Source Direct

Source Direct

Source Direct

Miers Porgan

Divine Feminine Christ Rose

Source Direct

EarAcoustic Audio Official

Source Direct

Jack Farley

Jack Farley

@JackFarley96

16 Dec 2025

The $48 Billion is over the past three quarters, the full figure for this year is going to be higher let's say $70 Billion. This is total CapEx. Conservatively saying 10% of that is for specifically GenAI training (Llama & Emu), that's still $7 Billion being spent to build out a product whose public display is getting only 10x more engagement time than a niche finance podcast (admittedly moderately popular) Huge caveats: Llama and Emu are powering tons of stuff on instagram, Meta glasses, business agents on WhatsApp, etc. Also, Llama is on other platforms not just Meta.ai. Lastly, not apples to apples, of course, LLMs are far newer and totally different to podcasts. My questions with w/ Gemini below what percentage of META CapEx is dedicated to AI There is no single line item on Meta’s balance sheet for "AI Spending," but we can derive a highly accurate estimate based on their 2024 financial filings and CFO statements. For all practical purposes, effectively 100% of Meta’s growth in spending is dedicated to AI. While Meta does not officially split the percentage, here is the breakdown of how much of that $39.2 billion CapEx is actually for AI versus "keeping the lights on." 1. The "Incremental" Estimate: ~40–50% of Total CapEx A standard method to isolate AI spending is to look at the increase over their historical baseline. 2023 CapEx: ~$27 billion (The "maintenance" baseline before the massive AI ramp). 2024 CapEx: ~$39.2 billion. The Difference ($12.2 billion): This jump is almost entirely attributable to new AI infrastructure (H100 GPUs and new data centers). This suggests that roughly 30% to 40% of the total 2024 budget was exclusively for new AI capacity, while the base $27B maintains their existing massive network (though even that base now involves replacing older servers with AI-ready ones). 2. The CFO’s "Substantially All" Confirmation During Meta’s recent earnings calls, CFO Susan Li has explicitly stated that the increase in capital expenditures is driven by investments in AI infrastructure. She noted that "substantially all" of the capital expenditure growth is for servers (GPUs) and data centers. Unlike Google or Microsoft, which spend billions on cloud services for other companies, Meta’s spend is almost entirely for its own products (consumer AI and Ads AI). 3. The "Hidden" AI Spend (Ads) It is misleading to separate "AI" from "Core Business" because Meta’s core business is now AI. The "Fungible" Cluster: The billions spent on servers for Instagram Reels recommendations are arguably "AI spend." Meta deliberately builds clusters that can switch between training Llama 4 (generative AI) and ranking ads (predictive AI). If you count the AI used to rank your Feed and target ads, then the vast majority—likely over 80%—of their infrastructure bill is now AI-related. Summary Table: Where the Money Goes Expenditure CategoryEst. % of CapExIs it AI?Generative AI Training~20–30%Yes (Llama 4, Emu, etc.)Core AI (Ads/Ranking)~40–50%Yes (Reels, Feed Algorithms)Legacy / Maintenance~20–30%No (Office space, network cables, non-AI servers) The Bottom Line: If you define AI strictly as "Chatbots and Llama," it’s about 25–30% of the budget. If you define AI as "Machine Learning hardware used for any purpose (including ads)," it is likely over 70% of their total Capital Expenditures. Next Step: Would you like to know how the market reacted to this spending—specifically, why investors punished Meta for this in late 2024 but are rewarding them for it now? what are the use cases of Llama and Emu besides powering Meta.ai Meta’s AI investment isn’t just for chatting; it is slowly becoming the operating system for everything you do on Instagram, Facebook, and WhatsApp. Here is how Llama (the brain) and Emu (the artist) are being used right now (December 2025) beyond just the chatbot. 1. Tools You Use (Consumer Features) If you use Instagram or WhatsApp, you are likely using Llama and Emu without realizing it. AI Stickers (Emu): When you type a prompt to generate a custom sticker in WhatsApp or Messenger, that is Emu running in real-time. "Restyle" & "Backdrop" (Emu): On Instagram Stories, the features that let you change the background of your photo or apply a completely new artistic style (e.g., "make it look like a watercolor painting") are powered by Emu's image editing capabilities. Ray-Ban Meta Glasses (Llama): If you wear the smart glasses and say, "Look at this ingredient list, is it gluten-free?", Llama’s multimodal vision capabilities are what "see" the text and analyze it for you. Feed Summaries (Llama): On Facebook, the "Comment Summaries" feature that condenses thousands of angry comments into a neat "Pro vs. Con" list is purely Llama. 2. Tools for Businesses & Creators (Where the Money Is) This is the most critical financial use case. Meta uses these models to automate the work of an ad agency. Advantage Creative (Emu): If a business uploads one photo of a shoe, Emu automatically generates hundreds of variations—changing the background from a beach to a mountain, adjusting the lighting, or expanding the image ratio—to see which version makes you most likely to click. Creator AI (Llama): Big influencers on Instagram use a "Llama-powered twin." It reads their past DMs and comments to learn their "voice," then automatically replies to fans on their behalf (marked as AI) so creators can engage with millions of followers at once. Business Agents (Llama): Small businesses on WhatsApp use Llama to handle customer service—booking appointments, answering FAQs, and taking orders 24/7 without a human. 3. Behind the Scenes (Internal Efficiency) Meta uses its own AI to fix its own code, saving millions in engineering time. Llama for Code (Code Llama): Meta’s internal developers use a specialized version of Llama to write and debug code. It acts like a "spellcheck" for software engineering, speeding up how fast they can fix bugs in the Instagram app. Content Moderation (Llama Guard): Instead of relying solely on human moderators to catch hate speech or violence, a version called Llama Guard reads millions of posts per second to flag policy violations instantly. 4. The "Open Source" Ecosystem Because Llama is open source, it is being used by other massive companies to power their products, effectively making Llama the "Android of AI." Zoom: Uses Llama to power its "AI Companion" meeting summaries. Salesforce: Uses Llama to power parts of its "Einstein" customer service bots. Goldman Sachs: Uses specialized versions of Llama for internal document analysis and research. Next Step: Would you like to know about "Project Avocado"—the rumored secret AI model Meta is building that (unlike Llama) they plan to keep completely private? where can people interact with the LLM of Llama other than meta.ai Since Llama is "open source" (unlike ChatGPT or Gemini), it is available on dozens of other platforms. Many users actually prefer these alternatives because they are often faster, more private, or offer features that Meta's own official chatbot blocks (like uncensored conversations or coding modes). Here is where you can interact with Llama 4 (and older versions) without touching a Meta app: 1. The "Privacy-First" Option: DuckDuckGo Where: duckduckgo.com/chat (or the "Duck.ai" button in their browser). Why use it: It is completely anonymous. DuckDuckGo routes your chat through their servers, stripping your IP address before it hits Meta. Models Available: They offer Llama 4 Scout (free) and Llama 4 Maverick (for subscribers). This is arguably the safest way to use Llama if you don't want Meta building a profile on you. 2. The "Speed" Option: Groq Where: groq.com Why use it: This is the viral "instant" AI demo. Groq uses specialized chips (LPUs) that make Llama generate text effectively instantly (hundreds of words per second). Experience: It feels like the AI is reading your mind because the answer appears as fast as you can read it. It is significantly faster than the official Meta AI app. 3. The "Aggregators" (All-in-One Apps) These platforms let you swap between Llama, GPT-4, and Claude in a single window to compare answers. Poe (by Quora): One of the most popular ways to access Llama 4 Maverick. It allows you to create "bots" with custom system prompts (e.g., "You are a Llama model that only answers in Python code"). Perplexity: When you use Perplexity's "Pro" search, you can go into settings and select Llama 4 as the underlying model that powers your search results. HuggingChat: The open-source community's version of ChatGPT. It usually hosts the very latest "base" versions of Llama before anyone else. 4. The "Browser" Option: Brave Leo Where: Built directly into the Brave web browser sidebar. Why use it: It reads the webpage you are currently looking at. You can open a long article, hit the "Leo" button, and ask Llama to summarize it without ever copying/pasting text. 5. The "Local" Option (For Geeks & Privacy) Tools: Ollama or LM Studio. How it works: You download the Llama file to your own computer (Mac/PC). Why do this: It runs offline. You can chat with Llama on an airplane with no Wi-Fi, and zero data leaves your device. This is how many developers and companies use Llama to ensure no secrets leak to Meta. Next Step: Would you like to know how the "Maverick" version of Llama (found on these paid sites) differs from the "Scout" version you get for free? so tell me more about meta.ai, are they trying to drive lots of traffic to meta.ai? seems like they are not getting very much traffic You are spot on—if you look strictly at the traffic to the website meta.ai, it looks like a ghost town compared to chatgpt.com. But this is by design. Your observation highlights the fundamental difference in strategy between OpenAI and Meta. 1. The "Iceberg" Strategy OpenAI needs you to go to their website or app because ChatGPT is their product. Meta does not want to force you to go to a new website; they want to put the AI in the app you are already addicted to. ChatGPT (The Destination): You have to stop what you are doing, open a new tab, and type chatgpt.com. It gets billions of visits because it is a destination. Meta AI (The Feature): The vast majority of "traffic" isn't counted on meta.ai. It happens inside the WhatsApp, Instagram, and Messenger apps. The Stats: While meta.ai (the website) might only get ~50–100 million visits, Meta claims over 500 million monthly active users for the AI. Where are they? They are the teenager asking the blue circle in Instagram to "write a caption for this photo," or the mom in WhatsApp asking for a recipe. They never visit the website. 2. Why the Website Exists (If no one uses it) If the strategy is "in-app," why spend money building meta.ai? The "Power User" Valve: Typing on a phone is slow. For Meta AI to compete as a coding assistant or a serious writing tool (like ChatGPT), it needs a desktop interface with a keyboard. Corporate/Enterprise: You can't ask employees to use their personal WhatsApp for work tasks. The website offers a neutral ground. The "Google Killer" Ambition: Meta eventually wants to replace Google Search. To do that, they need a URL that functions like a search engine. 3. Comparison: The "Destination" vs. "Omnipresence" FeatureOpenAI (ChatGPT)Meta (Meta AI)Primary Accesschatgpt.com (Website/App)Blue Circle in WhatsApp/IG/FBTraffic SourceDirect (People typing the URL)Integrated (People clicking while browsing)Goalspecific "AI productivity" sessionsIncrease time spent in social appsRevenue Model$20/month SubscriptionBetter Ad Targeting (Data) 4. The "Secret" Metric: Ad Data The reason Meta doesn't care if meta.ai has low traffic is that the website isn't where they make money. They make money when you chat with the AI in Instagram. The Feedback Loop: If you ask Meta AI on Instagram, "Show me cool hiking boots," Meta doesn't need you to visit a website. It just notes that you are interested in hiking and immediately sells that data to advertisers to show you North Face ads in your regular feed 10 minutes later. Next Step: Would you like to know how to turn off the AI training on your Instagram photos, or is it too late for that? ___

Meta AI

Use Meta AI assistant to get things done, create AI-generated images for free, and get answers to any of your questions.

meta.ai

3,191

Source Direct (@source_direct) • Instagram reel

Meta AI

Source Direct Music available now via Bandcamp. #SourceDirect #Short...