OpenAI just closed $122 billion in funding at an $852 billion valuation. But the number that should keep every competitor up at night isn’t the valuation. It’s buried in the investor letter: their APIs now process more than 15 billion tokens per minute.

Let that sink in. That’s not a model benchmark. That’s production throughput at global scale.

The Infrastructure Flywheel Nobody Else Can Spin

Most people read OpenAI’s announcement and focused on the headline funding number. I read the investor letter and saw something far more important — a compounding infrastructure flywheel that’s almost impossible to replicate.

Here’s how it works. More compute trains more capable models. More capable models produce better products. Better products drive adoption. Adoption drives revenue. Revenue funds more compute. Each turn of the wheel makes every token more intelligent and cheaper to serve.

OpenAI now generates $2 billion per month in revenue. Enterprise makes up more than 40% of that and is on track for parity with consumer by end of 2026. That’s not a research lab burning cash. That’s an infrastructure company with operating leverage.

Multi-Cloud, Multi-Silicon, Multi-Everything

What caught my attention most was how aggressively OpenAI has diversified its infrastructure stack over the past 15 months. This isn’t a company tethered to a single cloud provider anymore.

Their infrastructure now spans Microsoft Azure, Oracle, AWS, CoreWeave, and Google Cloud. On the silicon side, they’re running NVIDIA, AMD, AWS Trainium, Cerebras, and their own custom chip designed with Broadcom. Data centre partnerships include Oracle, SBE, and SoftBank.

This is a deliberate strategy. No single architecture can efficiently serve the full range of AI workloads — from real-time inference for 900 million weekly ChatGPT users to massive training runs for frontier models. By diversifying, OpenAI reduces dependency on any single vendor while optimising cost and performance across different workload types.

For anyone building enterprise AI strategy, that’s the playbook to study. Not the model, not the chatbot — the infrastructure portfolio.

Why 15 Billion Tokens Per Minute Changes the Competitive Landscape

At 15 billion tokens per minute, OpenAI has crossed a threshold that fundamentally changes competitive dynamics.

Consider what that throughput means in practice. It means OpenAI can serve latency-sensitive enterprise workflows, agentic multi-step chains, real-time coding assistance through Codex, and consumer chatbot sessions — simultaneously, at global scale. The API pricing tells the story: GPT-5.4 at $2.50 per million input tokens with cached inputs at $0.25. GPT-5.4 nano at $0.20 per million input tokens.

When you can process 15 billion tokens per minute and your cost per token keeps falling, you’re not just offering a model. You’re offering an infrastructure layer that enterprise customers can build on with confidence.

Anthropic, Google, and every other frontier lab now face a structural disadvantage. It’s not about who has the best benchmark score on a given Tuesday. It’s about who has the infrastructure to serve intelligence reliably, cheaply, and at scale. OpenAI has planted a flag that’s going to be extraordinarily expensive to match.

The Superapp Play Is an Agent Operating System

OpenAI used the word “superapp” in the investor letter. I’d call it something more precise: an agent-first operating system.

They’re unifying ChatGPT, Codex, browsing, and agentic capabilities into a single product surface. The logic is clear — when models become capable enough, the bottleneck shifts from intelligence to usability. People don’t want five disconnected AI tools. They want one system that understands intent, takes action across applications, and maintains context across workflows.

Codex now has over 2 million weekly users, up 5x in three months, with usage growing over 70% month over month. That’s not a developer toy. That’s a fundamental shift in how software gets built.

For enterprise architects, this convergence creates both opportunity and risk. The opportunity is a powerful, unified platform for agentic workflows. The risk is deep platform dependency on a single vendor’s ecosystem. If your AI strategy doesn’t account for both sides of that equation, you’re not doing architecture — you’re just shopping.

What This Means for Everyone Else

The uncomfortable truth is that infrastructure scale compounds in ways that model innovation alone cannot overcome.

Google has massive compute but hasn’t translated it into comparable API throughput or enterprise adoption at this pace. Anthropic produces excellent models but doesn’t have the infrastructure scale or the consumer distribution flywheel. Meta is open-sourcing models, which builds ecosystem but doesn’t generate the revenue needed to fund infrastructure at this level.

The remaining moat for competitors isn’t model quality — it’s specialisation, data sovereignty, regulatory positioning, and the reality that many enterprises will never put all their eggs in one basket. Open-source alternatives, European sovereignty requirements, and multi-vendor procurement policies will keep the market fragmented.

But fragmentation isn’t the same as competition. OpenAI is building the AWS of intelligence. Everyone else is competing for the workloads that don’t go to the default choice.

The Architect’s Takeaway

When I evaluate infrastructure strategies, I look for compounding advantages — systems where each investment makes the next one more effective. OpenAI’s flywheel of compute, capability, adoption, and revenue is the clearest example of that pattern I’ve seen in AI.

If you’re building enterprise AI architecture today, the question isn’t whether OpenAI’s models are better than Anthropic’s or Google’s on any given benchmark. The question is whether anyone else can match the infrastructure throughput, cost structure, and product integration that makes 15 billion tokens per minute possible.

That’s not a model question. That’s an infrastructure question. And right now, OpenAI has the most convincing answer.

• OpenAI’s New Prompt Injection Defences Are the Most Important AI Security Work This Year
• Enterprise AI Agents Need Standards Before They Need Scale in 2026
• What NVIDIA NemoClaw Signals About the Future of Enterprise Agent Architecture
• OpenAI Hosted on Azure What Microsoft Really Means for Enterprises
• How NVIDIA Is Expanding Into the Enterprise Agent Stack