The On-Package Delivery Stack: Technical Companion
IVR magnetics, thin-film substrates, high-density capacitors and the qualification-gated architecture of on-package power delivery.
IVR magnetics, thin-film substrates, high-density capacitors and the qualification-gated architecture of on-package power delivery.
How 48V becomes 0.8V at the chip die: multi-phase POL, IVR, on-package power delivery, and why the foundries hold the strategic value pool.
Where the market essay named the vendors and quantified the lead times, this piece explains the physics. Why voltage steps up, how a 400-tonne transformer moves 300 MW at 99% efficiency, why SF6 became standard and why it is being phased out.
Grid interconnect is the binding constraint on AI. Transformer lead times doubled since 2023. Hyperscalers now build own generation to escape 4-7 year utility queues.
Where the market essay named the vendors, this piece explains the physics. What a farad actually is, why HSCs beat batteries on 10-second pulses, why MLCCs derate under DC bias, and why a fifty-cent film cap decides whether an 800VDC rack survives its own noise.
Where the market essay named the vendors, this piece explains the physics. What 1,000 W/cm2 at the die actually means, how a hundred-dollar coupling either saves or destroys a rack, and where cooling is going next.
Where the market essay named the players, this piece explains the physics. What wide bandgap actually means, why SiC and GaN outperform silicon at the same voltage, and how a boule of vapour-grown crystal becomes a switch that runs at 800V and 200kHz. Cross-sections, process flows, roadmaps.
Every watt delivered to a GPU comes back as heat, and at 600kW per rack there is nowhere for it to go but water. Layer-by-layer analysis from die to dry cooler.
SiC and GaN are the enabling materials of 800V data centres. Boom, 60% price collapse, the Wolfspeed Chapter 11, and what comes next. Crystal boule to gate driver.
Four capacitor categories are the binding AI supply constraint: HSCs, EDLCs, MLCCs, silicon caps. Layer-by-layer on pricing, capacity, and the qualified vendor list.
The AI data centre power train is six stacks, not one. Where each stage lives, who builds it, what is standardised. Reference map for architects speccing new AI campuses.
The trade press writes about 800 VDC as one architecture. There are two, they behave differently in retrofit, and hyperscalers are choosing between them right now.