Part of: this research — view all 7 rows
Status Quo and Required Conditions For this outcome to materialize, a top-five lab (OpenAI, Anthropic, Google, xAI, or Mistral) must make access to its absolute best model tier strictly conditional on customers granting training rights over their sessions. No lab currently does this. Existing data-for-access programs operate entirely as optional incentives. Google’s invite-only Contributor Tier offers steep discounts in exchange for training rights, but the supported models remain available at standard API prices aistudio.google.com, and its strongest model, Gemini 4 Carbon, is listed for Standard access only aistudio.google.com. Similarly, OpenAI offers free token allowances and xAI provides monthly credits for data sharing, but neither uses data collection as an access gate for their flagship tiers.
The Commercial Barrier: Enterprise Revenue The primary obstacle to a mandatory data gate is intense commercial pressure from enterprise and API buyers, who demand data privacy and zero data retention (ZDR). These customers constitute the vast majority of lab revenue; for example, customers spending $100,000 or more account for 74% of Anthropic's revenue. We recently saw a stark demonstration of enterprise bargaining power when Anthropic mandated a 30-day safety retention period for its top "Covered Models" support.claude.com. Following pushback from over 100 enterprise customers anthropic.com, Anthropic rolled out Enterprise Frontier Safeguards, allowing eligible customers to keep monitoring data in their own cloud environments and retain ZDR cnbc.commlq.ai. OpenAI has also moved to protect enterprise privacy, introducing Private Safety Processing precisely to keep ZDR compatible with frontier capabilities openai.comopenai.com. With multi-homing on the rise, gating a top model behind mandatory training rights would likely drive highly lucrative enterprise clients directly to competitors.
Near-Term Pathways Despite the commercial headwinds, there is a plausible path to early resolution (reflecting the early 2028 10th percentile). A lab could launch a temporary, limited preview or an experimental release of a new flagship model—perhaps a highly capable agentic tier—that is initially restricted to design partners or a free tier where training rights are mandatory. Because the criteria require only one lab to act and do not specify a minimum duration for the restriction, a staged rollout designed to gather critical deployment data before broad commercial release could trigger a relatively swift resolution.
The Technical Catalyst: Continual Learning In the longer term, the calculus could change if continual learning from deployment data becomes technically necessary to maintain frontier capabilities. Currently, no publicly known frontier model updates its weights from live sessions; adaptation relies entirely on context injection, retrieval, or slow periodic retraining. If the paradigm shifts and real-world session rollouts become the bottleneck for advancing models, the immense technical value of deployment data might eventually outweigh the commercial friction of forcing enterprise clients to provide it.
The Long Tail The timeline stretches significantly (with a median in mid-2033 and a 90th percentile extending into 2052) because this strict data-gating paradigm shift may never actually occur. Labs might find that steep discounts, consumer-tier defaults, and voluntary licensing agreements provide sufficient real-world data without ever needing to coerce their most profitable clients. Furthermore, even if continual learning arrives, labs could deploy per-customer private adaptation that improves a model for a specific enterprise without pooling their proprietary data into the foundation model. These alternatives create a heavy upper tail, reflecting a high probability that top-tier models remain commercially accessible without mandatory training-rights concessions for decades to come.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited