Hundreds of millions of dollars. Trillions of tokens.

On August 20, Bloomberg dropped a figure that left the AI community with mixed feelings: Meta spends hundreds of millions of dollars annually purchasing third-party AI model services through Microsoft Azure, consuming trillions of tokens every week. Among those purchased models, according to reports, are OpenAI's — Meta engineers use GPT via Microsoft Foundry to do what? To benchmark, evaluate, and score Meta's own Llama models.

The company that champions open-source AI most aggressively, that shouts "AI democratization" loudest, is paying OpenAI hundreds of millions a year in "evaluation fees." It's like Tesla buying a fleet of BYDs to tear down and study, then telling the world electric cars should be free.

More fascinating still is Microsoft's role. It's OpenAI's largest investor, the operator of Azure, the builder of the Foundry model marketplace — and it sells model access to everyone, including its investee's competitor Meta. Meta, meanwhile, is spending $130–145 billion building its own AI infrastructure while simultaneously paying hundreds of millions a year for its rival's models to compare against. Both companies declined to comment on specific figures.

AI industry "coopetition" reaches its absurd peak in these numbers.

What Exactly Is Meta Buying? Not Dependency — A Yardstick

Let's be clear: Meta isn't running its core business on OpenAI models, nor is it outsourcing Llama training to GPT. Cross-referencing Bloomberg and subsequent reporting from multiple outlets, Meta's core use case for external models is benchmarking and evaluation.

Here's what that means. Every time Meta's engineers ship a new Llama version, they need to know how it performs across various tasks. That requires a "yardstick" — running the same tasks against the acknowledged frontier models (GPT-5.6, Claude Opus 5, Grok 4.6), comparing output quality, identifying gaps, and guiding the next training round. Every question, every comparison, every scoring run translates into token consumption. When you have hundreds of engineers doing this daily, token burn becomes astronomical.

Meta CTO Andrew Bosworth actually foreshadowed this in a public comment in July, acknowledging the company rents leading external models as part of development work. Nobody imagined then that this "renting" had reached hundreds of millions annually and trillions of tokens weekly.

This also explains why Meta chose Azure Foundry over direct contracts with OpenAI. Foundry is Microsoft's model marketplace — it hosts not just OpenAI but Anthropic, Google Gemini, and multiple open-source models. Meta can call multiple competitors' models on one platform for head-to-head comparison without integrating separate APIs. For Meta it's the most efficient approach; for Microsoft it's pure arbitrage — it doesn't train models, it just plays "AWS for AI models" and takes a cut of every token transaction.

Microsoft: The Switzerland of AI, Selling Shovels to Everyone in the Gold Rush

Pan over to Microsoft's side of the table, and you'll find Satya Nadella playing a far bigger game than simply "investing in OpenAI."

Microsoft's latest earnings show that as of July 29, 2026, the Foundry platform boasted 100,000 customers with revenue more than doubling year-over-year. Customers using multiple model providers grew fivefold since the start of the year; customers running at a trillion-token annual run rate quadrupled. Azure's full FY2026 revenue topped $100 billion, up 41% — and CFO Amy Hood noted on the earnings call that "there are still constraints in the system" and "demand continues to exceed available supply."

Microsoft's sharpest move: it doesn't pick sides. It's OpenAI's largest shareholder (~49%), but Foundry sells more than OpenAI — Anthropic, Google Gemini, Meta's own Llama, and various open-source models all have a place. It has deep product integration with OpenAI (Copilot, 365 Copilot), but that doesn't stop it from selling OpenAI tokens to Meta. OpenAI generates roughly 70% of Microsoft's AI revenue, but Microsoft's ambition clearly isn't "be OpenAI's channel" — it's becoming the "utilities company" for the entire AI industry.

This mirrors Amazon building AWS — Amazon is an e-commerce company, but AWS sells cloud services to every e-commerce competitor, including Netflix (once AWS's largest customer and a direct Prime Video rival). Nobody refused AWS because Amazon was a competitor; it was simply the best cloud platform available. Microsoft wants to be the AWS of the AI model era.

That Meta is willing to spend this much is itself revealing: no matter how much you bash closed-source models or champion open source, when you need a "ground truth" to measure your own model against, you still buy the strongest closed model as your ruler. This isn't a question of faith — it's engineering.

Open Source vs. Closed Source: Not Binary Opposition, But Layered Coexistence

After news of Meta's OpenAI purchases broke, open-source and closed-source partisans went to war on social media. Open-source advocates accused Meta of "betraying the revolution." Closed-source loyalists declared it proof that "open-source models just aren't good enough." Both sides missed the point.

The truth is: open source and closed source play fundamentally different roles in the AI value chain. It's not a death match.

Closed frontier models (GPT-5.6, Claude Opus 5, Grok 4.6) serve as the capability frontier and evaluation yardstick — they define what AI can do today, the ceiling against which everyone measures. Every model builder, open or closed, needs to watch that ceiling to identify gaps. Training frontier models costs tens of billions; only companies with massive capital and compute (OpenAI, Anthropic, Google, xAI) can play at that table.

Open-source models (Llama, Qwen, DeepSeek, Mistral) serve democratization and deployment — they let enterprises and developers run models on their own infrastructure without data leakage fears, API price hikes, or vendor lock-in. Llama ignited the entire open-source AI ecosystem; Qwen and DeepSeek continue to squeeze American giants on value-for-money. All true.

Meta building open-source Llama is real. Meta buying OpenAI tokens for evaluation is also real. The two facts don't contradict each other. Toyota buys Teslas to tear down and study; that doesn't stop Toyota from selling hybrids and EVs. Apple buys Samsung phones to analyze; that doesn't stop Apple from making iPhones. Using a competitor's product to benchmark your own is the most normal engineering practice in tech.

Notably, Meta is also discussing launching its own model API service to compete directly with Foundry. That means Meta is Microsoft's customer today and could be Microsoft's competitor tomorrow. Role reversals happen that fast in AI.

The $135B Capex Context: Nobody Is Self-Sufficient in AI

Zoom out further for context: Meta just raised its 2026 capital expenditure forecast from $115–135 billion at the start of the year to $130–145 billion. It spent $31.08 billion on capex in Q2 alone. Most of that goes to data centers, custom chips (MTIA), and GPU purchases.

Even so, Meta still needs to spend hundreds of millions annually on external model services. That tells us a truth obscured by the "AI self-sufficiency" narrative: no company in this industry is truly self-sufficient.

Google has TPUs and Gemini, but Google Cloud sells NVIDIA GPUs and offers third-party models through Vertex AI. Amazon built Bedrock and its own models, but the bestseller on Bedrock is Anthropic's Claude. Microsoft invested in OpenAI but Foundry sells everyone's models. Meta built open-source Llama but buys OpenAI tokens for evaluation. Every company is balancing "build our own" against "use others'."

The logic is simple: AI moves too fast. You're the best today; you might be surpassed tomorrow. Any company that uses only its own models without watching what's happening outside quickly becomes the frog in the well. Bosworth said Meta aims to meet 2026 compute needs with its own infrastructure, but "third-party cloud contracts" are also part of the plan — that wording itself admits even a giant like Meta treats external models and compute as essential, not optional.

Endgame Judgment: No Iron Curtains in AI — Just a Trade Network

Many love framing the AI industry as "bloc warfare" — open vs. closed, US vs. China, big tech vs. startups. But Meta's nine-figure OpenAI token spend tells us: there are no iron curtains in AI, only an intricate trade network.

Microsoft sells OpenAI services to Meta; OpenAI reaches more enterprise customers through Microsoft's cloud; Meta uses OpenAI's capabilities to improve Llama; Llama's proliferation drives broader demand for AI infrastructure; Microsoft's cloud and Foundry business grows larger — it's a positive-sum loop, not zero-sum.

What does this mean for AI builders and enterprise users?

First, don't pick sides. Don't reject closed-source models because of open-source ideology; don't dismiss open source out of faith in closed providers. Choose tools based on use case, cost-performance, and data security needs — not ideology.

Second, model evaluation is essential and expensive. Meta burning trillions of tokens weekly on benchmarking shows that "knowing where your model ranks in the industry" is worth serious money. If you're building AI products, don't skimp on evaluation — if you don't compare against the best, how will you know where you fall short?

Third, Microsoft's marketplace strategy is a blueprint for all platform companies. Not picking sides, not locking into a single model, letting customers access all the best models on one platform — this "Switzerland" approach is actually the safest bet in an era of rapid AI iteration. Stripe's OpenRouter acquisition and Ramp's Router launch are essentially the same play.

The AI race isn't a 100-meter dash — it's a marathon. In this marathon, there are no permanent friends, no permanent enemies, only permanent compute bills and token consumption. The hundreds of millions in "tuition" Meta pays OpenAI ultimately returns as a better Llama — and a better Llama forces OpenAI to build a better GPT. This kind of "you in me, I in you" competition is how AI progresses fastest.

See you tomorrow.