An AI model distilled from 7 billion parameters consumes up to ten times less energy than a large model at inference — and rivals it on most common tasks. This data, documented in a study published in April 2026 in the journal Joule, should shake a narrative that has become accepted wisdom: artificial intelligence always requires more computing power, therefore always more electricity, therefore always more data centers.

The major technology platforms have built an unprecedented investment program on this equation. Microsoft, Google, Meta, and Amazon announced together approximately $725 billion in capital expenditures for 2026 (Amazon ~$200B, Microsoft ~$190B, Alphabet ~$175-185B, Meta ~$115-135B). Beyond these four American hyperscalers, global data center spending would exceed $788 billion according to Gartner’s projections. The race for nuclear power plants has followed in their wake.

This narrative may be correct. But it rests on an architectural choice, not a physical constraint. And this choice could prove to be a losing one.

The Essential Points

  • A model distilled from DeepSeek-R1 (7 to 70 billion parameters) achieves performance comparable to models 3 to 5 times larger for the majority of common benchmarks, according to work published on arXiv in early 2025.
  • Distillation reduces energy consumption at inference by a factor of 5 to 10; for certain specialized tasks, small models divide energy consumed by one to two orders of magnitude.
  • Major technology platforms maintain their bets on very large models on the grounds that the frontier of capabilities always requires more computation — a valid argument for fundamental research, questionable for large-scale deployment.
  • The $725 billion in spending announced for 2026 constitute an industrial bet on the high trajectory of compute demand, not a response to an inevitable technical constraint.

What Distillation Changes in the Energy Equation

Distillation consists of training a smaller model — the student — to reproduce the behavior of a large model — the teacher. The student model is not a degraded version: it learns to imitate the probability distributions of the teacher, not just its raw outputs. The result is a more compact model that concentrates much of the large model’s capabilities into a far smaller computational footprint.

The study published in Joule by researchers from Microsoft Research documents the scale of the phenomenon at inference — the phase where the model responds to user requests, and which represents the bulk of operational data center consumption. At training, consumption is concentrated in time and shared across many uses; at inference, it is continuous, distributed, and directly proportional to the volume of requests. This is where the energy bill for large-scale deployment is determined.

The figures are clear. A model with 7 billion parameters — the family of small open models, such as distilled versions of LLaMA or DeepSeek — consumes 5 to 10 times less energy at inference than a 70-billion-parameter model, all else being equal. On narrowly defined tasks — classification, information extraction, answering standardized queries — the gap can reach one to two orders of magnitude. These are not marginal gains. These are changes of scale.

Distillation from DeepSeek-R1, released in open access in early 2025 by the Chinese company of the same name, served as a real-world demonstration. Distilled models at 7 and 70 billion parameters achieve performance comparable to GPT-4o on mathematical and reasoning benchmarks. The 32-billion-parameter model exceeds significantly larger competing models. This result does not prove that large models are useless. It proves that for a large fraction of production uses, small distilled models do the job at a fraction of the energy cost.

Why Major Platforms Continue to Bet on Massive Computation

The argument of the major platforms is coherent, provided its scope is clearly defined. Large models remain necessary for two main reasons.

The first is exploring the frontier. GPT-5, Gemini Ultra, Claude Opus — these models are not deployed to answer the same queries as a 7-billion-parameter model. They serve to explore what AI is capable of doing when not constrained by size: complex multi-step reasoning, integration of multiple modalities, advanced coding tasks, computational science. For these uses, more parameters and more computation continue to deliver measurable performance gains. The scaling laws of Kaplan and his coauthors, published in 2020, established this relationship — even if their validity at very large scales has been debated since.

The second reason is strategic rather than technical. Possessing the most powerful infrastructure creates formidable barriers to entry, attracts enterprise customers who want to guarantee access to the frontier, and positions platforms to capture the value of the next generation of models. Microsoft invested in OpenAI. Google has Gemini. Amazon has Anthropic. These commitments do not unwind over a quarterly result.

What the platforms do not say, however, is that the share of their production requests that actually requires frontier models is probably low. The bulk of volume — augmented search, document synthesis, code generation, customer support — can be handled with much smaller models. The current trajectory mixes two distinct logics: investment in R&D to push the frontier, and over-investment in infrastructure for uses that distilled models could cover at far lower cost. It is difficult to disentangle the two from the outside, because the platforms have no interest in doing so.

Open Laboratories as an Alternative Model

The gap between the narrative of major platforms and the practice of laboratories working in open access illustrates the nature of the choice at stake. Meta, with its LLaMA family, made the opposite bet: publish intermediate-sized models, optimized for deployment, capable of being distilled and fine-tuned by anyone with the technical resources. Mistral AI, in Europe, follows a similar trajectory — models of 7 to 22 billion parameters that compete with much larger proprietary models on most practical benchmarks.

These choices are not philanthropic. They rest on a different economic calculation: if large proprietary models are out of reach for the majority of enterprises, open and distillable models create an ecosystem that accelerates adoption and generates other forms of value. But the byproduct of these strategies is a continuous demonstration that the high trajectory of computation is not the only viable one.

DeepSeek took this logic to its limit. In January 2025, the open release of DeepSeek-R1 and its distilled derivatives caused a shock wave through the sector. Not because the model surpassed GPT-4o on all criteria — it does not — but because it achieved comparable performance on reasoning benchmarks with a fraction of the declared training resources, and because its open distillations immediately offered high-level performance to anyone wishing to deploy them. The competition between Western venture capital and Chinese planning illuminated the geopolitical dimension of this episode: computation is not only an engineering choice, it is a competition terrain between models of industrial organization.

The Nuclear Bet and Its Hidden Assumptions

The most visible consequence of the major platforms’ investment program is the revival of interest in nuclear energy. Microsoft signed an agreement to restart a unit at Three Mile Island. Google concluded contracts with small modular reactor developers. Amazon invested in Kairos Power. These commitments are real and engage significant capital over horizons of ten to twenty years.

The logic is straightforward: data centers consume electricity continuously, regardless of weather, making them ideal customers for a controllable energy source like nuclear power. Renewable energies, solar and wind, cannot meet this base load demand alone. The reasoning holds — on one condition: that compute demand actually follows the high trajectory.

This is where hidden assumptions become visible. If distillation and inference optimization techniques continue to progress at the current pace, energy consumption per request will decrease significantly, even if the number of requests increases. The net effect on electricity demand depends on demand elasticity: if uses multiply fast enough to offset efficiency gains, the high trajectory remains valid. This is what economists call the rebound effect, documented for every major technological transition since the steam engine.

The rebound effect is real and likely. But its magnitude is uncertain. Betting $725 billion and decades of nuclear commitment on a specific demand trajectory is an industrial choice, not a mechanical reading of a physical constraint. This choice could prove right. It could also create overcapacity assets if lightweight techniques take hold faster than expected in production uses.

What Efficient Optimization Shifts in AI Economics

The stakes go beyond electricity consumption. If small distilled models become the standard for production deployment, several economic equilibria shift simultaneously.

The first concerns market concentration. Large proprietary models are a formidable barrier to entry: training GPT-4 or Gemini Ultra costs hundreds of millions of dollars and requires infrastructure that only five or six organizations in the world can assemble. Distilled models break this barrier. A company that fine-tunes a 7-billion-parameter model on its proprietary data can obtain performance comparable to a large model for its specific uses, with infrastructure costs within reach of an SME. This is what our analysis on organizations that actually capture AI value pointed out: the advantage does not lie in access to the largest model, but in the ability to integrate it into operational workflows.

The second concerns infrastructure geography. If compute demand is less concentrated in large centralized models, it can be more geographically distributed. Intermediate-sized data centers, closer to end users, with lower energy footprints, become economically viable. This is a very different trajectory from the race for mega-campuses that major platforms are currently building.

The third concerns the relationship with regulators. In Europe, both the directive on data center energy efficiency and the AI regulation apply to infrastructure whose consumption is known and measurable. If large models maintain their dominance, the regulator has concentrated leverage over a few identifiable actors. If distilled models become distributed, regulation becomes more complex. The question of who regulates what in a distributed ecosystem has no stabilized answer yet.

Neither Technological Determinism Nor Efficiency Illusion

Two symmetrical errors lurk in analyzing this issue.

The first would be to conclude that distillation solves AI’s energy problem. Efficiency gains are real, but so is the rebound effect. If efficient models make AI accessible to ten times more users and uses, total consumption can increase even if consumption per request decreases. Aggregate electricity demand depends on the breadth of adoption as much as on technical efficiency. The International Energy Agency projects a doubling of data center consumption by 2030 — from approximately 415 TWh in 2024 to roughly 945 TWh in 2030 — even accounting for efficiency improvements.

The second error would be to treat the $725 billion in investments as a rational decision constrained by physics, which no one should question. These investments are industrial bets, carried by companies that have an interest in the high trajectory of compute demand becoming reality. Their forecasts are not neutral. That does not make them false, but it requires reading them for what they are: projections constructed by actors whose interests are aligned with a specific conclusion.

The open question is that of regulators and infrastructure investors. If inference optimization techniques continue to progress at the rate of the last eighteen months, at what point should long-term infrastructure commitments be recalibrated? Nuclear contracts signed today commit capital over horizons that far exceed an AI technology cycle. This gap between the speed of algorithmic innovation and the duration of physical investments may be the true industrial risk of the decade.

Sources

  1. Oviedo F. et al. (Microsoft Research), “Energy use of AI inference, efficiency pathways, and test-time scaling”, Joule, April 2026 — https://www.cell.com/joule/fulltext/S2542-4351(26)00114-5
  2. Oviedo F. et al., arXiv preprint (September 2025) — https://arxiv.org/abs/2509.20241
  3. DeepSeek-R1 Technical Report, arXiv, January 2025 — https://arxiv.org/abs/2501.12948
  4. International Energy Agency, Energy and AI, official report 2025 — https://www.iea.org/reports/energy-and-ai/executive-summary
  5. International Energy Agency, Electricity 2025, annual report — https://www.iea.org/reports/electricity-2025
  6. Kaplan, J. et al., “Scaling Laws for Neural Language Models”, arXiv, 2020 — https://arxiv.org/abs/2001.08361
  7. 2026 Hyperscaler Capex — Statista / Yahoo Finance — https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
  8. DeepSeek-R1-Distill-Qwen-32B, Hugging Face (official) — https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B