On Meta's earnings call, Mark Zuckerberg said Llama 4 will need almost 10x more compute to train than Llama 3.1, Quest 3 sales exceeded expectations, and more
Meta had just positioned Llama 3.1’s 405B release as a frontier-level open model. Zuckerberg’s compute estimate for its successor makes clear that the Llama 3.1 launch was not an endpoint, but the baseline for a much larger training effort.
The later report that Meta was training Llama 4 on a cluster of more than 100,000 H100 chips gives the earlier estimate operational context. Quest 3’s stronger-than-expected sales also gives Meta a positive near-term signal in the hardware business alongside its AI investment.
First-order effects
Meta must plan substantially more training capacity for Llama 4 than it used for Llama 3.1, raising the execution and capital demands of its next model cycle.
Better-than-expected Quest 3 demand supports Meta’s current mixed-reality hardware strategy and its ability to keep investing in that product line.
Second-order effects
A larger Meta training run intensifies competition for advanced AI infrastructure, making capacity planning more consequential for rivals pursuing frontier-scale models.
Meta’s open-model strategy becomes more expensive to sustain: after making Llama 3.1 widely available, it must translate a far larger infrastructure commitment into a model that retains developer interest.
Third-order effects
If successive open frontier models require sharply more compute, access to large-scale infrastructure—not just model-release strategy—will increasingly determine which companies can maintain credible open alternatives.
The combination of AI model spending and headset demand points to a platform contest in which companies fund long-horizon hardware and AI ecosystems together, though commercial returns will depend on sustained usage rather than training scale alone.
The trend: Frontier AI development is shifting from a model-release race toward a compute-capacity race, even for companies pursuing open model distribution.
Training Llama4 in early 2025 will require 10x the compute of v3.1, which used perhaps 3 months of 32k H100 equivalent. The shocking statement from Zuckerberg on the Q2 earnings call today is a declaration of Meta's moat for “Best Digital Assistant” business, where Meta expects …
While the meme of “nobody uses Facebook” sort of evolved to “only old people use Facebook”, $META reminded us that it's just another urban myth. Full piece below.
Another telling bit from Meta's earnings call: Zuck said that there are more young adult using Facebook, contrary to the popular narratives about the company but later on when pressed for details, they said Marketplace was driving much of that growth (along with Groups)
Llama 4 is coming next year and Zuckerberg said on today's Meta earnings call it will need about 10x more compute than 3.1 that just came out. *Jensen cries tears of joy* https://www.theverge.com/...
Llama 4 is coming next year and Zuckerberg said on today's Meta earnings call it will need about 10x more compute than 3.1 that just came out. Jensen cries tears of joy https://www.theverge.com/...
$META capex - expect 2024 capex $37-40B, updated from our prior range of $35-40B. - expect significant capital expenditures growth in 2025 AI Investment continues. $NVDA [image]
Heard on $meta ER: “LLAMA4 will require 10X more compute than LLAMA3” Jensen invited his fastest and biggest customer. $nvda Whoever has the most cost efficient compute wins! [image]
Says aiming for Llama 4 to be the most advanced model in the industry next year. ‘The amount of compute needed to train Llama 4 will likely be almost ten times more than what we used to train Llama 3’
Mark Zuckerberg on the earnings call: 'I think we're going to look back at Llama 3.1 as an inflection point in the industry where open source AI started to become the industry standard, just like Linux is.'
Zuck said on the call the amount of compute needed to train Llama-4 will likely be almost 10x the compute needed to train Llama-3, and that future models “will continue to grow beyond that.” For context, $META used two clusters of 24K+ $NVDA H100 GPUs for Llama-3 training.