In 24 months, AI server racks go from 120 kW to 600 kW. The electrical supply chain — transformers, power semiconductors, cooling systems, substations — wasn't built for this.
This is the most concrete “the constraint is physics, not software” piece I’ve read, and it’s exactly why I think LatAm’s opening is real: when power infrastructure value per rack jumps 10×, transformers carry 4-year lead times, and the front end depends on grain-oriented electrical steel and high-purity copper, the jurisdictions with clean firm power and strategic minerals stop being peripheral and become part of the stack. Silicon Carbide rectifiers and liquid cooling are the new gating items, and Brazil sits on several of the upstream inputs. I made the case for the region capturing the infrastructure layer here: https://thiagopedicosaragiotto.substack.com/p/the-physical-layer-of-ai-why-latam
thank you for the note! I appreciate the shares. I hear a lot of good things about Brazil's electrical ecosystem -- some founders I am talking to have deep roots there. Happy to chat about that.
Thank you, Chris — really appreciate you taking the time to reply.
I was M&A manager at one of the largest electricity companies here for six years (doubled revenue through strategic acquisitions), and now advising multinationals and VCs.
Happy to share on-the-ground insights. Would Substack DM or email work for a quick chat?
Spot on — and the closer nails it: "selling into a confirmed equipment cycle," not speculating on demand. Physics over narrative.
Where I'd push one level further: the bottleneck isn't the edge — its durability is. Spotting a tight supply chain is easy right now; the capital-cycle question is how long each one survives incoming capital. By that test, these layers split hard.
Transformers/grid is the durable one — 15 years of flat capacity, 3–5 years to build a plant, inputs sourced abroad. Capital can't respond fast enough; the shortage is the moat. SiC is the opposite: Chinese substrate expansion, EV overcapacity spilling in, Wolfspeed reorganized rather than retired. Same tightness today, but a fast supply response — usually where pricing power erodes first. Tight now, looser in 18–24 months.
Detection vs durability. That's the layer I keep working through. Great piece.
Mostly the second one, I think — it commoditizes. On the "new materials replace it" side I'd actually push back a little: in your stack SiC and GaN aren't fighting for the same job — SiC handles the high-voltage end, GaN the high-frequency end. And nothing's shown up yet that can take SiC's high-voltage spot; whatever eventually does is still years out in the lab.
So the pressure that matters now is commoditization, the one you named. The Chinese substrate push (SICC, TankeBlue, Sanan) keeps dropping the floor — SiC keeps the high-voltage job, but the margin in it shrinks each cycle.
The way I look at it: the value doesn't go away, it moves — to whatever stays hard to copy fast. The supplier locked into an Nvidia design (which holds for a generation, then resets), or whatever input stays genuinely scarce. So I wouldn't bet on the material itself — I'd watch where the bottleneck moves next.
This is the most concrete “the constraint is physics, not software” piece I’ve read, and it’s exactly why I think LatAm’s opening is real: when power infrastructure value per rack jumps 10×, transformers carry 4-year lead times, and the front end depends on grain-oriented electrical steel and high-purity copper, the jurisdictions with clean firm power and strategic minerals stop being peripheral and become part of the stack. Silicon Carbide rectifiers and liquid cooling are the new gating items, and Brazil sits on several of the upstream inputs. I made the case for the region capturing the infrastructure layer here: https://thiagopedicosaragiotto.substack.com/p/the-physical-layer-of-ai-why-latam
thank you for the note! I appreciate the shares. I hear a lot of good things about Brazil's electrical ecosystem -- some founders I am talking to have deep roots there. Happy to chat about that.
Thank you, Chris — really appreciate you taking the time to reply.
I was M&A manager at one of the largest electricity companies here for six years (doubled revenue through strategic acquisitions), and now advising multinationals and VCs.
Happy to share on-the-ground insights. Would Substack DM or email work for a quick chat?
Best,
Thiago
Spot on — and the closer nails it: "selling into a confirmed equipment cycle," not speculating on demand. Physics over narrative.
Where I'd push one level further: the bottleneck isn't the edge — its durability is. Spotting a tight supply chain is easy right now; the capital-cycle question is how long each one survives incoming capital. By that test, these layers split hard.
Transformers/grid is the durable one — 15 years of flat capacity, 3–5 years to build a plant, inputs sourced abroad. Capital can't respond fast enough; the shortage is the moat. SiC is the opposite: Chinese substrate expansion, EV overcapacity spilling in, Wolfspeed reorganized rather than retired. Same tightness today, but a fast supply response — usually where pricing power erodes first. Tight now, looser in 18–24 months.
Detection vs durability. That's the layer I keep working through. Great piece.
thank you so much. Excellent comment. Do you think new materials will replace SiC in parts of the stack? Or will this become more commodity like?
Mostly the second one, I think — it commoditizes. On the "new materials replace it" side I'd actually push back a little: in your stack SiC and GaN aren't fighting for the same job — SiC handles the high-voltage end, GaN the high-frequency end. And nothing's shown up yet that can take SiC's high-voltage spot; whatever eventually does is still years out in the lab.
So the pressure that matters now is commoditization, the one you named. The Chinese substrate push (SICC, TankeBlue, Sanan) keeps dropping the floor — SiC keeps the high-voltage job, but the margin in it shrinks each cycle.
The way I look at it: the value doesn't go away, it moves — to whatever stays hard to copy fast. The supplier locked into an Nvidia design (which holds for a generation, then resets), or whatever input stays genuinely scarce. So I wouldn't bet on the material itself — I'd watch where the bottleneck moves next.
Great insights.
Coincidentally am looking at startups that are using a new material to replace SiC (Gallium Oxide). What do you think?