Eliyahu Goldratt’s The Goal is a novel about a plant manager trying to find the one machine that sets the plant’s output. Covering the data center buildout, the market seems to have found its machine already: power. Elegant, but too simple.
Power is a challenge, but silicon is where we see the system binding today. Advanced packaging is booked through mid-2026. NVDA has roughly 60% of global CoWoS (the packaging step that stacks logic and memory) capacity for 2026 and 2027. All three HBM (high-bandwidth memory) suppliers have sold their 2026 allocation. In a truly power-constrained world, I would expect to see stranded GPUs sitting in warehouses. I have not.
Old chips are also living longer. If power were the binding constraint, and new chips delivered materially better tokens per watt, older GPUs should be retired faster or repriced lower. Instead, they are being recontracted above original rates, and that is happening even as per-chip draw goes 6x, from 300 watts in 2017 to an estimated 1,800 by 2027. That is not the behaviour of a market drowning in silicon and starved only for electrons.
The power story is more complicated. Large-frame gas turbine order books are full, but bridge capacity is available in fuel cells, reciprocating engines and aeroderivatives. Those categories can produce roughly 25 GW/year, against about 7.5 GW/year of data center additions. Order a heavy-duty turbine from GEV today and it arrives in 2031. That does not make power easy. It does mean the bottleneck is narrower than “generation.” The hard part is deliverable, permitted, AI-ready power at the right site by the right date.
Land tells the same story. Hyperscalers control about 54,000 buildable acres, less than 40% developed. Raw land is plentiful. Useful land is scarce. Our powered-land ladder climbs from $75,000/acre raw to $1.67 million/acre energized, and only 5% of parcels entering entitlement get there. That explains the land banks. It also explains the bloated load queues.
Which brings us back to The Goal. Relieve one bottleneck and another appears. Silicon should ease as manufacturing capacity expands. But delivering power, conditioning it for AI loads and putting it into shells built to use it is getting harder.
Start with the economics. Large-frame capex has moved from $0.9 million/MW in 2023 toward $2.4 million/MW by 2027. A large-frame combined cycle can take up to 80 months to reach commercial operation. Recips and fuel cells are faster, often 18 to 24 months, but faster is not the same as frictionless. They still need permits, fuel, interconnection, emissions approvals and a customer willing to pay for speed.
The politics add another drag. Batch Zero rationed access. The price was $50,000/MW, with 80% forfeited if the customer walked. Expensive, but at least it came with rules and a timeline. Governor Abbott’s moratorium and queue audit may change few outcomes while burning months. That delay could matter more than the money. Tariff reform and cost allocation will add more friction. The case for private-use networks gets stronger with every month lost in the queue.
In a system growing this quickly, no constraint stays binding for long. Chips bind today. The edge is spotting the handoff before it is obvious, because obvious constraints get priced. Goldratt’s plant manager did not find one permanent machine. He learned to watch where the line backed up.
Thoughts, questions or things we missed? Send me a note (or hit reply) - I would love to hear from you. Thanks for reading!
Morning Energy is a syndicated note published through Enverus Intelligence. My contributions will also be distributed here. Please note that links frequently lead to content available only to subscribers of Enverus solutions. Please reach out if you have any questions. Thanks! - Ian.