rationale:

Update August 5, 2026: Weighed against established Alphabet-backed and gigawatt-scale lab forecasts, the estimate was maintained to account for standard data center energization delays limiting late-2028 capacity.

Current Status & Strategic Focus Discovery Loop was announced today, August 5, 2026, as a public benefit corporation by prominent founders Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le 2 sources. While its seed round—co-led by Radical Ventures and Khosla Ventures, with Alphabet participating—commands significant attention, the terms are undisclosed and the round had not officially closed at launch 2 sources. The company's stated goal is to automate experimental loops in ML, science, and engineering 2 sources. Google Cloud is supplying compute for at least the first year 2 sources, but neither the company nor its partners have disclosed any long-term dedicated capacity figure, chip count, or multi-year infrastructure roadmap 2 sources.

Comparables & Capital Constraints Elite-founder "neolabs" in the 2026 environment are landing massive commitments within 1–2 years of founding, which sets the ceiling for Discovery Loop's potential. Thinking Machines committed to at least 1 GW of Nvidia Vera Rubin systems 2 sources, Safe Superintelligence (SSI) secured a reported $5B Nvidia investment to scale compute 10× 2 sources, and Reflection AI contracted GB300 capacity at $150M/month 2 sources. However, achieving gigawatt scale requires staggering capital. Using standard cost models, a 1 GW AI data center implies up to $38B in up-front capex and ~$8.5B per year in total cost of ownership epoch.ai. Reflection AI's massive $1.8B annual spend equates to roughly 0.1 GW of all-in capacity. Discovery Loop would require one or two mega-rounds to sustain a dedicated footprint significantly beyond the low hundreds of megawatts by 2028.

Alphabet Relationship & Resolution Criteria The most significant driver of upside is the founders' deep history building Google's TPU stack and Alphabet's dual role as investor and cloud partner digitalapplied.com. A large, dedicated TPU reservation announced in 2027 is highly plausible. However, strict measurement criteria require this capacity to be exclusively contracted and physically live by December 31, 2028. Shared, on-demand cloud usage—a likely setup for the company's first-year supply arrangement—does not qualify as exclusive capacity and would register near zero. Furthermore, widespread data center delays and power constraints in 2026 mean that even if a major TPU deal is signed in 2027, it may only be partially energized by the target date.

Final Assessment The median estimate of ~0.16 GW expects the company to close a substantial seed round and convert its Google relationship into an exclusive, dedicated TPU or GPU reservation over the next two years. This reflects a successful trajectory akin to Reflection AI’s scale, rather than an immediate jump to a multi-gigawatt mega-cluster. The 10th percentile (~0.017 GW) accounts for downside risks where the company remains a leaner research shop, relies heavily on shared non-exclusive cloud, or faces severe energization delays. The 90th percentile (~1.08 GW) captures the formidable upside tail: a scenario where Alphabet provides its most prestigious alumni with a massive, dedicated TPU allocation comparable to the 1+ GW deals seen by peers like Thinking Machines, fully live before 2029.