Y CombinatorOpen Models Change The Economics of AI
CHAPTERS
- 0:00 – 1:36
Why enterprises are shifting to open models (and what use cases drive it)
Jeffrey Morgan opens with the core trend he’s seeing: enterprise adoption of open models is accelerating. He attributes the shift to practical, high-volume applications—especially coding agents and AI assistant/co-worker workflows.
- •Enterprise momentum toward open models is rising
- •Adoption spans US- and China-origin model ecosystems
- •Coding agents are a primary driver of demand
- •AI assistants/co-worker agents broaden usage beyond developers
- 1:36 – 2:13
Who’s using open models on Ollama Cloud—and why Chinese models show up so strongly
The conversation moves to what Ollama’s token flow reveals about real-world model usage. Jeff notes that Ollama began as local execution but cloud usage is now dominated by Chinese-origin models accessed by companies globally.
- •Ollama started local; later added Ollama Cloud
- •Cloud token consumption is currently heavily Chinese-origin models
- •Usage is global, with notable demand in the US and Germany
- •Ollama’s position in the stack gives unique visibility into real usage trends
- 2:13 – 3:31
Cost vs. control: the enterprise reasons to adopt open models
Diana pushes on whether cost is the main driver. Jeff frames cost as the immediate pain point, but emphasizes a longer-term enterprise “North Star”: control, customization, and owning model behavior for business-specific needs.
- •Cost reduction is the fastest win for open models
- •Strategic goal is control and customization for business workflows
- •Lower cost enables experimentation and broader internal rollout
- •Open models become a pathway to tailored AI capabilities
- 3:31 – 5:32
Token usage explosion: coding agents, OpenClaw, and million-token contexts
Jeff shares a striking per-developer token consumption curve showing two inflection points. First came coding agents enabled by better open models; later, agentic ‘co-work’ tools (like OpenClaw/Hermes) and larger context windows drove explosive growth.
- •Per-developer tokens rose dramatically (not just aggregate growth)
- •Inflection #1: coding agents as open models improved
- •Inflection #2: OpenClaw/Hermes-style agents increased tool-using workloads
- •Context windows jumped from ~128k to 1M+, increasing token burn
- 5:32 – 7:27
Fine-tuning’s comeback: faster open-model release cadence vs. better tooling
The group debates whether fine-tuning is a hype cycle. Jeff argues the open-model release cadence is speeding up—making post-training harder to maintain—while improved tooling simultaneously makes fine-tuning more feasible for teams determined to keep up.
- •Open model iterations are arriving faster (e.g., repeated DeepSeek Flash updates)
- •Rapid releases make long-lived fine-tuned variants harder to maintain
- •Tooling improvements reduce friction for fine-tuning and post-training
- •Frontier-lab safety slowdowns could widen opportunities for open models
- 7:27 – 8:26
Where open models beat closed models: security testing, governance, and ‘refusal’ limits
They explore where open models can be a step-change better than closed frontier models. Security testing is highlighted: closed models often refuse offensive security tasks, while open models (including specialized variants) can support defensive testing and governance workflows—if safety screening is handled.
- •Security testing is a major use case where open can outperform closed
- •Closed models may refuse pen-testing; open models can be more permissive
- •Specialized ‘security researcher’ models exist alongside safer general models
- •Enterprise blockers center on safety/security assurance for adoption
- 8:26 – 11:31
Launching open models at scale: the day-zero playbook (engines, harnesses, benchmarks, capacity)
Jeff describes what it takes to support a new model release on day zero. Success requires coordinating inference engines, model-specific harnesses/tool-calling mechanics, benchmarking against references, and ensuring cloud capacity and hardware readiness—often under extreme time pressure.
- •Day-zero releases produce the largest traffic spikes
- •Inference engine support and performance tuning can take weeks—until it doesn’t
- •Harness readiness (SDKs/tool calling/multimodal changes) is critical
- •Final pre-release benchmarking ensures correctness vs. lab references
- •Capacity planning and provider coordination are constant constraints
- 11:31 – 13:58
Ollama as an OS layer: drivers, runtime glue, and developer experience
Harj frames Ollama as an operating-system-like layer connecting models, hardware, and applications. Jeff expands on the combinatorial complexity of aligning drivers, inference stacks, harnesses, and runtimes into a consistent developer experience.
- •OS analogy: integrate hardware ‘drivers,’ inference, and application runtimes
- •Goal is a common runtime that can pair any harness with any model
- •Team’s background (Docker/VMware) maps to inference-era systems problems
- •Open models need a ‘five-layer cake’ experience comparable to frontier labs
- 13:58 – 17:40
The hidden layers between model and app: knowledge, coordination, execution (and unbundling opportunities)
The discussion identifies ‘missing’ platform layers that become valuable as tokens commoditize. Jeff lists knowledge/context integration, coordination across sub-agents, and execution/sandboxing as major emerging categories likely to unbundle into separate best-of-breed companies.
- •As tokens become abundant, the scarcity moves ‘above the tokens’
- •Key platform layers: knowledge (data/context), coordination (multi-agent), execution (sandboxes/compute)
- •Stateful components (memory/storage/credentials) can’t be fully pushed into models
- •Open ecosystems tend to unbundle into best-of-breed providers over time
- 17:40 – 20:42
Open vs. closed steady state: most tokens open, hardest tasks closed, routed together
Diana asks about the eventual budget and usage split. Jeff predicts most enterprise tokens will run on open models (80–90%), even if spend is lower due to cost collapse; frontier models remain for the hardest tasks, with routers/orchestration mixing both.
- •Prediction: supermajority of enterprise tokens go to open models
- •Budget share may remain smaller because open models drive costs down
- •Frontier models stay important for the hardest problems
- •Routing/orchestration layers will blend open and closed effectively
- •Analogy: partner/associate structure in human organizations
- 20:42 – 22:57
Local vs. cloud execution: hybrid routing, model sizes on desktops, and what runs where
They shift to local models, noting modern hardware can run 20B–40B (sometimes 120B) parameter models effectively. Jeff describes a hybrid pattern: easier document/workflow tasks run locally for latency and cost, while difficult coding-agent work tends to stay in the cloud, mediated by a router.
- •Local+cloud is the likely enterprise pattern, not either/or
- •Modern hardware enables strong local inference at meaningful model sizes
- •Coding agents often need large cloud models; simpler workflows run locally well
- •Routers decide when to escalate from local to cloud based on task difficulty
- •Local usage mixes US/Europe/China models more evenly than cloud
- 22:57 – 25:12
Why Chinese models dominate cloud tokens—and NVIDIA’s open-source push to rebalance
The hosts react to graphs showing cloud consumption skewing Chinese, while local is more balanced. Jeff points to early signs of US-side resurgence via NVIDIA’s Nemotron, and Harj highlights NVIDIA’s incentives to open-source models/software to expand hardware demand.
- •Cloud-hosted open-model usage is currently dominated by Chinese-origin families
- •Local usage appears closer to parity across regions
- •Nemotron Ultra is discussed as an early wave of US/Western competitive large models
- •NVIDIA benefits by growing an open ecosystem that drives GPU demand
- 25:12 – 27:15
The return to local: DGX Spark/Station, Apple MLX, and desktop-scale ‘mini datacenters’
They dig into next-generation desk-side compute, including DGX Spark and Apple Silicon advances. Jeff argues that as hardware catches up, low-latency local coding loops will return, and organizations will combine chained local devices with cloud resources.
- •DGX Spark provides large unified memory on a desk; DGX Station aims for high performance without extreme noise
- •Devices can be chained via fast interconnects to run larger models
- •Apple’s MLX stack is described as mature and highly competitive
- •Trajectory: started local → moved cloud for coding agents → returns local as hardware improves
- 27:15 – 28:43
GPU scarcity and running a cloud service: pooling providers, placement, latency, and forecasting
Operating Ollama Cloud requires navigating volatile GPU supply and pricing. Jeff explains they mitigate constraints by partnering with multiple providers and solving scheduling problems—model placement, region, latency—so developers avoid negotiating GPU procurement themselves.
- •B200/B300 access is difficult for startups; market volatility is high
- •Ollama uses multiple provider partnerships to pool capacity
- •Hard problems: routing, placement, regional latency, performance targets
- •Products like Ollama/OpenRouter simplify access for developers without GPU deals
- 28:43 – 33:37
Flash models and orchestration: ultra-cheap workhorses vs. the ‘God model’ myth
Jeff describes a new efficiency wave—Flash-style models that are cheap per token and per task—making ‘unlimited tokens’ plausible. They argue orchestration of many smaller models will often beat reliance on a single massive model for everyday enterprise tasks.
- •Flash models drive ‘low cost per task,’ not just low cost per token
- •DeepSeek Flash is cited as a major growth driver on Ollama Cloud
- •Cheaper models enable chaining/orchestration to solve complex workflows
- •Most use cases hit ‘good enough’ intelligence; frontier models unlock edge cases
- •Open models are moving from ‘gap’ to head-to-head competition in some domains
- 33:37 – 36:14
Geopolitics and trust: model origin vs. where it runs, supply-chain analogies, and safety screening
They address concerns about using China-origin models for critical work. Jeff distinguishes between concerns about model origin and concerns about secure deployment, and compares model risk to long-standing open-source dependency risk—arguing robust enterprise security teams can mitigate it with screening.
- •Customers split: some care most about where the model runs; others about where it was trained
- •Concerns include ‘booby-trapped’ models and opaque behavior
- •No widely known ‘Manchurian Candidate’ cases cited in the discussion
- •Enterprises already manage open-source supply-chain risk at scale
- •Safety/security screening is positioned as the practical mitigation path
- 36:14 – 47:02
Ollama’s founding story: from ‘wrong idea’ YC application to the local-LLM pivot and explosive adoption
Jeff recounts the multi-year search for the right problem, including early Kubernetes/security directions and a period of uncertainty. The decisive pivot came from introspecting the team’s strengths and shipping quickly around the pain of running Llama locally—leading to rapid product-market fit and massive open-source adoption.
- •Applied to YC with a Kubernetes/Docker Desktop-adjacent concept
- •Spent ~2 years ‘lost in the wilderness’ trying multiple directions
- •Key pivot: make open models easy to run; bias to action (two-week launch)
- •Ollama branding positioned open models as approachable (mascot/character)
- •Adoption jumped from hobbyists to Fortune 500 due to low friction and permissionless use
- 47:02 – 49:43
Monetizing open source: why the cloud offering arrived later, and what risks come with waiting
They unpack the gap between viral adoption and revenue, similar to Docker’s history. Jeff explains they waited for open models to be ‘good enough’ for high-consumption workflows (coding agents), while warning that ignoring monetization can harm culture; he emphasizes staying close to real customers rather than treating users as an internet blob.
- •Ollama grew to huge GitHub adoption before meaningful monetization
- •Cloud monetization aligned with open models powering major workloads (coding agents)
- •Risks of waiting: culture that deprioritizes monetization and customer intimacy
- •Fortune 500 adoption was enabled by permissionless, local-first trialability
- •Customer conversations are essential to define enterprise-ready offerings
- 49:43 – 57:15
Why do YC as a second-time founder—and which old DevOps rules break in the AI era
Jeff explains choosing YC for community, peer support, and learning without repeating fatal mistakes. The closing discussion contrasts classic infrastructure-era assumptions (e.g., ‘layer on top is doomed’) with AI realities, and notes new operating principles when systems are nondeterministic and AI changes staffing and support needs.
- •YC helped reduce founder isolation and provided high-signal peer learning
- •Seeing ‘what good looks like’ improves execution standards and judgment
- •Old rule ‘being a layer is vulnerable’ is less true in AI; being closer to customers can be stronger
- •LLMs are nondeterministic by nature—systems practices must adapt
- •AI changes staffing needs and support operations; new playbooks are emerging
