Open Models Change The Economics of AI

Open models are solving the cost pain point for AI, enabling businesses to gain control and customize AI for their specific needs.

Interest in fine-tuning custom AI models experienced a dip but is now resurging, with the rapid pace of model releases posing a challenge.

The biggest trend in AI is the shift towards open models, driven by coding agents and AI assistants, with models originating from both the US and China.

Chinese-origin open models are being accessed globally through Olama's cloud, with significant usage from the US and Germany.

Cost is a primary driver for enterprise adoption of open models, but the desire for better control and customization is the ultimate goal.

AT&T has transitioned 40% of its token consumption to open models, primarily for coding agent workflows, and is evaluating Chinese models.

Token usage per developer on Olama's cloud shows significant inflection points driven by coding agents and the rise of OpenClaw.

The explosion in open model usage is enabled by improved context windows, now exceeding one million tokens.

While custom-trained open models were prevalent, the serving of out-of-the-box open models has recently gained significant traction.

The increasing speed of open-source model releases makes post-training and customization more challenging, though tooling is improving.

AI safety concerns are becoming more prominent, which may lead to a slowdown in frontier model development while open-source models continue to improve.

The GLM-53 model's cybersecurity capabilities are impressive, presenting opportunities for security and governance startups.

Security and safety concerns are the main blockers for enterprise adoption of open models, but overcoming these can open doors to Chinese-origin models.

Open-weight models are being used to detect hacks, highlighting their utility in security testing, a task proprietary models may decline.

Out-of-the-box open models have some safety training but can better discern between good and malicious use cases.

Olama's role as a distribution channel allows model developers to coordinate launches, leading to significant growth spikes when new models are released.

The rapid release of new models necessitates a well-developed playbook for successful day-zero launches, involving support, performance optimization, and use case identification.

Packaging models with appropriate harnesses and ensuring cloud availability is critical for handling the immense demand upon release.

Open-source harnesses like Codex and OpenCode are crucial for developers to effectively utilize new models.

Hardware and provider collaboration, including optimization for platforms like Apple Silicon, is essential for delivering a good user experience.

The tight integration of drivers, hardware, and applications is key to creating a seamless experience for developers.

Olama's role is to provide a common runtime that matches any harness to any model, simplifying the developer workflow.

Olama's deep technical expertise spans from hardware optimization to complex software stacks, drawing from backgrounds in AI research and distributed systems.

The ultimate focus is on the developer experience, ensuring that API calls return tokens efficiently and reliably.

A challenge with open models has been the lack of coordinated development across the stack compared to frontier model labs.

The "five-layer cake" model of AI development is evolving, with more opportunities for innovation in the layers between models and applications.

The unbundling of services in the AI space, similar to cloud computing, is likely to lead to specialized companies for knowledge, coordination, and execution.

Open source principles are fostering best-of-breed solutions for different aspects of AI, contrasting with bundled offerings.

Frontier model labs aim for walled gardens, while open-source developers seek more flexibility and open ecosystems.

An abundance of open model tokens is being created, shifting the scarcity to orchestration and complex engineering problems.

As coding agents improve, the traditional moat of well-maintained software may diminish, leading to more accessible and interchangeable AI components.

There is a move towards open standards and protocols, reducing vendor lock-in and enabling interoperability.

Core model functionalities like memory might be integrated into models, while staple problems like storage will likely remain separate.

Open models lack the out-of-the-box safety tooling of closed-model providers, presenting a challenge for businesses.

The future enterprise landscape will likely see a balanced allocation of budget between frontier closed models and open-source solutions.

The majority of token consumption is expected to be from open models, though closed models will remain relevant for specific use cases.

Open models are driving down costs, enabling widespread adoption and the development of new orchestration layers.

The collaboration between open and closed models, facilitated by platforms like Open Router, is creating new possibilities.

The analogy of a law firm with partners and associates illustrates how different AI models can work together.

A blend of proprietary and open-source software, as seen in cloud computing, is a common and successful pattern.

Local models are becoming increasingly powerful, with advancements in hardware making it feasible to run larger parameter models on personal devices.

Benchmarks suggest that some local open models are now competitive with top-tier cloud models for specific tasks like coding.

A mix of US and Chinese-trained models are used for local execution, while cloud-hosted models for coding agents are predominantly Chinese.

The stark difference in model origin consumption between local (balanced) and cloud-hosted (Chinese dominant) models suggests a need for more US-based large model development.

NVIDIA's open-source approach and hardware support are crucial for the growth of the open-source AI ecosystem.

NVIDIA's DGX Station and DGX Spark offer powerful on-desk computing solutions for running large AI models at competitive price points.

Next-generation hardware like the DGX Spark, with its high unified memory, enables the running of large parameter models locally.

Both Apple Silicon and NVIDIA hardware are driving a renaissance in personal desktop computing for AI workloads.

Olama's journey started with local models and is now expanding to the cloud, with the expectation that powerful local AI capabilities will return to desktops.

The GPU market is characterized by rapid price changes and supply-demand volatility, posing challenges for startups.

For startups with limited budgets, adopting ultra-low-cost, high-efficiency models like DeepSeek Flash is recommended for maximizing token usage.

Open models have bridged the gap with frontier closed models in terms of intelligence and are now focusing on achieving extreme efficiency.

The DeepSeek Flash model is leading the charge in price-effectiveness and adoption, enabling widespread use of open models.

Orchestrating multiple smaller, efficient models can yield results comparable to larger models, offering a more repeatable and trustworthy approach.

The idea of a single "god model" for all AI tasks has not materialized; instead, a combination of specialized and simpler models, along with orchestration, is proving more effective.

For most use cases, models that are "good enough" combined with architectural improvements are sufficient, though the most powerful models will continue to unlock new possibilities.

Open models are now capable of handling difficult, though not necessarily the hardest, problems, enabling wider adoption.

The possibility of an open-source "god-tier" model exists, with developments in models like ZPoAI and GLM showing frontier capabilities.

The competition between open and closed models is intensifying, moving from a gap to head-to-head parity in certain areas.

Geopolitical considerations in AI often revolve around model origin and where the models are run, with some customers prioritizing US-trained models.

The origin of AI models matters for mission-critical tasks, as it influences the data used for training and the model's behavior.

Concerns about Chinese models, even if hosted in the US, being "booby-trapped" are raised, highlighting the importance of understanding model origins.

Robust IT and security teams in Fortune 500 companies are accustomed to managing dependencies and security risks, as is common with open-source software.

While models are less transparent than traditional software, proper screening and safety checks can mitigate risks.

Olama's origins in the Docker ecosystem and focus on developer experience have led to its strong brand and market position.

Olama's founders, with previous experience building Docker Desktop, spent years searching for the right problem to solve with their developer tools expertise.

Olama pivoted from an initial focus on Kubernetes security to a broader developer security platform before finding its niche with local LLMs.

The name "Olama" was chosen for its character and association with open models (LLM + animal mascot), not directly from a specific model.

Olama's Series A pitch emphasized building a great developer experience and solving security problems, leveraging the founders' previous relationship with investors.

The explosive growth of Olama, beginning in early 2026 according to cloud token data, reflects the rapid adoption of open models.

The initial years of searching for product-market fit were challenging, but the focus remained on solving real customer problems.

A significant pivot involved shifting from security for Kubernetes to developer security on the desktop, which then informed the approach to local LLMs.

The crucial pivot to locally hosted LLMs was driven by a realization of the difficulty in running open-source LLMs and the opportunity to create a seamless gateway.

A bias towards action and rapid iteration led to the launch of Olama within two weeks, quickly gaining significant user traction.

Olama gained initial traction through organic discovery on Reddit, where users praised its ease of use for running local models.

The transition from hobbyist users to 85% of the Fortune 500 in a short period mirrors the speed of PC adoption after the homebrew computer club era.

The ease of getting started with open models, coupled with their ability to run anywhere, was a key factor in their adoption by enterprise developers.

Olama's monetization journey began with a privacy-focused AI product and evolved as open models became capable of handling complex tasks like coding agents.

The monetization strategy for open models involves offering privacy-focused AI products and capturing value as these models mature and handle more demanding use cases.

Waiting for the market to mature allowed Olama to identify and address the most critical problems for customers using open models for challenging tasks.

Y Combinator's support was crucial for Olama, providing a peer network and lessons learned from previous generations of startups.

The AI world presents new challenges and opportunities that differ from traditional infrastructure, requiring new lessons learned.

The experience of working on teams that have shipped successful technology with clear product-market fit provides valuable muscle memory for building new ventures.

The AI landscape is breaking some traditional DevOps and infrastructure rules, necessitating new approaches to building and scaling services.

Olama provides a curated and seamless experience for end-users, abstracting away the complexities of underlying inference providers and their potential issues.

Curation of the fragmented open-source AI ecosystem, including models, inference technology, and services, is a valuable contribution to developers.

Platforms like Open Router and OpenCode simplify model selection, payment, and integration, empowering developers to focus on building applications.

In an abundance of open models, scarcity shifts to bringing together these fragmented components into a cohesive and functional solution.