Investment Memo

A full investment memo on Bloom Energy: how its solid-oxide fuel cells turn gas into always-on onsite power, why data centers care, what Q1 2026 earnings really showed, how Brookfield financing changes the accounting, where turbines and reciprocating engines still win, and what BE must earn to justify a $69B market cap.

A grouped investment thesis on the AI memory hierarchy: HBM, server DRAM, LPDDR, flash, networking, KV-cache offload, and custom HBM. The memo separates near-term scarcity from the longer-term cHBM option, ranks the likely public-market winners, and defines the cycle risks and exit signals.

Vertiv is one of the cleanest public-market ways to underwrite AI data-center power, thermal management, and mission-critical infrastructure. Fundamentals are exceptional, but at roughly 50x FY2026 guided adjusted EPS midpoint, 9.0x sales, and 56x adjusted FCF, the base-case fair value is only around $340/share versus the recent $315.71 quote.

A critical data-center memo on Leopold Aschenbrenner's Situational Awareness 13F sequence and the post-quarter Nebius 13G: why the trade moved from chips to power, delivered clusters, goodput and AI-cloud unit economics. Compares CoreWeave, Nebius and IREN across live clusters, owned power, financing risk, customer quality and what is already priced in.

A deep dive on Meta's valuation, business structure and AI plays after Q1 2026: Family of Apps remains an elite ad cash cow, Reality Labs is still a loss-making option, and the $125-145B 2026 capex guide is the swing factor. Includes scenario valuation, AI revenue paths, infrastructure read-throughs and public references from Jefferies, Evercore, JPMorgan and Damodaran.

60–80% of tech equity value sits in terminal value, the assumption that moats compound. AI breaks that assumption for software first. 2025 inning: long silicon (AVGO, SKH), short SaaS (Duolingo −67%, Adobe −25%). What survives, what dies, and what it means for growth equity, private credit, and $300–500B/yr AI capex.

Samsung is printing ₩57T/quarter in memory OP (8x YoY). Three structural forces extend the DRAM cycle: AI inference DDR5 demand, NVIDIA SOCAMM/LPDDR5X pull (13x jump GB200→VR300), and HBM4 redemption after passing NVIDIA qual. At 2.2x annualized OP, the Korea discount is extreme.

XLI is a concentrated bet on three multi-year spending cycles: defense rearmament post-Hormuz, $2T+ infrastructure legislation entering peak disbursement, and AI-driven electrification. At 27x forward P/E it's not cheap, but the structural tailwinds justify scaling in via cash-secured puts.

Tesla is priced as a declining EV company. The market ignores autonomy (robotaxi live in Austin, Cybercab production started), robotics (Optimus Gen 3 summer 2026), and the SpaceX $1.75T IPO catalyst that could force a $3T+ merger. Sum-of-parts: $285B bear to $1.5T+ bull.

Palantir's Ontology creates infinite switching costs. AIP Bootcamps convert at 75% in 5 days. Maven is now a Program of Record. But at 60-80x revenue, you're paying for perfection. Best played as a put seller.

NVIDIA's $20B Groq buy marks the training-to-inference pivot. 92% AI training share, 18-year CUDA moat. $255B inference market by 2030.

Google trades 29% below peers. They invented transformers. TPUs save $3B/yr. GCP growing 44% CAGR. Cash-secured puts yield 40% annualized.

ASML owns 100% of EUV lithography. Every advanced chip runs through their machines. AI supercycle drives demand. High-NA EUV is next.

AI's real bottleneck isn't compute. It's memory. Three companies control all HBM supply. $4B market growing to $91B by 2030 at 56% CAGR. 118% utilization. Seller's market. MU is the play.

GPUs aren't the bottleneck. Power is. ERCOT has 78 GW requested, 1.1 GW approved. That's a 70:1 ratio. Bring-your-own-grid creates a $5.2B speed arbitrage over 36 months.

Unitree shipped 10,000 robots at $16K each. 10x cheaper than Boston Dynamics. $145B TAM by 2030. China controls the rare earth supply chain. The West has no answer.