2026-07-12 · 6 min read · Raw Drafts / AI / Organizations
Open Models and the Rest of the Work
What Japan's one-month AI startups reveal about the difference between building a model, deploying a product, and changing an organization.
In February 2024, Yann LeCun shared a Nikkei article about Japanese startups launching generative AI services in as little as one month. He introduced it as an example of the power of open language models. The article described companies taking foundation models developed elsewhere, adapting them for Japanese users, and releasing new services within one to three months.
It captured one of the more persuasive ideas of the early generative AI period. A small company no longer had to train a general-purpose model from the beginning. It could start with the work of Meta or another model developer, add Japanese-language training, apply specialised data, and build a service around it.
The effect was real. Open models significantly reduced the cost of starting. Work that once required a large research laboratory could be attempted by a startup with a much smaller team. Countries and languages outside the immediate priorities of American AI companies gained another way to participate. Local developers could adapt a model to their language, industry, and regulatory environment without waiting for a foreign platform to do it for them.
Two years later, the Japanese companies featured in this period offer a useful view of what happened after development became faster.
From model development to social implementation
ELYZA, a company founded by researchers connected to the University of Tokyo, developed Japanese-language models using open foundation models. Soon after the Nikkei article appeared, KDDI and one of its subsidiaries acquired a combined 53.4 percent stake in the company. The partnership combined ELYZA's model-development capabilities with KDDI's computing infrastructure, corporate customer base, networks, and implementation resources. Their plans included general Japanese models, models adapted to specific industries, and AI software for companies and local governments.
The acquisition does not contradict the case for open models. In some ways, it confirms it. ELYZA had been able to build credible technology quickly enough to become strategically valuable to one of Japan's largest telecommunications companies.
It also shows where the work moved.
Developing the model was one part of the system. Putting it into a contact centre, financial company, or government office required access to customers, industry data, computing capacity, security procedures, implementation teams, and someone willing to take responsibility when the system failed.
KDDI's announcement repeatedly referred to social implementation, a common phrase in Japanese technology policy. The concern was that generative AI projects would remain at the proof-of-concept stage. The companies needed tools and support that could carry a model into everyday operations, where it would be used by employees and judged through actual outcomes.
ELYZA eventually released ELYZA Works in September 2025, describing it as its first product business. The service allows employees to create AI applications for their own teams. The interval is revealing. A model or technical demonstration could appear within a month. Building a packaged product for organisational use followed a longer timeline.
When generation becomes production
A similar shift can be seen at Fotographer AI, another company associated with the original Nikkei story. By the end of 2024, it had raised a cumulative ¥324 million. Its stated focus had moved toward consistency and control in image generation, particularly for industries where the output had to represent a real product accurately.
Generating an attractive image had already become widely accessible. The commercial problem was making sure the model respected the product's materials, proportions, lighting, and identity across many images. An advertising agency or furniture company needed more than a surprising result from a prompt. It needed a repeatable production process.
This is a recurring pattern in generative AI. As the foundation models improve, the first demonstration becomes easier. Expectations then rise. A capability that seemed remarkable becomes an assumed feature, while the unresolved problems around it become more visible.
Accuracy matters. So do permissions, evaluation, integration, latency, provenance, cost, and the ability to escalate a decision to a person. The interface has to fit the existing work. The organisation has to decide whose job is being changed and who remains responsible for the result. Someone has to maintain the system when a new model appears three months later.
The model becomes replaceable
The underlying model may also change. A company can begin with Llama, move to Qwen, or use several models for different tasks. This flexibility is one of the benefits of an open ecosystem, but it also limits the model itself as a source of lasting differentiation. When competitors can access similar capabilities, value accumulates elsewhere: in data, customer relationships, distribution, workflow knowledge, and the accumulated experience of operating the system.
The language around openness also deserves some care. Models such as Llama were often described as open source, including by their advocates, while their licences did not always meet the stricter definitions used by the open-source software community. Open model is a useful broader term because it captures the practical change that mattered to these startups: they could obtain model weights, adapt them, and run them without building the foundation from zero.
Japan's government appears to have reached a similar conclusion about where support was needed. Its GENIAC programme began in 2024 by supporting the computing resources required to develop foundation models. It later expanded its emphasis toward data, matching developers with user companies, and bringing models into practical use. By 2026, the programme was still funding model development while explicitly promoting social implementation.
This suggests that the next constraint could not be solved by releasing more models alone. Developers needed access to organisational data and concrete problems. User companies needed the ability to evaluate and absorb the technology. The government began supporting the connections between them.
The rest of the work
LeCun's original interpretation was broadly correct. Open models created space for Japanese companies to build faster. They enabled local experimentation and reduced dependence on a small number of closed APIs. The services described by Nikkei could genuinely be developed in a fraction of the time previously required.
The phrase “developed in a month”, however, described the beginning of the process more accurately than the end.
Software development once occupied a larger share of the journey from idea to product. Generative AI has compressed some of that work. It can produce a prototype, interface, or adapted model before an organisation has decided how the result will be governed or used.
This creates a strange imbalance. The technical artefact arrives quickly. The organisation around it continues to move through meetings, budgets, procurement, integration, training, and gradual changes in behaviour.
The model can be ready before the company is ready.
That may be one of the more important effects of open generative AI. It has made the gap between technical possibility and institutional capacity easier to see. Building has become faster, while deciding what should be built, fitting it into real work, and maintaining responsibility for it remain comparatively slow.
Open models changed the cost of beginning. The rest of the work remained.
AI authorship note
This essay began with Kris Haamer's question about a Nikkei article shared by Yann LeCun and whether its prediction about Japanese startups proved correct. AI supported research, follow-up analysis, drafting, editing, and implementation. Kris is responsible for the editorial direction, the decision to revisit the claim two years later, and the framing around technical speed and institutional capacity.