🚀 The Great Migration: Why 85% of Fortune 500s Are Running Open AI Models

📊 The Shift Is Real—And It's Accelerating

A quiet revolution is underway in enterprise AI adoption. According to Jeffrey Morgan, co-founder and CEO of Ollama—the platform used by 9 million developers and 85% of the Fortune 500—businesses are rapidly pivoting from closed, proprietary models to open-source alternatives. The driver? It's not just cost. It's control, customization, and increasingly, geopolitical pragmatism.

While cost remains the most immediate pain point that open models solve, the real North Star for enterprises is gaining better control over AI and customizing models for their unique use cases. Morgan notes that cost is the gateway: solve it in the short term, and businesses can invest in fine-tuning models tailored to their specific needs.

"Cost is by far the largest pain point that open models can jump in and solve. But every business has a vision of getting better control over AI and customizing it for their business. That's really their North Star."

🌍 Chinese Models Dominating Cloud—But Not Locally

One of the most striking findings from Ollama's token flow data: cloud-hosted inference is now predominantly powered by Chinese-origin models, particularly from labs like DeepSeek, Qwen, and GLM. Meanwhile, US and European models remain competitive in the local deployment space, running neck-and-neck with their Chinese counterparts.

This bifurcation reveals an important dynamic. For coding agents and large-scale cloud tasks, Chinese models have become the go-to choice due to their combination of capability and cost efficiency. For privacy-focused, on-device workloads—like document processing or lighter assistant tasks—US models like Meta's Llama and Google DeepMind's Gemma continue to shine.

Geographic Demand: Interestingly, Germany has emerged as a major source of open model token consumption alongside the US, signaling broad international appetite for alternatives to closed frontier labs.

💼 AT&T and the Enterprise Playbook

A recent article in The Information highlighted that AT&T has already shifted 40% of its token consumption to open models. Right now, that's largely US and European models—but they're actively evaluating Chinese alternatives.

The primary use case? Coding agents. Morgan explains that the explosive growth in per-developer token usage has been driven by agentic workflows that write, test, and iterate code autonomously. Tools like OpenClaw (a Claude-style agent framework) and Hermes (an open-source agent orchestration project) have extended this capability beyond developers to finance, support, marketing, and sales teams.

The result: token consumption increased roughly 5x from early 2024 through April, driven first by coding agents and then by OpenClaw-style agentic workflows. In aggregate, Ollama Cloud has seen 150x growth since the start of the year, with individual developers moving from around 15 million tokens per week to north of 75 million.

⚡ The Flash Model Era: Cheap, Fast, and Good Enough

A new class of ultra-efficient models is reshaping the economics of AI. DeepSeek Flash leads this charge—offering frontier-adjacent performance at a fraction of the cost. These "flash models" are designed for high-volume, low-latency tasks and are becoming the workhorse for 80% of enterprise use cases.

Morgan believes we're entering an era of "unlimited tokens"—a return to the early ChatGPT experience where users didn't need to think about consumption limits. As flash models proliferate, token budgets will become less of a constraint, enabling widespread adoption across teams without the need for careful rationing.

"This new class of flash models—where they're good enough for 80% of the tasks, they're really fast, and they're ultra cheap—will enable widespread adoption within a team. You won't have to be thinking about how many requests you're making or how many tokens."

Moreover, these cheaper models can be chained together through orchestration, allowing smaller, specialized models to collaborate and solve problems that previously required a single, massive "god model."

🔐 Security, Safety, and the Geopolitical Elephant

One of the biggest blockers to open model adoption—especially for Chinese-origin models—is security and governance. However, Morgan reports that European and US enterprises are surprisingly pragmatic: if you can solve the safety problem, adopting Chinese models is completely on the table.

The key insight? Where the model is run matters more than where it's from. Customers care deeply about secure execution environments, data residency, and the ability to audit model behavior. Interestingly, Morgan draws a parallel to open-source software supply chain security—a problem enterprises have managed for decades across thousands of dependencies.

Still, model provenance matters for certain use cases. Models need to "speak" in a way that aligns with corporate culture and regulatory requirements. For mission-critical applications—like the Llama-powered analytics system monitoring a Finnish power plant to detect surges and keep the lights on—understanding the full data lineage of a model becomes essential.

🖥️ The Renaissance of Local Compute

While cloud models dominate today's headlines, Morgan predicts a pendulum swing back to local inference, particularly as next-generation hardware matures. NVIDIA's DGX Spark and Apple's M-series silicon are enabling developers to run models in the 20B to 120B parameter range—locally, with low latency, and at competitive price points.

The vision: a hybrid execution model where easier tasks run locally (for speed and cost) and harder, mission-critical tasks are routed to large cloud models. This mirrors the classic best-of-breed approach in cloud infrastructure, where customers prefer specialized tools over monolithic, bundled solutions.

Morgan sees this as the future of coding agents in particular. The GitHub Copilot experience—autocomplete in 100 milliseconds—could return to the desktop, powered by a GB300 on your desk. Ollama's journey, he notes, started local, moved to the cloud for coding agents, and will eventually come back local as hardware catches up.

📈 From Reddit Cranks to Fortune 500 in 18 Months

Ollama's growth story is remarkable—and instructive. The company spent two years searching for product-market fit, pivoting multiple times from Kubernetes security tooling to desktop developer security before finally landing on local LLM inference in July 2023, just as Meta's Llama 2 launched.

The initial user base? Hobbyists on Reddit. But within months, Ollama's frictionless developer experience—no permissions required, just download and run—translated seamlessly to Fortune 500 IT teams. The product went from zero to 100,000 GitHub stars faster than Docker or Kubernetes, and enterprise adoption followed almost immediately.

"What was really good for a hobbyist user translated very quickly to a developer within a business. Because LLMs are stateless, it made for such an easy transition."

Ollama didn't monetize until early 2024 with the launch of Ollama Cloud, waiting for the market to mature and for open models to reach true frontier-adjacent capability. The bet paid off: coding agents and agentic workflows finally reached the performance threshold where open models could compete head-to-head with closed labs.

🧩 The Curation Layer: Making Chaos Usable

In a world of abundant models, inference providers, and agent frameworks, curation becomes the scarce resource. Ollama acts as an operating system for the open model ecosystem—integrating tightly with hardware drivers (Apple Silicon, NVIDIA, AMD, Intel), inference engines, and application runtimes (Codeex, OpenClaw, Hermes).

Every model launch is a "fire drill," often coming together in the last 24 hours before release. The playbook: ensure the model is optimized for speed and accuracy, package it with the right harness or SDK, and coordinate with hardware and cloud providers to guarantee sufficient capacity on day zero.

This combinatorially complex problem—matching any harness to any model across any hardware—is what Ollama solves for developers who just want to build stuff, not debug obscure inference errors or navigate a minefield of undocumented API quirks.

🔮 What's Next: Orchestration, Memory, and the Unbundling of Frontier Labs

Morgan identifies three emerging problem spaces ripe for innovation:

• Knowledge: Connecting company data and context to models in a seamless, secure way
• Coordination: Orchestrating sub-agents across cloud and local environments
• Execution: Solving the compute and sandbox problem for cloud-based agents

As the open model ecosystem matures, these layers will likely unbundle—just as cloud infrastructure evolved from monolithic platforms (Heroku, Google App Engine) to best-of-breed components (databases, queues, storage). Developers will choose specialized tools for each problem rather than accept a bundled, walled-garden experience from a single frontier lab.

The question is whether AI coding agents themselves will accelerate this unbundling. If agents can maintain interoperability across memory systems, orchestration layers, and inference providers—constantly syncing and adapting—then lock-in becomes nearly impossible.

✨ The Big Picture

The open model revolution isn't just about cost. It's about control, customization, and competition. Chinese labs are pushing the frontier on efficiency and capability. US and European labs are doubling down on safety, governance, and developer experience. And enterprises are pragmatically adopting both, depending on the workload.

The steady state, according to Morgan: 80-90% of enterprise tokens will flow through open models, even if those models represent only 10-20% of the budget. The rest will go to frontier closed models for the hardest, most mission-critical tasks.

In other words: the future isn't "god model vs. swarm of small models." It's orchestration—best-of-breed components working together, with humans (and increasingly, agents) as the conductors.

The question is no longer whether open models will win. It's how fast—and who will own the layers in between.