Information Technology · Thursday, 6 August 2026

01 · Briefing · what happened

The AI build-out hits the wall money cannot buy past

Information Technology 2 min 13 sources

Cloud giants are pouring a record $600 billion into AI compute this year, and still say it is not enough. The ceiling has moved to the parts you cannot just buy more of - memory, and the wiring between chips.

$600bn

combined AI capex

Amazon, Google, Microsoft, this year

$220bn

Amazon's 2026 plan

raised from $200bn on memory costs

30%

optics-material shortfall

below customer need, Lumentum says

~$500bn

Apple value lost

after a weak forecast amid the memory crunch

At a glance

Amazon, Google, and Microsoft are on track to spend near $600 billion on AI data centers this year.
They still say they cannot add capacity fast enough - the money is not the limit.
The ceiling has moved to memory and to the wiring between chips, which you cannot just buy more of.
A memory shortage is now bleeding into consumer laptops and even Apple's next iPhone.
Chip-to-chip data movement is the new slow step, pushing the industry toward optics.
Meta joined the AI coding-agent race; China keeps pressing on cheap, open models.

Forces in play

AI capex High

near $600bn, still called not enough

Memory shortage High

shortage now reaching laptops and iPhones

Interconnect strain Building

data movement between chips is the slow step

Chip supply Easing

accelerators are plentiful; the squeeze is elsewhere

In play Amazon (AWS) — raised capex to $220bn, cites memory costs Google — supply-constrained; could out-build Nvidia on accelerators Memory makers — Samsung, and a shortage pulling in Chinese suppliers Meta — entered the coding-agent race with Muse Code

How it unfolded

Q2 earnings cloud giants lift capex again, still cite shortages
This week memory crunch hits laptops and Apple's forecast
This week Meta launches Muse Code; DeepSeek posts cheapest run cost

Full briefing

Amazon, Google, and Microsoft are on track to spend close to $600 billion on data centers this year, almost all of it for AI [1]. And they keep saying the same strange thing: it is not enough.

Amazon raised its 2026 spending plan from about $200 billion to $220 billion, and its chief Andy Jassy blamed “the higher cost of memory” for the bump [1]. Even then, he said, “we will still not have enough capacity to meet all the demand we have in 2026,” and expects the same in 2027 [1]. Google lifted its own guidance to $195-205 billion and called the market “supply-constrained,” with demand still outpacing the build [1].

Notice what is capping them. Not chips - Google alone could build more AI accelerators than Nvidia sells by 2028 [8]. The squeeze is on the parts every chip depends on. An “unprecedented memory shortage” has pushed HP, Asus, and Acer to slip small amounts of Chinese memory into some laptops [2]. Apple is struggling to source enough memory for its iPhone 18 Pro [3], and lost nearly $500 billion in market value after a weak forecast [4]. Samsung, meanwhile, is racing out next-generation AI memory to catch the wave [5].

The other ceiling is the wiring. As you pack thousands of chips together, moving data between them becomes the slow step. So the industry is shifting to optics, sending data as light instead of electrical signals over copper [7]. But Lumentum’s chief warns the material that optics depends on, indium phosphide, faces a shortfall that “will become worse than memory,” already running about 30% below what customers need [6].

The coding-agent race adds a new player

Meta released Muse Code, a terminal-based AI coding agent, plus an updated Muse Spark 1.2 model, putting it head-to-head with Anthropic’s Claude Code and OpenAI’s Codex [9][10]. Mark Zuckerberg said it “fans out to separate sub-agents working in parallel” on big jobs, and built six game features at once “with no collisions” [9]. Meta’s pitch is cost: AI chief Alexandr Wang called it “an incredibly good option, especially from a cost perspective” [10]. Yet firms already running these agents report the bills climbing fast - the hard part, they say, is still the human review the agents cannot do [13].

China keeps pressing on cheap, open models

A version of DeepSeek’s latest model is by far the cheapest well-known model to run. On tests it costs about 105 times less than one Anthropic model, though it still trails the top US labs on quality [11]. Hugging Face’s chief executive Clement Delangue said China is “clearly dominating on open models right now” and could reach the frontier “by the end of this year or next” [12].

02 · Lesson · why it matters

You cannot buy your way past the step that will not split

Parallel horsepower only speeds up the part of a job that can split - the step that will not split sets the ceiling money cannot lift.

How it works

A big job splits into a parallel part and a serial part
Add more workers and the parallel part shrinks fast
But the serial part does not - one worker still owns it
Its share grows until it dominates the total time
So the whole speedup hits a hard ceiling, set by the serial slice

The twist

You cannot buy your way past the step that will not split - past a point, adding parallel horsepower barely moves the finish line.

Where you've seen this

A dinner party

ten cooks chop fast, but one oven still bakes the roast in its own time

A road trip

more lanes speed the highway, but a single toll booth caps how fast everyone arrives

A newsroom

many reporters file at once, but one editor's final read sets the pace

The catch

The ceiling only bites once the serial slice is real - shrink or remove it and parallelism pays off again, which is exactly what the optics and memory race is trying to do.

Full lesson

The strange complaint of the richest companies on earth

Three companies will spend close to $600 billion this year, and their message is not triumph. It is frustration. They cannot go fast enough. Amazon’s chief said flatly that even at $220 billion, “we will still not have enough capacity.”

Read that twice. The bottleneck is not money. It is not even chips - Google could out-build Nvidia on AI accelerators within two years. The wall is somewhere else, and it is worth understanding, because it is one of the oldest facts in computing.

The half of the job that refuses to split

Picture any big task as two parts. One part can be split among many workers at once. The other part must happen in sequence - one thing after another, no matter how many hands you have.

Ten cooks can chop vegetables ten times faster. But the roast still bakes in the oven for its own hour, and a hundred cooks will not shorten it. The oven is the part that will not split.

Computing calls this Amdahl’s law, after the engineer who wrote it down in 1967. The rule is simple and unforgiving. The speedup you can win by adding parallel capacity is capped by the fraction of the job that stays serial.

Why the ceiling is lower than it looks

Here is the part that surprises people. Say a job is ninety percent splittable and ten percent serial. It feels like you should get close to a tenfold speedup with enough machines.

You do not. As you add workers, the splittable part shrinks toward nothing - but the serial tenth stays exactly where it was. Soon that tenth is almost the entire runtime. Even with infinite machines, you cap at ten times faster. The serial slice alone decides the limit.

That is why the cloud giants keep spending and keep saying it is not enough. Past a point, another rack of chips barely moves the finish line. The work those chips are waiting on - fetching data from memory, shuttling it between chips - is the serial slice, and you cannot add your way out of it.

The bottleneck moved, and the whole industry turned with it

This is why the frontier of the race has quietly shifted. Not to more chips - to memory and to wiring. A memory shortage has grown so tight it is now bleeding into laptops and phones. Engineers are racing to send data between chips as light instead of electrical signals, because the copper links became the slow step.

None of that is glamorous. It is plumbing. But the plumbing is where the ceiling lives now, so the plumbing is where the smartest money and the hardest problems have gone.

Who is standing under this ceiling

It is easy to read this as a story about a few trillion-dollar firms and their spending. It is not only theirs. The same memory shortage that caps their data centers is the reason a laptop or a phone costs more this year. The serial slice does not respect the size of the buyer.

And the arrangement underneath is worth naming. When a company reports that it “cannot add capacity fast enough,” that sounds like a fact about the world. Part of it is a choice about where to point the spending. Pouring money into the part that scales, rather than the part that will not, is a choice every organization makes in its own small way. We all do it. We hire more hands for the work we can divide, and quietly hope the one step that will not divide speeds up on its own. It rarely does.

The humbling part is how little any single seat can see of the whole. The company sees its capex. The laptop buyer sees a price. The engineer sees a copper link that will not go faster. Each is standing on a different part of the same ceiling, and none of them can lift it alone.

03 · Lab · your turn

Buy Past The Wall

Rehearse how adding parallel workers hits a hard ceiling set by the serial slice of a job.

04 · Hope · carry this

The wall sends our best minds to the unglamorous hard part - memory, light, the plumbing nobody claps for. That is the kind of problem people have always, eventually, solved.