Current State of Deployment Data in Training
The share of total frontier-lab training compute primarily driven by production deployment sessions is extremely small, with a central estimate of 1.0%. While deployment data does enter training, it acts as a secondary input rather than the primary driver. Consumer tiers at major labs do train by default — Anthropic retains Free/Pro/Max chats for up to five years, and OpenAI trains on Free/Plus/Pro by default. However, API and enterprise tiers, w
Set against related questions on lab margins , this estimate remains centered at 1.0%, reflecting that current lab profitability is driven by scale rather than by training on live user data.
My median estimate is 4.7%, anchored by a currently negligible baseline but leaving room for a more substantial right-skewed tail if frontier labs shift toward deployment-time weight updates by 2028. As of October 2026, research into models updating weights from deployment data remains a small niche, making up roughly 1-3% of alignment effort. The operational pressure to align such systems is very low: no frontier lab has shipped a model that updates its weights from live sessions, and the indus
Weighing this question against estimates for deployment-data training compute and commercial reliance on live updates thickened the upper tail of the forecast to account for a potential sudden reallocation of safety resources if continuous learning breaks out.
The most reliable baseline for this measure is Menlo Ventures' enterprise survey, which placed Anthropic at 40% and OpenAI at 27% in December 2025 2025: The State of Generative AI in the Enterprise. Historically, the leader's share has fluctuated widely, dropping from OpenAI's 50% in 2023 to 32% in mid-2025 before Anthropic took the lead [https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-en
The forecast was adjusted slightly outward to ensure the overall market leader's share mathematically accommodates independent estimates for specific competitors while factoring in the potential lock-in effects of enterprise continuous learning.
Anthropic is the clear target for this assessment, having overtaken OpenAI in revenue. Leaked financials indicate Q2 2026 revenue of $11.5B against OpenAI’s $6.7B, with Anthropic’s Q3 revenue reaching $16.8B OpenAI's September annualized revenue nears $50 billion, ... [Anthropic Q3 2026 revenue jumps 45% to $16.8B, reports ...](https://www.exchange4media.com/digital-news
Set against related projections for 2028 , this forecast was maintained near a 74% median, as the broader picture confirms that strong structural efficiencies support high margins today even without continuous-learning moats.
I am pricing Anthropic as the most likely candidate for the largest AI lab by revenue in FY2028. Recent unconfirmed reports indicate Anthropic's Q2 2026 revenue reached $11.5B, surpassing OpenAI's $6.7B, with Anthropic's annualized run-rate exceeding $65B in July and on track for $100B by year-end OpenAI's September annualized revenue nears $50 billion, ... [OpenAI's Sep
Set against related estimates of 2026 profitability , this forecast was slightly compressed to a 70.5% median to reflect expectations of gradual margin erosion from price competition by 2028.
Current State: The Dominance of Adapters Managed per-customer model customization is currently overwhelmingly adapter-based, supporting a median estimate of 12% for full-weight updates. Major platforms uniformly favor parameter-efficient methods: Google Vertex's Gemini tuning is LoRA-based and billed at the base model's rate 2 sources, Azure defaults to LoRA learn.microsoft.com, Thinking Machines' Tinker restricts training to LoRA adapters on shared base models beam.cloud, and Together AI position
This estimate remained largely unchanged, as the profound economic efficiencies of multi-tenant adapter serving are expected to offset technical pressures pushing for full-weight updates.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited