James McMillan III

Showing some hustle

My key takeaways of the $BE (Bloom Energy) earnings call. 1. Became the standard in nine months. All major U.S. hyperscalers plus over a dozen neo-clouds, AI labs, and colocation operators have validated and approved Bloom’s technology. KR Sridhar refused to break out active deployments vs. backlog vs. definitive agreements, but stressed the combination is already there. Context: nine months earlier they announced their first direct hyperscaler deal and said they wanted to become the standard the way they did in C&I (which took 10 years). “This is not a faster horse. This is a car. That is why this is happening, and this is not reversible.” 2. Capacity is explicitly not the constraint. They plan around the 30–40 GW of new AI data center capacity expected to come online in 2027 using a sophisticated algorithm on project timing. Units are fungible, once on a truck they can be redirected. “Capacity is not going to be our constraint as we see right now.” They will keep the promise on everything in the order book and the visible pipeline. 3. Hyperscaler diligence is extreme and strategic. Customers are not doing one-off transactions. They share confidential expansion plans under NDA and force Bloom to walk through committed capacity, scaling ability, and long-term partnership economics in detail. Validation only happens after that process. Same rigor (plus extra layers on long-term performance and availability) applies to financial partners. 4. Brookfield expansion to $25B total is a “financial shelf.” The additional $20B came after they watched the original $5B perform, spoke to 15-year customers, and stress-tested execution. Timing of drawdown depends on customer uptake, not a fixed window. Sridhar explicitly compared the speed of this expansion to how fast they became the AI standard, both faster than he would have predicted. 5. Project delays have contractual flexibility. MSAs + “Copy Exactly!” manufacturing allow equipment to be redeployed across sites. Financiers take title and need the same protections. Management does not comment on specific headline projects but stressed the structure protects them. 6. Time-to-power is existential math, not marketing. A 1 GW full-stack AI data center can generate $12–24B of revenue per year. Pulling power in a month (vs. multi-year grid/transmission delays) is worth $1–2B of high-margin revenue that would otherwise be lost. Inference demand will be even more distribution-constrained (urban sites where you cannot put a gas turbine). Bloom is positioned for both. 7. Competition is acknowledged but reframed. Near-term every technology that can deliver power quickly has a role because the shortage is so severe. Long-term, when a customer chooses, the decision is total cost of power-to-token, not LCOE. Bloom’s differentiators (800 V DC, reliability without overbuild, no NOx/SOx/water, permitability, ability to locate in cities) have no commercial equivalent today. On other fuel-cell technologies targeting the same market, Sridhar put current data-center share in the “very high 90s” and said he welcomes competition because it makes them hungrier. 8. Jevons paradox on token efficiency. Cheaper, more efficient Chinese or open-source models do not reduce power demand, they increase total token usage. “Whatever we are predicting on AI is an underestimate, not an overestimate.” 9. Service margins and customer stickiness are the quiet story. Service gross margin hit +22% this quarter (from –21% at IPO). 80% of 2025 orders were repeats from existing customers. Sridhar spent the last part of the call giving a rare, extended shout-out to the service team and framing happy customers + service economics as a major driver of enterprise value. Bottom line: The narrative they want is “standard for AI onsite power,” backed by fungible manufacturing, value-based (not LCOE) pricing, and a financing partner that just quintupled its commitment after watching execution.