Frontier Generalists or Specialised Models, Where Thought Leaders Think AI Is Heading

Narrative

Investor and analyst coverage points to a bifurcated rather than winner-take-all market. CB Insights forecast in January 2025 that closed frontier providers would dominate sophisticated applications, while smaller open models would serve edge and specialised work; its February interviews found 94% of organisations using at least two LLM providers. McKinsey's 2025 survey tempers both camps' claims: 88% reported AI use in at least one function, but only about one-third had begun scaling programmes enterprise-wide.

a16z's 2024 enterprise interviews favoured RAG or fine-tuning an open model over training from scratch, including one reported case where tuned Mistral and Llama were nearly as capable as OpenAI at lower cost. By 2025, the same investor's CIO survey said stronger long-context frontier models had made fine-tuning less necessary for many use cases, although highly specific workloads still justified it. Sequoia's Sonya Huang, whose firm invests in model infrastructure, similarly framed production as a mix: a large closed model for a prototype, then lower latency, lower cost, or fine-tuned models where production economics demand them.

The analyst record also exposes a tension around Chinese open weights. Sequoia's 2026 analysis reported that Qwen's share of new open-model adaptations rose from 1% in January 2024 to 69% by February 2026, while warning that weights alone do not make training provenance or trigger behaviour auditable. Forrester therefore treats openness as a multi-dimensional procurement issue, and Gartner cautions that open coding models can require substantial infrastructure and customisation. The evidence supports conditional model choice, not a settled forecast that either frontier APIs or specialist weights will absorb the other.

Named voices (follow-up interview sweep): Mike Cannon-Brooks, Atlassian CEO, 20VC, published 12 October 2026 (verified from the episode transcript). Confidence: hedged. He said 'the jury is really out on the benefits of building specialized models at the moment', granting real cases (large proprietary datasets, robotics, world models) while arguing most specialised-model builders are not in them. A specialist may ship with a price-performance edge, but 'will that price performance hold up over time? I'm not sure it will' as general models improve. He distinguishes fine-tuning from building a model: Atlassian runs 'a bunch of very specialized openweight models that we fine-tune' but does not claim them as its own, and says most 'own model' claims are a Chinese open-weights base with domain tuning on top. What would change his mind: specialists continuing to show gains over the latest foundation model on a price-per-watt basis; he notes harness design matters but 'is not model progress'. On open weights: customer push-back targets Chinese models, not open models; Atlassian uses Nemotron and Mistral and no Chinese models in production. Commercial position: application vendor that benefits from interchangeable model suppliers.

VC and analyst coverage increasingly separates the frontier-model race from the question of where enterprise value accrues. Sequoia's David Cahn described five model-race "finalists" in December 2024, but treated their strategies as differentiated bets in vertical integration, construction capacity and open source rather than proof that one general model will absorb applications. His July 2024 analysis put annualised Big Tech capital expenditure at $229 billion, up from $138 billion year on year, while explicitly leaving the revenue outcome as a test for late 2025 to early 2026.

The application-layer thesis is more conditional. Sequoia's Sonya Huang and Pat Grady revised their earlier data-flywheel thesis: specialised, difficult-to-obtain data can matter, but ordinary application data may be erased as foundation models improve, leaving workflow and user-network positions more durable. a16z's 2024 enterprise survey found 46% of respondents preferred or strongly preferred open models and nearly 60% would increase open use when fine-tuned open models approached closed-model performance; its later Harvey interview supplies a concrete company response, using OpenAI models with fine-tuning, post-training, prompting and retrieval rather than training a foundation model.

The reports therefore do not support a clean generalist-versus-specialist split. CB Insights describes buyers assessing small language models alongside Big Tech advantages, while McKinsey's decision framework says companies should buy services for commodity capabilities and invest where proprietary advantage exists. a16z's later investment argument goes further: model improvement can increase the value of hard-to-copy data and embedded professional process, citing Harvey's firm-specific workflows, yet this is an investor thesis rather than independent causal evidence.