notes on how a company should run

There is a version of this company that does not exist any more, and it is worth explaining why.

A few years ago, finding out whether a software idea was any good cost six months and a couple of salaries. That price shaped everything. You picked one idea, committed hard, and defended it long past the point where the evidence had stopped agreeing with you, because the alternative was writing off half a year and starting again.

That price has collapsed. Building something real enough to put in front of a real customer now takes weeks. Not a prototype. A working thing that either earns its keep or does not.

When the cost of finding out drops by an order of magnitude, the right number of things to find out goes up. That is most of our strategy in one sentence.

Now the obvious wrong conclusion. Cheap to build is not cheap to sell.

Nothing about the last few years has made electricity procurement a faster sale. It has not made a business owner more willing to hand a stranger their site contracts, or a retailer quicker to return a call. A distributor's connections queue does not move faster because somebody shipped something. Trust in this industry moves at the speed it always did, which is slowly, and mostly through people who already know you.

So anyone running five businesses at once is running none of them. Building is the cheap part now. Selling, supporting and earning the right to be in the room are not, and they never will be.

What we do about it. We start more things than we keep. Each one is small enough that killing it costs weeks rather than the company. Nothing we begin is allowed to get big enough that its failure ends the story.

Then, when something pulls, that one gets everything. Not a fifth of the attention. All of it. Pull means customers turning up unprompted, and the thing being talked about when we are not in the room.

Many small questions. One serious answer at a time.

What this means if you work here. You will build things that get shut down. Not because you did the job badly, but because the job was to find out, and it found out. We would rather learn that in five weeks than defend it for five years.

So killing your own work well is a skill we take seriously. Saying "this one is not it" early, in writing, with your reasoning attached, is worth more to us than a heroic rescue of something that should not have been rescued. Nobody is penalised here for a clean answer to a good question.

The upside is that you will take more swings in two years here than you would get in ten somewhere with a roadmap committee.

The cost is real. If what you want is to spend a decade making one product excellent, this will frustrate you. Some of what you make will be switched off, and you will be the one switching it off.

We think that is a fair trade. We would rather find out than be sure.

Two people deliver the same project.

Person A commits to $100 and 30 days, and comes in at $110 and 40 days. Person B commits to $125 and 42 days, and lands exactly on $125 and 42 days.

Person B is commended for accuracy and prudence. Person A gets a talking-to about being optimistic and underinformed.

Look at what actually happened. Person A delivered the thing cheaper and faster than Person B. To get there they questioned every line item and refused overruns that Person B simply absorbed. Person B's precision was not judgement. It was room. The work expanded to fill the container it was given, which is what work does.

We have this backwards, and it is worth saying why.

Error is legible. Waste is invisible.

Person A's miss produces a number you can point at: ten dollars, ten days. It appears in a variance report. Somebody has to explain it.

Person B's padding produces nothing. No artifact, no signal, no investigation. The fifteen dollars never needed and the twelve days never required leave no trace anywhere in the system.

So we audit deviation from plan rather than distance from what was possible. And because the estimator writes their own exam, setting the number they are then measured against, the incentive is obvious. Run that for a few cycles and only Person Bs are left. Nobody decided it. It selects itself.

The industry has already run this experiment. Since 2010, solar developers have bid aggressive fixed-price offtake contracts into competitive tenders. Over the same period, regulated network and transmission businesses, buying steel, copper, transformers and labour from the same suppliers, have watched costs climb.

The difference is not the technology. It is who eats the error. A developer who bids low and gets it wrong wears the loss. A regulated business earns a return on its asset base, so an overrun does not go unpunished so much as enlarge the base the return is calculated on.

The uncomfortable follow-on is that those aggressive bids did not merely track falling module prices. They caused them. Solar costs fall per doubling of installed volume rather than per year, which is Swanson's law, and somebody has to buy the volume before the price falls. A developer signing a fixed price today against a cost they will only reach tomorrow is personally financing the trip between the two.

They capture a thin margin on one project. Everyone downstream, including their competitors, inherits a permanently lower cost base.

What we are asking for is not optimism as a vibe. Anyone can forecast aggressively when they will not wear the consequences. What matters is the costly commitment: signing up to a number, holding the exposure, and not being able to pass it down the line. That is the behaviour that generates information nobody had before.

Three things follow.

Separate the forecast from the target. A forecast should be accurate. A target should be motivating, and will therefore sometimes be missed. A commitment to a customer should be safe. Those are three different numbers, and we should stop asking for one and grading it against all three.

Judge against the possible, not against the promise. The question is not whether you hit your number. It is whether this was the best available outcome, and what we know now that we did not before. Person A's $110 becomes the benchmark everyone quietly measures against next cycle. The organisation free-rides on it and then reprimands them for it.

Be open when someone takes a swing and it does not land. Not all of them will. The returns show up across a portfolio of attempts, not in every individual one. If the only safe career move is to pad, we lose the upside on all of them.

None of this argues for sloppiness. A swing is worth taking where the downside is survivable, the assumptions are written down, and we find out early whether they are holding. Aggression with a stop-loss is a strategy. Aggression without one is hope.

But the failure mode to worry about is not the person who tried for $100 and got $110.

It is the room where nobody bids $100 any more.

Tobi Lütke, the founder of Shopify, describes every working relationship as a battery. It starts half charged. Every promise kept adds a little. Every one broken takes some away.

Electricity people will recognise the shape. A battery charges slowly and on its own schedule. It discharges as fast as physics will let it.

Organisations have one too. It is years of calls answered, prices that did not move without warning, a supplier kept because the work was good rather than because the quote was low. Founders who care about this make sure the people doing the charging know it is their job.

Then the business is sold.

The new owner is not a villain. They have bought a full battery. Accountants even have a word for it. They call it goodwill, and it only appears on a balance sheet when somebody buys it. So the new owner has paid for the battery and needs the money back.

Spending it looks like profit. Raise prices and most customers stay, for a while, because they trust you. Cut the service team without an improved replacement and the calls still get answered, eventually. Switch to a cheaper supplier and nobody notices for a quarter. Every one of those moves shows up as margin. The drain shows up nowhere, until it shows up all at once.

Cory Doctorow named the pattern in 2022: enshittification. The word is crude because the experience is. He was writing about technology platforms, but it is not a technology problem. It happens anywhere trust is an asset nobody is required to account for.

The protection runs from small to large. At one end it is a customer service call. If the person answering is measured on how quickly they hang up, the company has told them to drain the battery and is paying them to do it. At the other end it is the constitution. A purpose the next owner can delete with an ordinary vote lasts exactly as long as the current owner does.

Being conscious of the trust battery is one thing. Designing the organisation to protect it is another.

We are designing ours to protect it. The purpose is being written into the constitution, where changing it is meant to cost more than changing anything else. The people who charge the battery will be paid for charging it, not for how quickly they get off the phone. Neither is in force yet.

The battery does not need protecting from the people who charged it. It needs protecting from whoever comes next, including whoever we become. That is why the protection has to be built before they arrive.

General Electric. The name is a statement of intent: a company built to make things for the electricity market, generally.

The fastest-growing part of that market now runs on power electronics. Solar inverters, home batteries, EV chargers, heat pumps. The leaders are Huawei, Sungrow, SMA and Enphase. GE is not in the top ten.

Welch ran GE from 1981 to 2001 and took its value from $14 billion to a peak of nearly $600 billion. He ranked managers on a curve and let the bottom 10 percent go. In 1995 he made Six Sigma company-wide: measure every step, squeeze out every variation.

For a turbine blade, that is exactly right. If you make the same thing month in, month out, year in, year out, variation is the enemy and the method is close to mandatory.

Invention is the opposite job. It is mostly failure. A lab that hits its targets every quarter has stopped trying anything hard.

Watch what happened when the method left GE. Bob Nardelli took it to Home Depot, a business built on staff in orange aprons who would spend 20 minutes on your plumbing. Costs fell. By 2007 its customer satisfaction ranked last among major US retailers, and its share price went nowhere while Lowe's doubled. Jim McNerney took Six Sigma into the labs at 3M, home of the Post-it Note. The share of sales from new products fell from a third to a quarter. At Boeing, he chose in 48 hours to re-engine the 737 rather than design a new plane. That plane became the 737 MAX.

Larry Bossidy took the same tools to AlliedSignal, a maker of aerospace and car parts, and did well. That is the point. The tool fits a business that repeats itself.

Nardelli and McNerney were not fools. Each delivered the numbers they were hired for. The numbers improved because something unmeasured was being spent.

GE spent it too. The next grid is not a handful of big spinning machines pushing power one way. It is millions of rooftops, batteries, chargers and heat pumps, coordinated in software to act as one system. The culture that perfects a turbine is not the culture that invents what replaces it.

Gas turbines are booming again, and GE Vernova is largely sold out for 2026 and 2027 on demand from data centres. Welch had the same luck. Cheap gas and efficient combined-cycle plants set off a gas boom, and the profits looked like proof of the management system. A tailwind is real. It is not a strategy.

It has taken more than a century for the path to electrification to become clear. The last 100 years electrified parts of our homes. The next 100 will electrify nearly everything else. That is the biggest wave in the set, and GE is not leading it.

Judged over 20 years, Welch was the manager of his era. Judged over 250, he found a century of stored energy in GE (engineers, patient research, the standing to take big bets) and converted it into earnings. Then he took the credit for the output.

Efficiency is not the problem. It is a tool for one kind of job. Before reaching for it, ask whether your value is in doing the same thing better or in doing the next thing first. If it is the second, protect the parts of the business that look wasteful on a spreadsheet.

GE had the name, the history and the engineers to lead the electrification of everything. It caught the wave in front of it instead.

Rory Sutherland argues there are two ways to create value. Make something better, or make it more desirable.

Most companies pick the first by default. Not because it is more powerful, but because it is easier to defend in a spreadsheet. An engineering improvement has a business case. Desirability rarely does.

That is the wrong trade, because the two are not additive. They multiply.

Sutherland's own example is the map in a ride-hailing app. It does not make the car arrive any sooner. It changes what the waiting is like, and that turns out to be worth more than shortening the wait. The engineering answer was expensive and marginal. The other one was cheap and enormous.

It runs the other way too. A genuinely good product with flat, apologetic framing underdelivers on its own merits, because expectation shapes experience. Sharp framing on a weak product gets you to disappointment faster. Neither works alone.

So we do both. Not as a fifty-fifty split, but as a discipline: we pay attention to whichever side the organisation is quietly starving. Usually that is the illegible one, the part that cannot produce a number to justify itself and so loses the argument by default.

A sealed-bid marketplace is not a better quote. It is the same installers quoting the same job at the same prices, arranged so that nobody has to sit at a kitchen table being sold to. None of the engineering changed. What it is like to buy changed completely.

What that means if you work here. You will be asked to make things work properly. You will also be asked why anyone should want them. Neither question is someone else's job, and "the numbers did not support it" is not a complete answer to the second one.

We would rather back a judgement call that might be visibly wrong than a cautious one that is invisibly wasteful. Error is legible. Waste is not. We try to price both.

If that sounds like how you already think, we should talk.