Meta's Scorched Earth Moment: Power, Not Models, Is the Moat
Chamath says Meta should have gone scorched earth two years ago but power constraints make the timing better now. The Q2 earnings call and a $50 billion Louisiana bet show why he is half right.
Meta spent $31.1 billion in one quarter, about half its revenue, to build data centers it will not fully own. At the same time its flagship open source model has not had a major release since April and its enterprise share is under pressure. That tension is the whole debate around Chamath Palihapitiya's claim that Meta should have played scorched earth two years ago but is in an even better position now because of power and compute constraints.
The Q2 2026 earnings call shows Meta agrees with the second half. Zuckerberg is no longer selling a model lead. He is selling power, compute, and distribution.
Chamath's argument
On August 6, Chamath posted: "Tactical Game Theory: Meta Scorched Earth. This should have been Meta's play two years ago. That said, they are in an even better position to do it now considering the power and compute constraints that are emerging." (@chamath)
Two years ago scorched earth meant open source. Give away Llama, commoditize the model layer, and win on distribution. Today it means infrastructure. If power and transformers are the bottleneck, owning gigawatts behind the meter is a moat that a better benchmark score cannot replicate. Others say it is too late because Llama has lost momentum and Meta has no cloud to rent the capacity to.
Both can be true.
What Meta actually said
The July 29 Q2 2026 call was the clearest pivot yet. The press release and Q1 transcript frame it together.
On Q2 results: revenue $60.8 billion, up 28% year over year and 8% quarter over quarter, with Family daily active people at 3.60 billion, up 3% year over year. Ad impressions up 14%, average price per ad up 12%.
On AI progress: Zuckerberg said "AI is accelerating our core business today, powering our next generation of products, and opening the door to entirely new enterprise opportunities." The Q1 call introduced Muse Spark as the first model from Meta Superintelligence Labs, with Meta AI sessions per user up double digits and the standalone Meta AI app near the top of app stores. Management said teams are already training successor models beyond Spark.
On infrastructure: The Q1 call raised 2026 capex guidance to $125 to $145 billion from $120 to $135 billion, citing elevated component pricing and data center costs, and disclosed multiyear commitments up $107 billion in a single quarter. The hedge is custom silicon with Broadcom plus AMD alongside Nvidia systems, with more than one gigawatt of own silicon rolling out. The Q2 update narrowed capex to $130 to $145 billion and raised the expense outlook to $165 to $169 billion to include $2.4 billion in legal charges.
For context on how Meta got here, see Meta's Llama Is Gone. Muse Spark Is the New Bet, which traced the April 8 decision to close the model and move to a paid API.
The capex bet: Hyperion and power
Hyperion in Richland Parish, Louisiana is now described as a 5 GW supercluster costing over $50 billion, up from $27 billion when the Blue Owl Capital joint venture was announced in October 2025. That is higher than the $27 billion figure revealed in October, when Meta and Blue Owl formed a joint venture to help with the buildout, originally planned as a 2 GW facility (CNBC).
The structure matters. Hyperion sits in a special purpose vehicle, with Blue Owl putting in $7 billion and Meta putting in $2 billion plus several billion already spent, according to the New York Times. That converts capex to operating expense and keeps the build going without putting the full $50 billion on Meta's balance sheet. The same off balance sheet pattern is appearing across hyperscalers, as detailed in AI data center debt, the $7 trillion question.
Power is the point. Louisiana tax rebates help, and hyperscalers plan $660 to $690 billion in 2026 capital spending, mostly for AI data centers, with Amazon about $200 billion and Alphabet $175 to $185 billion, according to Who Is Paying for All These Data Centers?. That article also notes PJM capacity prices rising from $28.92 per MW day for 2024/25 to $329.17 for 2026/27, a direct signal that grid power is getting more expensive and harder to secure.
Q2 numbers show the cost
Revenue was $60.8 billion, up 28% year over year and 8% quarter over quarter, with gross margin 81.4%. Operating income was $18.8 billion, margin 31% versus 40.6% in Q1. Net income was $15.8 billion, down 14% year over year, and EPS of $6.18 missed the $7.10 estimate by 13% after beating by 7% in Q1. Capex including finance leases was $31.08 billion, or about 51% of revenue on that basis, versus $19.0 billion in Q1. On the database definition ex leases, capex was $30.1 billion, or 49.5% of revenue. Free cash flow was $784 million on the official non-GAAP basis, down sharply from $12.4 billion in Q1. Cash and marketable securities were $90.3 billion against long term debt of $83.7 billion and total debt of $112.3 billion.
The chart makes the pivot visible. Revenue grew steadily. Capex inflected sharply in Q2. That is scorched earth in dollars.
How hyperscalers are turning that spend into an asset class is explored in Compute Is Becoming a Standalone Asset Class. Blackstone Just Bet $185 Billion on It..
Llama stalled, Muse Spark takes over
Llama remains widely adopted in open weight enterprise use, but momentum has stalled. No major release since Llama 4 in April. One widely cited enterprise survey put open source share falling and Meta's share within open source down from 76% in December 2024 to 69% in June 2025, with Qwen and DeepSeek taking share, though methodology varies.
Llama 4 Scout and Maverick were described as not materially better than DeepSeek V3, with a smaller expert count and a rushed release. The reasoning version of Behemoth is still pending while competitors have shipped.
Distribution still helps. Meta AI crossed 1 billion monthly actives on May 28, up from 400 million in September 2024, and the company reports 8 million advertisers using at least one GenAI creative tool. WhatsApp is the top surface for Meta AI queries. The new attempt to monetize is Meta Launches Muse Code, A New AI Coding Agent Powered By Spark 1.2, priced at $1.25 per million input tokens and $4.25 per million output tokens.
The broader open versus closed debate is covered in 60% of Enterprises Use Open-Weights. That Does Not Mean Closed Models Are Losing, which finds closed models still win spend while open wins volume.
How Meta stacks up vs Google and OpenAI
Google is training Gemini 3 largely on its own TPU, giving it a cost edge on inference. OpenAI still leads on mindshare but relies on hyperscaler balance sheets to build.
Meta's edge is different. It owns the customer. With 3.60 billion daily actives it can auction compute like it auctions ads, and it owns a large share of the AI glasses market via EssilorLuxottica, a hardware moat competitors cannot shortcut quickly. Its weakness is also different. It has no cloud like Azure or Google Cloud to rent capacity to, so every dollar of capex must be monetized through ads, subscriptions, or agents.
Valuation reflects that skepticism. At $591.06 on August 6, up 0.39% on the day, Meta trades at 22.2 times trailing earnings and 16.8 times forward earnings, with a PEG of 0.83 and EV to EBITDA of 13.9 times. That is cheaper than Alphabet at 24.3 times forward and Microsoft at 21.2 times. The Street remains constructive, with 57 analysts at Strong Buy and a mean price target of $759, about 29% above the current price, with a range of $580 to $1000. The comparison to Amazon's print is instructive, as we noted in Amazon Proved AI Capex Pays. Meta Is Still Asking for Patience: the market rewards capex when revenue is visible.
Pivot or desperation
It is a strategic pivot executed with desperation timing.
Strategic, because if power and grid interconnection are the binding constraint, owning 5 GW behind the meter is a moat you cannot buy in 2027. Auctioning compute, selling business agents, and charging for subscriptions gives Meta its first cloud like revenue line. Custom silicon hedges cost.
Desperation, because the timing was forced. Llama 4 underwhelmed, Behemoth is delayed, and open source share is under pressure. Capex at about half of revenue with no cloud to absorb it means Meta must invent enterprise demand while free cash flow is near zero. The execution risk of building multiple gigawatt campuses at once keeps the bar high.
Chamath is right that constraints help Meta now. Two years ago power was abundant and model quality mattered more. Today power scarcity makes Hyperion defensible. Critics are also right that on models Meta is no longer setting the frontier, it is chasing it.
What to watch
- Prometheus energization in 2026 and Hyperion milestones into 2027, plus lease terms when the Q3 10-Q is filed
- Whether Q3 capex moderates from about 51% of revenue including leases and free cash flow recovers from $784 million
- Muse Spark enterprise API traction and the Behemoth reasoning release, which needs to close the gap on benchmarks
- PJM and MISO capacity prices and transformer lead times, which determine whether power scarcity stays a tailwind for owners of generation
More research at bargo.ai/research.
Sources
- Meta Q2 2026 Press Release, Jul 29 2026 — investor.atmeta.com
- Meta Q1 2026 Earnings Call Transcript, Apr 29 2026 — The Motley Fool
- CNBC, Jul 13 2026 — Meta's Louisiana data center investment to reach $50 billion — cnbc.com
- Worth, Aug 6 2026 — Who Is Paying for All These Data Centers? — worth.com
- CNBC, Aug 5 2026 — Meta Debuts First AI Coding Agent to Take On Anthropic and OpenAI — cnbc.com
- Chamath Palihapitiya on X, Aug 6 2026 — Tactical Game Theory: Meta Scorched Earth — @chamath