Locked doors are easy to spot. The harder question is what happens when a powerful institution throws its gates open rather than bolting them shut — whether openness itself can become a more elegant form of control. Meta's AI strategy asks exactly that question.
The company has spent years building what it calls the "Linux moment" of artificial intelligence. Llama, its family of open-weight models, passed 350 million downloads by the end of 2024. Developers from Lagos to Bogotá fine-tune it, ship products on it, and treat it as a public good. It is an extraordinary narrative, and it contains a truth worth defending. It also contains a structure worth examining.
In my years translating Ethereum Classic documentation for Spanish-speaking readers, I learned that "open code" is less a legal fact than a living relationship of power. Sometimes walls keep people out. More often, in our industry, they simply make the inside harder to see.
The Architecture Behind the Generosity
Meta enters 2025 with roughly 350,000 H100-class accelerators under management — a compute reserve second only to Microsoft's estimated 500,000, and several times larger than Anthropic's or xAI's. It plans to push that toward the equivalent of one million H100s during 2025. Capital expenditures are slated to reach $60–65 billion, roughly 35–40% of revenue, compared with a historical range of 20–25%. This is not the budget of a research lab. It is the budget of a nation-state.
The technical strategy has three concentric layers. First, Meta open-sources its frontier models rather than selling API access. Second, it builds custom silicon — the Meta Training and Inference Accelerator, or MTIA — currently aimed at inference workloads rather than large-scale training. Third, it wraps both in a distribution layer no AI lab can replicate: WhatsApp, Instagram, and Facebook, together touching more than three billion human beings.
The conventional reading is that Meta is behind OpenAI and Google on raw model quality and is compensating through ecosystem. Llama 3.1 405B, the strongest open-weight release to date, sits within 5–10% of closed frontier models on complex reasoning and code generation benchmarks. The gap is real but narrowing, and on the LMArena leaderboard, Llama already stands as the best open model available.
But raw capability was never the point. The point is what happens around the model.
The Moat That Calls Itself Open
Let's be precise about the license that governs Llama. It is generous for most developers — free use below 700 million monthly active users — but it is also unilateral and conditional. Cross that threshold and you need Meta's written permission. Use Llama to train a competing large language model and you are explicitly barred. The steward can change terms, add restrictions, or withdraw access at the next version boundary. This is not a conspiracy; it is just how infrastructure wars are waged in the age of artificial intelligence.
The old Ethereum Classic slogan — code is law — was always incomplete. If code is law, then licenses are its treaties, and these particular treaties have a single signatory.
None of this negates the real value Meta has created. By flooding the market with capable open weights, Meta has compressed the pricing power of closed labs and given thousands of startups a foundation they could never have afforded. For small teams in the Global South, Llama is genuinely liberating. That is not marketing. It is material.
But liberation and decentralization are different categories. A city with open gates is still a city. It has a center, a tax system, and a police force. The question is who built the roads and who decides when they close.
Where the Money Actually Lives
Reading Meta's numbers, I keep returning to a single phrase from its Q4 2024 earnings call: AI-driven recommendation improvements lifted time spent on Facebook by 8% and on Instagram by 6%, while lifting ad conversion by roughly 10%.
Apply even a conservative fraction of that to an advertising business generating around $160 billion annually, and you arrive at tens of billions in incremental value — before considering a single enterprise API sale. This is the part of Meta's AI story most analysts get backwards. The company is not trying to become OpenAI. It is using AI to make its attention engine more precise, more predictive, and more profitable. The model is the means; the recommendation system is the end.
This is also where the ethical question sharpens. The same machine-learning pipeline that optimizes ad conversion optimizes for retention, and the same ranking systems that keep users scrolling can amplify outrage, misinformation, and algorithmic dependency. Meta's AI is not neutral infrastructure; it is the nervous system of a surveillance economy serving three billion participants. Regulators in Brussels have already opened investigations into Meta's "pay or consent" model, and the European Digital Services Act casts a long shadow over its content-ranking decisions. None of these risks appear in the trillion-dollar narrative, but they are embedded in its architecture.
The Trillion-Dollar Timeline
The headline claim — that Meta's AI initiatives could drive the next trillion-dollar phase by 2027 — deserves a colder look. Meta's market capitalization sits near $1.5 trillion. Adding another trillion implies a $2.5 trillion company within roughly two years. What would have to be true?
In the optimistic scenario — perhaps a 25% probability — AI-driven ad gains continue compounding at double digits, Meta's cloud business reaches $10 billion in annual revenue, and Ray-Ban Meta smart glasses, already past two million units sold, become a genuine hardware category. That path supports roughly $220 billion in 2027 revenue and, at 25 times earnings, a $2.5 trillion valuation.
In the base case, which feels more honest, ad gains moderate, cloud revenue lands between $3–5 billion, and glasses remain a promising experiment. Revenue reaches $190 billion, the multiple compresses slightly, and Meta crosses the trillion-dollar threshold somewhere between 2028 and 2030. The 2027 date is less a forecast than a hope.
What the analysis rarely mentions is the cost of the journey. Free cash flow is projected to fall from roughly $50 billion to $30–35 billion as capital expenditures consume the business. Buybacks will shrink. Dividends may pause. The market is being asked to fund a war chest today in exchange for a payoff that Meta itself cannot date with confidence. During a crypto bear market, we learned what happens when investors stop funding long-dated promises. Public equities have longer patience than cryptocurrency, but not infinite patience.
The Contrarian Test
There is a counterargument worth taking seriously. Meta's open-source strategy might be the most effective restraint on AI centralization that we currently have. Every startup that builds on Llama is a startup that does not depend on OpenAI's API. Every researcher who audits Llama's weights is a researcher who can see inside a frontier model for the first time. If we demand ideological purity — refusing to accept openness from a giant because the giant remains powerful — we risk policing the very ecosystem we want to protect.
I have sympathy for this view. During my years auditing L1 consensus mechanisms, I learned to distinguish between the distribution of copies and the distribution of control. Many "decentralized" networks failed precisely because they confused replication with sovereignty. By that standard, Meta's open weights are far more honest than most blockchain projects I have encountered.
But honesty at the center is not the same as safety at the edge. The meaningful test for any allegedly decentralized AI system is threefold. First, can it survive a change in its author's strategy? Second, can inference run outside the author's hardware? Third, can the data that shaped it be inspected and contested by the communities it affects? Meta currently passes the first test conditionally, fails the second almost entirely, and has no answer for the third.
Open weights without open infrastructure are a gift with strings attached. The strings are not visible in the immediate experience of the developer. They become visible only when the steward's incentives shift — and incentives always shift.
The Path Beyond the Gate
Perhaps the real lesson is not that Meta is malevolent, but that centrality is structural. Compute at this scale does not decentralize on its own; it concentrates by physics and by capital. A genuinely open AI future requires not just downloadable weights but portable inference, verifiable computation, auditable training data, and user-controlled identity — the very primitives that crypto has spent a decade building.
Meta has built the most impressive open-source castle in the history of software. The gates are wide. The courtyard is crowded. But the keys remain in a single hand, and a gate, no matter how hospitable, is still a gate.
We chart the code, but the soul chooses the path.
