For years, the most powerful AI systems in the world were available only through closed APIs. You paid per token, received an output, and had no visibility into — or control over — what was running underneath. That era is now ending. Open models, where the actual weights are publicly released, are rapidly closing the capability gap with the best proprietary systems. For anyone building AI infrastructure at scale, this shift changes the economics, the control dynamics, and the strategic calculations in fundamental ways. The Closed Model Playbook The business logic behind closed AI was straightforward and, for a time, extremely effective. A company would train a frontier model that no one else had. They would keep the weights private. Then they would charge customers for access through an API — typically priced per token or per request. The model itself became the moat. Because competitors could not easily replicate the weights, they also could not easily replicate the performance. This model powered much of the early commercial success of companies like OpenAI, Anthropic, and Google’s AI efforts. It aligned incentives cleanly: the better the model, the more customers were willing to pay for access. It also gave providers significant control over pricing, usage terms, and feature availability. However, this approach created structural dependencies for users. Your product’s intelligence was rented, not owned. Your costs scaled linearly with usage. Your uptime and capabilities were subject to the provider’s decisions. And if the provider changed pricing, modified terms, or experienced an outage, your system was directly affected. The Shift: Llama, DeepSeek, and the Erosion of the Moat The turning point came when Meta released the Llama family of models with publicly available weights. For the first time, a frontier-grade model could be downloaded, run on private infrastructure, fine-tuned for specific use cases, and optimised without any ongoing per-token fees or dependency on a third-party provider. Then came models like DeepSeek, which demonstrated that open-weight models could match or even exceed closed models on key benchmarks — while being developed at a fraction of the cost previously assumed necessary for frontier performance. These releases changed the strategic landscape. The assumption that proprietary always meant meaningfully better no longer held as strongly as it once did. More importantly, the economic model flipped. The Economic Inversion With closed APIs, you are renting intelligence. Every additional request costs money. Your marginal cost per inference remains relatively constant (or even increases) as you scale. You have limited ability to optimise latency, cost, or behaviour beyond what the provider allows. With open weights hosted on your own infrastructure, the economics invert: * There is a high upfront cost to acquire hardware and set up serving infrastructure. * Once that infrastructure is running, the marginal cost of an additional inference call approaches zero. At sufficient scale, this difference becomes decisive. Companies running large open models in production are already reporting inference costs that are significantly lower — in some cases by a factor of 5–10x — compared to equivalent usage through closed APIs. This is not a minor efficiency gain. It fundamentally changes unit economics for any product where AI inference is a meaningful portion of operating costs. Control vs. Dependency Beyond cost, open weights shift the power dynamic from dependency to ownership. A closed model is a dependency. The provider controls: * Pricing * Uptime and reliability * Model updates and deprecations * What the model will and will not do (through system prompts or policy layers) * Your ability to customise or optimise An open model becomes infrastructure you own. You decide: * Which hardware it runs on * What optimisations and quantisation techniques to apply * How to fine-tune or adapt it for your specific domain * What it costs per call (after the initial infrastructure investment) * What level of privacy and data control you maintain This shift — from renting intelligence to owning it — is the deeper strategic significance of the open model movement. It moves the model from being a recurring operational expense on an API bill to becoming a capital asset on the balance sheet. The Current Trade-off The gap between the best open and closed models has narrowed dramatically. In many domains, open models are now competitive enough that the tradeoff has changed. The question for many teams is no longer “Is the closed model significantly better?” but rather “Are the remaining capability advantages worth the ongoing dependency and linear cost scaling?” For some high-stakes or highly specialised use cases, closed models may still hold meaningful edges. For many others — particularly at scale — the combination of lower cost and greater control is becoming the more rational choice. What This Means Going Forward The moat that proprietary AI built over the last decade is eroding. Not because closed models suddenly became worse, but because open models became good enough that the economic and strategic calculus shifted. For inference engineers and AI infrastructure teams, this is the most important structural change happening in the industry right now. The decision between closed APIs and open weights is no longer primarily a technical one. It is a strategic decision about cost structure, control, and long-term dependency. The era of simply paying for intelligence is being challenged by the option to own it. How that tradeoff evolves will shape the economics of AI products for years to come. KOKASYNC is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit kokasync.substack.com/subscribe
