Research on power, markets and the deals connecting them.
Trader and macro researcher
London, UK
The feed
Short and unfiltered. A tape read, a print that did not make sense, something worth writing down before I forget it.
Triton Partners is looking to sell Trench Group, or take it public. Morgan Stanley is running a dual track, with an exit as early as next year.
Trench makes bushings, instrument transformers and coil products, the parts a grid cannot be built without. Triton carved it out of Siemens Energy in 2023, and revenue was €906m in 2025, up a third on the year.
Any fund that wants to own the bottleneck of the AI buildout should be looking at this.
When supply is capped, extra demand can't turn into extra volume, not in the near term. It turns into price. Price turns into Fed hikes, and hikes turn into a higher real discount rate.
So the upside is capped by capacity, and the discount-rate hit isn't.
Siemens Energy, June quarter. On gas it booked 15GW and shipped 6GW, a 2.5x book-to-bill, with lead times now three years or more.
The grid division is sitting on a €51bn backlog, and plans to add about 50% to its transformer and switchgear capacity by 2030.
Things to watch: GE Vernova's third quarter in late October, then Siemens Energy's full year in November.
Is anyone actually talking about the contractors doing the building for this AI buildout and grid enhancement?
The binding constraint on AI buildout is power delivery, not chips
CoreWeave borrows through separate SPVs, each holding chips plus one customer contract. DDTL 4.0, $8.5B in March 2026, priced at SOFR plus 225bps and rated A3 because it sits on a $14.2B Meta contract, so lenders are really underwriting Meta. DDTL 5.0, $3.1B in May, SOFR plus 450bps, rated Ba2, on two non-investment-grade customers. DDTL 5.5, $2.6B in August, SOFR plus 550bps, where the contracts average three years but the debt runs five, so lenders are funding two years of hardware with no tenant.
Lol fed rate hikes
Future research:
Hardware doesn't make money
Compute: chip fab → HBM and packaging → GPU owners (hyperscalers, neoclouds) → rental price indices (Silicon Data, Ornn, COIL) → exchanges (CME, ICE) → buyers (labs)
A GPU loses roughly half its resale value within three years, and rental rates for older chips are down 50 to 70%. A chip bought in 2025 and installed in 2027 has two of its best earning years gone before it earns a dollar.
US data centre electricity demand is expected to go from about 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027, with Goldman forecasting 95 GW by 2030. But of roughly 16 GW slated for 2026 delivery, only about 5 GW is under construction.