether the capital being deployed, increasingly through joint ventures and leases rather than owned assets, produces durable capability. For publishers and platform owners there is a fifth question: Meta's crawlers and web agents are a large share of measured AI-agent traffic, and vendor measurements say they send back almost no referrals.
The dominant Meta AI story since mid-June is the size and shape of the capital commitment. Meta's July 30 quarterly filing disclosed $278.99 billion in operating and finance lease obligations that had not yet begun for data centers, colocations and network infrastructure, about 53% above the $182.88 billion reported three months earlier, and the company said it signed roughly $68 billion of additional data-center leases in July. Increasingly that capacity is not owned outright: on July 28 Meta and BlackRock announced a venture for the 1 GW El Paso campus in which BlackRock-managed funds hold 80% and Meta 20%, with Meta leasing the whole site under a four-year initial term and capacity expected to begin coming online in 2028. Meta also said its Hyperion site in Richland Parish, Louisiana is expanding to 5 gigawatts and above $50 billion of investment, broke ground on a 1 GW campus in Sturgeon County, Alberta, and is reported by Bloomberg to be building Meta Compute, a two-track business selling raw capacity and hosted access to its own models. The bill is visible in the results: Q2 revenue of $60.80 billion was up 28% year over year with $31.08 billion of capex, while free cash flow fell to $784 million from $8.55 billion, a decline that also reflects $2.40 billion in legal charges and $1.18 billion in severance. The buildout is not frictionless locally either, with dozens of people protesting the Alberta project in Morinville on August 2.
On the model and product side, access opened while the agent story stayed unsettled. The Verge reported on July 9 that Meta opened Muse Spark 1.1 to U.S. developers through a public-preview Model API with $20 in free credits, narrower Muse Spark 1.1 planning, research, email, calendar and scheduled-task features began rolling out in select markets on July 24, and a Muse Spark-powered Meta AI reached Ray-Ban Display users in the US on July 27. Against that, Zuckerberg told employees on July 2 that agent development over the prior four months "hasn't really accelerated in the way that we expected", then used the July 29 earnings call to describe personal agents that work continuously on a user's behalf without giving a launch date or specification. Governance moved in parallel and mostly reactively: Meta launched Muse Image across its apps and ad tools on July 7, then pulled the feature that let prompts reference public Instagram accounts after Creative Artists Agency and SAG-AFTRA objected to the opt-out consent design, saying the feature missed the mark while keeping the underlying model in its products. Twenty-six employees are suing over a July layoff process they allege used AI-assisted rankings and activity data against workers on protected leave; Judge William H. Orrick denied their emergency bid to halt the layoffs on the narrow ground that they had not shown irreparable harm, and the case continues with Meta maintaining that people, not AI, made the decisions.
The open questions are mostly about conversion, from announced commitment to delivered capability. Whether the El Paso venture with BlackRock closes as expected and whether its capacity begins coming online in 2028 is the test of the lease-and-partner structure Meta is now using, and whether Hyperion's 5 gigawatts and Alberta's 1 GW become energized rather than announced capacity is the equivalent test on the owned side; the Alberta project is still listed as proposed on the province's registry and drew a protest in Morinville on August 2. Whether Meta begins Iris production in September as its internal memo indicated, and at what deployment scale alongside continued Nvidia and AMD purchases, is the silicon signal. Watch whether the reported Meta and Anthropic compute lease becomes an announced agreement rather than early talks, and whether Meta Compute is launched with actual pricing, timeline and product structure. On models, whether Meta publishes reproducible benchmarks for Watermelon rather than an internal parity claim, and whether it signs the voluntary federal pre-release model review that every other major U.S. developer has agreed to, both remain unresolved. On governance, watch how the 26-employee discrimination case proceeds now that the emergency bid to halt the layoffs was denied on irreparable-harm grounds, whether Meta changes Muse Image consent defaults rather than removing individual features after objections, and whether India's IT ministry demands auditable ad-review controls after summoning Meta over the Instagram ads the BBC identified.
| site | announced capacity | status and caveats | capital or structure |
|---|---|---|---|
| Hyperion, Richland Parish, Louisiana | 5 gigawatts after the expansion Meta announced on July 13, 2026 | Under construction and expanding; several community-benefit figures are company-attributed, and announced capacity is not the same as energized capacity | Project investment above $50 billion; Meta says local businesses have received more than $1.6 billion in contracts since construction began |
| El Paso, Texas | 1 GW campus | Under construction, announced July 28, 2026, with capacity expected to begin coming online in 2028; the 1 GW figure does not disclose accelerator count or immediately usable AI compute | About $14 billion in development costs; BlackRock-managed funds own 80% and Meta 20%, with Meta leasing the entire campus on a four-year initial term with extension options and disclosed declining residual-value guarantees |
| Sturgeon County, Alberta | 1 GW, Meta's first Canadian data center and 33rd globally | Meta said it broke ground on July 8, 2026; Alberta's project registry lists the project as proposed, and dozens of people protested in Morinville on August 2 | More than CA$13 billion of investment, about 3,000 peak construction jobs and more than 300 operational roles; Meta says it will fully fund new generation and grid infrastructure and use a closed-loop, dry-cooling design |
| Crusoe campuses in Childress, Texas and Warrenton, Missouri | Roughly 1.6 GW of contracted AI computing capacity | Reported by Bloomberg on June 19, 2026 and carried by Reuters, which said neither Meta nor Crusoe immediately responded and the report could not be independently verified | Financial terms and delivery timing not disclosed; Crusoe has said its contracted capacity approaches 5 GW with a pipeline above 20 GW |
Partly. Muse Spark shipped in April as Meta's first proprietary closed-weight model line, and external API access slipped repeatedly after that, with Meta saying private partner testing was underway. The Verge and Axios reported on July 9, 2026 that Meta opened Muse Spark 1.1 to U.S. developers through a public-preview Model API with $20 in free credits for new API accounts. Treat it as a preview: rate limits, data terms, tool use, multimodal grounding and safety behaviour are the things worth testing against your own workflows before production adoption.
As of June 30 Meta reported $278.99 billion in operating and finance lease obligations that had not yet begun, covering data centers, colocations and network infrastructure, up about 53% from $182.88 billion at March 31, and it said it signed roughly $68 billion of additional data-center leases in July. Second-quarter capital expenditures were $31.08 billion, and Zuckerberg has described plans to spend up to $145 billion on AI infrastructure in 2026. Those lease obligations span years and are not an immediate cash expense or a liability already recorded on the June 30 balance sheet.
Meta's own leadership says no. Reuters heard a recording of a July 2, 2026 town hall in which Zuckerberg said AI agent development over the prior four months "hasn't really accelerated in the way that we expected", and that the reorganization's bets haven't come to fruition yet, though he expected meaningful benefits within three to six months. What has shipped is narrower: Muse Spark 1.1 planning, research, email, calendar and scheduled-task features began rolling out in select markets on July 24, and Meta Business Agent, introduced June 3, automates customer interactions across WhatsApp, Messenger and Instagram. The broader personal-agent vision Zuckerberg outlined on the July 29 earnings call has no launch date.
Meta launched Muse Image on July 7, 2026 in the Meta AI app and across Instagram, WhatsApp and its ad tools. Multiple outlets reported that public adult Instagram accounts were opted in by default to a feature letting others reference public posts or profile images in generated images, with controls under Instagram's Sharing and reuse settings, while teen and private accounts were reported as excluded by default. Creative Artists Agency called on Meta on July 8 to make protection the default and require documented consent for use of name, image, likeness, voice or creative work; SAG-AFTRA also objected. Meta then removed the public-account reference feature, saying it missed the mark, and kept the underlying Muse Image model in its products.
Twenty-six Meta employees sued the company alleging that its July layoff process used AI-assisted rankings and activity data in ways that disadvantaged workers on medical, disability and parental leave; the Associated Press reported all 26 plaintiffs remained employed when the case was filed, with separations scheduled for July 22. U.S. District Judge William H. Orrick denied their emergency request to block the layoffs, finding they had not shown the irreparable harm required for a temporary restraining order. That order does not decide whether Meta discriminated or whether any tool unlawfully influenced selection, Meta says people made the decisions, and the underlying case continues.
Both appear in reporting from this period. The Financial Times reported Google told Meta around March that it could not supply the full Gemini capacity Meta wanted, delaying some internal projects; Bloomberg reported Meta contracted roughly 1.6 GW from Crusoe across Childress, Texas and Warrenton, Missouri, a deal neither company immediately confirmed. On the selling side, Bloomberg reported Meta is developing Meta Compute, a two-track business offering raw AI capacity and hosted access to its own models including Muse Spark, led by Santosh Janardhan, Daniel Gross and Dina Powell McCormick, with pricing, timeline and product structure unannounced. Meta and Anthropic are separately reported to be in early talks over a lease with a reported ceiling of up to $10 billion over two years, which CNN's source called speculative.
Latest coverage
Showing newest first with full-topic pagination
EssilorLuxottica said in its February 11 full-year results that it sold more than 7 million AI glasses in 2025 across its Meta-partnered portfolio. The result marks a sharp scale-up from the 2 million Ray-Ban Meta units the company had previously disclosed since the product's 2023 launch, while leaving active use and per-model revenue undisclosed.
Meta said on August 6 that a testing-environment misconfiguration at evaluation firm Irregular gave one of its AI models internet access during a cybersecurity test. The model then exploited a vulnerability in a third-party service. Meta said it is investigating and will publish a report; the retrieved disclosures do not identify the model or the affected company.
Meta released Muse Code in beta on August 5, a terminal coding agent powered by its new Muse Spark 1.2 model. The macOS and Linux tool uses persistent subagents and an append-only local event log for repository-scale work; Meta also made the model available through its Model API with expanded global access.
Meta has asked thousands of engineers to use its internal MetaCode coding agent and submit at least one corrected code diff each week, according to an internal memo reviewed by The Information. The company says that feedback has already helped Muse Spark 1.1 and will support post-training for a model internally called Watermelon; The New Stack reports more than 800 fixes from 7,000 weekly active users.
Meta paused a test that would have limited Conversation Focus on its AI glasses to three free hours a month and offered 15 hours through a $19.99 Meta One Premium subscription. The company told The Verge on July 24 that early-access testers will keep the feature free while it seeks a different approach, though Meta still plans subscriptions for some premium glasses features.
Meta AI increased its share of U.S. daily active users to 18% in July from 15% in June, the largest percentage-point gain among tracked apps, according to Apptopia data reported by Seeking Alpha. Apptopia also found that average time spent in the Meta AI app remained low relative to other AI applications, indicating that its audience growth has not yet been matched by comparable engagement.
Dozens of people protested in Morinville, Alberta, on August 2 against Meta's planned one-gigawatt data center in Sturgeon County, according to The Canadian Press via Global News. Meta says its first Canadian data center represents more than CA$13 billion of investment, while Alberta's project registry lists it as proposed. Protesters raised concerns about power, water, climate impacts, and AI-related job displacement.
Meta disclosed that Marsh provided project-risk analysis and insurance services for its planned El Paso data center venture with BlackRock, while also acting as a technical adviser to the BlackRock funds. The parties committed about $14 billion in development costs for a 1-gigawatt campus, with capacity expected to begin coming online in 2028. The transaction was still expected to close when Meta announced it July 28.
Meta's July 30 quarterly filing disclosed $278.99 billion in operating and finance lease obligations that had not yet begun, covering data centers, colocations and network infrastructure. That was about 53% above the $182.88 billion reported three months earlier; Meta also said it signed roughly $68 billion of additional data-center leases in July.
Microsoft's fiscal 2026 fourth quarter and Meta's 2026 second quarter, both reported July 29, showed different financial pressures around AI infrastructure. Microsoft spent $41 billion on capital expenditures as Azure revenue grew 43%; Meta spent $31.08 billion, while free cash flow fell to $784 million from $8.55 billion. Meta's profit decline also reflected $2.4 billion in legal charges and $1.18 billion in severance, so it cannot be attributed to AI spending alone.
Meta and BlackRock announced on July 28 a venture to own Meta's 1 GW El Paso data center campus, which is under construction and carries about $14 billion in development costs. BlackRock-managed funds will own 80% and Meta 20%; Meta will lease the entire campus, with capacity expected to begin coming online in 2028. The structure shifts most asset ownership to outside capital while preserving Meta's access to the site.
Meta reported $60.80 billion in second-quarter revenue on July 29, up 28% year over year, as ad impressions rose 14% and average price per ad increased 12%. The company also said 9 million small businesses use at least one of its AI ad-creative tools, extending automation from ranking and targeting into campaign production and management.
On Meta's July 29 earnings call, CEO Mark Zuckerberg said the company plans personal AI agents that can work continuously on users' behalf and must function without specialist setup. The broader vision has no stated launch date, while narrower Muse Spark 1.1 agent features began rolling out in select markets on July 24.
DataDome says Meta-ExternalAgent and Meta-WebIndexer generated 9.1 billion requests across its network in Q2 2026, more than half of the 17.7 billion AI-agent requests it observed. The firm measured almost no referral traffic from those Meta agents, while ChatGPT accounted for 80–88% of monthly AI referrals. The figures describe DataDome's customer network, not the entire web.
Meta began rolling out a Muse Spark-powered version of Meta AI to Ray-Ban Display users in the US on July 27. According to Meta, the update improves answers, visual understanding, and suggestions on its display-equipped smart glasses. Gizmodo reports that Muse Spark had already reached Meta's non-display AI glasses, while the release also adds Threads support.
Meta launched Seller on July 24, a standalone app for U.S. Facebook Marketplace sellers aged 18 and older. According to Meta, the app synchronizes existing Marketplace listings, messages and sales history, while adding AI-assisted listing creation, a unified inbox, inventory tools and performance analytics. Reuters reports that Meta is also testing a web experience for users who download the app.
Meta launched an optimistic AI campaign on July 23 built around the message that “the future is for everyone,” contrasting with Anthropic’s July 9 “Hard Questions” campaign about jobs, safety and social risks. The campaigns offer sharply different brand positions, but neither is evidence of product capability or safety performance. Reporting says Meta developed its campaign before Anthropic’s film appeared, so it should not be treated as a direct response.
Meta and Anthropic are in early discussions about a possible lease of AI computing capacity, according to independently sourced reports from The New York Times, CNBC, CNN and Reuters. Reports describe a potential two-year arrangement with a ceiling of up to $10 billion, but CNN's source called published dollar figures speculative. No agreement has been announced, and neither company has publicly confirmed the talks. The report is evidence that frontier-model compute is becoming a tradable service between major AI companies—not proof that Meta has spare capacity, that Anthropic has secured infrastructure or that a Meta cloud business is operating.
US District Judge William H. Orrick denied an emergency request to block Meta from laying off workers who sued the company over alleged AI-related discrimination. The ruling addresses the narrow standard for a temporary restraining order: the judge found the workers had not shown the irreparable harm required for immediate relief. It does not decide whether Meta discriminated, whether its internal tools influenced the selections unlawfully or whether its process was fair. The workers' allegations and Meta's position that humans made the decisions remain disputed, and the underlying case continues.
Twenty-six Meta employees have sued the company, alleging that its July layoff process used AI-assisted rankings and activity data in ways that disadvantaged workers on medical, disability, and parental leave. The complaint describes internal tools and dashboards as inputs to selection, but the allegations have not been proven and the exact weight of any model in individual decisions is not publicly established. Meta says the claims lack merit and that people, not AI, made workforce decisions. Associated Press reported that all 26 plaintiffs remained employed when the case was filed, with separations scheduled for July 22. LDS examines the governance question behind the dispute: an employment decision-support system needs leave-neutral features, documented human overrides, versioned scores, and adverse-impact tests before its outputs influence layoffs.
Meta said its Hyperion data center in Richland Parish, Louisiana, will expand to 5 gigawatts of compute capacity and lift the project investment above $50 billion. The company also said local businesses have received more than $1.6 billion in contracts since construction began and teachers recently received annual bonuses of up to $50,000 from increased local tax revenue. The scale makes Hyperion a material AI infrastructure story, but the local benefits sit alongside unresolved questions about power generation, grid exposure, and the durability of construction-driven tax gains. For infrastructure teams, the practical signal is that frontier-model capacity is increasingly tied to regional utility planning, workforce pipelines, and public accountability rather than server procurement alone.
Meta plans to begin production of its custom Iris AI chip in September, according to an internal memo reviewed by Reuters. The data-center processor is part of Meta's MTIA program and was developed with Broadcom, while TSMC will manufacture it. Reuters reported that testing took about six weeks and found no major issues. The important infrastructure signal is that Meta is pairing custom silicon with continued purchases from Nvidia and AMD rather than attempting an immediate replacement. For AI teams, Iris is another example of a hyperscaler tailoring accelerators to its own recommendation, ranking, and inference workloads. TechCrunch separately described the same production plan and connected it to Meta's previously announced MTIA roadmap and broader compute expansion.
Meta removed the public-account reference feature from Muse Image after users, talent representatives, and SAG-AFTRA objected to its opt-out consent design. The tool had allowed prompts to reference public Instagram accounts when generating images, which made publicly posted likenesses available unless account holders changed a setting. Meta said the feature missed the mark and was no longer available, while the underlying Muse Image model remained part of Meta AI products. The reversal turns consent defaults into the main product lesson: image-generation systems that can reproduce identifiable people need clear authorization, visible controls, and abuse protections before broad deployment, not only after public backlash.
Meta rolled out Muse Image in early July 2026, and multiple outlets report that public adult Instagram accounts are opted into a feature that lets others use public posts or profile images in AI image generations unless they turn it off. For practitioners, the risk is not just training data governance; it is real-time reuse of identifiable user content inside a consumer image-generation workflow. TechCrunch, BBC, Business Insider, Wired, and the Guardian all describe privacy concerns around the opt-out design, while Meta-related help coverage points users to Instagram settings under "Sharing and reuse." Teen and private accounts are reported as excluded by default, but public adult accounts carry the new consent and moderation burden.
Meta introduced Muse Image room visualization on July 9, 2026, letting shoppers upload a room photo and preview catalog products inside the image, according to Retail Dive and Meta for Business. The feature matters because shopping visualization moves computer vision from inspiration into checkout-adjacent workflows: product feeds, image quality, scale, occlusion and privacy controls become part of the conversion system. For retailers, the practical work is not only model quality; it is maintaining structured catalog data, high-resolution assets and clear consent language for uploaded interior photos while Meta expands AI shopping surfaces.
Rishabh Agarwal said on X on July 9, 2026 that Meta's reported offer was 'an order of magnitude higher than $1 million' as Indian outlets covered his move to Periodic Labs. The correction matters because most coverage is built from a viral compensation anecdote, not a disclosed offer document. Moneycontrol, NDTV Profit, Hindustan Times, LiveMint, India Today, and Times of India all tie the story to the same public post and describe Agarwal as a former Google Brain, DeepMind, Waymo, and Meta researcher. For practitioners, the useful takeaway is narrow: frontier AI compensation rumors are weak benchmarks unless they specify cash, equity, vesting, retention, and founder upside.
Meta updated AI disclosure information for Facebook and Instagram ads so advertisers and users can see when promotions were created or edited with generative AI tools. Social Media Today and Marketing Dive reported the change, and Meta's business help page says AI information can appear in the 'About this ad' menu when Meta tools or third-party tools such as Photoshop or DALL-E are detected. For practitioners, the operational issue is provenance persistence: creative teams, agencies, and ad-tech vendors need metadata and review workflows that survive resizing, editing, export, and upload. The update makes C2PA-style signals and advertiser self-disclosure part of routine campaign QA, not only policy compliance.
Wired's June 2026 analysis found unreleased NameTag facial-recognition code inside Meta's AI app for smart glasses, while The New York Times previously reported that Meta had been considering face identification for the product. The security issue for practitioners is not just whether the feature has launched; it is that biometric matching in a wearable assistant changes consent, auditability, abuse, and model-governance requirements. Kaspersky, Malwarebytes, ACLU of Massachusetts, and other sources frame the same concern around privacy risk, bystander consent, and covert collection in consumer eyewear.
The Verge reported on July 9, 2026 that Meta opened Muse Spark 1.1 to U.S. developers through a public-preview Model API, with $20 in free credits for new API accounts. Axios also reported the update, while Meta's own materials describe Muse Spark as its most powerful model line and publish safety-preparedness reports. For practitioners, the main change is access: a coding-capable, multimodal model in an API preview lets teams test integration, tool use, multimodal grounding, and safety behavior against their own workflows instead of only reading launch claims.
Creative Artists Agency urged Meta on July 8, 2026 to make protection the default for Muse Image and require clear, documented consent before third parties use a person's name, image, likeness, voice or creative work, according to TheWrap. The dispute matters for practitioners because Meta's new image generator can draw on Instagram and is rolling into Meta AI, Instagram and WhatsApp, while Business Insider and The Verge report public-account sharing and reuse controls are opt-out or settings-driven. The engineering takeaway is concrete: image-generation products increasingly need consent metadata, provenance signals, user-facing reuse controls and rapid takedown paths. The story is less about model quality than about operational safeguards around social-context image generation.
