Why This Moment Is Different
Compound Inflection Points
Previous AI waves peaked and crashed when a single enabling factor ran out — funding, compute, or credible results. The current era rests on several curves that inflected together: compute cost per unit of training kept falling on a better-than-Moore trajectory, the internet supplied a training corpus no previous generation had, the transformer made that compute and data efficiently exploitable, and scaling laws made the payoff forecastable enough to underwrite. Each factor amplifies the others — cheaper compute makes more data usable, better models justify more compute spend. This compounding is the structural difference from past waves, which had enthusiasm without an engine. It does not make a slowdown impossible; it means a repeat of the historical boom-crash pattern would require several independent curves to bend at once.
- Past winters: one enabling factor failed and the wave collapsed
- Current era: compute, data, architecture, and forecastability inflected together
- Compounding curves are harder to break than a single trend
Foundation Models Changed the Economics
The deepest structural shift is the foundation model pattern: train once at enormous cost, then serve every use case through an API. Pre-transformer ML required a bespoke model — and a scarce ML team — per task, so capability scaled with headcount, echoing symbolic AI's bottleneck. Now the marginal cost of adding an AI capability to a product has collapsed from an ML-engineering project to an API call and a prompt. That is why adoption spread through the software economy in a few years rather than decades: the entire developer population became the addressable market, not just ML specialists. General capability also compounds in a way task-specific models never did — one model's improvements flow to every downstream use simultaneously, and techniques like distillation push that capability into ever-cheaper tiers.
- Train once, deploy everywhere — fixed cost amortised across the whole economy
- Marginal cost of an AI feature fell from an ML team to an API call
- Improvements propagate to all use cases at once, unlike per-task models
The Flywheel and Its Physical Limits
AI revenue now funds the next generation of models, which unlock more revenue — a flywheel earlier waves lacked because they never reached commercial escape velocity. The capital deployed into datacentres, accelerators, and power is at a scale comparable to historical infrastructure buildouts, and that spend is itself a structural difference: this wave is pouring concrete. The binding constraints have shifted accordingly, from algorithms toward atoms — electricity generation, grid interconnection, cooling, HBM supply, and construction timelines. These physical limits are the most credible brake on the current era. The honest failure scenario is no longer "the technology stops working"; it is "returns on capability fail to keep pace with the cost of the buildout" — an economic stall rather than a scientific one.
- Revenue → research → capability → revenue: the flywheel earlier waves never closed
- Constraints moved from algorithms to power, cooling, memory supply, and construction
- The realistic risk is an economics stall, not a capability dead end
Structural Risks Worth Naming
Different does not mean safe. Frontier capability is concentrated in a handful of labs and hyperscalers, so the ecosystem inherits their outages, pricing decisions, and strategic pivots as systemic events — a dependency profile closer to cloud infrastructure than to a normal software vendor market. The supply chain runs through single points of failure in advanced fabrication and packaging that geopolitics could disrupt. And expectations have run ahead of deployment reality in many enterprises, where pilot-to-production rates remain the honest measure of the era. Open-weight models (Llama, DeepSeek, Qwen and their successors) are the meaningful counterweight to concentration — capable, self-hostable, and improving fast. Builders should treat model portability as an architectural requirement, not a nice-to-have: the abstraction layer between your product and any single provider is cheap insurance against every risk on this list.
- Concentration risk: a few providers' decisions propagate through the whole ecosystem
- Fabrication and packaging chokepoints are genuine geopolitical exposure
- Open-weight ecosystems provide the credible exit option — design for portability
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.