Open Models Change The Economics of AI

Why does it matter? Because the economics of AI are inverting — and most companies are optimizing for the wrong layer.

AT&T has already moved 40% of its token consumption to open models. Across Ollama's cloud, those tokens flow almost exclusively through Chinese-origin weights. The story isn't really about cost — it's about what becomes possible when AI tokens get cheap enough that you stop rationing them entirely.

• 80–90% of enterprise AI tokens will run on open models at 10–20% of the cost, eliminating the rationing that's currently blocking entire categories of use cases
• Per-developer token usage jumped 10–20X in a single year — not from new users, but from each individual going dramatically deeper
• A consumer MacBook now runs a model that matches Claude Opus on coding benchmarks
• Chinese-origin models already power virtually 100% of enterprise cloud AI traffic through Ollama

Open models won't just cut your AI bill — they'll let you build things closed-model pricing made impossible

"The super majority of tokens — call it 80, 90% — will be open models within a business." Jeffrey Morgan, CEO of Ollama, is explicit on the asymmetry: most tokens, a small fraction of the budget. Because the open model community keeps compressing costs, enterprises end up with token abundance they don't currently have — and abundance changes what you build, not just what you spend.

The conventional frame is cost-cutting. Morgan inverts it: cheap tokens eliminate rationing, and rationing is what's currently blocking agentic workflows, long-horizon automation, and high-frequency document processing. Cost gets you in the door. Control and customization — fine-tuning models for specific workflows — is where enterprises actually want to land.

AT&T has already made the shift. Per a recent Information article Morgan cited, 40% of their token consumption has moved to open models, predominantly US and European for now, with Chinese models under active evaluation. The workload driving almost all of it: coding agents.

Per-developer token usage jumped 10–20X in a year — proof that agents crossed from experimentation into workflow dependence

The graph Morgan showed looks like aggregate platform growth. It isn't. It tracks one developer over a year, measured in tokens consumed per week. The shape is the same hockey stick.

Two inflection points drove it. Early 2026: coding agents finally cleared the open model capability bar. Kimi, GLM, Minimax launched. Per-user consumption spiked. Then April: OpenClaw went mainstream. Anyone — not just engineers — could "take a hard problem, give it to an open model and let it go completely free." Context windows stretched from 128K to a million-plus tokens. The April jump alone was 10 to 20X per user. Ollama's full-year cloud growth: 150X.

When evaluating whether AI has genuinely embedded in a workflow, total user counts are the wrong signal. Tokens per active user is what matters. Flat user counts with exploding per-person consumption is what genuine workflow integration looks like. That crossing happened in early 2026 — and the token volume data makes it unambiguous.

Chinese models already power virtually 100% of enterprise cloud AI traffic through Ollama — geopolitics is the constraint now, not capability

For cloud-hosted enterprise tokens — the coding agents, the Fortune 500 deployments, the AT&T-scale automations — it's essentially 100% Chinese models. Two graphs show the split clearly. For local models running on developer hardware, US and Chinese are roughly tied. But for cloud workloads, Morgan is unsparing: "For cloud hosted models, it's like 100% Chinese models. Basically we need more US labs to make large models."

This isn't an ideological position — it's a capability observation. DeepSeek Flash drives most of Ollama's high-growth cloud consumption. The performance case has been made and enterprises are acting on it. AT&T is actively evaluating Chinese models. Others are already running them in production.

NVIDIA's Nemotron Ultra is the first meaningful US response at scale. But today, any enterprise running serious open-model cloud workloads is almost certainly running through Chinese-trained weights. A regulatory shift requiring migration away from those models wouldn't be a graceful swap — it would be a capability regression. The time to audit token flows is before that pressure lands.

The Manchurian candidate risk of Chinese models is real — but enterprise security already knows exactly how to manage it

The question Morgan hears most often is some version of: what if the model is rigged? It's not paranoid. Llama models are running the analytics stack at a Finnish power plant, monitoring for electrical surges that could knock out the grid. Model integrity genuinely matters.

His answer cuts through the alarm: "Supply chain poisoning is insane. It's a thing. It's been a thing for decades." The average enterprise application runs on thousands of open source dependencies — one poisoned package can compromise the entire stack. Chinese model risk is that same problem with one additional layer of opacity. You can't grep the weights for malicious code.

But you can screen them. "If you screen the model properly with safety checks, by and large, at least what we're hearing from customers is that can be solved." European enterprises and US Fortune 500 security teams are already treating Chinese model evaluation the same way they've treated third-party library vetting for years — systematically, not as a blanket block. The absence of robust tooling to do this at scale is a startup opportunity. Not a reason to avoid open models wholesale.

When tokens cost almost nothing, the moat shifts from inference efficiency to orchestration quality

DeepSeek Flash introduced something genuinely new: not just low cost per token, but low cost per task. Cheap-per-token models can still be expensive if they burn tokens getting useful work done. Flash models compress both metrics simultaneously.

Morgan reaches for the original ChatGPT mental model: "You didn't really have to think about how many tokens you were using. You would just use it every day, and you had unlimited. Ultimately, I think we return to that." Flash models get you back there — not through subscription bundles, but through prices low enough that rationing the calls stops making economic sense.

The design shift that follows is significant. Stop optimizing for fewer API calls; start optimizing for task decomposition. "By having these cheaper models... you can start to chain them together and build new problems that are solved by orchestration on top." Coordination between model tiers, state management across long agent runs, routing decisions at inference time — that's the competitive surface where the next generation of infrastructure startups will win. Not inference cost.

A 38B-parameter model on a consumer MacBook now matches Claude Opus on coding — the frontier API default needs revisiting

Qwen 3.8B at 38 billion parameters benchmarks equal to Claude Opus for coding. Morgan confirms it live. Hardware required: "not the lowest memory MacBook, but the second lowest memory MacBook you can buy from the store."

Coding agents are what drove the per-user token explosion. The model class best suited to power them now runs locally, offline, with no data leaving your infrastructure, on hardware your team probably already owns. The open-closed capability gap Morgan estimates at "less than three months" — down from a distance that made frontier APIs the unambiguous default 18 months ago.

For any architecture currently defaulting to frontier API calls for coding or document processing work: the benchmark data makes a genuine conversation unavoidable. Matching performance, fraction of the cost, zero data egress risk. The default assumption has an expiration date.

Product-market fit arrived in two weeks — after two years of paralysis

Jeffrey Morgan and his co-founder sat in a Toronto room and threw out two years of work. The question: what would you build if you were joining YC today? Two problems surfaced — a gateway to LLMs and a way to make open models actually run on real hardware. They gave themselves two weeks. Llama 2 dropped at the end of those two weeks. They shipped.

GitHub stars hit 100,000 faster than Docker, faster than Kubernetes. Roughly 18 months later: 85% of the Fortune 500. Morgan is unsparing on the years before: "two years of just frankly overthinking the customer, the product, and just not getting something out there."

The pattern is replicable. If you've been building for over a year without breakout signal, the question probably isn't what to optimize next. It might be whether to wipe the board entirely and ask what you'd build today with the team you actually have.

The model stopped being the moat

Morgan names the emerging unbundled layers explicitly — knowledge, coordination, execution, security — and says the companies that will own them haven't been built yet. That's the real signal in the Ollama story: open models didn't just change pricing, they changed what the scarce resource is. The same unbundling force that broke monolithic cloud platforms is now running one layer up, faster than anyone expected.

The model is cheap. The stack above it is wide open.

---

Topics: open source AI, enterprise AI, LLM economics, Ollama, DeepSeek, Chinese AI models, coding agents, local AI, AI security, startup founding, product market fit, token economics, open weights, AI infrastructure