Meta’s Search for AI Clout Takes It Into New Terrain
The Takeaway
• Meta’s Llama model trails rivals on AWS
• Company relies on cloud providers to pitch models to businesses
• Meta is in talks with Snowflake about possible deal
Zuckerberg wants to turn Meta’s large language model, Llama 3, into the industry standard for AI. Initially he has relied mostly on other tech companies to handle selling the software to customers, with mixed results so far.
Llama has struggled to gain traction on Amazon Web Services, the No. 1 cloud provider, which offers a variety of LLMs to businesses. Anthropic’s Claude is the most popular model on the platform, according to a person with knowledge of the situation. Salespeople at Microsoft, meanwhile, typically pitch Llama only to customers with existing data expertise, such as companies with their own in-house engineers and data scientists, according to a Microsoft employee.
Those factors could put pressure on Meta to figure out how to build its own enterprise sales and marketing team to pitch its products directly to business, something it has shown little stomach for in the past. For a time it pitched its Workplace communications platform to businesses through a dedicated sales team, but Meta didn’t consider Workplace a business priority and this year said it would shut the platform down.
A Meta spokesperson said the company had worked with a large number of partners for the launch of Llama 3.1 in July. They included AWS, Google Cloud, Microsoft Azure, Nvidia, Databricks, Snowflake, Groq and Dell, among others.
Some in the industry are upbeat about Meta’s prospects. “Right now, people are excited because of [the] Lamborghini. That’s OpenAI,” said Laurent Gil of Cast AI, a company that helps businesses cut their cloud costs. “Everybody wants to buy the Lamborghini. But Llama 3 is creating the Toyota. And guess what? There are more Toyotas than Lamborghinis,” he said.
Gil said most businesses don’t know which models to use and get caught up in the “hype” around OpenAI. He argued that Meta will win out because companies will recognize Llama as a lower-cost, high-quality alternative to the proprietary models.
Meta is working on ways to get more businesses to use Llama. In recent weeks, Meta executives have been in discussions with Snowflake about expanding on an existing Llama sales arrangement by allowing customers of the data analytics giant to train custom versions of Llama using their own data on Snowflake’s platform, according to a person with knowledge of the situation.
It doesn’t help that Llama ranks behind rivals like OpenAI, Google and Anthropic, according to the Chatbot Arena, a project from research group Lmsys and the University of California that evaluates AI models.
Yet some businesses are embracing Llama. Goldman Sachs is on track to spend over $1 million this year running Llama on Microsoft’s Azure cloud, primarily to automate coding, according to someone familiar with the company. (Goldman has used a range of other LLMs, The Information has previously reported.) Amazon, meanwhile, has said Japanese bank Nomura and outsourcing company TaskUs are using Llama through AWS.
“Meta is teaching businesses how to build their own large language models,” said Vin Vashishta, an AI advisor to companies.
AWS Talks
Meta is spending tens of billions on new AI chips to expand its computing capacity for its AI push. Precisely how that effort will generate a return is hazy.
Meta isn’t planning on directly charging either consumers or businesses for its LLMs, which are released as open-source software, unlike the models developed by OpenAI, Google and others. (Zuckerberg has, however, hinted at the possibility of Meta sharing in the revenue of the cloud firms that sell Llama.)
Zuckerberg has said Meta can utilize AI to improve its advertising business. More broadly, if Meta can make Llama the industry standard, it can undercut the efforts of rival companies with the same goal.
In a blog post last month, Zuckerberg likened Meta’s open-source strategy to past successful efforts to develop the Linux operating system as the industry standard for both cloud computing and mobile devices.
To broaden the distribution for Llama on various other companies’ platforms, Zuckerberg has had to navigate the complexities of the tech industry, where companies frequently must partner with their competitors in various ways.
Google—whose cloud unit offers Llama to its customers—competes with Meta in digital advertising, for instance. Microsoft, whose Azure cloud unit also offers Llama, has partnered with OpenAI, whose large language models compete with Llama. Amazon, though, already has a close relationship with Meta.
Amazon has long been a major advertiser on Meta. Last year, Amazon partnered with Meta to allow people to buy Amazon products through ads on Facebook and Instagram without leaving the social media sites.
Meta is also a longstanding customer of AWS. It has been using Amazon for small cloud tasks for about a decade. In recent months, executives from Amazon and Meta have been negotiating a potential deal that would boost Meta’s spending on AWS, likely to more than $1 billion per year, making Meta one of the service’s largest customers, according to a person involved in the discussions. Meta currently spends between $250 million and $500 million on AWS and Azure cloud services, according to The Information’s Cloud Database.
As part of the deal under discussion, Meta is also pushing AWS for discounted access to graphics processing units, including those from Nvidia, the person said. GPUs are the specialized AI chips used in training LLMs. Meta and AWS did not comment on the talks.
Meta has already ordered 350,000 Nvidia H100 chips for delivery by the end of this year, Zuckerberg said on an Instagram live video in January. In the same video, Zuckerberg said Meta was building a “massive amount of infrastructure” to support its AI plans.
Business History
Meta and AWS have already worked together on selling an enterprise product.
Around 2019, the two companies partnered on selling PyTorch, an open-source software library Meta built to create AI applications. Over the course of several years, PyTorch ended up as the most popular research-focused product in a market that also included Google’s TensorFlow, a similar open-source product, three former Meta employees said.
Things between Meta and AWS didn’t always go smoothly. During the PyTorch efforts, employees at Meta grew frustrated with AWS’ sales operation, two former Meta employees said. Some felt AWS was too slow to push PyTorch out to enterprise customers, the former employees said.
AWS provides Llama through its Bedrock platform, which offers customers a variety of different LLMs. AWS’ pitch is that it offers several models and lets customers pick what they want. Other models available include Amazon’s own Titan and models from Anthropic and Cohere.
An Amazon spokesperson said the company believes having a choice of foundation models is important, and a single model will not be right for every use case.
A Meta spokesperson said the company is “very happy with the [AWS] collaboration,” but did not comment on the rates of uptake for Llama 3.1.