Open Models Drive AI Adoption in Enterprises

AT&T has already shifted 40% of its token consumption to open models as businesses seek to customize AI for unique use cases beyond initial cost savings, Jeffrey Morgan stated.

AT&T has already transitioned 40% of its token consumption to open models, demonstrating a tangible shift in enterprise strategy.

Enterprises are actively evaluating models from both US and Chinese origins, indicating a global perspective on sourcing AI solutions.

The release cycles for open models have dramatically shortened from previous six-month intervals, making it challenging for teams to keep pace.

An example of this accelerated pace is the three iterations of the DeepSeek Flash model released over a single summer.

While custom training becomes more difficult with the closing gap between open and frontier models, improved tooling is now enabling development teams to manage these faster release cadences.

New models demand comprehensive support in inference engines, a process that can take several weeks to implement.

Key tasks involve meticulous verification of accuracy, speed, and adherence to reference specifications for each new model.

Developing new harnesses is also crucial to support specific advanced capabilities, such as multi-modality.

Ollama collaborates with developers ahead of general release to effectively manage the anticipated surges in usage.

Anthropic's platform team identified three critical layers in the AI stack: knowledge, coordination, and execution/sandboxing.

Following the pattern of cloud computing's evolution from monolithic platforms to specialized services, developers are now seeking best-of-breed products for each specific technical challenge in AI.

While frontier labs often aim to create proprietary, 'walled-garden' ecosystems, open-source developers prioritize flexibility and interoperability.

By adopting open-standard runtimes such as MCP, developers allow coding agents to synchronize memory across disparate systems and bypass the traditional vendor moats that previously hindered maintenance.

Open-standard runtimes, such as MCP, enable agents to synchronize memory across disparate systems, thereby eliminating dependence on specific model providers.

Stateful data storage remains a significant challenge and is not expected to be fully integrated into model layers.

Open models are projected to manage the supermajority, an estimated 80-90%, of business token usage.

This economic shift is largely due to open models significantly lowering costs, which in turn allows businesses to provide broader token access to their employees.

Frontier labs are expected to retain responsibility for the most complex and high-difficulty reasoning tasks, maintaining their niche in advanced AI.

Mid-tier tasks will likely see a hybrid approach, utilizing a mix of both open and closed models to optimize for cost and capability.

Local model adoption shows a balanced distribution among models of US, European, and Chinese origins.

Cloud-hosted usage data currently indicates a heavy skew towards Chinese models, dominating this segment.

There is a growing demand for more US-based large models to establish a more balanced ecosystem in cloud deployment.

Initial demand for coding agents migrated to the cloud due to substantial compute requirements.

Advancements in hardware are anticipated to enable sub-100ms autocomplete and testing loops directly on local machines.

Future workflows will likely integrate the speed benefits of local processing with the extensive intelligence and scale offered by cloud infrastructure.

DeepSeek Flash is highlighted as a prime example of a model offering ultra-low cost per token and per task.

The industry is aiming to return to an 'unlimited tokens' experience, reminiscent of early ChatGPT usage.

Achieving extreme efficiency is the next significant challenge, following the successful reduction of the intelligence gap in AI models.

Flash models are crucial for broad team adoption by eliminating per-request cost concerns, fostering wider usage.

Customers exhibit varied concerns, with some prioritizing the model's runtime environment, while others focus on its country of origin.

Security teams commonly assess AI models similarly to how they evaluate open-source software dependencies.

Open models are already being deployed for mission-critical applications, such as power plant analytics.

Rigorous screening and safety checks are crucial to mitigate risks associated with less transparent models.

The Ollama team initially applied to Y Combinator with a tool designed to be similar to Docker Desktop for Kubernetes.

Despite its YC participation in 2021, the product officially launched in July 2023, two years later.

The name Ollama was chosen to emphasize a focus on open models, extending beyond just the Llama family.

The founders spent two years struggling to identify a problem that resonated effectively with users, describing it as being 'lost in the wilderness'.

Maintaining a team of over 10 individuals during this search for product-market fit was a significant source of psychological stress.

The team attributes their persistence during this challenging period to a shared 'north star' of enhancing the overall developer experience.

The Ollama team decided to clear their existing roadmap, dedicating their efforts to building a functional gateway for AI models.

They committed to a rigorous two-week deadline to ship the initial version of Llama, emphasizing a rapid deployment strategy.

This pivotal period marked a significant shift from extensive overthinking to a bias for immediate, decisive action.

The swift launch generated a user base larger than all their previous product iterations combined, validating their expedited approach.

Ollama operated for two years without implementing a formal revenue model, prioritizing product development and user acquisition.

The company's monetization strategy is centered on developing a privacy-focused AI product, aligning with its core values and user needs.

The Y Combinator community fosters transparency among founders, allowing them to openly discuss challenges and share insights.

Leveraging prior experiences from ventures like Docker and VMware has provided valuable lessons and perspectives for Ollama's development.

The team emphasizes the importance of 'muscle memory' derived from successfully building technology that achieves genuine market fit.

The current AI ecosystem is highly fragmented, encompassing diverse models, inference technologies, and cloud services.

Ollama's mission is to harmonize these disparate tools, creating a cohesive and functional experience for end developers.

Through effective curation, Ollama transforms an overwhelming abundance of choices into a streamlined and usable developer workflow.