The cloud, without the data centre.

Every frame above was rendered on a Mac, not in a data centre. Kynd Share is about what happens when one Mac is not enough.

The honest part

The pull of the cloud is real.

Anyone who says otherwise has not sat watching a progress bar at 2am.

Creative work runs on momentum. You have the idea now, and you want to see it now—not after your Mac has spent forty minutes thinking about it. The big AI services understood that before anyone else did. Enormous power, instantly, on tap. Ideas into pixels, fast.

It is genuinely useful, and pretending otherwise would be silly. It also feels, quite a lot of the time, like everyone else is a step ahead and the only way to catch up is to rent someone else’s machines.

But that convenience is not free, and it is not paid for by the person typing the prompt.

Your Mac is not slow. It is alone.

The cost, and who carries it

Somebody lives next to the building.

Not an abstraction. A neighbourhood, a water table, an electricity bill.

It is easy to talk about compute as though it happens nowhere. It happens somewhere, and the somewhere has residents. Here is what the record actually shows—every figure below is from a regulator, a national laboratory, a peer-reviewed paper or a company’s own report, and every one is linked.

The hum that never stops

Virginia hosts the densest concentration of data centres on earth, so its legislature studied the problem properly. Its audit commission found that 29% of operational data centre properties sit within 200 feet of residentially zoned land, and that at least fifteen sites—one in ten—produced noise nearby residents regard as a problem.

The detail that matters most is this one. Councils measure noise in A-weighted decibels, a scale the commission notes was “designed to target excessively loud noise from sources such as parties and barking dogs”. The low-frequency drone of cooling plant is not fully captured by it. So the noise “rarely exceeds the allowable limits set in ordinances, despite the constancy of the noise”. It is legal, it is permanent, and it is inside people’s houses. One county’s acoustical consultant traced a 90 Hz tone into homes beside a complex carrying 424 rooftop exhaust fans.

Source: Virginia JLARC, Report 598, “Data Centers in Virginia” (2024), chapter 6; Prince William County Board of Supervisors, Noise Ordinance Update (2025).

The water

Google publishes site-level water use, which most operators do not. Its Council Bluffs campus in Iowa consumed 1,346 million gallons in 2025—about 5.1 billion litres, or roughly 14 million litres a day, averaged across the year. Microsoft reports that 48% of its water consumption came from areas under water stress.

Across the United States, direct on-site data centre water use reached 66 billion litres in 2023, roughly tripling in a decade.

Sources: Google 2026 Environmental Report; Microsoft 2026 Environmental Data Fact Sheet (FY25); Lawrence Berkeley National Laboratory, 2024 US Data Center Energy Usage Report. Daily averages are our arithmetic from annual figures; real use peaks in hot weather.

The air, and the health

A study accepted by Communications of the ACM modelled the public health cost of US data centre air pollution at $6.67 billion and an estimated 433 premature deaths in 2023, rising to $20.9 billion and 1,262 deaths by 2028 on a high-growth path—a 213% increase against 17% growth in America’s stationary fuel-combustion health costs overall.

We will quote the authors’ own caution rather than paraphrase it: these estimates “should not be interpreted as assigning responsibility for any specific individual health outcome”. They are modelled, not counted. They are also, on the authors’ own numbers, about 1.3% of America’s total fuel-combustion health burden. It is a real cost and it is not the largest one in the world.

Source: Han, Wu, Li, Wierman & Ren, “Health-Informed Computing”, arXiv:2412.06288, accepted by Communications of the ACM.

The bill, on everyone’s electricity

This is the one that surprised us. PJM runs the electricity market for thirteen US states. Its independent market monitor calculated that including data centre load in the 2027/2028 capacity auction increased what everyone on that grid pays by $6.5 billion—a 65.5% increase in total capacity market revenues.

That is not a bill sent to the data centres. It is spread across every household on the network.

Underneath it: global data centre electricity demand reached 485 TWh in 2025 and is projected to roughly double by 2030, with gas and coal expected to meet over 40% of the additional demand. In the US alone, serving it implies 148 GW of new grid connection capacity by 2030.

Sources: Monitoring Analytics, Analysis of the 2027/2028 RPM Base Residual Auction (January 2026); IEA, Energy and AI; Lawrence Berkeley National Laboratory, US Data Center Energy Usage Report (2026).

What sharing a Mac does and does not fix.

We would rather be straight about this than sell you something we cannot back.

Here is the part most people in our position would leave out. Roughly 92% of that modelled health cost comes from grid electricity, not from generators humming on site. And the Berkeley figures put the water embedded in data centre electricity at about 4.52 litres per kilowatt-hour—against a US average of 4.35. Close to identical.

Which means the honest conclusion is this: a kilowatt-hour costs roughly the same wherever it is drawn. Your Mac is on the same grid. Rendering locally does not make electricity free, and we are not going to tell you it does. Nor is an M-series chip more efficient per render than a rack of purpose-built accelerators—it is not.

What changes is everything around the electricity.

The argument for Kynd Share is not that computing is weightless. It is that the machines have already been built, the buildings they sit in are called houses, and the capacity is going spare. Meeting demand with hardware that exists is the one option that requires nobody to pour concrete, and puts nobody’s bedroom next to a wall of fans.

Most of the good evidence here is American, because the US has the densest concentration of these facilities and the most public reporting. We have not found UK data of comparable quality, and have not assumed the numbers transfer.

What Kynd already does

No AI data centres. That has never changed.

Every model Kynd Agent runs—video, image, text, voice—runs on your Mac. Your prompts, your footage and your ideas do not leave the machine to be processed by anyone. That was true on day one and it is still true today.

Today, Kynd runs one heavy render at a time on each Mac. In our M1 Max 64 GB tests, two LTX renders together finished slower than running them one after another. That result applies to the hardware and workload we measured. Extra simultaneous jobs on other Macs need benchmarks showing a real throughput gain and sufficient memory. For now, independent clips can run together across several hosts.

Which raises an obvious question. Whose?

Kynd Share

The machine you need is already on someone’s desk.

A marketplace for idle Mac time, brokered blind, paid at published rates.

Kynd Share lets people rent out the GPU time their Mac is not using, and lets other people buy it when their own Mac and their own network are busy. The clip renders on a stranger’s Mac and lands in your library like a local render.

Three things make it different from renting a server.

1 · The hardware already exists

Nothing is built for this. No new racks, no new cooling, no new grid connection. The capacity is Macs people already own, doing nothing between about midnight and eight in the morning. The marginal cost of a render is the electricity that machine draws—roughly 2.5p an hour—because everything else was already paid for years ago.

2 · It is blind by construction

Every job is encrypted for one specific Mac before it leaves yours. The host’s machine can open it; our broker cannot. Nothing is shown on the host’s screen, nothing is stored, nothing is logged. They are never asked to review what they render, because they cannot see it. We built it that way deliberately—screening other people’s creative work is not a job we want.

3 · Nobody sets their own price

The same duration and quality cost the same across host speeds. Every host earns the same rate per completed minute: standard video pays 720 credits, about £0.55. KyndBoost sends jobs to the fastest eligible available host, using measured pace and reliability. Faster hosts earn more by completing more work, and the chip name alone does not determine their speed.

The arithmetic

What an idle Mac is actually worth.

Spending £10,000 on a Mac Ultra so it can sit in the dark is a strange kind of ambition.

These examples show how measured output changes earnings at the same rate of 720 credits per completed minute of standard video. They are illustrations of throughput, not promises for a particular Mac.

Gross earnings before electricity, rounded from credits at the fixed book rate. Figures assume the stated output throughout every sharing window, with 30 eight-hour nights per month. Real earnings depend on completed jobs, render settings and demand. On the day we open there will be very little.

That last sentence matters, so we will not bury it. A marketplace with no buyers pays its hosts nothing. Early hosts are helping to build the thing as much as earning from it, and we would rather say that plainly than quote you a number and let you discover it later.

Each job is quoted before it runs. The rate is for completed video; hourly earnings depend on how much your Mac actually delivers.

And what KyndBoost costs.

Priced from the £5 starter pack, the dearest way to buy credits; larger packs work out cheaper per minute. Every job is priced before it runs and you set a ceiling it cannot pass without asking.

You are buying someone’s spare capacity, not a slice of a data centre with a building to pay for. The share that does not go to the host runs the broker, the ledger, the payments, the tax and the support—and none of it goes into concrete.

Why it matters

The money goes to the person whose Mac did the work.

This is the part we care most about. When you render in the usual way, your money buys shareholder returns for a company building the next data centre. When you render on Kynd Share, it goes to a person—a filmmaker in Leeds, a designer in Lisbon, someone who bought a serious Mac for one project and watched it idle ever since.

Hosting is free and needs no licence. You do not have to buy anything to take part, and you can stop at any moment. If cash payouts do not reach your country, your earnings convert to Kynd credits and buy your own renders instead—so nobody is locked out of the useful half.

We would rather build something where a good machine earns its owner money than something where a good machine makes a distant company money.

Coming soon

What is true today, and what is not yet.

Kynd Share is built and tested end to end—encryption, matching, escrow, ratings, payouts. It is not open yet. It arrives in the next Kynd Agent release, and we will say so here when it does.

At launch it will do video and images. Sharing text models works too, but it is staying switched off until it has had more real use—we would rather ship two things that behave than three that mostly do.

If you own a Mac that spends its nights doing nothing, that is exactly the machine this needs.