Behind-the-meter is not free.
There are only a handful of operational U.S. data centers running off-grid or partially grid-connected right now, and several have already reported cracked turbines, forced diesel operation, and accelerated equipment failure. Schedule value has to exceed the higher all-in cost and reliability risk. Most models still underweight that trade-off.
1
32
ERCOT’s queue is sitting at roughly 474 GW. Record peak demand is a fraction of that.
When the ratio gets this distorted, the rational response is an audit and conditional classification, not more speculative study slots. Firm power path becomes the filter.
28
Replying to @Filecoin
Correct ranking. Capital is abundant relative to deliverable firm power and interconnect positions. Sites and operators that control a credible path to electrons on a defined timeline will clear the next cycle. The rest will wait.
1
49
Replying to @SemiAnalysis_
This is the correct governance response to a massively oversubscribed queue. When large-load requests reach multiples of system peak, the default of “first to file gets study slots” collapses. Conditional classification and audit of who is actually bringing capacity (and who is speculative) is how you separate real projects from queue pollution. Firm power path becomes the filter.
327
Replying to @MelvinInvests
The revenue-per-MW stack is real. The binding constraint is still whether the megawatts are actually deliverable on the required schedule.
A lot of the 38 GW “gap” is still soft, queue positions and optimistic energization dates rather than firm power. Neoclouds that control real time-to-power (behind-the-meter, contracted generation, or accelerated interconnection) will capture that upside. The ones sitting on paper capacity will not.
Own the requirements. Own the capacity.
19
Power architecture and density requirements are still moving faster than most site underwriting assumes. Sites that control both the land and a credible path to firm capacity (on-site, hybrid, or contracted) will clear the next 24–36 months. The rest will wait on someone else’s queue.
27
Hyperscalers do not evaluate on-site generation on LCOE. They evaluate it on recovered compute revenue. A month of earlier power on a 1 GW campus is worth more than most people model. Schedule is the scarce resource.
34
Replying to @PolymarketMoney
Schedule is the scarce resource, not the cents-per-kWh. Anything that reliably compresses time-to-power (on-site, hybrid, or accelerated interconnect) has higher economic value than most LCOE models capture. Hyperscalers already price it that way
46
Replying to @PhotonCap
Power availability is already re-ranking locations in real time. When interconnect timelines and resource-adequacy rules tighten in the primary markets, sites with surplus firm capacity stop being secondary. They become primary. NVIDIA is simply formalizing the matching.
1
1
67
Replying to @christopherrufo
Communities are reacting to real grid stress, water use, and cost allocation, not just messaging. A nuclear plant is firm capacity. A data center is large incremental load. When the grid is already short, people correctly prefer the former. The solution is not better PR. It is sites that bring or secure their own firm power so the load does not externalize the shortfall.
195
Replying to @UnoMasReactor
On-site and hybrid generation deals are the rational response when interconnection queues run measured in years.
22
Replying to @AndyMasley
Most capacity discussions still under-specify the full power and resource chain. Local vs. upstream impacts are frequently conflated.
23
Unix (incl. Linux) now runs ~92% of websites with a known OS. Windows is under 9%.
On servers - including most of the ones powering the internet and AI - the argument was settled years ago.
Yet in large parts of corporate IT and especially energy/critical infrastructure, Windows remains the default on desktops and operational systems.
The technical fight was won. The operational and commercial one was not.
Where did we go wrong, Unix people?
48
Replying to @NVIDIAAIInfra
Power architecture and density requirements are moving faster than most site underwriting still assumes.
2
468
Replying to @effiekav
Power and site control are the scarce assets. Compute follows the electrons that can actually be delivered on schedule.
16
Traditional real estate brokers and AI infrastructure teams still speak different languages.
One sells location and acreage. The other needs density, liquid cooling, and power that actually shows up on schedule.
That gap is expensive.
36
Most AI compute capacity marketed as “available” is still soft on the only things that matter: firm power and real energization dates.
Own the requirements. Own the capacity.
43
Replying to @scottbraddock
Political and grid constraints are now material variables in site selection. Ignoring them is how timelines slip.
2
150
Replying to @novogratz @GregAbbott_TX
Grid reliability and firm energization dates are now first-order underwriting criteria. Soft power assumptions are no longer viable.
91
Replying to @LibertyLynx
Most capacity marketed as available still assumes interconnection and energization timelines that do not exist. The physical constraints are the actual critical path.
23

