Information Technology · Saturday, 15 August 2026

01 · Briefing · what happened

A memory chipmaker just passed Tencent as China's most valuable company

Information Technology 5 min 16 sources

ChangXin Memory is worth $524bn, more than the internet giant that buys its chips. Flash and DRAM makers are now the best-paid layer in computing - and the bill is showing up in phone prices, PC parts and the cost of an AI answer.

$524bn

CXMT market value

ahead of Tencent's $511bn

48%

of world flash to data-centre drives

up from 26% a year earlier

12GB

memory in the Pixel 11 Pro

down from 16GB, at $100 more

$3.96

DeepSeek's new peak price

per million chunks of output, up from $0.87

At a glance

ChangXin Memory, a Chinese maker of the working-memory chips that sit beside a processor, passed Tencent to become China's most valuable company at $524 billion.
It got there partly because Tencent fell - the internet giant disclosed computing spending up 176% and free cash flow going negative.
Tencent buys memory from CXMT, so investors punished the buyer and rewarded the seller over the same orders.
Flash storage is the other winner: enterprise SSDs took 48% of every flash bit shipped worldwide last quarter, up from 26% a year earlier.
Much of that storage holds nearby copies of data an AI model has already read, so it does not have to fetch them again.
Intel's chief executive says memory has stopped being a commodity and hints at stacking it directly onto a processor.
Consumers pay the bill: the Pixel 11 Pro ships with 12GB of memory instead of 16GB and costs $100 more.
DeepSeek is raising the price of an AI answer more than fourfold, with peak and off-peak rates from August 16.

Forces in play

Memory scarcity High↑

Kioxia's entire 2026 flash output is already sold out

Buyers locking supply High↑

four big cloud firms have committed roughly $2 trillion, much of it memory

Consumer squeeze Building↑

Pixel prices up $100 with less memory; older Intel chips selling again

New supply coming Easing↑

SK hynix restarting Dalian expansion, Intel eyeing a return to memory

In play CXMT — Chinese DRAM maker, now China's most valuable company Tencent — its customer, and the company it just overtook YMTC — broke into the world's top three flash makers Intel — hinting at a return to the memory business Google — shipping a phone with less memory at a higher price

How it unfolded

Jul 27 CXMT lists in Shanghai, closes day one up 466%
Mon analyst estimate puts big-cloud purchase commitments near $2 trillion
Wed Tencent reports computing spending up 176% and shares fall; YMTC enters the global flash top three
Thu CXMT passes Tencent to become China's most valuable company
Aug 16 DeepSeek's price rise takes effect

Where this points

Watch whether the new supply lands before the shortage does - SK hynix's Dalian expansion and Intel's hinted return are both years out, so the near-term test is whether phone and PC prices keep climbing into next year.

Full briefing

The supplier passed the customer

ChangXin Memory Technologies makes DRAM - the working memory chips that sit next to a processor and hold whatever it is using right now. On Thursday it became the most valuable company in China, worth $524 billion against Tencent’s $511 billion [1].

CXMT listed in Shanghai on July 27 and closed its first day up 466% [1]. It is the first chip company to top mainland China’s stock market in the exchange’s 35-year history [1]. Its market value is now roughly half of Micron’s $1 trillion and about 60% of SK hynix’s $880 billion [1].

The crossover was not a rally. CXMT’s own shares fell 1.2% that day [1]. Tencent fell further, after reporting that its spending on computing rose 176% to 52.8 billion yuan in the quarter, pushing free cash flow negative [1]. Investors punished the buyer and rewarded the seller over the same purchase orders. Tencent is a CXMT customer, having signed a $3 billion server memory deal in June [1].

What the machines actually eat

Yangtze Memory Technologies shipped 14% of the world’s flash storage last quarter, entering the global top three for the first time, according to Counterpoint Research [2]. Samsung led with 25%, SK hynix followed at 22% [2].

The demand behind that is specific, and it is worth understanding. Enterprise SSDs - the fast storage that sits just under a data centre’s memory - absorbed 48% of every flash bit shipped worldwide, up from 26% a year earlier [2]. A large share of that capacity holds nearby copies of things an AI model has already read, so the machine does not have to fetch them again [2]. Kioxia’s entire 2026 output is sold out [2]. Samsung has capped flash production to make higher-margin memory chips instead [2].

The money is following. Alphabet, Microsoft, Meta and Amazon have purchase commitments totalling roughly $2 trillion, a large share of it memory, by one analyst’s estimate [3]. Meta’s commitments sit near $349 billion and Amazon’s near $130 billion [3]. Apple - until recently the industry’s biggest memory buyer - is flat at about $57 billion [3].

Sandisk shares are up more than sixfold this year [4]. It has pre-sold about half its fiscal 2027 storage output to eight customers, three of them big cloud firms, and expects gross margins near 80% through 2030 [4]. SK hynix is reportedly restarting expansion at its Dalian flash plant to lift output 50% by 2027 [5]. The plumbing is catching up too: the first PCIe 6.0 drives are reaching market, four years after the standard was ratified [6].

Intel’s chief executive, Lip-Bu Tan, has changed his mind about a business he used to call a commodity. “Now it has become different,” he said, hinting at stacking memory directly on top of a processor [7]. He has hired Seok-Hee Lee, who used to run SK hynix [7].

The bill reaches your phone

Google’s new Pixel 11 phones all start $100 above last year’s equivalents [8]. The Pixel 11 Pro and Pro XL ship with 12GB of working memory, down from 16GB in the two previous generations [8]. Google blamed “a severe, supplier-driven RAM shortage” and said it is engineering around it by changing how on-device models load [8].

PC buyers feel it too. Intel says a “sudden inrush of demand” has hit its older Raptor Lake processors, which high DDR5 memory prices have made attractive again [9]. The company says that line stays a “core part of the portfolio” for years [9].

Cheap answers stop being cheap

DeepSeek, the Chinese lab that built its name on undercutting Western prices, is raising them. From August 16 its V4 Pro model costs $3.96 per million output tokens at peak hours, up from $0.87 [10]. A token is a chunk of text a model produces, a bit under a word on average. Off-peak is half that price. The cheaper V4 Flash goes from $0.28 to $1.32 at peak [10]. The company said the split is meant to “allocate resources more reasonably” [10]. It released V4 Pro formally on Thursday, saying it improves the model’s ability to run tasks on its own [11].

Google went the other way, for now. Gemini 3.7 Flash launched Thursday at $0.75 per million input tokens and $3.75 per million output tokens - a discount that doubles on January 1, 2027 [12].

Capacity keeps being booked ahead of the demand. IBM and Together AI signed a $240 million multi-year deal for an Nvidia-powered cluster running open models [13]. Elon Musk told SpaceX staff that xAI’s data-centre power will grow sevenfold, to 10 gigawatts, by late 2027 [14]. He claimed that would bring $300-500 billion in annual revenue - a company projection, not a result [14].

Two ways a nearby copy goes wrong

Researchers at Germany’s Helmholtz Center for Information Security found a flaw in processors made by China’s Loongson [15]. The chips leak data out of their L1 cache - the smallest, fastest memory on a chip, where whatever is in active use sits [15]. Because that cache is not walled off between programs, an attacker can read another program’s data, including full-disk encryption keys held by the operating system [15]. The leak crosses out of a virtual machine to the host [15]. There is no software fix; owners must replace the chips or keep private data out of that cache [15].

At the far end of the same hierarchy, Nine PBS says it lost access to 50 terabytes of archive - about 70 years of television [16]. Its cloud storage vendor went out of business without warning [16]. It is suing the data-centre company that houses the material to make sure it survives [16].

02 · Lesson · why it matters

Why your computer is only fast when it already has what you need

A computer is fast because what it needs is usually close by - and this year's machines need things that never are.

How it works

Memory is a pyramid: on-chip, then main memory, then storage, then the network
Each layer down is far slower and far bigger
A machine keeps a nearby copy of what it just used
That bet works because most programs reuse the same small set of data
An AI model has almost no reuse - it reads everything, every time
So the bet fails, and the whole pyramid has to be bought bigger and closer

The twist

A computer is fast not because the processor is quick but because what it needs is usually already close - and the moment that stops being true, no amount of processor speed can save it.

Where you've seen this

A kitchen

salt on the counter, flour in the cupboard, the rest at the shop - you arrange it by how often you reach for it

A hospital ward

the trolley holds what is used hourly; the pharmacy holds the rest, minutes away

A library desk

the books you are working from stay open in front of you, not reshelved after each glance

The catch

A nearby copy is only worth having while it is still true - the hard part was never filling it, but knowing when what it holds has gone out of date.

Full lesson

A ladder, not a box

Ask most people where a computer keeps its data and they picture one place. It is a ladder.

At the top sit a few dozen slots inside the processor itself, holding the numbers it is working on this instant. Below that, a small pool of very fast memory on the same chip. Below that, the main memory sticks. Below that, storage. Below that, the network, and whatever is sitting in a building in another country.

Every rung down is roughly ten times slower than the one above and vastly bigger. The gap from the top to the bottom is not a matter of a few percent. It is the difference between reaching for the pen in your hand and posting a letter.

The bet everything is built on

Nothing would work if a machine had to walk the whole ladder for every piece of data. So it does something simpler: it keeps a nearby copy of whatever it just used.

That trick only works because of a habit programs have. Whatever a program touched a moment ago, it will probably touch again; whatever sits next to that in memory, it will probably want next. Reading a file, drawing a window, running a spreadsheet - all of it circles a small pool of the same data, over and over.

So the nearby copy is a bet: the past predicts the future. For fifty years the bet has paid off so reliably that most people who use computers, and plenty who build them, never think about the lower rungs at all. Speed was something you bought from the processor. The ladder was hidden.

What happens when the bet loses

An AI model does not have the habit. It does not circle a small pool. To produce one word, it streams billions of numbers it will not reach for again until the next word, when it streams them all once more.

There is no small hot set to keep nearby. The bet fails on almost every step. And when the bet fails, the machine runs at the speed of the slowest rung it has to reach. The processor’s speed stops mattering, because it is standing still, waiting.

That is why the industry stopped trying to hide the lower rungs and started buying them. It could not make the bet work, so it made the ladder shorter: more memory, closer in, faster underneath. A DRAM maker is now worth more than the internet giant that buys its chips. Storage drives that hold what a model has already read swallowed nearly half the world’s flash last quarter, up from a quarter a year ago. The lower rungs became the product.

The ladder is a purchase, not a law

It looks like physics. It is a budget.

How much fast memory sits beside a chip is not fixed by nature - somebody decided how much to fit, and what it costs to fit it. Right now a small number of very large buyers have committed to roughly two trillion dollars of hardware, much of it memory, years in advance. That commitment is what sets the price of the memory in everyone else’s machine.

Which is how a phone gets less memory than last year’s model and costs a hundred dollars more. And how a five-year-old processor becomes desirable again, because the newer memory it would need has got too expensive. Nobody made a decision about that phone buyer. The shape of the ladder was set well above them, and it reached down.

The hard part was never filling it

Keeping a copy nearby is easy. Knowing when it has stopped being true is the difficult half.

Both ends of the ladder showed the same failure this week. On a Chinese processor, the tiny store of data in active use turned out not to be walled off between programs. One program could read another’s secrets out of it. The copy was near, fast, and not as private as everyone assumed. At the other end, a broadcaster found that fifty terabytes of archive, kept safely far away, had become unreachable. The company holding it had quietly stopped existing.

A near copy you cannot verify and a far copy you do not control fail in the same way. You trusted something that was no longer what you thought it was.

Where everyone is standing

The ladder used to be an engineer’s concern. It is now a price everyone pays, mostly without seeing it.

The person choosing a phone this month is standing on it. So is the one buying an older chip for a build, and the developer whose cost per answer more than quadruples on Sunday. So is the broadcaster whose seventy years of tape sat on a rung it did not own. All of them are being moved by decisions taken several rungs above where they stand.

And almost nobody can see the whole of it from where they stand. The phone buyer does not know that a listing in Shanghai three weeks ago is part of why the price moved. The engineer optimising a model does not see the archive going dark. Every seat on the ladder gets a clear view of one rung and a rumour about the rest.

03 · Lab · your turn

Keep it near

Rehearse how a machine decides what to hold close, and feel what happens when nothing is ever asked for twice.

04 · Hope · carry this

Every era of computing has hit a wall like this one, and every time the answer came from someone deciding to build the boring layer nobody wanted. Memory is having that moment now.