(Preview) Meta’s Plans to Spend $135 Billion, The ‘AI Bubble’ Bubble?, Why Hyperscalers Should NOT Invest in TSMC
Summary
- Meta’s rally rested on renewed proof that its enormous advertising base can still compound, not on a suddenly credible roadmap for $135 billion of AI spending. Revenue grew 24%, after 26% the prior quarter, while guidance called for 26%-30% growth—including roughly 400 basis points from currency—on a roughly $200 billion-scale business. Ben called that “their best growth in like five years.”
- The core business appears to have been under-monetized: Meta is raising engagement and ad load simultaneously without driving users away. AI recommendation capabilities improve engagement, while the transformer-based GEM ad model scales with compute; that validates more infrastructure spending, though Ben’s blunt caveat was, “I don’t think that takes $135 billion.”
- At the top of Meta’s $115 billion-$135 billion CapEx range, essentially every projected dollar of 2026 free cash flow goes into data centers and chips. Ben described Zuckerberg as “burning the boats”: CapEx was roughly $12 billion-$15 billion only a few years ago, and the increase alone exceeds what Reality Labs lost across 13 years.
- Apple’s mobile constraints may have created Meta’s advertising machine by forcing Facebook to become “just an app” instead of a platform. The 3.5-inch screen turned ads into full-screen feed content rather than apologetic banners, producing an experience users preferred. In Ben’s telling, Apple is “arguably the most important company” to Facebook’s success, however unwilling Meta would be to admit it.
- The strongest defense of Meta’s AI investment is already visible inside advertising, even without superintelligence. Content understanding, longer-context targeting, just-in-time ad generation and commerce links could make “every single pixel on Instagram” monetizable; even 1%-5% improvements in effectiveness become enormous against a roughly $200 billion-scale business.
- The unresolved question is whether those tangible ad gains justify wagering the entire cash engine on an undefined future product. Andrew reduced the strategy to “CapEx plus super intelligence equals profit” and asked where differentiation comes from; Ben answered that truly transformative applications are definitionally hard to specify in advance, and waiting until AI becomes existential means deciding too late.
- Meta and Apple are running opposing experiments whose outcome may not be legible until 2036. Apple is effectively renting AI capabilities and may be betting that AI will not disrupt its core business; Zuckerberg is using founder control to ensure Meta owns its technological destiny. As Ben put it, “What do you want them to save the money for?”
Deep dive
1. Meta’s advertising engine keeps outrunning its scale
Ben thought Wall Street overreacted negatively to the previous quarter and may now be overreacting the other way. Meta’s latest 24% growth was actually below the prior quarter’s 26% and benefited more from currency, but Ben thought the next-quarter guide was 26%-30%, with roughly four percentage points of foreign-exchange help.
The remarkable part is the base: on a roughly $200 billion-scale business, Meta is still producing what Ben called “their best growth in like five years.” This reprises the company’s pattern around Stories and Reels—investors repeatedly rediscover that it can grow dramatically from a much larger starting point.
The earlier concern was where another Stories- or Reels-sized inventory surface would come from. The answer turned out to be the existing products: Meta can “shove that many more ads” into them while engagement also rises, suggesting the business may have been under-monetized throughout its entire history.
2. Compute is helping ads, but $135 billion is a different proposition
Meta can plausibly connect AI investment to current results. Better recommendations support engagement, while GEM—described as “not an LLM per se” but with transformer-based aspects—improves as more compute is applied, creating a direct link between infrastructure and better advertising. Ben nevertheless drew the line clearly: “I don’t think that takes $135 billion.”
The upper end of Meta’s $115 billion-$135 billion CapEx forecast approximately equals its projected free cash flow for the year. Ben’s interpretation: Zuckerberg is “burning the boats” by directing essentially every available dollar toward chips, data centers and related infrastructure.
The scale dwarfs Reality Labs’ former status as the symbol of founder excess. CapEx was roughly $12 billion-$15 billion only a few years ago; Ben estimated that the increase to this year’s ceiling exceeds Reality Labs’ cumulative losses over 13 years, making that spending a “footnote” by comparison.
3. Zuckerberg offered conviction rather than an ROIC model
The first analyst asked for examples of returns on invested capital over three, five and 10 years. Andrew summarized Zuckerberg’s response as a “paragraph-long ‘no’”; Ben appreciated the honesty, even as the subsequent list of possible applications required “a lot” of salt.
Andrew’s skepticism tracked a listener’s question: is Meta’s extraordinary ad revenue merely funding Zuckerberg’s recurring desire to be more than a collection of apps? His shorthand for the AI strategy was “CapEx plus super intelligence equals profit”—an enormous commitment without a clearly differentiated product at the end.
Ben’s defense began with the premise behind Zuckerberg’s spending: AI could displace today’s dominant interaction model. If just-in-time intelligence arrives through glasses, earpieces or handheld devices, Meta must build the infrastructure and capabilities now or accept irrelevance later.
That makes Meta’s contrast with Apple unusually sharp. Apple is outsourcing AI to Google, which Ben said either retreats from the doctrine of controlling critical technology or implies AI will not disrupt Apple’s core business. He conceded that Apple’s judgment “might be true.”
4. Apple’s constraint accidentally built Meta’s moat
Facebook originally wanted to become a platform for third-party apps, Facebook Credits and products like FarmVille. Its HTML5 mobile approach failed, forcing a 2012-2013 reset around building an excellent native phone app—and, as Ben put it, “on the phone, you don’t get to be a platform.”
The 3.5-inch screen forced Facebook into an unusually powerful ad format: for some fraction of the time, every pixel could become a full-screen advertisement. Instead of relegating ads to banners, Facebook made them part of the content stream; users preferred the clean alternation between full-screen content and full-screen ads.
Reels extends that advantage. A user may watch a compelling advertisement for 20 seconds despite retaining complete control to skip it, because “it’s all content.” The latest results revalidated Ben’s 2013 argument that Meta had even more capacity to place immersive ads in front of users than it realized.
Ben still considers Apple’s ATT policies and preferential device APIs problematic. Yet across Facebook’s larger history, he argued that Apple “saved Meta from itself” and is arguably the company most responsible for its success.
5. Today’s ad upside and tomorrow’s unknowability support the wager
Ben argued that AI as it exists today already offers “astronomical” opportunity: just-in-time ad generation, richer targeting, longer context windows and on-screen object recognition could make every Instagram pixel monetizable. His concrete example was clicking Andrew’s microphone in a podcast image and buying it directly.
Against a roughly $200 billion-scale business, even 1%-5% improvements in targeting, click-throughs or advertising effectiveness matter enormously. Ben’s critique was that Zuckerberg undersells this case because he “doesn’t really like ads”; Andrew agreed on the opportunity but still questioned why capturing it requires roughly $130 billion.
Ben answered with an analogy to regulation: harms are visible, while innovations foreclosed by regulation cannot be named because they never happened. Likewise, the more transformative AI becomes, the less anyone can specify its future applications. Andrew trusted Meta’s existing moat; Ben countered that by the time existential risk is provable, action is too late—“Let’s schedule our podcast for 2036 and see who was right, Apple or Meta.”