WatchReal vs Hype14 Aug 20261:56 watch4 min readSemiconductors: chips, fabs & yield

The Chip So Big They Can't Throw the Broken Part Away

This AI Chip Is the Size of a Dinner Plate — And It Breaks Every Rule of Yield

A full wafer-sized silicon die lying on a bench, one corner region catching warm gold light, the rest cool

The picture for this piece. Made for it: no people, no brands.

In plain words

Every chip factory throws the broken chips away. Cerebras built a chip too big to throw anything away.

The normal rule is simple: cut the wafer into hundreds of chips, test them, and discard the bad ones. That rule is what yield means.

Cerebras makes a processor that is the whole wafer, about the size of a dinner plate. There is nothing to cut off, so there is nothing to discard.

So they flipped the rule. Instead of designing against defects, they designed for them.

The chip carries about 900,000 tiny cores. If a defect kills one, the fabric routes around it and the wafer keeps working.

Going wafer-scale also let them skip the memory stacking and packaging steps that are the real bottleneck for other AI chips.

Picture it. Instead of throwing away the apples with a bruise, they baked a pie that tastes fine with a few bruised ones in it.

Why it matters to you. When a rule is a hundred years old, ask what it was protecting. Sometimes the constraint moved and the rule did not.

Watch on the channelOpen on YouTube

The number

900,000

CORES ON ONE WAFER · CEREBRAS WSE-3 · 2026

What to do Monday

Claude Code v2.1.232 (shipped 13 Aug 2026): your sessions can now message each other. Type @ in a prompt and pick another live session by name. Subagents also fork by default now — they inherit your full conversation and prompt cache and run in the background rather than blocking you. Then set /config → "Messages from your other sessions" to HOLD on whichever session is doing deep work, so inbound messages queue instead of interrupting. Full walkthrough in the video.

The honest caveat

Cerebras built for one kind of workload, and routing around defects costs area and power that a small chip does not pay. Design for the defect is a luxury of scale. It is not a rule most fabs can copy.

Your turn

Genuine question for the yield people: is "design for the defect" a wafer-scale luxury, or should more of us be building redundancy in instead of screening it out?

Argue with me on LinkedIn

Sources

PrimaryCerebras Systems Q2 2026 results, announced 12 August 2026 (GlobeNewswire; investors.cerebras.ai; SEC Form 8-K) — GAAP cloud and other services revenue $126.0m, up 281% year on year; core cloud and other services revenue $127.7m, up 287%; GAAP total revenue $180.1m, up 74%; core gross margin 41%, an improvement of roughly 940 basis points year on year; $6.4bn gross proceeds from its IPO; more than 600 MW of data-centre capacity live or under contract for delivery by the end of 2027.
PrimaryCerebras Q2 2026 release, same document — "through its wafer-scale architecture, Cerebras avoids many components currently in short supply, including HBM memory, CoWoS packaging, and 3nm fabrication technology." This sentence is the thesis of the video.
PrimaryCerebras engineering blog, "100x Defect Tolerance: How Cerebras Solved the Yield Problem" — redundant compute cores, redundant routing and a fail-in-place fabric; approximately 900,000 cores remain functional after defect mapping and redundant routing; each core is roughly 0.05 mm²; the company claims about 100× the fault tolerance of a GPU measured by silicon area affected per defect. ⚠ THESE ARE CEREBRAS' OWN PUBLISHED FIGURES. No independent third-party verification was located, and the video attributes them as company claims rather than as measured results.
PrimaryCerebras product page — WSE-3 die area approximately 46,225 mm², about 56× the area of the largest GPU.
PrimaryAnthropic, Claude Code changelog v2.1.232, 13 August 2026 — subagents fork by default and inherit the full conversation and prompt cache, with non-teammate spawns in interactive sessions running in the background; "@" mentions another live session by name; SendMessage resolves bare session names; new /config rows "Dialog expiry" and "Messages from your other sessions" (accept / hold / refuse). This is the Hack of the Week. No vendor endorsed. Cerebras is named because it is the subject of the story; NVIDIA is never named and "a GPU" appears only as the size comparison Cerebras itself publishes. Nothing here characterises any company negatively.

Read the full script (279 words)

Every chip factory runs on one rule: cut the wafer into hundreds of chips, bin the broken ones. So what happens when your chip is the whole wafer? This week Cerebras reported cloud revenue up two hundred and eighty-one percent — on a processor the size of a dinner plate, fifty-six times a GPU. Nothing to dice. So what do you throw away? So how do you get yield on a chip you can't cut? You stop designing against defects and design for them: nine hundred thousand cores, each five-hundredths of a square millimetre. A defect kills one core; the fabric routes around it and the wafer keeps going. But here's what actually matters: wafer-scale let them skip H-B-M memory, CoWoS packaging and three-nanometre capacity — the three shortages everyone else is fighting. They didn't beat the bottleneck. They built something that never touches it. So where does your line still assume nothing breaks? Now, your FabSpeak Tip of the Week. This shipped on the thirteenth and almost nobody noticed: in Claude Code, your sessions can now talk to each other. Type the at sign, name a live session, and you message it directly. And subagents now fork by default — they inherit your whole conversation, so you stop re-explaining context, and they run in the background instead of blocking you. Still sitting there babysitting long jobs? Kick off the review in one session, keep working in another, message it when you need the answer. Then in slash config, set messages from your other sessions to hold. I checked the tooling in this build — the messaging is live and addresses sessions by name. That's FabSpeak. One story a week, taken apart properly.

Semiconductors Semiconductor Manufacturing Wafer Fab Yield AI Chips Advanced Packaging HBM CoWoS

Keep reading

All pieces

NVIDIA's Chip Is Finished — It Still Can't Ship (The Real AI Bottleneck)

NVIDIA's fastest AI chip has been finished for months — and it still can't ship. The reason isn't the 2-nanometer transistor everyone's racing toward; it's the very last step, advanced packaging.

16 Jul 20261:36tip

AI Ate Your RAM — And There Was Never a Shortage

Your RAM got about twice as expensive this year, and the reason has almost nothing to do with a shortage — the memory fabs are running flat out at record output, with $52B going into 300mm memory equipment in 2026 alone, up 29%. What actually happened is a reallocation, and it comes down to physics…

03 Aug 20262:11tip

Samsung Made $75 Billion in One Quarter. The Reason Is 12 Chips Stacked Into One.

[Website catch-up note, added 19 Oct 2026 under Step 0.0: this run was produced on 11-12 Oct outside the website toolchain, so its description, diagram and site page were completed afterwards from the run's own script and the sources below. The delivered script and caption were not changed.]

11 Oct 20261:43tip

OlderYour Fab's AI Goes Blind on the 2nm Node — Here's Why NewerChip Factories Everywhere. Nobody Left to Run Them.