Own the bottlenecks, not the apps. Finding moats in compute, photonics, memory, energy, grid, digital scarcity. @joinautopilot portfolio managed by @haztan
Los Angeles, CA
Joined May 2026
- Tweets570
- Following39
- Followers69
- Likes36
Pinned Tweet
Underneath every AI app is a supply chain that can't expand fast enough: compute, HBM, packaging, photonics, EUV, power, cooling, grid.
We invest in these bottlenecks and identified four clear sleeves:
• AI Semiconductors 57%
• Energy 17%
• Grid & Thermal 14%
• Digital Scarcity 12%
Full strategy and live portfolio on @joinautopilot ⬇️
marketplace.joinautopilot.co…
Hitachi Energy is putting C$479M (US$330M+) into a plant for transformer pressboard, the insulation inside large power transformers. Canada has no pressboard line today. Its own exec calls insulation one of the most critical bottlenecks in transformer making. Target: 2028.
Everyone counts transformer orders. The squeeze can sit one layer down, in a material most people never name. This line lands in 2028. Hitachi's new Mississippi transformer plant starts in 2029. Scarcity clears in factories, not headlines. $HTHIY
ASML says its planned 2027 and 2028 output won't fully meet customer demand, even after a 30% capacity add in 2027. Some customers already want to place orders for 2029 and 2030.
The scarce asset in AI isn't the model. It's the only EUV supply chain on earth. $ASML
Bloom just ordered $138.5M of 38 kV switchgear and distribution transformers from LS Electric for a Pennsylvania fuel-cell data center. Ship window: April to July 2027.
That's the quiet half of on-site power. Fuel cells make MW. Switchgear and transformers make them usable. $BE
TSMC handed part of CoWoS to GlobalFoundries. $2B, five years, silicon interposers out of Malta, NY. First US source.
The interposer is the silicon floor the GPU and HBM sit on. Bigger accelerators need bigger floors. That area is the scarce asset, not the logo on top. $GFS
The catch is the clock. Volume ramp is H1 2028. Until then the floor, and the deep trench capacitors built into it, still come from offshore.
Packaging capacity now gets contracted like fab capacity. Watch who signs the next interposer and substrate deals. $TSM
The day we flagged PJM's ~6.8 GW firm-capacity hole, DOE offered Vistra a conditional $4.2B loan to uprate its PJM nuclear fleet: +433 MW, ~4 GW kept running 20 more years.
The cheapest firm MW is the one already built.
$VST
Micron's Taoyuan union authorized a strike 1,994 to 14.
Taiwan is its critical DRAM/HBM hub. Memory is sold tight through 2027. A surprise walkout can be called by text, so the fab can't prep.
The scarce asset isn't just wafers. It's the people running them.
$MU
The quiet bottleneck inside an AI rack is voltage, not watts.
A 1 MW rack on a 54V bus pulls roughly 18,500 amps. At 800V DC it's about 1,250. Same power, far less copper and heat.
That's why Monday's news was all power conversion: Navitas and Microchip, Renesas, Infineon.
The AI power bottleneck just hit a regulatory wall.
PJM, the largest US grid and home to Data Center Alley, scrapped its emergency reliability backstop auction one day after FERC called the plan deeply flawed. The ~6.8 GW hole remains. No new bid window.
The AI power bottleneck moved upstream of the plug.
Major gas turbine OEMs report slots sold out through 2028, with books into 2030-31. Roughly 20% of GE Vernova's gas power backlog is now data-center and AI work.
AI training doesn't just consume megawatts. It beats transformers to death.
GPU clusters swing load by tens of MW in minutes. Documented ramps hit ~1.9x base power in <250 ms. That thermal cycling ages insulation 2-3x faster than seasonal loads they were designed for.