Space · Sunday, 23 August 2026

01 · Briefing · what happened

LandSpace landed a rocket in the Gobi Desert. The fix after December's crash was fewer engines, not more.

Space 3 min 18 sources

China's second rocket recovery in six weeks, and its first on land, came from a startup that made its booster ask less of itself. It arrives as Western satellite firms run out of rides.

engine lit for the final braking

down from several on the flight that crashed [1]

390 km

from the pad to the landing site

in Minqin County, Gansu, about eight minutes after liftoff [2]

100

SpaceX Falcon flights this year

against a record 165 in 2025 [7]

flights LandSpace wants from each booster

no Chinese booster has flown twice yet [1]

At a glance

LandSpace landed the bottom half of its Zhuque-3 rocket in the Gobi Desert on Tuesday evening, 390 km from the launch pad. [2]
It is China's first rocket recovery on land, and the first anywhere on landing legs by anyone other than SpaceX and Blue Origin. [2][5]
The same rocket's first flight reached orbit last December and then crashed while trying to land. [3][4]
The fix was subtraction: one engine lit for the final phase of the landing instead of several, plus a redesigned heat shield. [1]
China's other recovery, a Long March 10B caught in a net at sea on 10 July, used a completely different method. [5]
SpaceX flew its 100th Falcon mission of 2026 on Saturday and now has more than 11,000 Starlink satellites working in orbit. [7][8]
Satellite firms cannot buy shared SpaceX launch slots past late 2028, and one analyst describes an industry in panic. [9][15]
Washington's answer was a target of more than 1,000 launches and re-entries a year by 2030. [11]

Forces in play

Chinese reuse push Building↑

Two Chinese boosters recovered in six weeks, with at least five more reusable rockets close behind [1]

Western launch supply High↑

SpaceX's shared slots are sold out past late 2028 and Blue Origin's New Glenn is grounded after an explosion [9][10]

Money for satellites High↑

Muon Space and Starcloud each raised $250 million this week, for satellites that still need a ride [13][15]

Setback from a failure Easing↓

A Long March 7A exploded on 10 August; China was launching again six days later [6]

In play LandSpace — the Beijing startup that landed the Zhuque-3 booster CASC — China's state rocket maker, which caught a booster at sea in July SpaceX — still flies most of the world's launches, but is winding the Falcon 9 down Satellite operators — holding finished satellites they cannot get to orbit The White House — set a target of 1,000 launches and re-entries a year by 2030

How it unfolded

Dec 2025 Zhuque-3's first flight reaches orbit, then crashes on landing [3]
10 Jul A Long March 10B booster is caught in a net at sea, China's first recovery [5]
10 Aug A Long March 7A explodes in flight and is lost with its satellite [6]
18 Aug Zhuque-3 flies again and lands on legs in the Gobi Desert [2]
20 Aug Washington publishes a policy aiming at 1,000 launches and re-entries a year [11]
Next LandSpace inspects the booster and says it wants to fly it again [1]

Where this points

Watch whether either recovered Chinese booster actually flies a second time - a landing is a photograph, and reuse is only proved by the next flight [1].

Full briefing

The only trick that ever made orbit cheaper

Flying the same rocket twice is the only change that has ever made reaching orbit meaningfully cheaper, and until this summer only American companies could do it [5]. That is why LandSpace’s landing is not really a national-pride story. Western satellite firms are running out of rides. SpaceX has signalled it means to stop selling commercial launches after 2028, and shared launch slots are already unbookable past late that year [9][15]. A second country that can land and refly boosters changes how much room there is in orbit for everyone.

Subtraction, not power

The interesting part is how LandSpace got there. Its first Zhuque-3 reached orbit last December and then crashed while trying to land [3]. The company did not answer with more thrust. It cut the number of engines relit for the final phase of the descent to one [1]. It also redesigned the heat shield, and taught the onboard safety system to work out where the booster would come down [2][3]. Dai Zheng, who runs the programme, called the result a “one-shot reverse-parking” move at high speed [1]. Video afterwards showed a fire in the rocket’s tail once it was down, which the company has not explained [3].

The rocket’s own numbers disagree

Reports differ on the machine itself. SpaceNews and Ars Technica put the rocket at 66 metres tall; Spaceflight Now says 76.6 [2][1][3]. The payload figures split the same way. Ars describes the current design lifting a little over 8 tonnes to low orbit when the booster is saved, while SpaceNews quotes 18.3 tonnes [1][2]. Ars attaches that larger figure to a planned upgrade with more fuel and stronger engines, so both are likely right about different rockets [1].

A target is not a rocket

Washington answered the week with a goal rather than a vehicle. The new National Space Transportation Policy, the first rewrite since 2013, asks for more than 1,000 launches and re-entries a year from American soil by 2030 [11][12]. Phil Larson, who helped write the previous version, called it “evolution rather than a major policy reset” [11]. He also pointed out that most of it needs laws and money Congress has not yet provided [11].

Money is not the part that is short. Muon Space raised $250 million this week at a $1.5 billion valuation, and Starcloud raised $250 million with Nvidia joining in [13][14][15]. Both are buying factories for satellites that currently have nowhere to go [10].

The rest of the week

China’s Chang’e 7 was due to leave for the Moon’s south pole this weekend. It is aimed at the rim of Shackleton Crater, where permanent shadow may have trapped water ice for billions of years [17][16]. In the same week, NASA scientists published a study saying microbes carried by astronauts could survive in those shaded pockets [18]. A human carries about a million bacteria on each patch of skin the size of a pencil eraser [18]. The worry is that the ice we travel to read may end up carrying our own signature.

02 · Lesson · why it matters

Why nine engines are a safety net going up and a hazard coming down

An extra part is either a spare or one more thing that must not fail. Which one it is depends on the job.

How it works

A landing needs several things to work in a row
Each one has its own chance of working
The chance of all of them working is those chances multiplied
Multiplying numbers below one always shrinks the total
So removing a required step lifts the odds without improving any part

The twist

Reliability is a product, not an average - which is why asking a machine to do less can help more than making any single part of it better.

Where you've seen this

Getting a prescription filled

doctor, pharmacy, insurer and delivery are each usually fine, and all four are required

A recipe with fifteen steps

cutting two steps helps more than perfecting one

Sign-offs at work

every extra signature is one more chance the thing stalls

The catch

It flips when spares are involved: nine engines that must all fire is a risk, but nine where you only need eight is a safety net.

Full lesson

The fix that sounds backwards

LandSpace’s rocket reached orbit last December and then crashed trying to land. Eight and a half months later the same design came down on its legs in the Gobi Desert.

Between those two flights the company did something that reads like a typo. For the last and most delicate part of the descent, it lit fewer engines than before. One, instead of several.

That is not a Chinese trick, or a rocket trick. It is arithmetic, and it is one of the least intuitive things about how machines and organisations actually behave.

Chances multiply, they do not average

A landing is a sequence. The booster flips over. It fires to slow down. It rides through the heat of re-entry. It steers to the pad. It fires again to brake. Its legs unfold.

Every one of those has to work. Not most of them. All of them.

So the chance the landing works is not the average of those steps. It is the product: each chance multiplied by the next. And multiplying numbers smaller than one always shrinks the total.

Take six steps, each with a nine-in-ten-plus chance of going right - say ninety-five percent. That feels close to certain. Multiplied out, it is seventy-four percent. Stretch the same quality of work across ten steps and you are down to sixty. Nothing got worse. There was just more of it.

Why deleting beats perfecting

This is where the counterintuitive bit lands.

Suppose you have those six steps at ninety-five percent. Overall, that is a 73.5 percent chance of getting down in one piece. Now you get one improvement.

Pour everything into making a single step nearly flawless - ninety-nine percent - and the whole rises to 76.6. Redesign so that step is not needed at all, and it rises to 77.4.

Removing a step will always beat perfecting it. No amount of engineering gets a step past a hundred percent, and not needing it is exactly what a hundred percent means. Every engine you do not relight is an engine that cannot fail to relight.

The same parts, the opposite meaning

Now the flip, and it is the part worth carrying.

On the way up, a rocket with nine engines is safer than one with a single engine, because it can lose one and still reach orbit. The job needs any eight of nine. Run the numbers with each engine at ninety-five percent and the climb succeeds about ninety-three times in a hundred.

Coming down, if all nine have to relight, the job needs all nine. Same nine engines, same ninety-five percent each. Now it works about sixty-three times in a hundred.

Nothing about the hardware changed. What changed is whether the task needs all of them or any of them. Extra parts are a safety net in the first case and a pile of new risks in the second, and they look identical sitting on the pad.

The chains nobody counted

Most people never light an engine, and sit inside these chains all the time.

A prescription that needs a doctor, a pharmacy, an insurer and a courier. A visa with seven offices. A benefits claim, a mortgage, a repair booked through three companies. Every one of those steps is run by someone who can truthfully say that their part almost always goes through.

Nobody in the chain is looking at the product. Each office sees its own link, and its own link looks fine. The arrangement itself - how many gates there are, and who added each one - was rarely decided by anyone in one go. It accumulated, one reasonable-looking approval at a time, and it usually serves whoever added the gate more than whoever has to pass through it.

And the count is invisible from inside. You find out how many steps there were only when one of them closes.

What the engineers were counting

The team that landed the booster in August were not better at engines than the team that lost the first one. Nothing in the day’s reporting suggests they were.

They were counting a different thing. Not how good each part was - how many parts had to be right.

That is a small change of view, and it does not make anyone clever. If anything it should make the chains we all live inside look less like carelessness and more like arithmetic that nobody was standing far enough back to do.

03 · Lab · your turn

Land the Booster

Rehearse why a chain of required steps gets less likely with every step, and why removing one beats perfecting one.

04 · Hope · carry this

LandSpace turned its own crash into a landing in eight and a half months. Hard things get faster once somebody has shown they can be done at all.