🌿 Friday's Climate Infra Brief: Inside the Park Home Run
This news caught my eye: In Feb, TeraWulf paid $200MM for a shuttered aluminum plant in Hawesville, Kentucky. Just for the grid connection sitting on the property. In July it signed a 20-year lease with Anthropic on the site worth roughly $19B in contracted revenue, for about 401 MW of critical IT load. This is a great developer play, an inside-the-park home run. I’d love to finance this deal. Please call me if you have a similar one.
A couple of observations:
Hybrid power and conditional grid service is the new normal. The old days that you just connect to the grid and pay someone else to figure out 24/7 power are long gone for large load. Firm grid service has become a limited edition product that the pre-orders are extremely long. So regulators are stepping in and invented a compromise product. The weird thing is, they all independently invented the same one. Look at the Batch Zero framework the PUCT approved for ERCOT in June. A large load gets a firm number it can draw no matter what. If they want more, they have two options: agree to let the grid cut their power when things get tight, or build their own generation to cover the gap. FERC approved SPP’s conditional high impact large load service that same month. It gives you up to 7 years of non-firm transmission while you wait for the real thing, requires you to have backup generation, cuts you off if that backup fails, and bans you from selling demand response. FERC also dropped regulatory orders in June pushing non-firm contract demand across every RTO it regulates.
Read all these together, the pattern is pretty clear. The new norm is conditional grid service, plus your own generators/power (BYOP), plus batteries or something else to absorb the curtailments. Hybrid power is here to stay. You run islanded prime power on day one, and then when the grid connection finally comes through, you switch to grid-parallel mode for resilience and flexibility. The supply chain is already reacting. GE Vernova was quoting roughly 3-year lead times for turbines. Bloom Energy is scaling up to 2 GW a year of fuel cell capacity by the end of 2026. They reportedly took an Oracle order up to 2.85 GW, and just signed a partnership with Brookfield. Bloom’s own survey says the share of data centers using onsite generation as primary power will jump from 13% to 38% by 2030.
Capital isn't flowing into grid-enhancing tech fast enough, because it sits outside the data center's fence. The grid has four layers: generation, storage, transmission, and load. Everyone knows tech companies will pay pretty much anything for power, but that money doesn’t reach all four.
Generation, storage, and load control all have data centers as customers, and we’re watching assets get repriced and companies attract capital at a pace the sectors have never seen. Transmission tech is selling to utilities that earn a regulated return on capital deployed, asking them to buy equipment designed to avoid deploying capital.
Which is a shame, because it works (shout out to dynamic line rating tech, grid optimization, advanced conductor materials). Great River Energy lifted average line capacity 25% with dynamic line rating, and 42.8% on one line. National Grid’s DLR saved £21 million in constraint costs last regulatory period. Salt River Project got 80% more capacity out of an 8.5-mile Phoenix line without touching a single 1970s tower, 12% under budget.
My guess is the fix comes through the tariff, not the utility: give the large load a path to pay for the upgrade that shortens its own queue. FERC’s June show cause orders push in that direction, requiring RTOs to justify why their large-load studies don’t evaluate alternative transmission technologies. SPP’s answer is due August 17. Please, let them pay. It will make the grid a better place :).
Compute is starting to look like a network rather than a set of campuses. Deloitte estimates inference will make up roughly 2/3 of AI compute in 2026, with inference-focused cloud infrastructure spend passing training for the first time at around $20.6B. JLL expects inference to overtake training as the dominant requirement around 2027. Inference means serving users, which requires geographic distribution, pushing servers out to regional edges and embedded systems. Build.inc breaks this down nicely: the metro edge is 1 to 5 MW, the regional edge is 5 to 20 MW, and they argue that regional tier is where the real institutional development money is going to flow. The size matters also for reasons beyond latency. A 15 MW site is something a county will permit and onsite generation can comfortably serve. A 500 MW campus is neither.
We’re already seeing pilots of this. Nvidia and EPRI are pushing a fleet of micro data centers designed to sit directly next to substations, with construction targeted by the end of this year. They estimate workloads will only need to be relocated about 0.1% of the time.
Taking over old brownfield industrial sites is the fastest way to build that 5-20 MW distributed network, because what you’re really acquiring is a load position. Which brings us right back to TeraWulf in Hawesville.
So who actually finances everything inside the fence? The tech buyer absolutely needs a hybrid setup. They can buy everything inside the fence, and definitely do not want to own energy infrastructure, as they do not want to be in the energy business. That creates a massive 3rd-party ownership opportunity for infra investors.
PS: If you enjoyed this, share it with a colleague or friend who might too. And if you want to chat, I’m at meng@fridaymorninglabs.com
Sources: TeraWulf and Anthropic disclosures and trade coverage (Feb and July 2026); ERCOT Batch Zero framework and PUCT approval (June 2026); FERC order approving SPP CHILLS, Docket ER26-1323 (June 5, 2026); FERC show cause orders to six RTOs (June 18, 2026) and Akin analysis of the SPP order; Bloom Energy 2025 and 2026 Data Center Power Reports; GE Vernova Q1 2026 earnings commentary via Tech Fund; Microgrid Knowledge and Data Center Dynamics on bridge power; DCD on Amazon fuel cell siting; Deloitte TMT Predictions 2026; JLL 2026 Global Data Center Outlook; Build.inc on edge development and brownfield conversion; IEEE Spectrum on Nvidia/EPRI distributed inference; National Grid RIIO-T2 DLR reporting; Watt Coalition / AMP Coalition / Grid Strategies analysis of California advanced transmission filings; company disclosures (LineVision, Heimdall Power).