Strategic Posture and Background AMI Labs, co-founded by Yann LeCun, raised a $1.03B seed round in March 2026 at a $3.5B pre-money valuation 2 sources. Rather than joining the brute-force LLM scaling race, the lab explicitly pursues an alternative research agenda focused on world models and modular, domain-specific systems artificialintelligence-news.com. LeCun has publicly stated that scaling LLM compute is the "slow way" to human-level AI and that there will be no "country of geniuses in a data center next year" observer.com. This fundamental anti-scaling philosophy strongly suggests AMI will not pursue the multi-gigawatt infrastructure buildouts seen at incumbents like OpenAI or Anthropic.

Budget Realities and Compute Anchors A strict budget check constrains baseline expectations. AMI's only known compute arrangement is a cloud contract with Nebius, tied to its 240 MW Béthune campus in France, but specific capacity figures remain undisclosed usinenouvelle.com. The $1.03B seed must cover both elite talent and compute over multiple years. Assuming standard all-in power conventions (roughly 3.2 kW per GB300-class GPU, as seen with Mistral's 44 MW build crn.com) and current market rates of $6-8M per MW-year for dedicated capacity, dedicating a realistic portion of this funding to infrastructure yields a steady-state footprint of roughly 15-40 MW. This perfectly aligns with a median estimate of approximately 42 MW (0.04167 GW) for a serious, dedicated research cluster.

Downside Risks and Strict Resolution Criteria The resolution criteria strictly require compute to be "owned or under signed contract, exclusive-use, and live by 31 Dec 2028." Because AMI is a very small, research-stage organization of around 50 people nebius.science, much of its usage might rely on on-demand or shared cloud resources rather than dedicated, reserved blocks. If shared capacity is excluded under the strict definition of exclusive use, or if AMI's non-LLM models require significantly less compute than anticipated, the resolvable capacity could easily fall into the single-digit megawatts or near zero.

Upside Tail and Future Expansion While the median estimate sits in the tens of megawatts, a fat right tail accounts for plausible expansion scenarios. AMI boasts deep-pocketed strategic backers, including Nvidia and Samsung 2 sources. It is entirely possible that AMI raises a multi-billion dollar Series A in 2027 and signs an anchor-tenant deal at a European AI gigafactory. Should Nvidia or a sovereign-AI consortium structure a massive strategic compute deal, dedicated capacity could surge into the 100-300 MW range by late 2028. However, a footprint exceeding 0.5 GW remains highly unlikely given the massive capital requirements and the founder's public rejection of the infrastructure scaling race.