At the Sohn Montreal conference on June 4, KKR's Brandon Freiman — head of North American infrastructure at one of the world's largest infrastructure investors — told Bloomberg that AI data center demand is creating a natural gas "renaissance" after decades of flat U.S. plant development. The projection behind that call: 148 gigawatts of additional data center power capacity needed by end of decade, against roughly 42 gigawatts consumed globally today. Bloom Energy reported Q1 2026 revenue of $751 million, up 130% year-over-year. Bergen Engines signed 750 megawatts of on-site gas generation to Crusoe's AI datacenter sites on June 3. This is not a forecast — the on-site power buildout is already being contracted and funded.
The investment implication is structural in a way the GPU shortage is not. GPU shortages can be optimized around: better batching, quantization, inference efficiency gains. Power shortages cannot be compressed. Utility grid interconnects take 3-to-7 years from application to energization. Behind-the-meter solutions — on-site gas, battery, nuclear — are the only path to acceleration, and they require capital commitments made years before the compute comes online. For infrastructure investors, contracted power capacity is now a scarcer and more defensible asset than GPU reservation agreements. The companies that locked in power — not just chips — in 2024 and 2025 are setting the AI capacity ceiling for 2027 and beyond.
The Department of Commerce finalized a $150 million CHIPS Act award to xLight on June 2, funding the world's first free-electron laser (FEL) EUV prototype at Albany NanoTech Complex. xLight's FEL approach delivers EUV light to up to 16 scanners simultaneously — against ASML's laser-produced plasma one-source-per-scanner constraint — with a prototype targeted for 2028. Former Intel CEO Pat Gelsinger is a backer. The EUV light source is the last critical link in the leading-edge supply chain not controlled by a U.S. or allied entity; this is the first federal commitment to closing that gap domestically.
Bergen Engines and Crusoe signed a ~750 MW power agreement for on-site gas generation across Crusoe's U.S. AI datacenter sites — 438 MW under contract plus 310 MW letter of intent — phased through 2027. With 45 GW in its power pipeline and a $1.375 billion Series E at a $10 billion+ valuation, Crusoe has positioned itself as a power infrastructure company first and a GPU cloud operator second. Grid-constrained hyperscalers competing for utility interconnects on 3-to-7 year timelines cannot replicate contracted behind-the-meter capacity on short notice; the Bergen agreement is evidence of the structural moat this creates.
Q2 AI revenue of $10.8 billion grew 143% year-over-year and Q3 AI chip guidance of $16 billion was still up sharply — but guidance came in below the $17.2 billion analyst consensus, and management reiterated rather than raised the full-year target. In a market where every guidance raise has been rewarded, a hold was treated as a miss. The sector rotation was broad: ARM fell 6.2%, AMD fell 5.4%, Micron fell 7.7% — while power infrastructure plays moved in the opposite direction, GE Vernova rising 0.7% as investors rotated from chip-cycle exposure toward the longer-duration buildout.
GE Vernova's electrification segment booked $2.4 billion in AI data center equipment orders in Q1 2026 alone, already exceeding its full-year 2025 total. Total backlog reached $163 billion, up 71% year over year. Gas turbine slots are reserved through 2026 across major data center operators. The one-day counter-move against a semiconductor selloff matters less than what it implies: when chip guidance misses, capital is rotating not toward software but toward the physical power infrastructure that semiconductors require.
SemiAnalysis publishes their space datacenter cost model. Timed to SpaceX's IPO roadshow, SemiAnalysis published To Boldly Go: The Case for Space Datacenters — a cost model for orbital compute economics. The core finding: space-to-earth compute cost parity is achievable in the late 2030s under aggressive launch cost reduction scenarios, but requires solving thermal regulation and orbital debris management at scale. Some light reading before the expected June 12 SpaceX IPO.
Europe's AI datacenter ambitions are stalling. Bloomberg reported June 2 that the EU's flagship AI datacenter initiative is stumbling over funding shortfalls and timeline delays — a sharp contrast to the pace of U.S. federal commitments (CHIPS Act, xLight, TSMC Arizona) and private capital deployment (Alphabet's $80 billion equity raise, Crusoe's $1.375 billion Series E). The divergence compounds: power permitting in Europe runs 5-to-10 years versus 3-to-7 in the U.S., and European sovereign cloud mandates add further friction for hyperscaler buildout. Where advanced compute capacity concentrates geographically over the next decade is being decided now, in the gap between who can permit and fund and who cannot.
Computex 2026 closes in Taipei. Final-day announcements from AMD, MediaTek, and Taiwanese supply chain partners. Watch for any closing remarks on CoWoS expansion timelines and HBM4E packaging ramp.
SuperAI 2026 conference in Singapore. AI infrastructure panels featuring Weka, Crusoe, and others from the compute stack. Likely venue for announcement activity from Asian hyperscalers and sovereign AI programs.
SpaceX IPO pricing. SPCX expected to price at $135/share, raising $75 billion at a $1.75 trillion valuation. Morningstar has called the valuation nearly twice fair value. The pricing will be the first real test of whether institutional investors underwrite the AI infrastructure thesis at orbital scale.
SpaceX (SPCX) begins trading on Nasdaq. At target pricing, the largest IPO in history by both raise size and opening market cap — and the most direct public market bet on AI compute infrastructure as a multi-decade physical asset class.