Ep. 018 - Stop Saying Half of 2026 US Datacenter Capacity Is Canceled (Datacenter, Energy) | Jeremie Eliahou Ontiveros, Reyk Knuhtsen, Ellie Holbrook, Jordan Nanos

Summary

The hosts push back hard on the Bloomberg-linked narrative that half of 2026 U.S. data center capacity has been canceled, arguing the underlying report mixes announced projects with executable ones and ignores hyperscaler scale. They contrast the headline number with concrete examples like Amazon’s 2025 buildout and expected 2026 additions, plus CoreWeave and other signed demand. The second half focuses on behind-the-meter power as a practical response to grid constraints, with Oracle’s New Mexico Project Jupiter used as a case study for permitting, pipeline, and local opposition bottlenecks. The team also lays out a power-supply thesis: gas remains the near-term answer, solar and batteries are constrained by interconnection and land, nuclear is too slow, and OEMs are racing to expand turbine and fuel-cell supply as data center demand keeps climbing.

Notes

• The episode opens by framing two recent SemiAnalysis pieces: the rebuttal to the 'half of 2026 capacity is canceled' story and the behind-the-meter power forecast to 40 GW by 2028.
• The guests say the viral cancellation headline came from a single report that others repeated without checking the underlying methodology.
• Their core critique is that scheduled capacity is being treated as if it were already executable capacity.
• Reyk argues the mismatch is obvious from hyperscaler disclosures, especially Amazon’s build pace.
• The team treats Amazon and other hyperscalers as evidence that the denominator in the cancellation story is broken.
• They emphasize the scale of requested load, saying U.S. operators have asked for over a terawatt of data center power.
• That requested load already exceeds the current U.S. system peak load, underscoring why power supply is now the gating variable.
• They argue the market boom began in late 2023, when early developers realized power scarcity was coming and started signing deals.
• A central distinction is between early-stage projects and projects that are actually under construction.
• The team says many of the 'delays' are simply projects that were never realistic in the first place.
• Ellie uses Oracle’s Project Jupiter in New Mexico as a concrete example of local and regulatory resistance.
• Project Jupiter is described as a behind-the-meter site that needs new gas pipeline infrastructure.
• The routing problem matters because the pipeline itself is not yet approved.
• Ellie notes that trucking CNG or LNG is a fallback, but only at limited scale.
• She says the infrastructure challenge in New Mexico is especially unfavorable for gas delivery projects.
• The site reportedly moved away from turbines and toward Bloom fuel cells.
• The team uses Project Jupiter to illustrate how permitting, route approval, and local comments can derail timelines.
• Reyk says the bottom-up forecast for behind-the-meter capacity is over 40 GW of net additions by 2028.
• He cites 3 GW of behind-the-meter deals signed in just two weeks as evidence the forecast is not crazy.
• Reyk argues behind-the-meter projects face lower delay risk than utility-tied projects because utilities are not bound to informal schedules.
• He says utility promises routinely slip by years, turning handshake timelines into moving targets.
• He points to Colossus 1 and 2 as proof that large behind-the-meter builds can be executed at scale.
• The data center power thesis is not just about avoiding delays; it is also about having many more build options than before.
• Reyk says grid-connected solar’s nameplate numbers overstate actual capacity value because of intermittency and ELCC derating.
• He says the real bottleneck is whether solar and batteries can be built on-site for data centers instead of waiting on the grid.
• He adds that utility-scale solar can be land-intensive, making siting itself a constraint.
• Reyk’s near-term view is that nuclear and SMRs are too slow for the current AI power cycle.
• He expects the 2020s to be dominated by gas-fired power for data centers, with more diversification later.
• Ellie suggests carbon capture could be co-located with BTM sites and data centers.
• Jordan describes a possible 'peak turbine' moment in 2026, when overbuying leaves some turbines underutilized.
• He says some turbines may spill into a secondary market if projects cannot execute quickly enough.
• The episode ends by speculating that a fully robot-built data center could arrive by 2029 or 2030 at the earliest.
• The robotics aside highlights labor intensity and supply-chain limits, including actuators and permanent magnets.