Meta는 오픈 소스 AI를 죽였습니다.

Meta Just Killed Open-Source AI Meta gave away the best AI infrastructure in the world, let everyone build on it, then pulled the ladder up. This isn't betrayal. It's the pattern. Meta just killed open-source AI. Not gradually. Not with a transition plan. One announcement: Muse Spark is proprietary, Llama is deprecated, and the 1 billion downloads that made Llama the default AI infrastructure for thousands of startups now lead nowhere. Mark Zuckerberg called open source "the path forward" in October 2024. Eighteen months later, the path has a dead end sign. If you built your company on Llama, you're now holding a dependency on abandoned infrastructure. This post is about why Meta did it, why it was predictable, and what you should have done differently. The Android Playbook Google gave away Android for the same reason Meta gave away Llama: to prevent a competitor from owning the platform layer. In 2008, Apple had the iPhone, a closed ecosystem, and the beginning of a monopoly on mobile computing. Google didn't sell phones. Google sold ads. So Google made Android free, flooded the market, and ensured that no single company could control mobile access. The strategy worked. Android has 70% global market share. Google makes $200B+ annually from the ad inventory that free Android creates. Meta's Llama playbook was identical. OpenAI was building the "Apple of AI" — closed, premium, platform-controlled. Meta couldn't compete on product (ChatGPT had the interface, the users, the momentum). So Meta competed on infrastructure: give away the model, let everyone build on it, create path dependency. The 1 billion Llama downloads weren't a metric of generosity. They were a metric of lock-in. The difference is that Android's economics were sustainable forever. Google's marginal cost of Android distribution is zero. Llama's marginal cost of frontier model training is now $100M+ per run and climbing. That difference is everything. Why the Economics Stopped Working In 2023, training a frontier model cost roughly $10 million. Meta could afford to give away the weights because the cost was marketing spend, not capex. The goodwill, talent acquisition, and ecosystem lock-in justified the expense. In 2026, Meta is spending $115–135 billion on AI infrastructure. Frontier training runs are 50–100x more expensive than they were three years ago. Every Llama improvement immediately accrued to competitors — including Chinese AI labs building commercial products on open weights. Meta wasn't just giving away infrastructure. It was subsidising its own competition. The equation flipped when Llama 4 flopped. If your open-source model is winning, the subsidy is strategic. If your open-source model is second-place, the subsidy is self-harm. CNBC's report that Llama 4 "failed to captivate developers" meant the strategy wasn't even buying market position anymore. It was just burning money. That's when you pivot. The Real Moat Was Never the Model Muse Spark's most interesting feature isn't its reasoning mode. It's Shopping Mode — an AI-driven commerce experience integrated into Facebook, Instagram, and WhatsApp's 3.3 billion users. The model weights don't matter. The proprietary user data graph matters. Your social connections, purchase history, behavioural patterns, conversational context — that's the moat. And it's only a moat if it's closed. This is the pattern every AI company is converging on. OpenAI doesn't sell GPT-5. It sells ChatGPT Plus subscriptions, enterprise APIs, and platform lock-in. Anthropic doesn't sell Claude's weights. It sells Cowork, Claude Code, and managed agents. The model is infrastructure. The product is the interface, the data, the workflow integration, the distribution. The companies that understand this are building moats. The companies that don't are arguing about whether Llama 3.1 or GPT-4 is better on MMLU. What This Means for Startups If you built on Llama, you have three problems right now. Problem 1: No migration path. Meta deprecated Llama without a forward path. Your fine-tunes, LoRAs, and custom adapters don't transfer. That work is sunk cost. Problem 2: The cost cliff. Self-hosted Llama was cheap. Moving to API-based GPT/Claude/Muse Spark raises unit economics by 5-10x. If your margin structure assumed free inference, your business model just broke. Problem 3: It will happen again. Meta's pivot is a pattern, not an exception. Any open-source model maintained by a single company is a contingent bet on that company's business model remaining stable. Google could do this with Gemma. Mistral could do this tomorrow. Open-source maintained by corporations is not a foundation. It's a temporary subsidy. The Architecture Lesson I run seven AI agents across five different model providers. None of them depend on a single model family. When Moonshot's balance runs low, I shift to Anthropic. When Anthropic's context window isn't enough, I use GPT-4. When I need reasoning depth, I use Claude. When I need speed, I use kimi-k2.6. The model was never the bottleneck. The architecture was. This is what Karpathy calls "autoresearch" — not using one model for everything, but decomposing work into stages where each stage gets the right intelligence in the right context. The research agent reads. The writing agent produces. The review agent checks. Each can swap models without the system breaking. The startups that survive Meta's pivot are the ones who built this way: model-agnostic pipelines where the value is in the workflow, not the weights. The Contrarian Take The "open-source vs. proprietary" debate is becoming a distraction. The real question is: who controls the distribution layer? Meta doesn't need to sell Muse Spark. It needs Muse Spark to make Instagram Shopping 15% more effective. OpenAI doesn't need to win on benchmarks. It needs ChatGPT to be the default interface for 900M+ users. For founders, this means three things: One: Don't build on someone else's free model as your permanent foundation. Treat open-source models as starting points, not strategies. Budget model-switching as a line item. The half-life of your current model dependency is shorter than you think. Two: Build moats in data, workflow, or user behaviour — things that don't transfer when you switch models. If your value is "we fine-tuned Llama 3.1," you don't have a business. You have a hobby. Three: Watch the Chinese labs. DeepSeek, Qwen, and GLM are now the most credible open-source frontier options. They're also geopolitically unstable — subject to sanctions, export controls, and data sovereignty concerns. Betting on them is betting on US-China relations remaining stable. Good luck with that. The Precedent In 1985, nobody could have told you that "UX designer," "data scientist," or "growth marketer" would be among the most common job titles in business by 2020. Those roles didn't exist. The people who worried about computers taking jobs weren't wrong that certain jobs were ending. They were wrong that new categories wouldn't emerge. The same thing is happening now with AI infrastructure. The categories are shifting. The companies that survive aren't the ones with the best model. They're the ones with the architecture that doesn't care which model wins. Meta pulled the ladder up. The question isn't whether they'll drop it back down. The question is whether you should have been climbing it in the first place. Utkarsh is the co-founder and CEO of AbleCredit and AbleWorks. He writes about AI, infrastructure, and the patterns that repeat across technological revolutions.