Efficient AI expands usage while reshaping infrastructure economics
StrengtheningDuring 2027–2029, smaller and more efficient models could shift AI spending away from exclusively frontier-scale training and toward broad, high-volume inference. The investment outcome depends on whether usage expands faster than efficiency reduces compute demand, and whether workloads remain concentrated on premium accelerators or migrate toward cheaper and more diversified hardware.
November 21, 2026 — The evidence supports both sides of the scenario. Distillation can preserve most model performance while improving efficiency, and lower inference costs may be necessary for broad enterprise deployment. At the same time, agentic workloads increase reasoning and inference intensity, while AI infrastructure spending, data-center construction, and hardware demand remain on strong growth paths. The most plausible outcome is a mixed transition: lower compute cost per task, much wider usage, continued aggregate infrastructure growth, and a changing mix of hardware demand.
Geography: Global
Portfolio implications
Your exposure
How this scenario reads against the holdings in your portfolio: which positions carry more exposure, and which look relatively resilient.
Log in to see how this scenario affects your portfolioDevelopments
Inference growth can offset efficiency gains
Strengthens the thesisAgentic systems can use more inference compute because they perform more extensive reasoning, while affordability and enterprise adoption are already identified as drivers of data-center demand. Inference-driven infrastructure spending is forecast to grow rapidly as AI becomes operationalized across applications and workflows.
Why it matters: A major reduction in compute per task does not necessarily reduce total demand if the number of AI-enabled tasks and the reasoning intensity of each workflow rise substantially.
Potential beneficiaries
- AI infrastructure providers
- Data-center operators
- Power and networking infrastructure providers
Potential losers
- Businesses that assume lower unit costs will automatically translate into lower aggregate infrastructure demand
The evidence favors the bullish adoption outcome, although it does not quantify whether usage growth will exceed efficiency gains over 2027–2029.
Large-scale infrastructure commitments create near-term oversupply risk
Weakens the thesisHyperscalers are projected to commit more than $700 billion to data-center capital expenditure in 2026, with the five largest allocating more than $800 billion to construction. AI infrastructure spending has also reached a high growth rate, and planned accelerator deployments remain substantial. If smaller models reduce demand for frontier training or premium GPU capacity faster than new applications absorb it, capacity utilization and rental economics could weaken.
Why it matters: The transition could be disruptive for infrastructure even if AI adoption remains strong, because the industry may have committed to expensive capacity before the model mix becomes more efficient.
Potential beneficiaries
- Buyers of compute capacity if rental prices decline
- Enterprises seeking lower-cost AI deployment
Potential losers
- Owners of excess premium data-center and accelerator capacity
- Suppliers dependent on uninterrupted frontier-scale expansion
A credible bearish branch, but current evidence still shows strong spending and demand rather than actual oversupply.
Specialized models lower the economic barrier to enterprise AI
Strengthens the thesisDistillation can transfer capabilities from advanced models into more efficient models while retaining most performance. Since enterprises may not be able to deploy inference at scale unless costs fall significantly, smaller models could unlock AI use cases that are currently uneconomic.
Why it matters: This supports a broad adoption cycle in which software and workflow deployment expand even as the cost of each individual AI task declines.
Potential beneficiaries
- Enterprise software providers
- Companies deploying AI into business workflows
- Inference-optimization and model-compression providers
Potential losers
- Providers whose economics depend primarily on expensive inference per task
Strong support for the efficient-model adoption thesis, but the evidence does not establish that smaller models will displace frontier models in all workloads.
Thesis assessment
| Component | Direction | Note |
|---|---|---|
| Specialized models achieve most required business value at lower cost | strengthening | Distillation is described as retaining most original-model performance while running more efficiently. |
| Lower unit costs broaden AI adoption | strengthening | Enterprise-scale inference may require a significant reduction in cost to serve, while affordability is already driving demand. |
| Aggregate compute demand declines | no decisive new evidence | Efficiency supports lower demand per task, but agentic reasoning, enterprise adoption, and infrastructure spending point toward continued demand growth. |
| Premium accelerator economics weaken | partially mitigated | A shift toward CPUs and inference-focused systems raises substitution risk, but current accelerator deployment and spending remain strong. |
Bottom line
Smaller AI models are not automatically bearish for AI infrastructure. The strongest scenario is a change in the composition of demand: cheaper models enable many more deployments, while agentic and inference-heavy workloads keep total computing demand growing. The main risk is that efficiency gains and hardware substitution arrive faster than usage expands, leaving existing premium capacity oversupplied.
Favor exposure to the broader AI deployment ecosystem rather than assuming that all AI infrastructure spending will remain concentrated in frontier-scale training. The key indicators are inference-cost declines, workload growth, CPU and specialized-hardware adoption, capacity utilization, and the pace of data-center commitments.

