• Zuckerberg says model sovereignty remains critical amid open-source growth
• Meta narrows 2026 CapEx guidance as free cash flow falls

Meta is weighing whether to use its rapidly expanding data center capacity internally or rent computing power to outside companies, with CEO Mark Zuckerberg saying the company has received numerous offers at substantial premiums but sees greater long-term value in selling AI-powered products.

Zuckerberg said demand for computing capacity far exceeds available supply, creating an opportunity for Meta to monetize some of the infrastructure it is building. A substantial share of that capacity will support model training, while the rest could be directed toward improving Meta’s advertising and recommendation systems, developing consumer and business agents, operating model APIs or renting compute directly.

“We believe that there will continue to be a significantly higher margin on selling intelligence rather than selling compute directly,” Zuckerberg said during Meta’s second-quarter earnings call.

Meta is developing personal AI agents that Zuckerberg said could work continuously on users’ behalf across areas such as health, relationships and finances. The company plans to share more about those products soon and expects its messaging platforms, particularly WhatsApp, to become important surfaces for interacting with multiple agents.

CFO Susan Li said industry infrastructure has been underbuilt relative to the wave of AI adoption, making near-term capacity particularly valuable. Meta’s current construction plans are designed to maximize capacity in 2026 and 2027 while laying data center and network foundations that preserve flexibility over server deployments in 2028 and beyond.

Meta also does not plan to rely on open-weight models developed by others instead of building its own frontier systems. Zuckerberg said existing open-source models are not as capable as frontier models and relying on another company’s technology would introduce strategic and policy risks.

“Having kind of sovereignty over building your own models is going to be an important part of that stack going forward,” he said.

Zuckerberg nevertheless argued that open-source models will remain important for companies seeking greater control over their technology and data. Their proliferation would not eliminate Meta’s API or infrastructure opportunity because the models still need to be hosted and run for inference. Efficient operation and access to compute will remain competitive advantages, he said.

Building its own models also allows Meta to optimize systems for specific uses, including Instagram recommendations, advertising, personal agents and small-business assistants, rather than depending on general-purpose models tuned by outside laboratories.

The strategy is consuming an increasing share of Meta’s cash. Free cash flow fell to USD 784m from USD 8.55bn in the same quarter last year as property and equipment purchases nearly doubled to USD 30.12bn.

Meanwhile, the company narrowed its 2026 capital expenditure guidance to USD 130bn-USD 145bn from USD 125bn-USD 145bn. Second-quarter capital expenditures reached USD 31.08bn, driven by servers, data centers and network infrastructure.