The AI industry's most consequential resource-allocation question reduces to four blunt directives: build foundation models, chase architectural breakthroughs, run experiments, and — if forced to ship product — ship only the components that sit closest to the source. Everything downstream is someone else's problem.
The logic is specific to this moment in AI's development cycle. Pre-training — feeding trillions of tokens through self-supervised learning to produce a base model — remains the binding constraint on what AI systems can do. The post-training stack (supervised fine-tuning, reinforcement learning from human feedback, preference alignment) adds polish, safety guardrails, and task-specific behavior. It does this cheaply enough that dozens of competing teams can replicate the work using open-source base models and off-the-shelf toolchains. The moat, such as one exists, lives in the base model itself.
Four Principles for Scarce Compute
Four operating rules, each structured as a priority filter, capture the logic.
Pre-training over post-training. A handful of organizations worldwide can independently train frontier-scale models. Post-training expertise is far more common and is drifting toward commodity status. A strong base model — the "soul" of the system — determines the ceiling for every application built on top of it.
Frontier research over popular applications. New architectures (mixture-of-experts optimization, test-time compute scaling, reasoning enhancements, self-evolving systems), novel data engineering, and theoretical interpretability work carry high failure rates. They also carry the only plausible path to durable competitive separation. Building vertical agents or fine-tuning models for industry-specific tasks generates near-term revenue, and near-term competition that erodes margins fast.
Exploration over delivery. Small-scale architectural experiments — testing data mix ratios, probing training stability boundaries, verifying novel ideas — produce disproportionate returns per dollar spent. Productization, compliance, customer support, and engineering maintenance consume resources at scale with diminishing returns. Research organizations that attempt full-stack delivery find themselves staffing operations desks when they should be staffing labs.
Head nodes over tail nodes. When shipping product becomes unavoidable, the priority belongs to upstream components: base models, training frameworks, data pipelines, tokenizers, mixture-of-experts routers, and foundational capabilities like long-context handling. Downstream work — user interfaces, industry plugins, customer service workflows, vertical fine-tuning — fragments into low-margin price competition. The analogy is manufacturing: produce core components, not final consumer devices.
The Supply-Chain Calculus
Each rule reinforces the same capital-allocation thesis. Compute, elite research talent, and high-quality training data are scarce inputs. Spreading them across the full AI value chain dilutes their effect. Concentrating them upstream maximizes the leverage of each GPU hour and each researcher's time.
The competitive math supports this. Hyperscalers have structural advantages in full-chain delivery: global sales teams, compliance departments, enterprise relationships, and operational budgets that dwarf anything a research lab can assemble. DeepSeek and Mistral gained outsized influence by publishing strong base models and letting the broader ecosystem handle application-layer work.
The Real Bet: Pre-Training Is Still the Bottleneck
The sharpest edge in this analysis is a read on where AI development actually stands today. Many recent capability breakthroughs — reasoning performance, planning, multi-step tool use — have traced back to pre-training advances and architectural changes, not to post-training refinements. If that pattern holds, the organizations controlling the strongest base models will set the terms for every downstream player, much as semiconductor fabricators set the terms for device makers.
This is a wager on sequencing. The current period — where pre-training quality remains the primary bottleneck and post-training cannot compensate for a weak foundation — must persist long enough to reward heavy upstream investment. Teams following this doctrine accept distance from revenue and slow monetization paths in exchange for accumulating intellectual property, proprietary training data, and institutional knowledge that compounds over time. The trade-off is captured with unusual clarity by one phrase: "do the hard things that others cannot or will not do." For investors evaluating AI companies, the question reduces to whether a given organization is building the engine or decorating the dashboard — and whether the engine still determines who wins.