All articles

You don't need a new data center to use good AI

AI companies are building huge data centers, and the neighbors often pay part of the cost. A lot of everyday AI can now run on one small computer on your desk.

You may have heard that AI needs more and more data centers. A data center is a huge building full of computers.

Some of what you've heard is true. Some of it isn't. Here is what we found, and where The Box fits.

Yes, they are building a lot

In 2024, data centers used almost 5% of all the power in the US. A US government lab, Berkeley Lab, made its best guess for 2030. It could be about 12%. [1]

The IEA is a world energy group. It said the world's data centers could use about twice as much power in 2030 as in 2025. [2]

The people next door pay part of the price

Data centers often bunch up in the same place. Around the world, they use only a small part of all power. But where they bunch up, they can use as much as 20 to 30% of the power there. [2]

You can see it on power bills. A group, PJM, runs the power grid for 13 states and Washington, DC. The grid is the web of power lines that bring power to your home. Each year, PJM has a sale to make sure there will be enough power down the road. A watchdog group goes over that sale. It found that data centers caused $6.3 billion of the cost of the sale PJM held in 2026. It said other people pay for that on their power bills. [3]

You can see it in the air. In April 2026, a civil rights group, the NAACP, took a company, xAI, to court. It said xAI is running gas turbines without the permits the law says it must have. A gas turbine is a big engine that burns gas to make power. The turbines sit in Southaven, Mississippi. They power an xAI data center just over the state line in Memphis. The NAACP said there were 27 when the case began, and 33 by May. [4]

There is another side. Mississippi has said the turbines can be moved, so they don't need that permit. The US government has asked the judge to throw out the case. When we last checked, in September 2026, no judge had ruled. [4]

You can see it in water, too. In 2023, US data centers used about 66 billion liters of water. Making their power took about 12 times as much. [5]

And towns and neighbors are fighting back. Look at the first three months of 2026. In that time, they blocked or held up at least 75 data center plans in the US. [6]

What is all that for?

The IEA asked a "what if." What if every web search in the world were a simple AI question instead? That would use less than 1% of the power those buildings use now. [2]

So what is the rest for? The IEA can only guess, since the companies haven't said. One use is training, which is how a new AI model is made. Another is making videos and pictures. Another is a helper that does one step after another, for a long time. Some are business uses no one can see yet. Some of these jobs can take hundreds or thousands of times more power than a simple question. [2]

We used to say, "It's a lie that we need more data centers for access to great AI." That was too strong. Some of that new space will train new models, and no desk can do that. But do you need a new data center to get good AI for your own work? Not any more.

Good AI now fits on a desk

Some good AI models are now free for anyone to download, like Qwen 3.6 and Gemma 4. They can run on one strong desk computer. [7][8] The Box has them. It also has Hermes, a free helper made by Nous Research. Hermes can work on a job one step at a time. [7][12]

The standard Box is small enough to sit on a desk. Its main chip is made to use about 120 watts. That's for hard work over a long time. It can use a bit more for a short time. Its power supply is made to give up to 400 watts. It plugs into a normal wall socket, and it has a fan to keep it cool. It doesn't need a new building or new power lines. [7][9]

What The Box doesn't fix

We'd rather you hear the down side from us.

The models were made in data centers. Big models are trained in a place like that. [2] Gemma 4 was made by Google DeepMind. Qwen was made by a team at Alibaba. [8] Running one on your desk doesn't change that. But that work is already done. It is shared once, and every box uses a copy of that same file.

You still use power from the grid. Your box will add to your power bill. Making that power will use water, too. Berkeley Lab looked at the water it takes to make power. For the same amount of power, making normal US power took about as much water as making power for data centers. [5]

We can't promise it will use less power for each question. A big data center can help many people at once, so each answer can use very little power. The IEA said a simple AI question now uses less power than a TV in the same time. [2] We haven't found a fair study of that against one box on one desk. So we won't say it is. Work with many steps will use more power, no matter where it is done. On your desk, it can only use as much as one small box can.

It doesn't stop the building. One more box won't stop a giant data center. And your own use is not what is behind all that building.

So why does it matter where it lives?

If your own use is so small, why care? Because of where your money goes.

Cloud AI lives on a company's computers, over the web. You often pay for it every month. Some of that money goes to the companies that build those buildings, again and again. With The Box, you buy the computer once. The work you do on it is done on your desk. You don't pay an AI company for it each month.

That doesn't mean no money goes their way. The chip in the standard Box is made by AMD, which also sells chips for data centers. [7][13] The free models came from Google and Alibaba, and they run data centers too. [8][14] Our own website is on Amazon's cloud. [11] So when you buy a box, some money still goes to the same kind of business, one time. It just doesn't keep going every month.

On our sales calls, people often bring up the cost of data centers to the planet. One person on our waitlist said they "don't want to support big companies," most of all with their own work. [10]

We think that's fair to want, and now you can choose it. Data centers aren't evil, and a box on a desk isn't free of cost. But for a lot of everyday work, it's no longer true that good AI needs more data centers. You get to choose where your AI lives, and who you pay.

If this is how you'd like to work

You buy The Box once. We set it up before we send it to you. When we put out an update, like a new model, you choose if and when to install it. We don't promise updates on any set schedule.

See pricing and tiers (or join the waitlist if the current batch has sold out)

Related reading:
Some data you can't put into AI. Some you'd rather not. (/articles/privacy)
What owning your AI means (/articles/sovereignty)
The case for paying once (/articles/cost-control)
What "set up for you" covers (/articles/setup)

Sources

  1. US data centers used 192 TWh in 2024, 4.7% of US electricity; central 2030 estimate 11.8% (range 9.5–15.3%) — Shehabi, A., United States Data Center Energy Usage Report: 2025 Update, Lawrence Berkeley National Laboratory, June 18, 2026, https://escholarship.org/uc/item/33m6w3x0 (PDF: https://escholarship.org/content/qt33m6w3x0/qt33m6w3x0.pdf). Checked in the PDF text.
  2. IEA, Key Questions on Energy and AI (World Energy Outlook Special Report, April 2026) — https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf (landing page: https://www.iea.org/reports/key-questions-on-energy-and-ai). Checked in the PDF text: 485 TWh (2025) to 950 TWh (2030); "largest technology companies are contributing to a surge in data centre investment, as their capital expenditure exceeded USD 400 billion in 2025 – and is expected to jump by another 75% in 2026" (Exec. summary); 1.5% globally but "as much as 20-30%" in clustered regions; <4 TWh if all searches were simple AI text queries, "less than 1% of total data centre consumption today"; "energy use per AI task dropping by at least an order of magnitude annually in recent years"; "simple text queries now typically consume less electricity than running a television over the same period of time"; video generation, reasoning and agentic tasks "can consume hundreds or thousands of times more energy per query than simple text generation" (Exec. summary); data centres are "the critical infrastructure for training and" using AI; "raises a question that the available disclosures do not yet answer… the bulk of planned capacity must be destined for other workloads, such as large-scale model training, video and image generation, autonomous agents running multi-step pipelines, or enterprise deployments not yet reflected in public usage figures", no company breakdown (box in ch. 2).
  3. PJM: data center load $6.3B of $16.4B (38.2%) in the 2028/2029 capacity auction; IMM says data centers should "not impose capacity costs on other customers" — primary: Monitoring Analytics, Market Monitor Report, PJM Members Committee webinar, July 27, 2026, slide "2028/2029 BRA Data Center Impacts", https://www.pjm.com/-/media/DotCom/committees-groups/committees/mc/2026/20260727-web/item-02---marketing-monitoring-report---presentation.pdf · "just-held" auction, 13 states + DC, "ratepayers in PJM are … paying capacity charges" — Utility Dive, July 20, 2026, https://www.utilitydive.com/news/pjm-data-centers-capacity-auction-imm-bowring/825626/
  4. xAI / Colossus 2: NAACP suit filed April 14, 2026, represented by SELC and Earthjustice; 27 turbines without an air permit in Southaven, MS, powering Colossus 2 in South Memphis — https://naacp.org/articles/naacp-sues-xai-illegal-pollution-data-center-power-plant · May 6, 2026 motion for emergency relief: 33 turbines (six added), plant in Southaven supplies Colossus 2 in Memphis, TN — Earthjustice, https://earthjustice.org/press/2026/naacp-asks-court-for-emergency-action-to-stop-illegal-air-pollution-from-xais-data-center-power-plant · DOJ moved to intervene and sought dismissal; MDEQ treated the turbines as mobile/portable sources — Utility Dive, June 18, 2026, https://www.utilitydive.com/news/doj-intervenes-xai-data-center-gas-turbine-lawsuit/823267/ · Background: separate March 10, 2026 permit for 41 permanent turbines — https://www.selc.org/press-release/mississippi-regulators-rubber-stamp-air-permit-for-xai-power-plant-ignoring-overwhelming-public-pushback/
  5. Water: direct use 66 billion liters in 2023 (p. 55); indirect ~800 billion liters; indirect water intensity 4.52 L/kWh for data center electricity vs 4.35 L/kWh US average (p. 57) — LBNL, 2024 United States Data Center Energy Usage Report, pp. 55–57, https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf (the PDF sits behind a browser check for some readers; the 2025 Update in [1] does not revise these water figures)
  6. "At least 75 data center projects worth approximately $130 billion were blocked or delayed in a single quarter" (Q1 2026, local opposition) — Data Center Watch, https://www.datacenterwatch.org/q1-2026
  7. Models the box runs (Qwen 3.6 27B / 35B MoE, Gemma 4 26B), Hermes Agent ("Nous Research's autonomous agent. Not ours."), Ryzen AI Max+ 395 in both memory options (64 GB and 128 GB), ~4.5 L chassis, 400 W FlexATX PSU, Noctua fan, custom hardware = Enterprise tier — internal: company-profile/current-facts.md (Models) and company-profile/product.md (stack list, platform specs)
  8. Gemma 4 "Developed by Google DeepMind", Apache 2.0 — https://ai.google.dev/gemma/docs/core/model_card_4 · Qwen3.6 authored by the Qwen Team (Alibaba), Apache 2.0 — https://huggingface.co/Qwen/Qwen3.6-27B
  9. Framework Desktop: "Processor power level: 120W sustained, 140W boost power"; 400 W FlexATX PSU, Gold (110V); 96.8 × 205.5 × 226.1 mm, 4.5 L — https://frame.work/desktop?tab=specs
  10. Buyer values ("environmental cost of data centers" under "Values that keep coming up"); waitlist lines "Don't want to support big companies esp. with my IP" and "Decoupled from the corporate machines that have taken previous generations of art" — internal: arx-site/docs/our-story-interview.md, "From the sales-call notes" and "From the waitlist, 10. Not feeding big tech / my IP"
  11. "Website hosting and storage — Amazon Web Services, Inc. Hosts the site and stores waitlist signups" — published privacy policy, twenty/legal/md/privacy-policy--web.md (§ service providers, line 90)
  12. Hermes Agent: "MIT — see LICENSE. Built by Nous Research." — https://github.com/NousResearch/hermes-agent (opened 2026-09-28)
  13. AMD Instinct GPUs "drive frontier AI and HPC in data centers"; page description: "designed to accelerate frontier AI, data center, Sovereign AI and HPC workloads" — https://www.amd.com/en/products/accelerators/instinct.html (opened 2026-09-28)
  14. Google: "We custom build servers exclusively for our data centers" — https://datacenters.google/ · Alibaba Cloud: "Alibaba Cloud operates 107 availability zones in 31 regions around the world", page section "Data Centers Around the World" — https://www.alibabacloud.com/en/global-locations (both opened 2026-09-28)