/QAble Weekly/Vol. 010 · 28 Aug 2026

● This week’s signal » Two backup systems, one maintenance job, and an entire company offline. Redundancy only counts if the copies can fail separately.

Weekly

Signal Over Noise

‹ PrevNext ›

For Engineering
Leaders

5-minute read

‹ PrevNext ›

Friday, 28 August 2026 · Vol. 010

In Brief

Proton knocked offline by Frankfurt cooling failureAug 27
Stability AI raises $76M from Universal, Sony, Warner and EAAug 25
Emerald AI raises $150M at $1.05B valuationAug 25
Claude logs elevated errors across four flagship modelsAug 24
Gatik raises $200M Series D for autonomous truckingAug 25
DeviQA formalises testing for AI-written codeQA Financial, Aug 26
Proton knocked offline by Frankfurt cooling failureAug 27
Stability AI raises $76M from Universal, Sony, Warner and EAAug 25
Emerald AI raises $150M at $1.05B valuationAug 25
Claude logs elevated errors across four flagship modelsAug 24
Gatik raises $200M Series D for autonomous truckingAug 25
DeviQA formalises testing for AI-written codeQA Financial, Aug 26

Story of the Week

Proton had two backup cooling systems. One maintenance job took out both.

Proton, the privacy company behind Proton Mail and Proton VPN, runs its Frankfurt datacentre with two redundant air compressors driving the cooling. Two, so that one can fail without consequence. Just after 11pm on August 26, the cooling failed completely. By 11:15pm the room was climbing from 21.8°C toward 51.9°C, thirty degrees in under half an hour. Some probes read 60°C, and network cards hit 105°C against a normal 45°C. At 00:09 essentially every Proton service stopped at once: Mail, VPN, Calendar, Drive, Pass, SimpleLogin, Wallet and its Lumo assistant. Service returned at 02:27, about two hours and twenty minutes later. Proton’s own investigation found the cause, and it is the part worth remembering: an air filter replacement on both redundant compressors. The backup was real. Its independence was not. Two machines maintained identically, by the same process, on the same day, are not two chances to survive. They are one. To Proton’s credit they published that detail, and confirmed the incident hit availability only, with no data exposed or lost.

Why it matters: Ask of every redundant pair: could one maintenance action, one config push or one expired certificate take out both? If yes, you have one system. Publishing the specific cause, rather than “a cooling issue,” is what lets everyone else check their own setup. More vendors should.

A cooling failure in Proton’s Frankfurt datacentre took every service offline for roughly two hours and twenty minutes. · Logo: Proton

WeeklySection 01 · This Week’s Launches

Product Launches

Someone finally wrote down how to test code that an AI wrote

What: The most useful thing published this week was not a product. It was a testing method written specifically for code that an AI wrote.

Reported in this week’s QA Financial roundup on August 26, the quality-engineering firm DeviQA has formalised a testing method built specifically for software written or modified by tools like GitHub Copilot, Claude Code and Cursor. Its starting point is that ordinary testing still applies but is no longer sufficient, because AI introduces particular failure modes: confident wrong assumptions, business logic that is incomplete rather than broken, changes nobody asked for, and hidden dependencies, all while the code itself looks perfectly correct. The sharpest rule in it is this: AI-written tests are not accepted as proof that AI-written code works. If the same assumption shapes both the code and the test that checks it, the same mistake sails through both. The method also sizes regression testing by what a change could actually affect across workflows, permissions and data, rather than by how many lines moved.

Why it matters: If the same model wrote your feature and its tests, you have one opinion checked twice, not a check. Scope regression by blast radius, not diff size. AI changes are often small in lines and wide in consequence.

Launch Log

DeviQA
A formalised testing methodology for AI-assisted development: independent verification, adversarial edge-case testing, and validation of AI-written tests.
Z.AI
GLM-5.3 Flash released on August 26, continuing the rapid cadence of cheaper, faster models in the open-weight field.

WeeklySection 02 · Frameworks & Failures

Frameworks

The next big AI constraint is not chips or code. It is the electricity bill.

Emerald AI raised a $150M Series A on August 25 at a $1.05B valuation, co-led by Energize Capital and DCVC, with NVIDIA, Samsung, Siemens, Salesforce Ventures and Aramco Ventures among the backers. What it sells is unusual and clarifying: software that makes a data centre flexible about power. When the grid is strained, it slows, pauses, caps or relocates selected computing jobs instead of demanding more electricity. Emerald says this could free more than 100 gigawatts of capacity on the existing US grid, and it is already running at multi-megawatt scale, including a roughly 100MW facility in Manassas, Virginia with Digital Realty and NVIDIA. The framing worth keeping: for two years the scarce resource in AI was chips. It is becoming power, and power is a physical negotiation with a utility rather than something you can buy your way past.

Why it matters: If your AI roadmap treats compute as purely a purchasing decision, add grid capacity and interconnect queues to the plan. Workloads that can be paused or moved are becoming cheaper to run than workloads that cannot. That is now an architectural choice.

Failures & Data

Claude went down again, and the numbers on AI code got worse

On August 24, Claude logged elevated errors across four flagship models, Mythos 5, Fable 5, Opus 5 and Opus 4.8, reaching past claude.ai into the API, Claude Code and Cowork, throwing 529 Overloaded errors for roughly three hours. Regular readers will recognise the shape of this: it is the third of the last five volumes to carry a Claude incident. Alongside it, research from DeviQA put numbers to the quieter problem. 65% of development teams now actively use AI coding tools, 52% report bug volume going up, and 58% say their testing workload has grown. The tools are working exactly as advertised. They produce more code, faster. The checking simply has not scaled with it.

Failures & Incidents

Proton offline for 2h 20m after cooling failure (Aug 26 to 27)
Air filters replaced on both redundant compressors; the room went from 21.8°C to 51.9°C in under half an hour. Availability only, no data lost.
Proton incident report
Claude logs elevated errors across four models (Aug 24)
Mythos 5, Fable 5, Opus 5 and Opus 4.8 affected, plus the API, Claude Code and Cowork; 529 Overloaded errors and timeouts for roughly three hours.
Multiple outlets, Aug 24
Bug volume rising for teams using AI coding tools
65% of teams now use them, 52% report more bugs, and 58% report a heavier testing workload.
DeviQA, via QA Financial

Hiring & Trends

The music industry spent years fighting AI. This week it bought a piece of one.

On August 25, Stability AI closed a $76M Series B, and the investor list is the story: Universal Music Group, Sony Music Group and Warner Music Group, plus Electronic Arts, AMD Ventures, Coatue, Greycroft, Sean Parker and Eric Schmidt. These are the same major labels that have spent years in public and legal conflict with generative AI companies over training on their catalogues. Universal and Warner already had partnerships with Stability; Sony is new. Stability is now the first AI company to hold direct equity from all three at once. Read plainly, the labels appear to have decided a seat at the table beats shouting at it from outside, and that licensing is a stronger long-term position than litigation.

WeeklySection 03 · Editor’s Note

By the Numbers · The AI quality gap, quantified

QAble Weekly analysis · sources per figure

51.9°C

temperature reached in Proton’s Frankfurt datacentre, up from 21.8°C in under half an hour

Source: Proton incident report, Aug 27

2h 20m

every Proton service offline: Mail, VPN, Calendar, Drive, Pass, Wallet and more

Source: Proton incident report, Aug 27

52%

of development teams report bug volume rising since adopting AI coding tools

Source: DeviQA research, via QA Financial, Aug 26

$150M

raised by Emerald AI to make data centres dial their own power use up and down

Source: BusinessWire, Aug 25

Editor’s Note

Viral Patel, Co-Founder of QAble — Viral PatelCo-Founder, QAble

“Proton did everything the textbook says. Two cooling systems, so one could fail. Then somebody changed the filters on both, and the textbook stopped helping.”

Your backup and your primary: could one mistake take out both?

This is the kind of failure worth studying, because nobody was careless.

Proton built its Frankfurt datacentre with two compressors running the cooling, precisely so that losing one would not matter. That is correct design. Then the air filters on both were replaced, and both stopped working. The room went from a normal 21.8 degrees to 51.9 degrees in under half an hour, and every service the company runs went dark for a little over two hours.

The lesson is not about cooling. It is that two of something only protects you if the two can fail separately. Same maintenance window, same procedure, same engineer, same config file, same certificate expiry: any one of those quietly turns your pair back into a single point of failure, while the diagram still shows two boxes.

And this week handed us the software version of the same mistake. A testing firm published a method for checking AI-written code, and its central rule is that AI-written tests do not count as proof that AI-written code is correct. The reason is identical. If one assumption produced both the work and the check on the work, you have not checked anything. You have asked the same question twice and heard the same answer.

Two boxes on a diagram. Two green ticks in a pipeline. In both cases the honest question is the same: if the first one is wrong, would the second still catch it?

Most teams have never actually tested that. It is uncomfortable to test, because the answer is often no, and finding out costs a maintenance window you would rather spend on features.

WeeklySection 04 · Briefing

Funding & M&A

Emerald AI $150M · Series A
Gatik $200M · Series D
Stability AI $76M · Series B

Research

Formal Conjectures
A benchmark of 2,615 problems formalised in Lean 4, including 1,029 still-open research conjectures, so a result counts only once a machine can check the proof.
Agentic Proving for Program Verification
Claude Code hit a 98.1% success rate end to end on the CLEVER Lean 4 benchmark, which mostly tells you the benchmark is now too easy to measure anything.

Quote of the Week

“No emails were lost, but email delivery in both directions was delayed during the incident.”

Proton, August 27 incident report

Market Signals

01Redundancy is being re-examined as a maintenance problem rather than a hardware count, after one filter change took out both of Proton’s cooling compressors.
02Testing methodology for AI-written code is starting to be written down formally, with independent verification as the central rule.
03The data on AI coding tools keeps pointing the same way: more code, more bugs, more testing work, not less.
04The music majors moved from litigating against generative AI to holding equity in it, which reframes the fight as licensing rather than legality.
05AI’s binding constraint is shifting from chips to electricity, and power flexibility has become a fundable software category.

Community & Debate

Is your redundancy actually independent?

Proton’s filter-change root cause set off a wave of engineers checking whether their own paired systems share a maintenance window, a config source or a certificate expiry.

Hacker News

Should AI be allowed to write its own tests?

DeviQA’s rule that AI tests cannot validate AI code drew agreement on the principle and complaints about the cost in practice.

Ministry of Testing

Labels as shareholders: truce or capture?

The Stability round split opinion between a sensible licensing settlement and the majors buying influence over how the tools treat their catalogues.

Music and tech forums

Sources: 01 August 27 outage: incident report (Proton) | 02 Proton goes down after cooling failure in Frankfurt data center (Data Center Dynamics) | 03 DeviQA, Cobalt and GitLab: the latest vendor and product news (QA Financial, Aug 26) | 04 State of AI-generated code 2026: the QA and testing gap (DeviQA) | 05 Stability AI raises $76 million backed by Universal, Warner, Sony and EA (Variety, Aug 25) | 06 Emerald AI raises $150M Series A at $1.05B valuation (Emerald AI, Aug 25) | 07 Claude status history (Anthropic) | 08 Formal Conjectures: a benchmark for verified discovery in mathematics (arXiv) | 09 Agentic Proving for Program Verification (arXiv)