What minimum annual committed spend will buy customer-specific continually-updated model weights from a top-five-by-revenue AI lab as of 2028-12-31?

The median estimate of $1,000,000 per year reflects a market transition away from self-serve fine-tuning and toward high-touch enterprise engagements. Currently, no top-five lab offers customer-specific weights that update continually from live deployment data. Furthermore, OpenAI is winding down its self-serve fine-tuning platform openai.com, and Anthropic has never offered a first-party fine-tuning API. With the existence of a qualifying offering by the end of 2028 sitting at roughly a coin flip, the distribution heavily weights what the entry price would be if a top lab does introduce this capability.

The most likely scenario is a sales-led enterprise product, anchoring the median and 75th percentiles ($1M to $3M). Mistral Forge is the closest borderline case today, built for "continuous adaptation" via operational feedback mistral.ai; third-party estimates for its platform license run between $500K and $2M annually, with a typical first year around $3.7M datacamp.com. Similarly, OpenAI’s Custom Model program—a high-touch service run with dedicated researchers—has historically required multi-million-dollar commitments techcrunch.com. If continual learning requires dedicated pipelines and forward-deployed engineering, an entry point around $1M aligns with existing enterprise AI spending floors.

The right tail ($3M to $10M+) is driven by the severe economics of serving full-weight forks. A model requires thousands of concurrent decoding sequences to overcome the ~103x compute efficiency penalty of batch-size-one inference. If multi-adapter serving fails to maintain quality under sequential fine-tuning, organizations will need full-parameter updates—a workload Shopify already runs internally shopify.engineering. Serving these forks efficiently mandates dedicated capacity on a massive scale, placing the minimum viable spend firmly in the millions and confirming the premise that personalized weights strongly favor large organizations.

Conversely, the left tail ($92.5K to $335K) reflects the possibility that low-rank adapters prove sufficient for continual learning. Multi-adapter serving incurs roughly a 5% throughput penalty, making the underlying economics highly efficient. Committed-capacity products set a low floor here; Google's Provisioned Throughput sells in increments that annualize to roughly $24K to $32K cloud.google.com. Even with Google now billing tuned Gemini 3+ inference at a 1.5x premium over base rates cloud.google.com, an automated, log-driven continuous tuning feature could easily enter the market as a self-serve add-on, dropping minimum commitments to the $100K range or lower.