Meta is the social network company behind Facebook, Instagram, Reels, Threads and WhatsApp. Billions of people worldwide use their products to stay connected and share photos and videos. In exchange for this service, Meta serve ads placed by advertisers. Their product is virtual, non physical and as a result is a high margin business.
For years they have invested in a high quality AI team. Their AI lab is lead by AI New York University professor Yann LeCun https://ai.meta.com/people/yann-lecun/ Yann LeCun’s research paper, “Backpropagation Applied to Handwritten Zip Code Recognition” in 1989 was the spark that showed the practical applications for deep learning using convolution neural networks (CNN). He won the computing’s highest honor, the Turing Award in 2018 which he shares with Professor Yoshua Bengio and Professor Geoffrey Hinton for deep learning. The winner’s lecture is available on YouTube: https://amturing.acm.org/vp/lecun_6017366.cfm
Recent News:
July 23, 2024, Meta blog released Llama 3.1 405 Billion parameter LLM model https://ai.meta.com/blog/meta-llama-3-1/ Meta also upgraded 70B and 8B. They modified their license to allow the 405B model to train other models. The large 405B model requires more computation and memory. Inference pricing is significantly higher for and it is posted on https://llama.meta.com for various cloud providers.
| Cloud | 8B input | 8B output | 70B input | 70B output | 405B input | 405B output |
| Azure | $0.30 | $0.60 | $2.65 | $3.50 | $5.33 | $16.00 |
| AWS | $0.30 | $0.61 | $2.68 | $3.54 | ||
| Databricks | $1.00 | $3.00 | $10.00 | $30.00 | ||
| Fireworks.ai | $0.20 | $0.20 | $0.90 | $0.90 | $3.00 | $3.00 |
| IBM | $0.60 | $0.60 | $1.80 | $1.80 | $5.00 | $35.00 |
| Octo.ai | $0.15 | $0.15 | $0.90 | $0.90 | $3.00 | $9.00 |
| Snowflake | $0.57 | $0.57 | $3.63 | $3.63 | $15.00 | $15.00 |
| Together.ai | $0.18 | $0.18 | $0.88 | $0.88 | $5.00 | $15.00 |
Meta releases meta.ai chat website, added chatbot to WhatsApp, Instagram and Messenger and Llama3: https://investinwhatyouknow.com/meta-has-released-8b-and-70b-llama3-llm-model/
Chatbot is freely accessible from https://meta.ai without an account.
Cost Cutting / Earnings Opportunity
During Covid, Meta like other high tech firms hired heavily. Post Covid as economy shrinks, Meta trimmed products and implemented big lay offs. In 2022 November 9th, Meta announced they will lay off 11,000 or 13% of its staff and implement hiring freeze through Q1 2023: https://about.fb.com/news/2022/11/mark-zuckerberg-layoff-message-to-employees/. On March 14th 2023, Meta announced second round of layoff of around 10,000 additional employees and eliminated 5,000 open positions not yet filled. Meta also flattened the management hierarchy and asked many managers to become individual contributors. Managers would have about 10 reports. Meta’s making these changes for more efficiency in execution and cost.
The layoff reduces expenses by a lot and as a result recent earnings in latest quarterly report is way up. More recently Meta has focused its attention on AI and reduced work on Metaverse. Cuts in Metaverse spending helps reduce cost. These savings enable it to invest in AI infrastructure.
AI Opportunity
Many Meta products can be enhanced with AI. Users can use AI to get things done. Businesses can use AI chat bots to interact with their customers. At the heart of what Meta does is advertising. AI can help accurately direct ads to the right user. When a meaningful ad is served, merchants win and the user is not annoyed. AI can also surface more relevant content for the user. In this respect, more of a user time is spent on the Meta platform and exposing user to more ads. Users of the platform need to carefully control their time on social media to avoid wasting too much time there. Meta has the Advantage+ portfolio of products to use AI for automating advertising campaigns. They can use generative AI for text and image.
Meta is right on top of the latest development in LLM AI. They’ve released their Llama 2 model and its performances is close to ChatGPT level. Llama2 with 70B parameters is trained on 40% more data than Llama1. Meta has been released it and made it available with research and commercial license. Performance of Llama2 is competition with other open source models but still lags GPT4. Like open source software, since the Model is open source, contributions for other researchers can speed up its rate of improvement. Sharing the model also reduces CO2 waste from everyone doing massive amount of AI training. In the Q4 2023 earnings webcast, Meta CEO Zuckerberg announced Llama3 is being trained. He also explained why Meta open source Llama. First is the model are safer and more secured. Second because of the high scrutiny and optimization from the open source community, the models are more efficient and require less compute and therefore cost. Third standardization makes other products created by others easier to integrate with Meta. Fourth, this helps with recruiting as the brightest want their work to have high impact.
Meta has a data advantage. On Facebook and Instagram, users can make their posts public. These public posts are in text, image, and video. Also because of its large user base, Meta can collect feedback on its AI products to further refine them. Products like AI recommended Reels get better quicker with feedback. A product may be perform well at the start, but through feedback, it becomes very good. Meta’ working an AI assistant call Meta AI.
Meta has built their first in house AI training and inference MTIA chip. They claimed is 2x more efficient than equivalent GPU. They have a lab of 16,000 A100 GPUs with 5 Exaflops. This cluster was used to train Meta’s 65B Llama model. https://arxiv.org/pdf/2302.13971.pdf.
Meta also has built its own video processor accelerator that is 9x faster than software encoders.
In Q4 2023 earnings webcast, Meta plans to build the world class AI compute infrastructure. By end of 2024, they will have 350k H100 GPUs and other GPUs. It will be the equivalent of 600k H100 GPUs. He sees a trend where each state of the art LLM is built with 10x the compute of the prior generation. They will continue to build specialized custom silicon. Meta’s 2024 capital expenditure will be in the range of $30-$37 billion in AI and non AI hardware and data centers. Smaller competitions will have a tough time matching this level of spending.
Metaverse
Meta’ Reality Labs division had over $1 billion in revenue. Their Quest 3 virtual reality goggles are popular. They are also selling smart Ray-Ban Meta glasses. They believe these devices make AI more accessible.
Scale Opportunity
Meta is the world largest social network. Retailers who need to reach a large audience prefer to advertise on Meta.
Strong Board of Directors
Tony Xu, CEO/Founder of Doordash joined Meta’s board January 11, 2022. Xu’s graduate of University of California, Berkeley in Industrial Engineering and Operations Research and Stanford Graduate School of Business MBA.
Marc L. Andreessen was cofounder of Netscape. He has a B.S. in computer science from the University of Illinois at Urbana-Champaign.
Sheryl Sandberg served as Meta COO. Sheryl has BA from Harvard University and an MBA from Harvard Business School. Prior to Meta, she was vice president of Global Online Sales and Operations at Google and had served under President Clinton in United States Treasury Department as chief of staff.
Q4 2023 Earnings
After posting strong Q4 2023 earnings, Meta shares on the next day finished at $474.99 up 20.32%. Helping lift the share price was Meta also announced they would start paying quarterly dividends of $0.50. In addition Meta has $30.93 billion available for stock buyback and it authorized an additional $50 billion increase for share repurchase.
Q4 2023 earnings transcript: https://s21.q4cdn.com/399680738/files/doc_financials/2023/q4/META-Q4-2023-Earnings-Call-Transcript.pdf
Health Stats
| Q4 2023 | Q4 2022 | Change | |
| Revenue (millions) | $40,111 | $32,165 | 25% |
| Costs and Expense (millions) | $23,727 | $25,766 | -8% |
| Income from Operations (millions) | $16,384 | $6,399 | 156% |
| Operating Margin | 41% | 20% | |
| Net Income (millions) | $14,017 | $4,652 | 201% |
| Diluted Earnings Per Share | $5.33 | $1.76 | 203% |
| Headcount | 67,317 | -22% | |
| Cash (billions) | $65.40 | ||
| Long Term debt (billions) | $18.39 | ||
| Capital Expenditure (billions) | $7.90 |