META (Minutes: No 2027 Capex guidance; racing to build compute capacity in the next two years.

Dolphin Research's transcribed notes on $ Meta Platforms.US FY26Q2 earnings call

I. Key takeaways

1) Guidance: 26Q3 total revenue guided to $61.0–64.0bn, assuming FX is a ~1ppt headwind to YoY growth. FY26 total expense outlook raised at the low end to reflect the $2.4bn legal accrual recognized in 26Q2; FY26 total expenses now $165.0–169.0bn, and OP still expected to exceed 2025. FY26 capex (incl. principal on finance leases) raised at the low end to $130.0–145.0bn vs. prior $125.0–145.0bn.Assuming a stable tax regime, the effective tax rate for the remaining quarters of 2026 is expected at 15–17% vs. prior 13–16%.

2) Results: 26Q2 total revenue $60.8bn (+28% YoY, +27% cc). Family of Apps (FoA) revenue $60.4bn (+28% YoY), incl. ad revenue $59.4bn (+27% YoY, +26% cc).Ad impressions +14% YoY with healthy growth across regions, driven by user/engagement gains and ad load optimization. Avg. price per ad +12% YoY on better ad efficacy, a more favorable macro vs. last year, and FX tailwinds, partly offset by faster growth from lower-monetizing surfaces/regions.FoA other revenue topped $1bn for the first time in a quarter (+73% YoY), led by paid messaging and subscriptions on WhatsApp. Reality Labs revenue $431mn (+16% YoY), with strong AI glasses offset by lower Quest headset sales.

3) Profit and expenses: 26Q2 total expenses $42.0bn (+55% YoY), including a $2.4bn legal accrual and $1.2bn severance from May 2026 layoffs. Growth was driven by comp, infra costs, legal, and third-party AI token costs.Ex-severance, comp growth reflects added tech hires over the past year, especially AI talent. Infra cost growth reflects D&A, data center opex, and third-party cloud.GAAP OP $18.8bn (-8% YoY) with OPM 31%; ex-legal and severance, OP +9% YoY. Tax rate 16%; NI $15.8bn; diluted EPS $6.18.

4) Capex and financing: 26Q2 capex (incl. finance lease principal) $31.1bn, focused on servers, data centers, and network infra; FCF $784mn. Cash and marketable securities $90.3bn; debt $83.7bn.A strong balance sheet enables access to broad capital markets to supplement operating cash flow. Debt mix has risen in recent years to lower WACC, with a tilt to cost-efficient long-duration funding for long-cycle investments (esp. AI infra).Meta is also broadening financing formats and just announced a strategic JV with BlackRock to develop a 1GW data center in El Paso, Texas.

5) Headcount and legal: EoQ headcount 75k+ (down 3% vs. 26Q1) still includes ~8k impacted by the May 2026 layoffs; most affected roles should drop from reported HC by end-26Q3. Meta continues to monitor material legal/regulatory matters, with ongoing teen-related scrutiny in multiple markets.Several U.S. teen-related cases are scheduled for trial this year and could result in material losses.

II. Call details

2.1 Management remarks

1) Users and community

a) 3.6bn people use at least one app daily. Several milestones this quarter: Instagram DAUs reached 2bn, Threads MAUs topped 500mn as the fastest-growing conversation app ever, and Facebook DAUs have been above 2bn for some time.b) WhatsApp hit a record in message volume, peaking at 30mn messages per second during the World Cup final. WhatsApp has a new lead who previously founded one of India’s most important payments firms.c) The scale and reach across the app family provide a powerful platform to deliver innovation to billions.

2) AI in the core biz: recommendations and content

a) Strong optimism on integrating LLMs into recommenders. LLMs provide first-principles understanding of what content conveys and why it appeals, and deepen understanding of user interests and intents, enabling more relevant and engaging recommendations.b) Instagram global time spent grew double digits YoY, driven by Feed and Reels ranking improvements. Facebook video time spent +9% YoY globally and >10% in the U.S./Canada, led by ranking gains.c) LLM capabilities in ranking keep improving: using content understanding and better synthetic training data to upgrade existing systems, while agentic LLMs assist engineering by assessing content quality, spotting trends, and testing ranking changes. Earlier this year, every public Reels and Feed post on Instagram began automatic LLM processing, analyzed on dimensions from topic to tone, with more surfaces rolling out on Facebook.These signals feed downstream ranking, recommendations, and policy enforcement, a key foundation for deeper personalization. The Muse family now handles tasks like video topic classification and summarization, with positive early results.

d) Reels shipped the largest single ranking upgrade ever. Faster inference and a new architecture that taps longer user histories improved prediction, lifting Instagram sessions by 15bp, with forwards and time spent outperforming as strong matching signals.This capability is migrating to Feed with comparable early results.

e) Content freshness improved meaningfully. Real-time infra and new video modeling let the largest rankers identify high-quality new Reels at creation time.Over half of recommended content in Instagram Feed is now new within a day, more than 2x a year ago.

f) More user control over recs: Instagram users can prompt Youralgo in natural language to tune recommendations. Facebook launched Shape Your Feed, and early retention among users of the feature exceeds 80%.

g) Next-gen recommenders are on track, incl. a foundation model that powers both organic and ads ranking, plus an LLM-native recommender. First research milestone this half: continual pretraining of large models on rec data with healthy scaling laws observed, with more progress expected in H2.

h) Muse Image and Muse Video will massively expand the universe of discoverable content. Historically content came from followed friends/creators and unfollowed creators; going forward, an almost unlimited, personalized content universe will make the service more useful and engaging.

3) Ad monetization efficiency and marketing tools

a) Part one is optimizing ad load within organic engagement, continuously improving when and where ads show. In 26Q2, Meta expanded inventory availability and completed the global ad rollout on Threads; on WhatsApp, more ad destination types and objective options were added, with global rollouts progressing as planned.b) Part two is better advertiser performance. Meta launched a generative recommender, a paradigm shift from scoring each candidate ad to an LLM jointly reasoning over ad content and user preferences to predict the best ad for each person, improving matching precision and compounding returns to advertisers.The first gen model is live in ads recall with clear performance gains.

c) Early LLM-based user preference understanding lifted Instagram in-app event conversions by 1%. In 26Q2, Meta advanced user understanding models that jointly analyze ad and organic behavior to improve UX and advertiser outcomes.Combined with GEM for ads ranking and sequence learning, results include +8.3% ad clicks and +15.7% conversions on Facebook.

d) AI-driven Advantage+ E2E solutions continue to scale, with a >$75bn annualized run-rate this quarter. Because stacking tools compounds performance, Meta is pushing deeper adoption.Case: an Indian online apparel brand had manually built campaigns on Facebook and Instagram; by adopting Advantage+ shopping ads plus audience/placement/budget optimization, purchases rose 13% and add-to-cart conversions +16%.

e) Gen-AI creative tools adoption is expanding, with 9mn+ SMBs using at least one AI creative tool. Image generation, now supporting batch creation from existing assets including video, more than doubled adoption this quarter.A new E2E creative solution turns real performance signals into next-step creative decisions via AI infra while preserving brand voice. It integrates Day 1 with agency workflows so teams can diagnose, generate, and scale efficient creatives without leaving existing processes.With Muse Image rolling out, advertisers should further scale high-quality, on-brand creatives.

4) New products and revenue lines: personal agents, business agents, subscriptions, and APIs

a) Three layers of opportunity: first, AI investment is accelerating core use cases—better UX, stronger advertiser ROI, and faster internal build/deploy. Second, personal agents under development lay the foundation for the next product and revenue wave.Third, large enterprise-facing opportunities exist across APIs, business agents, potential compute resale, and other enterprise services.

b) Personal agents: Meta will launch agents that work 24/7 on users' behalf across health, relationships, finance, and more. Code is the first domain to truly work end-to-end, but engineers are technical and willing to invest in tuning; at consumer scale, agents must be out-of-the-box simple.

c) Messaging and privacy: in a multi-agent future, WhatsApp and other messengers become even more important; WhatsApp is already the largest entry point to Meta AI. This quarter, Incognito mode launched on WhatsApp and in the Meta AI app so private chats with assistants are not visible to Meta.Strong privacy and safety are foundational to agent development.

d) Business agents: Meta opened business agents globally on WhatsApp and Messenger, with 1mn+ businesses weekly using them for conversations or sales, and is rolling out on Instagram. This month, Meta launched the Business Agent platform to build, customize, and deploy agents at scale on WhatsApp with enterprise-grade controls, guardrails, and measurement.Companies can define rules and deliver personalized experiences in the messaging apps customers already use. Case: Brazilian car-rental major Movita (~400 locations) deployed a WhatsApp agent to complete selection, quote, and payment in a single thread; returning customers can book in as few as three messages.Over one month, daily bookings via WhatsApp rose 44% YoY, and 85% of conversations were fully resolved by the AI agent without human intervention.

e) Agents will learn from daily conversations and feed insights back to merchants. Meta is building capabilities to summarize sessions, digest overnight changes, and surface customer needs, with plans to add growth suggestions, competitive intel, and real-time performance feedback.Longer term, the goal is a 'business-in-a-box' service so merchants can start and run an entire biz on Meta. Monetization will combine subscriptions and usage-based pricing, with more products moving toward an ads-like model where fees are paid only on delivered outcomes.Over time this enables an efficient auction on Meta’s own compute, similar to today’s ad system.

f) Subscriptions: Meta launched Meta One to bring more tools and AI features across the app family, with multiple tiers and price points as demand grows. Meta One evolves the subscription portfolio to create more value for consumers, businesses, and creators.

g) Models and APIs: Meta recently introduced competitively priced, high-intelligence model APIs with encouraging initial results. Muse Spark is open to U.S. developers to broaden distribution and lower adoption friction.Meta will expand to more channels and countries and open to enterprise customers, while building features that make Muse Spark easier for enterprises to adopt.

5) Models and Meta Superintelligence Labs (MSL)

a) MSL, launched just over a year ago, is on a solid trajectory. In the last month, Meta released Muse Spark 1.1 and Muse Image.b) After rebuilding Meta AI on Muse Spark, daily users engaging with the assistant rose 60%, with strong WoW growth.c) Muse Spark 1.1 is a highly efficient coding model, strong in computer use, tool use, and multimodal understanding. It is available via a new public API and will scale through partners and coding agents in coming weeks.The focus on strengthening Spark is driven by the opportunity to launch agents aligned to Meta’s mission and businesses.

d) AI is also accelerating internal product builds. Earlier this year Meta launched Instagram Instance and recently shipped Forum (groups app) and Seller (marketplace app).Shipping new apps should become much easier, so Meta plans to try more ideas and distribute them via the recommender to interested users, similar to Threads.

6) Hardware and glasses

a) Closer to personal superintelligence, seamless hardware matters more. Glasses are ideal: all-day wearable and assist without pulling people out of the moment.b) Glasses remain among the fastest-growing consumer electronics, with the lineup expanding. In partnership with EssilorLuxottica, Meta launched its own line, including designs co-created with Kylie Jenner, the first glasses shipping with Muse Spark onboard to understand what users see and respond more helpfully.Early sales are strong and ahead of expectations, and more details will come at Connect on Sep 23.

7) Compute, infra, and strategy

a) As AI usage scales across products and biz, Meta will keep investing heavily in infra. Capacity planning is shaped by key factors: the external build environment is highly dynamic near term and long term, and the industry historically underbuilt for this AI wave, making current capacity, including owned, extremely valuable.Supply chains must expand to support Meta’s and others’ AI capacity needs.

b) Meta is confident it can fully utilize capacity atop existing experiences while investing in foundation models that create new opportunities. The current plan aims to maximize 2026–2027 capacity; historically, each capacity add has proven highly valuable for scaling experiences, and Meta expects the same now.

c) Longer term, the exact usage ramp is hard to predict, but Meta’s distribution advantage creates opportunities to deliver valuable AI products to 3.6bn users and millions of businesses. That holds even without frontier models, but being at the frontier unlocks new markets and may require extra compute.Thus, the long-term capacity strategy is to retain flexibility to keep expanding in 2028+, laying data center and network foundations ahead of server decisions. These long-lived assets provide optionality so investment pacing can adjust with AI adoption.

d) Meta is also investing in in-house silicon and related areas for strategic flexibility and supply leverage, improving long-term returns.e) Management expects industry-wide compute to remain tight for the foreseeable future. Meta believes its models, consumer experiences, and enterprise products are the best ROI uses of its infra.Enterprise offerings could take multiple forms—tools, APIs, or direct compute monetization when demand is red hot. Flexibility helps fund builds more efficiently while preserving strategic control of compute, providing multiple return paths for invested capital.Current third-party bids for Meta’s compute are well above Meta’s own acquisition cost.

8) Strategy and values

a) Meta’s enduring thesis is to put power in users’ hands, keeping products affordable and accessible for all, benefiting both community and commerce. The same principle guides AI development in this new phase.b) Meta is the only large company prioritizing getting superintelligence directly into people’s hands: not centralizing it, but widely distributing it so everyone can aim it at what matters most.That is how society progresses over time.

2.2 Q&A

Q: Among new opportunities—consumer products, business agents, APIs, compute resale—which could scale first in 2026–2027 to show quantifiable ROIC?

A: A meaningful chunk of compute is allocated to model training to stay among the leading labs, which is critical investment. The rest supports a portfolio: upgrading core products, imminent consumer launches, APIs, business agents, dev tools on the roadmap, and direct compute sales—where current bids for compute are well above Meta’s cost.The working view: selling intelligence should carry structurally higher gross margins than selling raw compute, though compute sales are also a large opportunity. Management is optimistic across all vectors and expects tangible growth, with more details to come on several fronts.

Q: There is public debate about doubling capacity by 2027. Any early view on 2027 capex, drivers, or financing thoughts for this multi-year build?

A: No specific 2027 capex outlook yet. Infra planning remains highly dynamic, with a wide outcome range even for this year’s outlook.The current plan seeks to maximize 2026–2027 capacity while retaining flexibility to grow in 2028+ and to make server decisions when actual 2028+ demand is clearer. Capacity needs for the coming years are still being sized, and near-term capacity is judged more valuable than long-term. Planning remains very dynamic.

Q: For enterprise opportunities, how much can be addressed by today’s go-to-market built around ads and marketing vs. needing new sales motions?

A: Enterprise will be a mix. One part extends the existing business—selling to marketers and consumer-facing firms trying to reach and sell to customers, which is the bulk of Facebook and Instagram today.Messaging and business agents on other services can extend this, with pricing aligned to outcomes like the ad system. Think of it as an extension of relationships with millions of advertisers and hundreds of millions of SMBs—an inherently natural opportunity.Execution will prioritize maximizing user and merchant outcomes and building robust auctions over maximizing short-term sales, which is historically better for the biz long term.

The other part is broader enterprise customers. Meta is building coding and internal productivity tools partly because Meta needs them and can tune them for its workflows; now that these exist, there is significant opportunity to serve both SMBs and large enterprises.This is a new muscle relative to history, and Meta will share more build plans. Overall, while headlines often focus on compute, the enterprise opportunity spans compute, APIs, productivity, and business agents beyond marketing—a very large opportunity and an important new capability to build.

Q: How do you think about funding over the next few years—debt vs. equity—and balancing aggressive investment with capital needs?

A: Funding is a core element of financial planning. Strong operating cash flow already puts Meta in a favorable position to fund infra.Meta has been evolving its capital structure to increase the debt mix and lower the cost of capital. For long-cycle projects—esp. AI infra—adding cost-efficient, long-duration funding is prudent.Meta is also broadening financing avenues, including the BlackRock JV announced yesterday, and will continue to select sources prudently as projects progress.

Q: Consumers mostly use AI today as 'better search.' Will consumer adoption bridge the usefulness gap soon, or will it take more patience?

A: Some areas have already broken through. AI is unusual in that new capabilities emerge roughly annually (and possibly faster), spawning new product lines. There is an existing consumer assistant market, where Meta AI and competitors participate.Over the past year, coding agents grew very fast as the first truly agentic market. Coding worked first because customers are technical, willing to invest in setup, and coding is digitized and closed-loop, making it comparatively easier to train.The bet is that personal consumer agents will become a massive market. Five years out, it is hard to imagine a world without billions using personal agents that understand goals and act 24/7 across health, hobbies, personal finance, family operations, relationships, and careers.

Building for consumers differs from building for developers. At billion-user scale, products must 'just work' and be simple.Many agents, especially proto-agents, still require heavy tinkering or even device-level setup; a single use case may look magical, but reliability degrades with sustained use. Companies that deliver out-of-the-box personal agents will capture the next major AI opportunity.This effort runs in parallel with Meta AI, business agents, and continued model R&D that unlocks new capabilities and product lines. Launch is not far off, but details are limited pre-announcement.

Q: How far along is the roadmap for LLMs in ranking and recommendations—where are we on 'better models + more compute'?

A: There is a clear upgrade path through the rest of this year and into 2027 that should keep lifting engagement on Facebook and Instagram. A few points stand out.First, recommendations will become more personalized and interest-aligned by advancing models and architectures to better capture interests and respond faster to what users care about in the moment. AI helps expand LLM-based content understanding, deepen grasp of valued posts/creators, push high-quality, fresh, trending content forward, and reduce low-quality share.Second, data infra is improving to train on more data and use it more efficiently—richer descriptors for content users engaged with before, finer-grained content to enrich historical sequences, and longer sequences and more complex architectures across Facebook and Instagram to match larger datasets.Third, significant progress in LLM-based content understanding will embed further into both recs and policy enforcement, where LLMs excel at deep content comprehension. Meta is also applying agentic, LLM-based methods to revamp recommenders; the number of ranking agents shipped increased this half, boosting engineer productivity and opening an exciting path.

Q: You are fielding many bids to monetize compute while also buying from multiple third parties. How does that reconcile—timing and transition, or train vs. infer, or chip-task fit?

A: At a high level, aggregate compute remains far below aggregate demand. Hence many external bids for existing capacity while Meta has many high-value internal uses.Operationally, the balance is how much to monetize now vs. cultivating future assets. It is a portfolio: one cannot only bet long term without validating near-term market demand, nor sell all capacity for short-term profit and miss the multiplier from building intelligence atop compute.So Meta is scaling compute aggressively, aiming to monetize directly when appropriate while also pursuing many intelligence-monetization use cases—enterprise and some consumer, plus core ranking/recs/ads—which alone can soak significant capacity.There is also a timing gap: data centers under construction monetize only when online, but demand is very strong, so Meta is pushing all these tracks in parallel.

Q: One year into building the core lab team—how is the lab doing, when will we see faster cadence in models/chips, and what sustainable edge is being built?

A: The trajectory is compelling. Released models perform strongly on their early scaling rungs, and Meta is scaling to larger, more advanced models.There is an intelligence dimension and a data dimension. Each product category has a flywheel—learning from community behavior, improving with feedback, and getting better over time. While the industry focuses on raw model IQ, data and knowledge are equally critical to serve users.For personal superintelligence, building a clear model of individuals' lives and goals is crucial. Companies start from different strengths across markets and segments.Investing across multiple fronts early reflects that the technology is general-purpose—once built, intelligence can power many uses—but each area still requires its own flywheel to create durable advantage. Meta has proven its ability to scale working product experiences to billions, arguably best-in-class globally.Hence confidence in personal agents; even more in business agents given the existing advertiser/SMB base and rollout capability. These, together with the data flywheels being built, form lasting advantages.On the research side, culture matters—managing teams well, keeping friction low, and compounding over time. That is often hard to quantify but central to how businesses succeed.

Q: Muse Spark 1.1 is near the frontier but positioned for low cost. Do you need to compete at both low-cost and high-performance ends?

A: Scientifically, new behaviors emerge at each training stage, so Meta builds from smaller to larger models stepwise. The released models occupy specific points on the scaling ladder, with larger models to come.Given their sizes and the lab’s stage, Muse Spark 1 and 1.1 are strong. Meta also needs more advanced models—hence scaling up and building research infra—while maintaining efficient models to handle the bulk of consumer-scale requests.At billions of users, you need advanced models for hard problems and efficient ones for most prompts. Efficiency matters greatly, but solving the hardest problems for global enterprises and consumers matters too, so both ends are important.

Q: You said the lab will return to open source. How does that align with monetization for AI products and models?

A: Open source is a vital part of the ecosystem, beneficial globally and a positive feedback loop for Meta by drawing the community into its infra stack and work and contributing improvements. Meta will continue a mix of open- and closed-source, as always stated.Counterintuitively, shipping open models can be more work. If Meta only served closed, internal use cases, models could be more uneven; once open, they must be more rounded.Meta wanted to ensure MSL could pursue the strongest models without constraints. Expect some open models to be reintroduced soon. The stance remains pragmatic: open source is important and Meta will contribute, while keeping a hybrid approach.

Q: If open-weight models proliferate, does that change the need to build closed frontier models—could you rely on them instead?

A: Today’s open-weight models do not match frontier capabilities, so the basic answer is no. There is also policy uncertainty about relying on third parties. Meta believes it can do better in-house, and dependency carries risk.Stepping back, Meta is a full-stack tech company—own DCs, own infra, own chips, own low-level software. Facebook won early because it really scaled when others could not. Full-stack ownership enables more personalized, optimized, and efficient products.Owning the model layer will grow in importance—critical for Meta and explaining why others care about open source: even if they cannot build models, they do not want to depend on a few closed labs.There is room for open models in the world. That does not diminish API or other business opportunities because models must still be run and inferred, and efficient operations and compute access are commercial advantages. Discerning customers globally will want control, trust their models, ensure data safety, and avoid sending data to competitors. So open source matters, but owning core models remains crucial for Meta.Also, benchmarks compress differences into a few points, but real capability profiles vary—models have different strengths and 'personalities.' Building personal superintelligence or SMB business agents may require skill mixes different from other labs' tuning.As with end-to-end work on Instagram recs and ads, full-stack modeling will be a durable edge and the base for personal and business agents. Coupled with distribution and scaling prowess, these form lasting advantages.It is a large investment and a big bet, but the tech is working, the lab’s trajectory is strong, and Meta is excited about upcoming products. Participants in this wave should earn attractive long-term returns.

Q: You said 'maximize 2026–2027 capacity.' Is that demand- or supply-driven—do you plan to fully use capacity by 2027, or will 2028+ be sized by product/service demand?

A: Two factors. First, Meta is demand constrained now and expects to remain so for the foreseeable future, including in core businesses—additional compute still has many positive-ROI uses.Second, long-term capacity constraints are uncertain and supply chains must expand. Looking to 2027+ and especially 2028+, much will change externally and internally, with more clarity when those cards are on the table.So planning for 2028 prioritizes flexibility: secure land and power now and push large chip-purchase decisions later. For now, there are ample high-quality uses for 2026–2027 capacity, which current builds target.

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosure