Meta's Llama Is Gone. Muse Spark Is the New Bet

From open-source champion to closed API: why Meta killed its scorched-earth strategy and what $130 billion in capex needs to prove

Meta spent three years telling developers that open source AI was the path forward. On April 8, 2026 it launched a closed model and stopped investing in Llama. The pivot explains why Meta is now spending $130 to $145 billion a year on data centers while free cash flow collapsed, and why the market punished Meta while rewarding Microsoft and Amazon for the same capex surge.

The scorched earth that never happened

On Aug 6, Chamath argued Meta should have played scorched earth two years ago by flooding the market with open models, but is in an even better position now because power and compute constraints are emerging.

The logic was simple. If models are free, the value moves to distribution. Meta has 3.6 billion daily users across Instagram, Facebook, WhatsApp and Ray-Ban glasses. Commoditize the model, win on distribution.

Meta did the opposite. Instead of open-sourcing the frontier, it closed it.

That choice is the core of the strategy debate. Is this a disciplined pivot to monetization, or a forced retreat after Llama fell behind?

Llama is gone

For practical purposes, Llama is gone. Meta has not formally deleted the weights, but Meta abandons open-source Llama for proprietary Muse Spark (The New Stack, Apr 30, 2026) and Meta Ended Llama and Built Muse Spark (Miraflow, Apr 20, 2026) describe the same shift.

• Last frontier release was Llama 4 Scout and Maverick in April 2025. No major release since. Community threads now ask why Meta has not dropped any new Llama version lately.
• Meta's own line is that current Llama models will continue to be available as open source. That covers existing weights only. It says nothing about future development. The expectation inside the AI community is incremental maintenance, not frontier investment.
• There is no migration path. Llama was downloadable weights you could self host and fine tune. Muse Spark is cloud only, no weights, private API preview only. Different deployment model, different tooling.

The data was already moving before the decision. Menlo Ventures enterprise data shows total open-source LLM share fell from 19% in Dec 2024 to 13% in Jun 2025 to 11% in Dec 2025. Llama still holds about 70% of that shrinking open-source slice, but the pie is shrinking. Reviews of Llama 4 called it a bit odd, with 16 to 128 experts versus DeepSeek V3's 256 expert high sparsity, benchmark gaming rumors, and coding performance closer to Alibaba's Qwen QwQ 32B than to frontier.

This is the context for our earlier look at why 60% open-weights does not mean closed models are losing. Volume and spend are different games, and Meta was losing the volume game it invented.

Muse Spark: the new playbook

Muse Spark is not a Llama upgrade. It was built from scratch by Meta Superintelligence Labs, a new division formed in 2025 after Zuckerberg recruited Scale AI's Alexandr Wang. Wang said nine months ago we rebuilt our AI stack from scratch with new infrastructure, architecture and data pipelines.

What it is:

• Natively multimodal, trained jointly on text, image, video and audio
• Three reasoning modes: Instant, Thinking and Contemplating with multi-agent orchestration
• Claims 10 times less training compute than Llama 4 Maverick for similar capability
• First Meta model with a price tag: public API launched July 9, 2026 at $1.25 per million input tokens and $4.25 per million output tokens, with an OpenAI compatible endpoint. Muse Spark 1.2 followed Aug 5, 2026 focused on coding.

Distribution is the moat. Spark powers Meta AI across Instagram, Facebook, WhatsApp and Ray-Ban glasses, plus a new Shopping Mode for conversational commerce. The thesis is that owning the user surface matters more than owning the open-source goodwill that drove 1.2 billion Llama downloads.

This mirrors the broader shift we covered in Jensen Huang on why closed models are cheaper and open models are about control. Meta is now competing on closed-model terms.

The $130 billion question: capex vs cash flow

Meta's Q2 2026 print (ended Jun 30) shows why the market is anxious.

• Revenue $60.8B, up 28% year over year and 8% quarter over quarter. Still strong.
• Operating income $18.8B, margin 30.9% versus 40.6% in Q1. Net income $15.8B, down 13.6% year over year.
• Capex $30.1B in Q2 alone, 49.5% of revenue, versus $19.0B in Q1 and $16.5B a year ago. R&D $21.7B, 35.6% of revenue.
• Free cash flow $1.7B, down 86.8% quarter over quarter and 80.6% year over year. Fool flagged a 91% plunge to $784M on a slightly different FCF definition, same direction.
• Balance sheet flipped to net debt: $90.3B cash versus $112.3B debt, negative $22.1B net cash versus positive $20.7B a year ago.

Guidance raised, not cut. Meta guided 2026 capex $130 to $145B and opex $165 to $169B. Morgan Stanley expects global cloud capex to reach $1.2 trillion in 2027, up 30% year over year, with Alphabet, Amazon and Meta all raising 2026 guidance. Bernstein details Meta's Hyperion project jumping $27B to $50B plus a $14B BlackRock JV in El Paso and a $9B Canada site.

The market verdict was split. As we detailed in Amazon proved AI capex pays, Meta is still asking for patience, Microsoft disclosed 30 million plus paid Copilot seats and was rewarded, Amazon grew AWS 36.7% and was rewarded, Meta grew 28% but raised spend without quantifying AI revenue and was punished. That gap is the entire investment case.

META closed $588.77 on Aug 5 and traded $589.97 intraday Aug 6, down from a July 10 high of $669.21 after earnings. RSI 28.9 signals oversold, but the move reflects capex fear, not accumulation.

Valuation reflects that fear. At $589.97, Meta trades at 22.1 times trailing earnings, 16.8 times forward earnings, PEG 0.83 and EV/EBITDA 13.9 times, cheaper than Microsoft at 27.5 times trailing and Alphabet at 18.2 times. Consensus is strong buy 1.35 out of 5 from 57 analysts, mean price target $759, about 29% upside, range $580 to $1000. Barclays at $780 argues selling intelligence via APIs, tokens as a service and bare metal rental could be $65 billion plus in annual revenue in a few years.

Competitive position vs OpenAI and Google

Chamath is right that constraints help incumbents who already secured power and chips. Gas turbine queues run about 36 months and a handful of suppliers gate output. That favors anyone with locked in power, not just open-source players. See Power is the new GPUs for why 40% of announced data center projects are stalling.

But Meta now competes on OpenAI and Google's terms, and the scoreboard is mixed.

• Pricing: blended price per million tokens is about $0.44 for open-source, $1.13 for Google, $1.56 for xAI, $3.50 for OpenAI and $10.00 for Claude. Muse Spark at $1.25 input and $4.25 output undercuts OpenAI but not Google or open-source.
• Adoption: open-source still leads volume, but OpenAI reaccelerated. Last 30 days, open-source held about 70% of tokens, OpenAI 12% with triple digit growth, Google 8% and Claude 8% declining.
• Frontier: benchmarks are saturated. Only FrontierMath still discriminates. Meta is not setting the frontier. Muse Spark scores 51 on the Artificial Analysis Intelligence Index, competitive on cost per token, mid pack on pure reasoning.

In short, Meta traded a volume lead in open-source for a share fight in closed APIs where OpenAI is reaccelerating and Google has TPU cost advantage.

What to watch

• Llama maintenance versus Spark investment. Does Meta ship any meaningful Llama update, or is maintenance the final word?
• Muse Spark adoption. API usage, developer retention from Llama, and whether Shopping Mode and Meta AI surfaces convert to paid inference.
• Capex discipline. Q3 actual versus $130 to $145B guide, and any disclosure of AI revenue increment. The market rewarded specificity from Microsoft and Amazon and punished its absence from Meta.
• Power execution. Not announced dollars, but delivered gigawatts and turbine slots for Hyperion, El Paso and Canada.

Meta's strategy is now coherent: close the model, monetize the distribution, and outspend on power and compute while constraints last. Whether that is a pivot or desperation depends on whether Spark can close the frontier gap before the capex bill comes due.