SSI operates as a pre-revenue, deep-stealth research lab with roughly $3B in early funding, heavily reliant on Google Cloud TPUs as an anchor customer 3 sources. In July 2026, NVIDIA took a reported ~$5B equity stake in the company, granting priority access to the next-generation Vera Rubin platform 3 sources. SSI leadership explicitly stated this partnership will increase its compute "by an order of magnitude" within 12 months, or by mid-2027 3 sources. Crucially, however, no specific gigawatt figure, GPU count, power capacity, or data-center location was disclosed 3 sources.

To translate these PR statements into gigawatts, we must look at SSI's financial and physical constraints. A pre-deal baseline of tens of megawatts (20–70 MW) is highly plausible given the lab's initial funding, small headcount, and lack of inference revenue. A genuine 10x increase by mid-2027 therefore implies roughly 0.2–0.6 GW of dedicated capacity. Because leased AI capacity costs on the order of $10–15 million per MW per year, SSI's roughly $8B lifetime capital stack would be severely strained to sustain a 1 GW footprint without massive, ongoing vendor financing or an unprecedented new funding round 2 sources.

The median estimate of ~0.67 GW reflects a massive scale-up from SSI's mid-2026 baseline while accounting for an additional 18 months of growth beyond the mid-2027 "10x" target. Assuming SSI releases a successful initial model, it is well-positioned to raise additional mega-rounds at higher valuations, enabling further expansion. However, keeping the median below the 1 GW threshold acknowledges the steep financial and logistical hurdles—such as Vera Rubin supply constraints and data-center power delays—facing a lab without the immense commercial revenue streams of an OpenAI or Anthropic.

The wide distribution (ranging from a P10 of 0.16 GW to a P90 of 2.33 GW) is driven by two primary uncertainties. On the downside, the strict resolution criteria require capacity to be "owned or contracted for exclusive use." Since much of SSI’s compute relies on fungible cloud rentals (Google TPUs and NVIDIA-brokered capacity), it may not strictly qualify as exclusive, or SSI’s deep secrecy may simply prevent accurate public reporting by end-2028 2 sources. Conversely, the upper tail accounts for the possibility that the NVIDIA partnership operates similarly to the 1 GW-class Vera Rubin commitments seen elsewhere in the industry 2 sources, or that SSI secures an exclusive, multi-gigawatt hyperscaler allocation following a major research breakthrough.