Meta · Llama

Meta's open-weight LLM family — the only frontier-tier model line shipped by a hyperscaler advertising company, with no native API business and the largest disclosed GPU fleet in the industry.

1. Core Product / Service

• Llama model family — Llama 1 (2023-02, leaked then partially open), Llama 2 (2023-07, open weights with community license), Llama 3 (2024-04, then 3.1 / 3.2 / 3.3), Llama 4 (2025) generation, with reports of a Llama 5 / next-gen line in development as of 2026. Mixture-of-Experts at the top end; Llama 4 introduced "Behemoth" / "Maverick" / "Scout" tiered SKUs spanning small-on-device to frontier-class parameter counts.
• License: Llama Community License Agreement — open weights with Acceptable Use Policy + a 700M monthly-active-user threshold above which Meta requires a separate commercial license. Permissive enough for ~99% of the developer ecosystem, restrictive enough that AWS / Google / Microsoft / TikTok-class platforms cannot freely commercialize Llama as their primary commercial model without a Meta agreement.
• Meta AI — the consumer-facing assistant integrated into Facebook, Instagram, WhatsApp, Messenger, and the standalone Meta AI app. Powered by Llama. This is the distribution moat — billions of users reachable without any new app.
• Ray-Ban Meta + Quest — embodied AI surfaces using Llama-derived multimodal models.
• No 1P API business — Meta does not sell Llama tokens commercially. There is no "Llama API" endpoint at meta.com. Llama is monetized entirely indirectly: ad targeting + engagement on Meta's own platforms, plus third-party hosting via together-ai, DeepInfra, Groq, Fireworks, AWS Bedrock, Azure AI, openrouter — none of which Meta takes a cut of.

2. Target Users & Pain Points

• Open-weight ecosystem — anyone running self-hosted inference (vLLM / SGLang / TGI via ai-inference-engines) for cost / privacy / customization reasons. Llama is the default open-weight choice in the Western ecosystem (Mistral and DeepSeek/Qwen split the rest).
• 3P inference providers — together-ai, DeepInfra, Groq, Fireworks, etc. monetize Llama-the-product without paying Meta licensing for sub-700M-MAU usage. Llama is the commodity supply that the L3b layer competes on.
• Meta itself — billions of consumer users via Meta AI in Facebook / Instagram / WhatsApp.

Pain solved (for the ecosystem): a credible open-weight frontier alternative to closed labs without the Chinese-origin geopolitical concerns of DeepSeek / Kimi / Qwen.

3. Competitive Landscape

Meta's positioning is uniquely structural: it is the only frontier-tier lab whose business model does not require Llama to generate token revenue. Meta's $150B+ annual revenue comes from advertising; Llama exists to (a) improve Meta's own products, (b) commoditize closed-AI competitors' moats, (c) attract talent / dev mindshare.

4. Unique Observations

• Frontier training cost (Llama 4 / 5-class): Meta has been the most transparent of the closed-camp on capex. Mark Zuckerberg disclosed in 2024 that Meta would deploy roughly 350,000 H100-equivalent GPUs by end of 2024 and grow to ~600,000 H100-equivalents in the broader AI compute fleet. By 2025–2026 the fleet has crossed the 1M+ AI accelerator mark including B200 / Trainium-class additions. At those numbers, single-run frontier Llama 4-class training is plausibly multi-hundred-million to ~$1B compute-amortized, scaling further for Llama 5. This is frontier-bracket spend even though Meta does not publish the per-model number.

• API pricing — top SKU: N/A — there is no Meta Llama API. The closest thing to a "list price" is the cost to run Llama on third-party inference providers. Reference prices (2026-05): Llama 3.3 70B / Llama 4 large MoE on together-ai / DeepInfra / Fireworks land in the $0.20–$0.90/M input · $0.30–$0.90/M output range — the cheapest open-weight frontier option in the Western stack. Meta captures zero of that revenue directly.

• Pricing vs estimated unit cost — gross margin signal: this is the most distorted entry in the batch because Meta does not price Llama as a token product. The ~$0.50/M output 3P market price is set by competitive pressure between Together / DeepInfra / Fireworks / Groq — none of whom paid the training cost. Meta's effective "margin" on Llama is best measured as (ad revenue lift from Meta AI in FB/IG/WA) + (commoditization damage inflicted on OpenAI / Anthropic margins) − (training + inference capex). The first two terms are large and growing; the third is in the tens of billions per year. Whether the trade-off pencils out is a strategic, not unit-economics, question.

• Open vs closed strategy + rationale: fully open weights (community license). Stated rationale, repeated by Zuckerberg in shareholder letters and the open-source AI manifesto (2024): (a) commoditize the model layer to prevent any single closed competitor from controlling the AI substrate Meta's products depend on; (b) attract developer + research mindshare to keep Meta's ML talent pipeline strong; (c) leverage external scrutiny + community contributions on safety / capability. The strategic logic is the inverse of OpenAI / Anthropic — those labs need closed weights to monetize tokens; Meta needs open weights so that no one ELSE controls a closed token monopoly that Meta would have to pay for.

• No API biz — commercial implications: this is the cleanest test in the industry of whether distribution beats monetization in foundation models. Meta has spent more on Llama infrastructure than any Western lab except possibly OpenAI / Stargate, and earns no direct token revenue. The bet pays off if (a) Meta AI inside FB/IG/WA materially lifts ad revenue, and (b) a commoditized open-weight ecosystem prevents any single closed competitor from extracting Microsoft-style platform rents. Both are early-stage claims as of 2026.

• Vertical integration — own DCs, NVIDIA-anchored, Trainium experiments: Meta operates its own hyperscale data center fleet (one of the four largest in the world along with AWS / Azure / GCP), purpose-built for the Llama training + recommendation-system workload. Hardware: NVIDIA H100 / H200 / B200 anchor (the ~350k+ H100 buildout disclosed 2024), supplemented by Meta's own MTIA (Meta Training and Inference Accelerator) custom silicon for inference-heavy recommendation workloads (deployed at growing scale). This is the deepest L3 → L1 vertical of any frontier lab outside Google — Meta owns the building, the network, the power, deploys NVIDIA + custom silicon, runs the largest training jobs in the industry on its own DCs. Compare: OpenAI Stargate (in build), xAI Colossus (built fast, single site), Anthropic (no DC), DeepSeek (no DC, rents High-Flyer-bought H800s), Google (full TPU + DC + fab partner, the closest peer).

• Advertising vs research justification: Meta's full-year capex was ~$37B in 2024 and stepped up materially in 2025 (announced 2025 capex guidance was in the $60B–$65B+ range). The majority of that increase was AI infrastructure — both for Llama training + Meta AI inference, and for the recommendation systems that drive Reels / FB / IG ad revenue. Ad-revenue growth (Q4 2024 ~$48B+ Meta total revenue, 2025 continuing in the $40–$50B/quarter band) is what funds it. This is the single largest justification flywheel in AI: ads pay for compute, compute trains Llama + recommendation models, those models lift ads. No other frontier lab has this loop.

5. Financials / Funding

• Parent / structure: Meta Platforms Inc. (NASDAQ: META). Llama is not separately funded — fully internal R&D under FAIR (Meta AI Research) plus the GenAI product team.
• 2024 revenue: ~$164B (full year). 2025 continued the growth trajectory.
• 2025 capex guidance: $60–$65B+ range, AI-weighted.
• GPU fleet disclosures: 350k H100-equivalents by end of 2024 (Zuckerberg public statement), growing to 600k+ H100-equivalent across the broader AI fleet by 2025, with B200 / Trainium / MTIA additions through 2025–2026.
• Llama-specific revenue: $0 (no API). Indirect revenue (ad lift attributable to Meta AI) is not separately disclosed.

6. People & Relationships

• CEO: Mark Zuckerberg.
• Chief AI Scientist (FAIR): Yann LeCun — long-standing public advocate of open-source AI.
• GenAI product leadership: Ahmad Al-Dahle (VP GenAI), Joelle Pineau (head of FAIR Labs, prior).
• Notable alumni who left for AI labs: Guillaume Lample + Timothée Lacroix (founded mistral), many ex-FAIR researchers across the industry.
• Hardware partners: NVIDIA (H100 / H200 / B200 anchor), AMD (MI300X selected workloads), TSMC (MTIA fab partner), Broadcom (network).
• Distribution (3P, no Meta cut): together-ai, DeepInfra, Fireworks, Groq, AWS Bedrock, Azure AI, openrouter, Hugging Face.
• Competitors (closed): openai, anthropic, google-deepmind, xai.
• Competitors (open-weight): deepseek, kimi, qwen, mistral — the open-weight ecosystem Llama anchors but does not exclusively own.