I am going to make a five-year prediction about which layer of the AI stack becomes a commodity and which does not. Before I get there, I separate two things this kind of article almost never separates: the data I have verified today, and my own reading of that data towards 2031. The first carries a source. The second carries my name, not any report's.
THE FACTS THIS RESTS ON
Verified fact (Epoch AI): the price of inference for a fixed model capability falls between 9x and 900x per year depending on which benchmark you measure. a16z documented a 1,000x fall over three years for a capability equivalent to GPT-3. It is the fastest and most sustained price decline in enterprise technology today.
Verified fact (Epoch AI, May 2026 comparison): despite that fall, the best open-weights model still sits, on average, between 3.5 and 4 months behind the best closed model, with a 7-8 point gap on their capability index — and that gap widened slightly in 2026 relative to 2025, it did not close. On specific coding benchmarks, some open models already beat closed models from the previous generation: the distance is only sharp at the absolute frontier.
Verified fact (Menlo Ventures, survey of 495 enterprise AI leaders, Nov 2025): three providers — Anthropic, OpenAI and Google — hold 88% of enterprise model market share, with Anthropic rising from 24% to 40% in a year. Quoting the report directly: for frontier use cases such as coding, "users are quite price-insensitive and will pay more for performance".
Verified fact (same Menlo report): in the infrastructure layer (vector databases, orchestration, evaluation), incumbents retain 56% of spend. In the application layer, startups capture 63% of the market against 36% the previous year, despite incumbents having distribution, data and balance sheet.
Verified fact (SemiAnalysis): the rental price of H100 GPUs fell 58% between 2023 and 2025 — textbook commoditisation — and then rose close to 40% between October 2025 and March 2026, with on-demand capacity essentially exhausted. Nvidia holds more than 60% of installed AI compute worldwide (Stanford AI Index 2026) and appears in 91% of chip citations in research papers (Air Street, June 2026). The big exception is Google, which now covers close to 76% of its cumulative compute with its own TPUs rather than Nvidia — but it is the exception, not the norm.
Verified fact (McKinsey, 1,993 organisations, June-July 2025): only 13% of companies have hired AI regulatory compliance specialists, and only 6% AI ethics experts. Verified fact (Stanford AI Index 2026): the foundation model transparency index fell from 58 to 40 out of 100, and 80 of the 95 notable models released in the past year do not publish their training code.
Verified fact (Stanford AI Index 2026): employment of young software developers (22-25) fell almost 20% relative to 2024, concentrated in entry-level roles — while mentions of AI skills in US job postings rise to 2.5% of the total (+55% year on year) and the "agentic AI" cluster grows 280% in a year.
MY FIVE-YEAR PREDICTION (2031)
→ My prediction: the raw compute layer (GPUs, cloud rental) will keep oscillating between apparent commoditisation and episodic scarcity, but it will not become a real commodity. The chip, its manufacturing (concentrated in TSMC) and the software ecosystem around it will remain a moat, unless more hyperscalers replicate Google's move towards their own silicon — something only one player has done at that scale so far.
→ My prediction: access to "a given model capability" will keep getting cheaper at the historical rate — that is already a de facto commodity and will become more so. But the absolute frontier will not. It will stay expensive, stay concentrated in 2-4 providers, and that oligopoly will set the price at the top, not the market.
→ My prediction: the application layer will remain where most new value is created because it is the easiest to start in — and for that same reason it will be the fastest to commoditise at the individual level. A product that is "a well-written prompt on someone else's API" will have a competitive advantage measured in months, not years. What survives will be applications with proprietary data, existing distribution, or deep integration into a real workflow — not the ones that merely wrap a model.
→ My prediction, and the strongest of the four: the AI evaluation and governance layer is today the worst resourced of the four — 13% and 6% specific hiring, transparency falling — while regulation, though delayed, keeps pushing in that direction: the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026) postponed the EU AI Act's high-risk obligations (Annex III) from 2 August 2026 to 2 December 2027 (Annex I: 2 August 2028) — it did not remove them, and it did not touch the Article 5 prohibitions, the Article 4 AI literacy duty or the Article 50 transparency requirements, in force since 2025. In five years it stops being an optional extra and becomes a structural function. It will not be a commodity. It will be a shortage of qualified talent, precisely because almost nobody is training for it today.
WHAT IT MEANS FOR HIRING
Do not hire "for AI" in general — it is too wide a category to mean anything in 2031. Hire for the layer that does not commoditise: deep integration of the model into a real business process, judgement about what to build versus what to buy, and the ability to evaluate whether what an AI system returns holds up.
The profile of "knows how to use model X" has a short shelf life — model X itself will be different, or cheaper, or replaced within eighteen months. The profile of "knows how to verify when an AI system is wrong, and under what conditions it stops being reliable" does not expire with the next release of any model. It is, in fact, the role most likely to be in short supply.
Which layer of the stack are you hiring for right now — and does it match where you think the value will be in five years?