Information Technology · Wednesday, 12 August 2026

01 · Briefing · what happened

The AI industry is spending billions to kill the slowest step in an answer

Information Technology 3 min 16 sources

AMD, Cerebras, and Together AI are all pouring money into one thing - making an AI reply faster - and the real enemy is the rare slow moment, not the average one.

17,000

words per second

AMD's new etched chip, 48x an Nvidia GPU by its own claim

$240M

inference chips rented

Together AI's Nvidia capacity via IBM's cloud

$38bn

one memory factory

SK Hynix, feeding the chips that run AI

421

flaws patched

Microsoft's monthly fix; one already under attack

At a glance

AMD bought Taalas, a startup that etches AI models straight into silicon to cut the wait for an answer.
Cerebras (single-wafer chips) and Together AI ($240M of Nvidia capacity) are chasing the same goal: faster, cheaper replies.
The shared enemy is the cross-chip tax - big models split across many chips, and every word waits on the slowest hop.
Meta open-sourced Muse Glimmer and Nvidia released its first open model, Nemotron, reopening the open-weight fight.
SK Hynix will spend $38bn on a memory-chip factory; the big clouds have committed close to $2 trillion to AI capacity.
A rough week for uptime: a Ceva Logistics breach rippled to banks and retailers, Microsoft patched 421 flaws, LexisNexis went dark for days.

Forces in play

Speed race High↑

AMD, Cerebras, Together AI all spending to shrink the wait for an answer

Cross-chip tax Building→

big models split across many chips; every word waits on the slowest hop

Build-out cost High↑

$38bn for one memory plant, ~$2 trillion in cloud commitments

System fragility Building↑

Ceva breach, 421 Microsoft flaws, LexisNexis outage in one week

In play AMD — bought Taalas to etch models into silicon and speed up answers Cerebras — keeps a whole model on one wafer to skip the cross-chip tax Together AI — committed $240M of Nvidia chips for cheaper inference Meta / Nvidia — both released free, downloadable AI models this week

How it unfolded

Wed Cerebras signs Lovable onto its single-wafer chips for speed-sensitive work
Thu AMD acquires Taalas to etch models into silicon
Mon Meta open-sources Muse Glimmer; SK Hynix confirms $38bn memory plant
Tue Together AI signs $240M IBM deal; Nvidia releases open Nemotron model

Where this points

Watch whether AMD's etched-silicon bet actually ships at scale next year - the whole approach only pays off if it holds up under real, unpredictable load, not just a benchmark.

Full briefing

The loudest tech news this week was Meta open-sourcing a powerful AI model. The most consequential was quieter: a spending race to make AI answers fast. AMD bought a startup that etches models straight into silicon [1][2]. Cerebras signed a coding platform onto its single-wafer chips [3]. Together AI committed $240 million to rent Nvidia capacity from IBM [4]. All three are chasing the same thing - cut the wait between you and the model.

The race to make AI answer faster

An AI answer is not one step. It is a chain of steps, and each one waits on a chip. Today’s big models are too large for one chip, so they are split across many, which have to talk to each other for every word generated. That cross-chip chatter is a tax on speed, and it is where the wait hides.

On Thursday AMD acquired Taalas, a Toronto startup that bakes a model’s weights directly into a chip instead of loading them from memory [1][2]. Its first test chip served an older Meta model at nearly 17,000 words a second - 48 times faster than an Nvidia GPU, its maker claims [1]. AMD plans to pair the chips with its server racks, doing the heavy first pass on GPUs and the fast word-by-word generation on the etched silicon [1].

The same week, Cerebras signed Lovable, a tool where 50 million projects have been built, to run its speed-sensitive work [3]. Cerebras keeps a whole model on one dinner-plate-sized wafer, which avoids the cross-chip tax entirely. And Together AI rents out inference - the step where a trained model actually answers - to other companies. It locked in $240 million of Nvidia chips through IBM’s cloud, chasing the “lowest cost per word” [4]. The mechanism under all three is the same: fewer chips talking to each other means less waiting.

Meta and Nvidia reopen the open-weight fight

Meta released Muse Glimmer, a freely downloadable model, and promised to open its more powerful Muse Spark 1.2 within weeks [5][6]. Mark Zuckerberg paired it with a 6,000-word essay arguing that open models, which anyone can run and inspect, should be the default [5]. Nvidia - better known for selling the chips - put out its own first open model, Nemotron, the same week [7]. Open weights matter because they let companies run AI on their own machines instead of paying per query. LTX also shipped an open video model that turns a still image into a 10-second clip in under 7 seconds [8].

The build-out keeps eating money and chips

The speed race runs on hardware, and the hardware bills keep climbing. SK Hynix said it will spend $38 billion on a new memory-chip factory [9]. That memory feeds AI chips, and it is one of the tightest bottlenecks in the whole supply chain. The big cloud companies have now committed close to $2 trillion to lock in AI capacity [10]. South Korea is standing up a $3.5 billion fund to speed its own chip hubs [11]. And the US moved to wall off part of the supply chain. The FCC proposed banning imports of Chinese optical transceivers, the small parts that shuttle data between machines inside a data center [12].

A rough week for keeping systems up

The other half of running at scale is not breaking. A breach at shipping giant Ceva Logistics rippled outward to banks, retailers, and even Steam gamers whose data passed through it [13]. Microsoft’s monthly patch fixed 421 separate flaws, and North Korea’s hackers were already exploiting one of them [14]. LexisNexis pulled three services offline for days after “unusual server activity” on a third-party system [15]. And Mozilla had to revoke a Firefox signing key after an unencrypted copy leaked, a reminder that one exposed key can undo the trust in everything downstream [16].

02 · Lesson · why it matters

Why "usually fast" is not fast at all

When one thing needs many parts, its speed is set by the slowest part that shows up - and at large scale, a slow one always shows up.

How it works

One request needs many services, each usually fast
But 1-in-100 of them is randomly slow
String 100 together and almost every request hits a slow one
So the rare slow moment - the tail - sets how the whole thing feels
Fixing the average does nothing; you have to fix the worst case

The twist

At scale, 'usually fast' is not fast. The rare straggler, not the average, decides how the system feels - so you engineer the worst case.

Where you've seen this

Airport security

most lanes move; the one jammed lane sets your memory of the whole airport

A group project

the deadline is set by the slowest member, not the average one

A supply chain

the shipment arrives when the last, slowest part does

The catch

Chasing the tail has limits - past a point, shaving the worst case costs more than the speed is worth, and a rare stall you can't see is still there.

Full lesson

The number that lies

Ask an engineer how fast a system is, and you often get an average. The average is comforting. It says: most of the time, things are fine.

But you do not live in the average. You live in single moments - this click, this reply, this page. And the average hides the moment that ruins your day.

Say a service answers in a tenth of a second, almost always. One time in a hundred, something snags and it takes two full seconds. The average barely moves. Your experience does not run on the average. It runs on the two-second stall you actually hit.

Why rare slowness stops being rare

Here is the trap that the AI speed race is really fighting. A modern answer is not one step. It is a chain of them.

To reply, a big model is split across many chips, and they pass work between each other for every word. A single web page might call on a hundred separate services behind the scenes. Each one is usually quick.

Now do the arithmetic. If one call in a hundred is slow, and a page needs a hundred calls, then almost every page will hit at least one slow call. The rare event stops being rare the moment you stack enough of them. At scale, “one in a hundred” becomes “nearly every time.”

This is why the average is a liar. The thing you feel is not the typical step. It is the worst step in a long chain - the straggler. Engineers have a name for it: the tail. It is the slow edge of the distribution, the last few percent, and it is what decides how the whole system feels.

The whole race, seen through this

Now the week’s news reads differently. AMD did not buy a chip company for bragging rights. It bought a way to keep a model on fewer chips, so there are fewer hops where a straggler can hide. Cerebras keeps a whole model on one wafer for the same reason. Together AI is renting capacity so no request has to wait in a long line.

None of them are chasing a faster average. A faster average is easy and nearly useless. They are all hunting the tail - the rare slow moment that, multiplied across billions of requests, becomes the thing every user remembers. Billions of dollars, aimed at the last few percent.

Who chose the average

There is a quiet arrangement underneath this. Somebody decided which number counts as “the speed.” For a long time, that number was the average, because it is easy to measure and easy to make look good.

The average serves the person reporting it. It smooths away the bad moments into a tidy figure. The person living inside the system - waiting on the stalled page, the frozen app, the reply that never comes - feels the tail, not the average. The gap between those two numbers is the gap between how a system is sold and how it is lived.

You are already inside this

This is not only a data-center problem. You are the slow call in someone else’s chain, and you are the one waiting on someone else’s.

The group project finishes when its slowest member does, not its average one. The airport is remembered for the one jammed lane. The dinner is late because of the one dish that would not cook. A team, a road, a supply line - each is only as quick as the slowest part that turns up on the day. Add more parts, and the odds that one of them stalls climb toward certainty.

The lesson is not that slowness is everywhere. It is that scale changes the math. When many things must go right at once, the rare failure is no longer rare - it is the normal case, wearing a disguise. Planning for the average is planning for a world you do not live in.

The catch

Chasing the tail has no clean finish. Shave the worst one-in-a-hundred and a worse one-in-a-thousand waits behind it. Past some point, the effort to remove a rare stall costs more than the speed is worth.

And a stall you cannot see is still a stall. The person who hit it does not care that they were unlucky. From inside the system, no seat sees the whole shape of it - only the moment it landed on them. Seeing that shape, and how little of it any one of us watches, is most of the humility here.

03 · Lab · your turn

Chase the tail

Rehearse spending one upgrade on the average versus the rare straggler, and feel why the tail sets how a system feels at scale.

04 · Hope · carry this

Most of this work goes into moments no one notices - the stall that never happens, the reply that just arrives. There is a quiet decency in years spent smoothing an instant you will never think about.