eno reyes

Frontier model TAM is overweighted; margin compression is inevitable as open models commoditize commodity tasks. Value shifts to application layer (outcome pricing) and enterprises owning their intelligence (harness layer, continuous learning, on-prem). Model labs face existential debt service needs requiring historic free cash flow generation.

Open models will dominate 99% of AI workflows within 3 years as cost efficiency wins; labeling open models 'Chinese' is a scare tactic by frontier labs. Security risks are equivalent across all model providers and context-dependent — enterprises should evaluate per task, not origin.

The harness — where agent state, context compaction, and dynamic model allocation live — is the new application layer. Routing is commoditized; value accrues to systems that maintain stateful, project-level intelligence allocation. Continuous learning happens at harness layer, making enterprises sovereign over their accumulated workflow intelligence.

Enterprises fear model labs (OpenAI, Anthropic) will compete with them using their own data; on-prem deployment is about control, not tech. Winning enterprise sales requires discovery-mode selling — learning the customer's biggest unsolved problem — not persuasion, because the market is nascent and non-zero-sum.

Most AI application companies (neolabs) will cease as independent businesses within 18 months. Survivors own durable, proprietary workflows (e.g., legal) that won't be obsoleted by frontier model improvements and have high barriers to new entrants. General computer-use agents (Excel/Jira automation) are undifferentiated and vulnerable.

Microsoft wins by owning data centers (energy→intelligence) while staying model-agnostic on Azure. Nvidia must verticalize into chips to escape debt burden; model labs (OpenAI, Anthropic) moving to custom silicon will reshape vendor relationships. Debt cycle is existential for model labs, manageable for hyperscalers with durable cash flows.

Chinese models (e.g., DeepSeek) show no backdoors vs US models; all models reflect creator biases (Anthropic blocks recursive self-improvement content). Enterprises should assess censorship, task fit, and switching risk for every model regardless of nationality. National security use cases are the only clear exclusion.

Traditional SaaS (Airtable, etc.) faces cannibalization by AI-native tools; they must hit repeated blockbusters or sell to roll-ups (Bending Spoons). Market size expansion allows multiple winners (Linear + Atlassian both crushing), but agile workflow itself may be disrupted by AI-native development methodologies.