"A Script for Mark Zuckerberg" — @BenThompson

Why this is in the vault

Thompson's fictional earnings script is the clearest articulation of why Meta's AI capex is structurally defensible — through the lens of ad-inventory expansion, not productivity/platform ambitions — and introduces a "rent compute as discipline mechanism" proposal worth tracking.

The core argument

Thompson writes a fictional Q2 2026 earnings opening statement as Zuckerberg delivering a structured mea culpa paired with a bullish AI investment thesis. The structure:

The mea culpas (what Zuckerberg got wrong):

• Platform obsession: Repeated attempts to build developer platforms (Facebook Games, Facebook web platform) distracted from the core ad/entertainment flywheel. Apple forcing Facebook into "just an app" actually saved them.
• Reality Labs: Admitted as both "I thought it was cool" AND platform desire. Burned credibility. Labs tech may matter long-run but investment scale was wrong.
• Undervaluing the ad business: Core failure — didn't make the affirmative case for digital advertising while Apple successfully framed it negatively. Missed that entertainment is the ideal ad substrate.

The AI bull case (four-layer structure):

1. Existential defense — every digital business is in AI's crosshairs; not investing is worse
2. Ad machine amplification — ML/GPU-driven algorithms already improving targeting and recommendation; LLMs will do it further; ad business top line has no price elasticity ceiling
3. Inventory expansion — "AI makes every pixel monetizable" = largest inventory expansion ever; investors have consistently panicked (Stories, Reels) and been wrong both times; this is even bigger
4. Human-first differentiation — Meta serves entertainment/connection/shopping ("off the clock") not productivity/work; this is an advantage in the AI era, not a weakness; humans will want more human connection as AI proliferates

The novel structural proposal: Rent excess compute on spot market with claw-back rights. Two effects: (a) proceeds fund further build-out, (b) rental price creates a hurdle rate that forces Meta to only take compute back when their own use exceeds market rate — hard constraint that disciplines AI spending toward the ad flywheel, not platform dreams.

The meta-framework Thompson is deploying: Aggregation Theory + revealed preference data loop: Meta's best decisions came from data validating product intuition on human connection/entertainment; worst decisions came from imitative platform ambitions. The earnings script is Thompson arguing Zuckerberg should name this pattern publicly, commit to it structurally, and use compute rental as a mechanical enforcement mechanism.

Mapping against Ray Data Co

Investing thesis — strong connection. Meta is one of the largest demand-side drivers of the AI compute buildout (alongside Google, Microsoft, Amazon). Thompson's piece gives a bull case for why Meta's capex is structurally committed rather than optionally reversible — the inventory-expansion logic means Meta will keep buying compute for years. This is directly relevant to the chip-fab/memory capital cycle thesis: demand-side visibility from Meta is a signal for Phase 2/3 duration. Thompson implicitly validates that the hyperscaler capex wave is not a bubble (these companies have clear ad ROI paths for the infrastructure).

phData client conversations — medium connection. The "every pixel monetizable" inventory expansion framing is a useful translation layer for explaining AI ROI to ad-tech or media clients. The pattern (GPUs → better recommendation → more engagement → more ad inventory → compounding returns) is a template for how to frame AI infrastructure investment ROI to C-suites skeptical of capex.

RDCO agent strategy — weak but present. The "off the clock" vs productivity framing is an interesting lens: RDCO's COO agent is a productivity/work tool, but Thompson's point that humans want connection and entertainment as relief from AI productivity is a useful counter-signal for where AI assistant fatigue could emerge.

⚠️ Sponsorship

No sponsor detected.

Related

[[2026-01-29-stratechery-meta-earnings-zuckerberg]] — Thompson's prior Meta earnings coverage; direct predecessor, same company/beat [[2026-01-14-stratechery-meta-compute]] — Meta compute capex analysis; foundational context for this piece's AI investment case [[2026-05-04-stratechery-google-meta-earnings]] — Q1 2026 Meta earnings coverage; immediate prior quarter context [[2026-05-28-stratechery-eric-seufert-interview-models-ads-ai-upside]] — ads × AI upside thesis from Eric Seufert; complements Thompson's inventory-expansion argument