Meta used its second-quarter earnings call in 2024 to set out an AI strategy built around heavy investment rather than immediate generative-AI revenue. CEO Mark Zuckerberg said the company was planning the computing clusters and data it would require for several years, including infrastructure for its next major model, Llama 4.
The plan reflected a belief that leading AI systems would demand substantially more computing capacity. Meta expected Llama 4 training to require almost ten times the compute used for Llama 3, which was believed to have relied on 16,000 GPUs. Zuckerberg wanted the new model to become the industry’s most advanced during the following year.
Meta’s approach combined long-term model development with more immediate improvements to recommendations and engagement across its existing products. The company was prepared to increase capital spending before generative AI became a direct source of revenue because it believed the same infrastructure could support several strategically important uses.
Llama 4 Required a Much Larger Compute Plan
Planning for Llama 4 meant securing hardware, data and data-centre capacity well before the model could contribute to revenue. The projected increase over Llama 3 showed that Meta expected frontier-model development to become more resource-intensive rather than relying on efficiency improvements alone.
The scale of the programme made infrastructure a central part of Meta’s AI strategy. Model ambitions could only be realised if the company had enough GPUs, power and supporting systems available when training began. By planning clusters several years ahead, Meta sought to avoid having its product roadmap constrained by a shortage of computing capacity.
Zuckerberg’s objective for Llama 4 also placed the investment in a competitive context. Meta was not presenting the model as a modest upgrade. It wanted to challenge the most advanced systems in the industry, which required accepting high costs before the technical outcome or commercial return was certain.
Capital Spending Rose Before Generative AI Revenue
Meta projected full-year capital expenditure of between $37 billion and $40 billion, raising the range by $2 billion from its previous estimate. Investors were also warned that capital spending was expected to increase significantly again in the following year.
Chief financial officer Susan Li said Meta did not expect generative AI to produce revenue during 2024. That admission made the timing of the strategy clear: the company was funding infrastructure in anticipation of future products and capabilities rather than matching each investment to near-term generative-AI income.
Meta nevertheless entered this investment cycle from a strong financial position. It reported second-quarter revenue of $39 billion and net income of $13.5 billion. Those figures represented year-on-year increases of $7 billion and $5.7 billion respectively, giving the company room to pursue a costly programme while its established advertising business continued to generate substantial earnings.
Flexible Infrastructure Reduced the Investment Risk
Li explained that Meta was building AI capacity with flexibility in mind. Hardware acquired for model training could later be used for inference, and with modifications it could support ranking and recommendation workloads. The same infrastructure was therefore not tied permanently to a single model or development phase.
This flexibility reduced some of the risk associated with ordering capacity years in advance. If a particular generative-AI use case developed more slowly than expected, Meta could redirect computing resources towards systems already central to Facebook and Instagram. Training, serving models and improving content recommendations all provided potential uses for the investment.
The company’s “Core AI” work was already producing measurable results in its existing products. Zuckerberg pointed to a unified video recommendation system for Facebook that had increased engagement on Reels by more than the earlier move from CPUs to GPUs. That comparison suggested that AI infrastructure could create value before users paid directly for a generative-AI product.
Advertising Remained the Long-Term Commercial Prize
Zuckerberg expected AI eventually to change how advertising was created and personalised across Meta’s platforms. His vision was that advertisers could provide a business objective and a budget, leaving Meta’s systems to generate copy, personalise campaigns and manage more of the process.
That prospect connected frontier-model investment with the company’s core revenue engine. Better recommendations could increase user engagement, while automated advertising tools could make campaigns easier to create and more precisely targeted. Meta did not need to rely solely on subscriptions or direct model access to justify its AI spending.
The potential audience was already enormous. More than 3.2 billion people used a Meta application each day, and Threads was approaching 200 million monthly active users. Improvements distributed across that base could have a meaningful commercial effect even if each individual change appeared incremental.
Meta’s strategy therefore rested on several time horizons. Recommendation systems offered benefits in the present, flexible infrastructure supported multiple workloads, and Llama 4 represented an attempt to lead the next generation of models. More automated advertising promised a route to future monetisation.
The company was clear that the immediate cost would arrive before the full return. Its wager was that owning enough infrastructure and developing advanced models would create options across products, advertising and AI services. In 2024, Meta chose to build for that future rather than wait for generative AI to prove its revenue potential first.
