The AI economy in 2025: stability, model shifts, and the startup opportunity

Executive overview

After years of rapid, disorienting change, the AI economy settled into a recognisable structure in 2025: model layer, application layer, infrastructure layer. Each has a workable playbook. The ground has stopped shifting as violently, which means finding startup ideas has returned to normal difficulty.

The infrastructure build-out — GPUs, data centres, power — mirrors past technology revolutions. Founders are not Comcast; they are YouTube. Overinvestment in infrastructure creates the cheap abundance that makes application-layer startups possible.

The AI bubble fear is misplaced for startup founders: infrastructure overbuilds benefit the application layer, not threaten it.

Model dominance is shifting

• Anthropic overtook OpenAI as the most-used API among YC Winter 26 applicants, reaching over 52% — up from ~20–25% for most of 2024.
• Coding performance drove the shift; Anthropic made coding a deliberate internal north star.
• Familiarity bleed-through may amplify this: founders using Claude for personal coding default to it in their products.
• Gemini climbed from low single digits to ~23% in the same period; Harj switched to it as his primary tool for its grounding and real-time accuracy.
• OpenAI retains stickiness through memory — it knows users' personalities and preferences.

Model arbitrage is the new normal

• Mature AI companies (Series B+) are abstracting away model loyalty, building orchestration layers to swap models per task.
• One startup uses Gemini for context engineering, then feeds output to OpenAI for execution — swapping as rankings change.
• Proprietary evals on domain-specific data are the actual moat, not model loyalty.
• This commoditisation is good news for application-layer startups: competition between labs drives down inference costs.

The AI bubble argument

• The telecoms bubble of the 1990s created the cheap bandwidth that made YouTube possible; the AI infrastructure boom is the same dynamic.
• Carlota Perez's framework: technology revolutions have an installation phase (heavy CapEx, bubble feeling) followed by a deployment phase (abundance, application explosion).
• 2025 is the transition point — data centres are being built, but the application-layer equivalents of Google and Facebook haven't been founded yet.
• For a dorm-room founder, Nvidia's valuation is irrelevant. The overbuilding is their tailwind.
• Meta and the hyperscalers must overbuild — it's their CapEx risk, not startups'.

Infrastructure constraints are pushing to space

• Power, land, and regulation are all bottlenecks to AI data centre expansion.
• Boom Supersonic pivoted to generating power via jet engines for data centres; the supply chain for those engines is already backed up 2–3 years.
• Space data centres — laughed at 18 months ago — are now being pursued by Google and SpaceX.
• YC has a cluster of companies addressing the build-out: space data centres, Helion for energy, and Zephyr Fusion (space-based fusion, recent YC grad).

Domain-specific models are becoming viable

• Fine-tuning open-source models with RL on proprietary domain data can beat frontier models on narrow tasks.
• One YC healthcare startup beat OpenAI benchmarks using only an 8B parameter model trained on their data set.
• Risk: general model releases (GPT-4.5, 5.1) can erase fine-tuning advantages; domain specialists must keep iterating.
• Barrier to building models is falling — the rare combination of research, engineering, and startup skills is now more widely distributed.

Hiring and team size: the second-wave reality

• Early 2025 saw companies reach $1M ARR with no hires and raise Series A; this did not scale into leaner post-Series A orgs.
• Post-Series A, the hiring playbook looks largely the same as pre-AI — companies still bottleneck on execution talent, not ideas.
• The driver: rising customer expectations offset productivity gains, keeping headcount demand high.
• Gamma reached $100M ARR with 50 employees — a notable inversion of the traditional headcount flex.
• The "one person, trillion-dollar company" era is not 2026; sub-100-person companies doing hundreds of millions in ARR is the near-term reality.

First-wave vs second-wave AI companies

• First-wave vertical AI companies (e.g. Harvey in legal) raised large rounds early and spent heavily on fine-tuning that bought no durable advantage.
• Second-wave competitors (Legora, Giga) entered with better foundation models and no fine-tuning debt, and are closing the gap fast.
• Capital as a moat: some first-wave companies locked up enough institutional capital to crowd out Series A competitors — but that only works until well-funded challengers arrive.

Vibe coding became a category

• What YC partners observed as a founder behaviour in early 2025 became a major product category by year end.
• Winners include Replit, Emergence, and Anti-Gravity (Varun Mohan, ex-Codeium, backed by Google).
• Production-quality, 100% reliable vibe-coded output is still not achievable as of end 2025.