Context

Jeffrey Morgan is co-founder and CEO of Ollama, a Y Combinator company used by 9 million developers and 85% of the Fortune 500 to run open AI models both locally and in the cloud. Because Ollama sits directly in the token flow, Morgan has unusually direct data on which models businesses actually use and how that's shifting. He talks with YC's Lightcone hosts about the sudden explosion in open-model adoption, the "hidden layers" of infrastructure being built on top of raw model access, and Ollama's own multi-year search for the right problem before "Ollama" existed as a product. For PMs building anything on top of AI models, this episode is a grounded look at where real enterprise adoption is happening right now, not just in benchmark announcements.

The Big Idea

Open models crossed from hobbyist curiosity to enterprise default in under two years, largely because coding agents and always-on assistants created demand for token volume that made cost the dominant constraint, and now that open-model intelligence has nearly caught up to frontier labs, the real competitive scarcity has shifted to the orchestration, security, and curation layers built on top of raw model access.

Morgan says AT&T alone has shifted 40% of its token consumption to open models, predominantly through coding agents, and Ollama's own cloud token usage grew 150x since the start of the year. But he's explicit that cost is only the entry point: every business's real "North Star" is gaining control and the ability to customize models for its specific use case, something only open weights make possible.

Key Insights

1. Cost is the entry point, but customization is the real reason enterprises adopt open models

Morgan says cost is "by far the largest pain point" open models solve, but frames it as a short-term unlock rather than the end goal. Once a business has solved cost, open weights let it actually customize a model for its unique use case in a way a closed API can't, which he describes as every enterprise's real North Star. AT&T, cited in a recent industry report, has already shifted 40% of its token consumption to open models, predominantly for coding-agent workflows, while still evaluating Chinese-origin models for further adoption.

2. Two separate demand explosions, months apart, explain the shift

• What: Ollama's cloud token usage per developer shows two distinct growth spikes in the same year: an initial jump when open models like Kimi, GLM, and Minimax first became capable enough to power coding agents, and a much larger explosion in April driven by the OpenClaw and Hermes agent projects.
• Why it matters: the second wave mattered more because it extended AI delegation beyond developers to non-technical roles like finance, support, marketing, and sales, and it coincided with open-model context windows jumping from 128K tokens to over a million, letting agents work through much longer, more autonomous tasks.
• Example: aggregate token usage through Ollama's cloud grew roughly 150x since the start of the year, a scale of growth Morgan compares to the entire PC industry's decade-long path from hobbyist "homebrew computer club" adoption to mainstream business use, compressed into about 18 months for open models.

3. Security and safety, not capability, are now the main blocker to Chinese-origin model adoption

Morgan says that once security concerns are addressed, adopting Chinese-origin open-weight models is "completely on the table" for the U.S. and European businesses Ollama talks to, and that capability is not the limiting factor. He points to models like GLM-5.3 showing genuinely impressive cybersecurity capabilities, which creates real opportunity for startups building security and governance tooling around open models, since even Hugging Face reportedly had to use open-weight models to detect a hack against a frontier lab's own infrastructure. Open models are also more willing to perform security-testing tasks, like penetration testing your own product, that closed models like Claude will typically refuse outright.

4. Launching a new open model well is an "operating system" problem, not a model-quality problem

Ollama has developed a repeatable playbook for what Morgan calls a "day zero" model launch: getting the model supported and accurate in the right inference engines, making sure developer harnesses (like the open-source OpenCode or Codex harnesses) are updated to support the model's new capabilities, and running benchmarks against the final packaged product to confirm it matches what the lab specified. Much of this work happens in a "fire drill" in the last 24 hours before release, because labs often only give a few weeks of advance access. Morgan compares Ollama's role to a classic operating system: gluing together drivers (hardware and inference providers), the application runtime (harnesses), and the model itself into one working stack, a combinatorially hard problem that frontier labs solve internally but that open models need a separate layer to replicate.

5. The stack is unbundling, and the scarce resource has moved above the token layer

Morgan describes an emerging pattern similar to early cloud computing: bundled all-in-one platforms (the Herokus and Google App Engines of the 2010s) eventually lost out to best-of-breed point solutions for each layer of the stack. He expects the same to happen in AI, citing an Anthropic platform team talk that named three unbundled problem areas: knowledge (connecting a company's data and context to a model), coordination (managing subagents that spin off across cloud and local environments), and execution (the sandboxes and compute needed to actually run agent work). With open-model tokens now abundant and cheap, Morgan argues the new scarcity is in these orchestration layers above the tokens, not in raw model access itself, which is exactly where he expects new startups to form.

6. Enterprise AI spend is settling into a hybrid, not a winner-take-all split

Morgan predicts that within a business, the "super majority" of tokens, 80 to 90%, will eventually run through open models, but that this won't translate proportionally to budget, since open-model pricing may only claim 10 to 20% of total AI spend even as it carries the bulk of usage. He expects frontier closed models to remain reserved for the hardest tasks and for coordination or routing decisions, while the high-volume, repetitive "line item" work increasingly runs on cheap open models, similar to how enterprises historically blended proprietary cloud services (like AWS's DynamoDB) with open alternatives (like PostgreSQL) rather than choosing one exclusively.

Mental Models & Frameworks

The "five-layer cake" of AI infrastructure, and its hidden sixth layer

Morgan borrows a framing from NVIDIA's Jensen Huang describing AI infrastructure as five layers: applications, models, infrastructure and inference, chips, and energy. He argues frontier labs have all five layers pre-integrated and ready for developers on day one, which is exactly what Ollama is trying to reproduce for the open-model ecosystem without owning every layer itself. Crucially, he notes the "model" layer actually contains much more than raw weights: developer APIs, orchestration components, and harness-specific tooling are all hidden inside what looks like a single layer, and he expects several of these hidden sub-layers to eventually unbundle into their own standalone companies, the same way infrastructure layers have historically unbundled in cloud computing.

Two-week bias to action versus two years of overthinking

Ollama spent roughly two years as a company (originally focused on Kubernetes security and developer tooling, unrelated to what Ollama eventually became) failing to find real product-market fit despite deep technical talent, largely because the team kept refining an idea rather than shipping something and testing it against real usage. When they finally gave themselves a two-week deadline to ship a minimal local-LLM tool, right as Llama 2 launched, that product found more users in two weeks than their prior two years of work combined. The team's own retrospective lesson: a tight, forced deadline that produces a real, shippable artifact can surface product-market fit far faster than continued internal deliberation about the "right" idea, even for an experienced, technically strong founding team.

Trade-offs & Nuance

Local models versus cloud models is a hybrid, not a binary choice

Morgan describes local and cloud model usage as complementary rather than competing: easier, well-defined tasks like document processing run effectively and cheaply on local hardware (a mid-tier MacBook can now run 20 to 40 billion parameter models competitively with frontier cloud models for some tasks), while genuinely hard tasks like complex coding agent work still perform best on large cloud-hosted models. He expects this hybrid to deepen as new desktop hardware (Apple Silicon, NVIDIA's DGX Spark and Station) makes locally running larger models increasingly practical, potentially bringing today's cloud-first coding-agent workloads back toward local execution once local hardware catches up to the latency and speed developers want.

Security researchers face a real dependency-risk trade-off with open models, but it's not fundamentally new

Morgan acknowledges the theoretical risk that a Chinese-origin open model, even hosted domestically, could be subtly compromised (what the hosts call "the Manchurian candidate problem"), but points out no known cases of this have surfaced, and argues the underlying risk is structurally similar to the decades-old software supply-chain risk of any application depending on thousands of open-source dependencies. He notes this is "a little more opaque because you can't dig into the model," but that the same disciplined security screening businesses already apply to open-source software dependencies can address it, based on conversations with Fortune 500 IT and security teams already managing this.

Common Mistakes

Mistake: Treating an open-source user base as an anonymous blob instead of real customers

Morgan says one of Ollama's biggest regrets from the two years between launching its open-source tool and beginning to monetize was not staying closely enough in touch with its rapidly growing user base. He warns that treating a fast-growing open-source community as an undifferentiated mass, rather than actively learning what specific users need now and what they'll need in six months, is a real risk of open-source-led growth, and says the team is now investing heavily in exactly this kind of direct customer contact.

Mistake: Assuming infrastructure lessons from the pre-AI era still hold

Morgan describes several "muscles" the Ollama team had to actively unlearn from prior infrastructure careers (VMware, Docker). One is the old Platform-as-a-Service era belief that building a layer on top of someone else's infrastructure is a structurally vulnerable startup position; in AI, he argues going further up the stack can actually be an advantage because it puts you closer to the customer. Another is the assumption that software should behave deterministically and exactly as designed: LLMs are inherently non-deterministic, which is a feature of the technology, not a defect to engineer away, and teams applying old infrastructure rigor uncritically will fight the wrong battles.

Practical Application

Ship a minimal version on a hard deadline instead of continuing to refine the idea

If a team has spent an extended period (weeks, months, or longer) iterating on a product idea without clear signs of real user pull, consider setting a short, forced deadline (Ollama used two weeks) to ship the simplest possible version and put it in front of actual users. Morgan's account suggests the forcing function of a real deadline and real usage data can surface product-market fit far faster than continued internal debate about which idea is "right."

Route AI workloads by task difficulty, not by defaulting to the most powerful model

When architecting an AI-powered product, build in routing logic that sends easy, high-volume, well-defined tasks to cheap, fast open or small models and reserves expensive frontier models for genuinely hard problems or coordination decisions. Morgan's data shows this hybrid approach is already the emerging enterprise default, and the newest class of ultra-low-cost, low-latency open models (like DeepSeek Flash) is specifically built to serve as the high-volume "workhorse" layer under this kind of routing strategy.

Talk to your fastest-growing open-source or free-tier users directly, even before you monetize them

If your product is growing quickly through an open-source or free channel without an obvious monetization path yet, deliberately invest time in direct contact with that user base (calls, direct feedback loops) rather than treating rapid growth alone as sufficient signal. Ollama's retrospective view is that this direct contact, not growth metrics, is what would have told them earliest when the market had matured enough (with coding agents in 2026) to introduce a viable paid product.

Questions to Consider

• For our own AI-powered product, which tasks are genuinely hard enough to require a frontier closed model, and which are we routing to an expensive model out of habit rather than necessity?
• If we have a fast-growing free or open-source user base, are we treating that growth itself as sufficient signal, or are we doing the harder work of direct contact to understand what those users will need six months from now?
• Are there infrastructure or process assumptions carried over from a pre-AI era of our business that we haven't actually re-examined for whether they still apply, the way Ollama had to unlearn Platform-as-a-Service and determinism assumptions from their prior infrastructure careers?
• If a current initiative has been in a multi-year search for the right problem without clear traction, what would a two-week forced-deadline version of it look like, and what would we learn by shipping that instead of continuing to refine the idea?

Bottom Line

Open models went from a hobbyist curiosity to powering 40% of a major enterprise's token consumption in under two years because coding agents and autonomous assistants created enough token-volume demand that cost, and then customization, became the deciding factor over raw capability. As open-model intelligence closes in on the frontier, the real competitive opportunity has moved to the layers built on top of tokens: orchestration, security, curation, and routing, exactly the unbundled problems a new wave of startups is now forming to solve.

Concepts to Explore

Bundling and unbundling in infrastructure markets

Morgan repeatedly invokes the classic cloud-computing pattern where all-in-one bundled platforms (Heroku, Google App Engine) eventually lost share to specialized, best-of-breed point solutions for each layer of the stack. This concept matters for anyone evaluating AI infrastructure strategy: frontier model labs are currently trying to keep developers inside an all-in-one bundled ecosystem (their own agents, memory, and context management), while the open-model ecosystem is fragmenting into specialized providers for orchestration, security, and execution, a tension the episode frames as still actively unresolved.

Notable Quotes

"Cost is something they can solve in the short term, but it then enables them to then go and customize these models for their unique use case." (Jeffrey Morgan)

"There's just new lessons we're learning going from infrastructure 1.0 in the 2000s to cloud in the 2010s to now the AI space. There are a lot of rules that break." (Jeffrey Morgan)

"Now there's a scarcity in bringing that together into something that works." (Jeffrey Morgan)