SasoEquity
AI & Robotics Investor
Panasonic's rack-level backup architecture addresses the AI buildout's biggest near-term constraints, grid interconnection queues and power quality at 100kW-plus densities, an asymmetric play on power as AI's bottleneck. Hyperscalers can add protected kWh per rack instead of overbuilding central UPS, and peak shaving flattens spiky loads. Supercapacitor CBUs, targeted for FY27, handle fast transients for the 800VDC shift. Over 80% of FY29 is locked in design wins. x.com
Panasonic may be the cleanest under-the-radar AI data center power play, a dominant incumbent that benefits from rising rack power density. It holds about 80% of data center battery backup units, a moat built on years of qualification and 600M+ cells shipped without critical safety incidents. It is converting idle EV lines to triple Japan capacity at modest capex, has over 80% of FY29 revenue locked via design wins, targets ~4x revenue to ¥800 billion at 20%+ ROIC, and adds 800VDC products. x.com x.com
Per Truist, AI data center capacity grows from 24GW to 100GW by 2030, and every watt must be converted, stored and cooled before a GPU runs, paying firms between grid and chip. DC-DC conversion is the layer to watch, as rising density makes each efficiency point worth more: Monolithic Power, Vicor, Infineon, TI, Analog Devices, Renesas, onsemi. Panasonic appears in six layers and Delta in four, paid more than once per rack. Eaton, Vertiv, GE Vernova, Schneider and ABB bring power in. x.com
AI power-semiconductor suppliers stand to make generational money as AI racks outgrow data-center wiring. NVIDIA's 2027 rack pulls 600 kW, while wiring built before 2025 was designed for 40 kW, so 800V DC power sidecars are being bolted beside existing racks. Those need silicon carbide and gallium nitride chips, made at volume by only a handful of companies whose stocks fell 40-60% as EV demand slowed. Amazon, Microsoft, Google and Meta are now new customers that are not bargaining on price. … Show more
QuantumScape's solid-state batteries are positioned for high-density backup battery units in NVIDIA's 800 VDC power racks, which ship in the second half of 2026. Jensen Huang said the roadmap remains intact, with Navitas's GaN and SiC chips directly in the power conversion path. The power side of the AI factory trade just got another green light. x.com
A shortage of high-power cylindrical battery cells for data-center backup units is emerging, and cell makers are the main beneficiaries. Samsung SDI supplies cells to Simplo, which builds backup battery units for Meta and Amazon, and its output has risen sharply with data-center demand; industry sources say Panasonic and Samsung face supply shortages. Samsung SDI and Panasonic Energy look like the cleanest winners, with Murata and LG Energy Solution next. Not every shortage means a huge market. … Show more
Vertiv should be a net winner in power content per megawatt as AI racks move to 800V DC. The shift will cannibalize its legacy UPS and AC distribution, but it is building the replacements: centralized rectifiers, 800V busway, DC energy storage and rack-level DC/DC, while cooling remains a big opportunity either way. Nvidia targets 800V DC for MW-scale racks from 2027, though only 15-25% of data centers may adopt it by 2030, mostly as hybrids on an AC backbone, so AC isn't going away. x.com
Vertiv and Quanta Services get paid whether or not Nvidia's Rubin Ultra ships, because every rack upgrade is a power-infrastructure contract. As @KawzInvests explains, Rubin Ultra's 600-kilowatt racks cannot run on old wiring, so Nvidia moves power conversion to an 800V DC sidecar rack, raising semiconductor content per power unit from about $95 to $880. The rewiring itself is the trade. x.com
HVDC and 800V DC power delivery is the theme to focus on now, with STMicroelectronics, Navitas, Hyliion and Ideal Power as the preferred bets. Across the AI buildout, power and optics keep gaining narrative strength and attention, while packaging remains critical but is increasingly well understood by the market. x.com x.com
Powell Industries is the source-agnostic way to play AI power demand, because its custom heavy-duty switchgear is needed whatever generates the power and protects multi-million-dollar clusters from power surges. Per Bloomberg, AI-chip electricity demand could jump more than 1,100% by 2033, with the US taking 64%. Powell's $2.4B backlog, with visibility into 2028, shows data-center operators locking in gear early to avoid grid delays. x.com x.com
QuantumScape has a second growth leg the market is not pricing: battery-in-rack backup for AI data centers. Its non-combustible ceramic separator lets cells sit inside server racks, buffering millisecond power spikes faster than external UPS systems as racks pass 100 kW. In EVs, Honda signed a development deal after validating hardware. A hyperscaler design win would re-rate it as AI infrastructure. It has $904M of liquidity, but is pre-revenue and depends on scaling its Cobra process reliably. … Show more
Denser AI accelerators raise the value of efficient power delivery near processors, potentially making Navitas one of the most hidden AI winners in Nvidia's architecture. Its 24 kW board targets single-stage 800V-to-12V conversion at about 97% efficiency in Nvidia's 400 mm by 60 mm format. An 800V-to-6V design developed with Nvidia and Claros offer other paths into the power stack. This is a showcase, not a production order; shipment volumes, revenue and competition remain uncertain. x.com
