Meta’s Llama 4 family of LLMs is making “great progress in training” said CEO Mark Zuckerberg late Wednesday, with Llama 4 Mini done with pretraining. It's going to “unlock a lot of new use cases” he said – but admitted that AI won’t drive much monetisation at Meta until 2026.
In light of this “new competitor, DeepSeek” which released its models under an MIT licence, Zuckerberg said he believes there's “going to be an open-source standard globally. I think for our national advantage, it's important that it's an American standard. So, we take that seriously.
Open standard should be American
“We want to build the AI system that people around the world are using and… some of the recent news has only strengthened our conviction that this [open-source] is the right thing for us to be focused on,” he added.
“Our goal for Llama 4 is to lead. Llama 4 will be natively multimodal.
“It's an omni-model, and it will have agentic capabilities. So, it's going to be novel… and I'm looking forward to sharing more of our plan for the year on that over the next couple of months,” he added; comments that suggest a full-fat release of Llama 4 is not likely until at least Q2, if not later.
Capex, chips, inference…
Met's CapEx for full-year 2025 will be $60 billion - $65 billion.
Meta CFO Susan Li confirmed in a Q4 earnings call "servers will be the biggest growth driver [as the] largest portion of our overall CapEx budget [as we] continue to invest meaningfully in core AI, but we are also expecting growth in non-AI capacity."
Strikingly, in earnings presentation small print, Meta reveals that after an assessment of the useful lives of certain servers and network assets, it is increasing their estimated useful life to 5.5 years, effective beginning fiscal year 2025.
Based on the servers and network assets placed in service as of December 31, 2024, we expect this change in accounting estimate will reduce our full year 2025 depreciation expense by approximately $2.9 billion - Meta
"On the data center side, we're anticipating higher data center spend in 2025 to be driven by build-outs of our large training clusters and our higher power density data centers that are entering the core construction phase,” she told analysts, adding that "we're also investing in fiber to handle future cross-region training traffic."
(Meta is building a massive 2GW capacity data center, it confirmed. That's enough power for approximately five million homes.)
Meta developing custom silicon
“We're also very invested in developing our own custom silicon for unique workloads, where off-the-shelf silicon isn't necessarily optimal and specifically because we're able to optimize the full stack to achieve greater compute efficiency and performance per cost and power because our workloads might require a different mix of memory versus network, bandwidth versus compute, and so we can optimize that really to the specific needs of our different types of workloads,” Meta’s CFO added.
DeepSeek and CapEx views
Don’t DeepSeek’s innovations suggest that such extensive CapEx is potentially… wasteful, suggested one analyst on the January 30 call?
“There's a number of novel things that they did that I think we're still digesting. And there are a number of things that they have advances that we will hope to implement in our systems. That's part of the nature of how this works, whether it's a Chinese competitor or not,” Zuckerberg responded. “It's probably too early to really have a strong opinion on what this means for the trajectory around infrastructure and capex...
See also: DeepSeek left API keys, plaintext logs exposed in easily found database
“There are a bunch of trends that are happening here all at once. I continue to think that investing very heavily in capex and infra is going to be a strategic advantage over time. It's possible that we'll learn otherwise at some point, but I just think it's way too early to call that" -Zuckerberg
HE added that it "already seemed pretty likely even before this, that of all the compute that we're using, the largest pieces aren't necessarily going to go toward pre-training. But that doesn't mean that you need less compute, because one of the new properties that's emerged is the ability to apply more compute at inference time in order to generate a higher level of intelligence and a higher quality of service” he emphasised.
Meta’s Q4 net income was $20.8 billion, up 49% year-on-year. Q4 revenue was $48.3 billion, up 21%. Headcount grew10% year-on-year to 74,067 as of December 31, 2024.
