Grid and Compute @GridandCompute

A 100% renewable PPA doesn't power a 1 GW AI cluster at 3 AM. Hyperscalers are quietly shifting to hybrid power underwriting: • Solar/Wind for green accounting (VPPAs) • Co-located gas/diesel for 24/7 firming Speed-to-power requires baseload reality, not just ESG credits.

Read this week’s 3-minute executive brief and subscribe to Grid & Compute here: gridandcompute.substack.com/p/capital-stru…

Edition #4 is live: Capital Structure as a Moat in Physical Compute. Software margins no longer protect AI infra when power and hardware CAPEX dominate. We break down how top teams de-risk technical milestones to transition from VC equity to project debt. Link below 👇

The ultimate moat in AI infrastructure is capital compounding. As physical bottlenecks stall legacy deployments, the teams that master capital structure will outperform those with superior models alone. Edition #4 drops Sunday: 'Capital Structure as a Moat in Physical Compute.'

Regulatory walls in Tier-1 hubs like Northern Virginia are cost-of-capital multipliers. As NIMBYism and grid queues stall mega-campuses, developers who pivot to distributed 20-30MW inference nodes de-risk their interconnection timelines and unlock cheaper regional project debt.

Morgan Stanley CapEx per GW by chip architecture: • Rubin Ultra: $50B/GW • Vera Rubin: $49B/GW • GB300: $39B/GW • TPUv7: $27B/GW • Trainium3: $21B/GW Custom ASICs cut deployment CapEx by ~50% vs Nvidia flagship racks. Rack BoMs + non-rack power costs dictate chip choice.

AWS is dropping $21B across 34 data centers in MS (~2.2 GW) for Anthropic (via Measured AI). Underwriting mechanics: SB 2001 bypassed PSC review, Entergy ratepayers fronting rate riders, $5B local tax rev offsets state loans by '32 Speed-to-power is statutory fast-tracking.

The shift to 800V DC power architectures isn't just an engineering flex—it's a financial de-risking move. Standardizing on OCP blueprints and solid-state transformers removes equipment risk, making 1MW+ rack deployments underwriteable for debt and infrastructure capital.

@Polymarket This is the productivity yield justifying the $100B+ AI CapEx spend. Absorbing a 175% ticket surge without adding 200 support hires isn't just opex efficiency—it proves inference unit economics are scaling fast enough to underwrite massive power demand.

@flatkey101 Gas is effectively acting as the interim interconnect. The key question now is capital structure: how much of this mobile/temporary gas capacity gets refinanced into long-term project debt once grid upgrades actually catch up in 2028-2030?

Cleanview data: 56 GW (30% of US data center pipeline) is going off-grid to dodge 7-yr grid queues. Gas dominates capacity. When 1 GW yields $10-12B, thermal efficiency takes a back seat to speed-to-power. Off-grid gas is the new moat.

Anthropic's $65B+ run rate relies on 15 GW across 15 campuses, but 70% of its leased pipeline sits behind unbuilt grid/power infrastructure. From 20-yr smelter leases to 90-day SpaceX compute, paper capacity vs energized compute is the ultimate underwriteable risk.

@HopfJames The real signal here is the capital structure: an EPC giant like Samsung C&T taking $70M direct equity skin-in-the-game on a FOAK nuclear reactor. Bringing in Tier-1 engineering to de-risk balance-of-plant execution is how SMRs actually make it to 2030 energization.

@HopfJames This is the exact template for hyperscaler load addition: Google funding EPUs at Vogtle & Hatch creates ~96 MW of net-new firm power without cannibalizing existing generation. Using the NU-1 tariff to deliver $900M in rate relief eliminates the primary regulatory bottleneck.

Hyperscalers scaling CAPEX to $1T+ proves that computing is now an asset-heavy asset class. The real winning playbook isn't just taking on massive dilution—it's using equity to de-risk physical assets until project finance and debt can fund the remaining infrastructure expansion.

@PolymarketMoney Google funding nuclear expansion at Vogtle & Hatch is the blueprint for 24/7 firm power. Absorbing capital costs for off-take while generating $900M in local rate relief solves the biggest public backlash against data centers: local electricity price spikes.

@PolymarketMoney Qualcomm partnering with Lumentum & Corning on 10 Tb/s co-packaged optics hits Broadcom & NVIDIA right at the interconnect moat. Pushing optical bandwidth directly to chip-to-chip interfaces proves networking speed and power efficiency are the core bottlenecks of scale.

@Polymarket Macro rhetoric meets physical reality: calling AI the new Industrial Revolution is right on scale, but beating China isn't a software race—it's a grid race. China added 1,200 GW of capacity while US sites sit in 4-year queues. Power supply is the real geopolitical bottleneck.

@PolymarketMoney Huge move: Texas freezing ERCOT interconnects for 474 GW in backlog hits neoclouds hard. When 90% of requests are data centers, 'paper capacity' gets audited fast. If you can't bring self-generation or off-grid power, your energization date just blew out by years.

Eight years ago, Bloom Energy rang the opening bell at the New York Stock Exchange. Today, we join the S&P 500. This milestone belongs to the employees, customers, and partners who have helped Bloom become the company it is today. We’re just getting started.