In aggregate it’s a lot of electricity, enough to impact local grid markets.
But as a fraction of inference revenue, there is good reason to estimate that the direct electricity costs are perhaps 1-5% of inference revenue.
That’s a lot, but hypothetically if electricity prices double, it’s more likely to lead to angry residential consumers than a doubling of inference token prices.
> That’s a lot, but hypothetically if electricity prices double, it’s more likely to lead to angry residential consumers than a doubling of inference token prices.
But that’s a feedback loop that could lead to more supply being built. Or it could lead to regulations that put enormous surcharges and low limits on data center power consumers, which effectively shield consumers directly from the lack of supply. Which in turn removes the incentive for consumers to expand power generation capacity to keep up (or to allow data centers to expand it themselves).
There is enormous opposition to a new data center project 30 miles outside my city because they want to build a 1000MW natural gas power plant and a 500MW biomass plant (that will take millions of acres of forest to sustain).
If that opposition continues, I see little chance of anything but electricity cost rising as a percent of inference.
Biomass powerplants use waste material as fuelnot fresh trees. They use waste wood like sawdust, scrap pieces. The most common fuel type is trash. They are used to recycle not to burn fresh resources.
That’s partially true, for some plants but not for this one.
This would be the largest biomass plant in the US. 5x larger than the next largest plant.
A 500MW plant would require almost 50x the amount of available waste material from logging and sawmill operations in the entire state of TN (where the plant is being built).
The developer of the project said that the plant was designed to create demand for low grade pine timber following pulp mill closures in the area and that the plant will burn between 350-500 truckloads of logs per day.
That’s good to know and thanks for the links. I guess what I’m seeing is 5-7 year timelines on new construction, interconnect agreements, pushback on larger market rate agreements from FERC when states have new regulations, etc. In any case it will be interesting!
At this point it seems like data centers will not be able to share in their consumption and will be responsible for a lot of their generation costs. This might also impact the 5% number quite a bit if data center owners need to bankroll a 5 billion dollar plant then wait five years and then pay for its operation so investor-owned utilities can amortize the cost.
Timers also let you logical-AND elements within one simple declaration. Combine that with a random delay, and you can easily define a trigger such as "the 2nd Saturday of the month, anytime between midnight and 6AM":
You can tell a timer whether to re-use that same randomly-selected time during subsequent runs (FixedRandomDelay), or to re-roll the time for every trigger. They can also persist their last-run history across reboots, and trigger if they missed their previous time-slot.
The improved logic vs. cron is one of the features which makes my life easier, and so I started using systemd more often. IMHO some of the newer additions aren't quite as solid; but the basic units have proven to be very handy. Another example: The 'path' unit, which can trigger whenever a defined file is updated.
It's a human problem. Noone "expects" their own "minimal"
checks to compound. And everyone (re)implements periodic checks in ever-more-inventive ways.
BSD cron solves this ? Alright. Does prometheus use it to schedule its periodic polling ? Do OTel collectors use it to smear their backend metrics uploads ? Does your own code use cron jobs for its "periodic" heartbeating ?
If the answer to that is all-yes, then my compliments to you. Alas ... at least my experience says that's not how things are done.
The "forced to use systemd" meme is a weird inversion. No one complained about being "forced" to use sysvinit when their was no choice, but when distros started to use systemd, which still allows to use sysvinit style services, that's suddenly framed as an infringement of free choice.
journalctl is good idea executed terribly. It's builtin data format is massively wasteful and not even good at searching for data you want (often need to go thru all the files anyway)
Journalctl with SQLite or a columnar embedded database would be amazing for sysadmins. Current format is entirely useless
That's interesting as it is also visible with a crontab -l so that new scripts can avoid using the same second. However, if you have a farm of machines with the same hardcoded value, they will all still stampede at the same time just not at the top of the minute. By allowing a random value, their start times are slightly offset with in the minute. Is the precise 60s consistent interval really that critical or just that each machine checks once per minute-ish?
Yes, using prime numbers are a good way to reduce the risk of multiple timers coinciding, for example a 10s, 30s, 60s period on a freshly restarted node / process.
Not PCB layout, but I've been building CircuitLab Assistant (https://www.circuitlab.com/), a chat window 100% integrated with our schematic editor and simulation engine. Can build / modify a schematic, troubleshoot, answer questions, configure simulations, etc.
If anyone wants to try an early beta, reply with your CircuitLab username and I'll get you set up. We know have a lot more to do, but would love some feedback on this early version!
But as a fraction of inference revenue, there is good reason to estimate that the direct electricity costs are perhaps 1-5% of inference revenue.
That’s a lot, but hypothetically if electricity prices double, it’s more likely to lead to angry residential consumers than a doubling of inference token prices.
Lawrence Berkeley National Lab, November 2026: https://eta-publications.lbl.gov/sites/default/files/2026-11...
There are substantial economies of scale on both power generation (grid-scale) and inference power consumption (batching and utilization).
reply