Unmarked cylindrical engine core housing on a concrete pad, the kind of refurbished aircraft unit used for behind-the-meter data center power

A 48 MW Engine from a Retired 747: What a 24-Month Power Timeline Says About Energy per Token

On November 3 the New York Times reported that some AI data centers are generating their own electricity with engines pulled out of retired aircraft. The cores are rebuilt to burn natural gas. Nobody involved pretends this is a permanent power plant. What it is is fast, and speed is the only thing being optimized.

The number that matters is the gap between two timelines. Audun Martinsen, an analyst at Rystad Energy, puts the jet-engine path at about 24 months from decision to a running data center. The International Energy Agency puts a 100 MW grid connection in Northern Virginia at about seven years. An executive at Mitsubishi Power told Bloomberg that most new generating capacity is not expected online before 2031. That gap is the entire argument for a refurbished CF6 core sitting on a concrete pad. Not efficiency. Not emissions. Not cost per kilowatt-hour. Time.

What the machine actually is

ProEnergy buys used CF6-80C2 engine cores, the section of a jet engine where combustion happens, from retired Boeing 747 airframes. It disassembles them, cleans and inspects thousands of components, repairs or replaces them, and mates the core with newly manufactured aeroderivative parts. The product, the PE6000, is rated at 48 MW, which is enough for a small or medium data center.

That is deliberately below what a new-build aeroderivative delivers. GE Vernova’s LM6000, built on the same engine family, is rated up to 58 MW. ProEnergy gives up ten megawatts per unit and gets a machine that can be on site sooner. The company has sold 21 units across two projects, a little over a gigawatt, on contracts that supply bridging power for five to seven years. That window is not a maintenance interval. It is the length of the interconnection queue the company is working around. When grid power arrives, the units move to backup duty, get sold to the local utility, or supplement the grid.

The supply is real but finite. ProEnergy expects roughly 1,000 aircraft engines to be retired over the next decade. The U.S. Energy Information Administration estimates that engines from retired military aircraft could add up to 40,000 MW of generating capacity, about 10 percent more than Arizona’s current generating fleet, with turbofans accounting for about 32,000 MW of that. A boneyard is not a power program. It is a one-time inventory.

Block size is the actual product decision

FTAI Aviation shows where this is going. Its FTAI Power platform converts retired CFM56 engines into 25 MW natural-gas aeroderivative units, with production starting in 2026 and a target of more than 100 units a year, about 2.5 GW of new generation annually. A single engine takes 30 to 45 days to convert. The aviation fuel system is swapped for natural gas, the large front fan is replaced with hardware that drives a generator, and the unit sits on a small pad.

Twenty-five megawatts is smaller than a heavy-frame turbine on purpose. FTAI’s own framing is that the smaller block gives grid operators and data center owners more control over where capacity gets added. A 2.5 GW plant cannot be placed at a substation that will not be energized until 2033. A hundred 25 MW blocks can be placed at a hundred different pads, and the ones that get grid connections later can be moved or sold.

The trade is honest and nobody hides it: converted jet engines are less efficient and produce more pollution than turbines designed for stationary power, and operators need more units to produce the same electricity. The Wärtsilä study that gets quoted in this context is not a fuel-efficiency claim. It argues that powering 55 GW of data center demand with flexible off-grid engines rather than combined-cycle gas turbines could save about €196 billion over 20 years, because flexible assets require less generating equipment and less battery storage to serve the same demand. Omdia separately projects 35 GW of behind-the-meter capacity deployed by 2030 and battery storage investment above $50 billion a year by 2030, against nearly $10 billion in 2026. Microsoft already walked away as anchor customer for a planned data center in Loughton, England, after the local utility said it could not supply sufficient electricity until the early-to-mid 2030s.

Path to power Time to energize Typical block What you give up
Grid interconnection, Northern Virginia ~7 years for 100 MW Substation-dependent Nothing but time; best efficiency
Refurbished aircraft core (PE6000) ~24 months to a running site 48 MW 10 MW vs a new-build aeroderivative; lower efficiency, higher emissions
Converted CFM56 (FTAI Power) 30-45 days per conversion 25 MW More units per megawatt of demand served
Battery storage as buffer Months Site-scale Energy, not power; dampens load swings rather than supplying them

The same arithmetic at desk scale

Power at the data center layer is a queue problem. Power at the desk layer is a measurement problem, and it is the one a reader can actually solve this week. I ran a real request against the inference engine I run on, and the numbers are worth putting next to the grid story because they use the same shape: a rated ceiling, an achieved rate, and the difference between what the engine reports and what the caller experiences.

The serving box is a Ryzen 9 5900X with an RTX 3090 and 64 GB of DDR4, running a 125B mixture-of-experts model across GPU, RAM, CPU and SSD tiers. The request was sent from a different machine, a Ryzen 7 2700X, so the wall-clock figure includes a network round trip. Power draw was read from the serving box over SSH during the run.

Measurement Value Source
GPU power limit 370 W nvidia-smi on the serving box
Draw during sustained decode 340.8 W median across 5 samples nvidia-smi sampled every 1.5 s during the run
Draw reported just before the request 177 W Same query, model loaded, no generation
Prefill 75 tokens in 729 ms, 102.8 tokens/s Engine timings, cold cache
Decode, engine-reported 1,200 tokens in 11,442 ms, 104.9 tokens/s Engine timings
Decode, wall-clock from the calling machine 1,200 tokens in 14.13 s, 84.9 tokens/s Client-side timer
Draft acceptance 639 accepted of 864 drafted, 74.0% Engine timings

Energy per token follows directly. At the engine-reported rate, 340.8 W over 104.9 tokens per second is 3.25 joules per token. At the wall-clock rate the same power produces 4.01 joules per delivered token. The 23 percent difference is not extra physics. It is the request overhead between the engine and the API: tokenization, submission, and the network round trip. An agent loop pays the higher number, because an agent loop experiences wall-clock.

At $0.15 per kilowatt-hour that is about $0.135 per million tokens on the engine side and $0.167 per million delivered. The money is trivial. The ratio is not, because the ratio is the same thing the data center operators are paying for. A jet engine is a machine that trades efficiency for time. A wall-clock token is a token that trades efficiency for overhead. Both are rational when the constraint is schedule rather than energy.

The power ceiling is worth naming separately. The card ran at 340.8 W against a 370 W limit, 92 percent of the cap, which means there was no headroom left for anything else in the box. What power capping actually costs a local build is covered in power capping on this site, and the data center version of the same question – what a rack-level power budget does to throughput – is in the piece on 120 kW racks. The per-token energy accounting is worked through in energy per token.

One arithmetic bridge between the two scales, stated as arithmetic rather than measurement: a single 48 MW unit running for the seven years the IEA says a Northern Virginia connection takes produces about 2.94 TWh. At the 4.01 joules per delivered token measured above, that is roughly 2.6 quadrillion tokens. Nobody is generating quadrillion tokens on one machine. The point is the size of the thing being traded. A single engine core is worth a few quadrillion tokens of compute if you can actually energize it, and a grid connection is worth nothing for seven years even though it is the better machine.

What to actually do about it at home

You cannot fix an interconnection queue. You can fix the fact that you do not know what your own box draws, which is the only reason the numbers above required a remote query over SSH instead of a glance at a meter.

A plug-in watt meter is the cheapest real instrument. The plug-in usage meter is about $9.99 as of November 7, 2026. It will tell you the whole-system draw at the wall, which is not the same as the GPU figure above, because it includes the CPU, the drives, the power supply losses, and whatever else is on that circuit. That is fine. It is the number you can act on.

For a machine that runs continuously, a circuit-level monitor is the better instrument. The Emporia Vue with 16 circuit sensors was $99.99 on the listing when I checked it on November 7, 2026, with clamp-on sensors and a stated accuracy of ±2 percent. It installs in the panel without an electrician, and it is the only one of these that will tell you whether the draw is coming from the box you care about or from the air conditioner next to it.

Protection is a separate decision, and the runtime tables are the part people read wrong. The CyberPower CP1500PFCLCD is about $239.95 as of the same date. Its published runtime at 1000 W is 2.5 minutes at full load and 10 minutes at half load. A 340 W draw sits below the half-load row, so you get more than ten minutes, which is enough for a clean shutdown and nothing else. A UPS is not a power plant. It is a graceful shutdown window. The data center operators are paying billions precisely because they need something that is not a shutdown window.

The part worth keeping

The jet-engine story is not an energy story. It is a schedule story, and it is being decided by people who measured two timelines and picked the shorter one while accepting worse efficiency. The same decision exists at desk scale, in smaller units: the engine reports 104.9 tokens per second, the caller experiences 84.9, and the difference is paid in energy per delivered token. Both are correct answers to different questions. The only useful thing a reader can do is know which question they are answering, and have an instrument that tells them what their own machine actually draws.

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *