An $80bn memory quarter, a free 40-language speech model, and experts who defer

Samsung Electronics published its third-quarter results on 29 October, and says they make it the first technology company anywhere to pass 100 trillion Korean won in quarterly operating profit. Operating profit came in at KRW 107.4 trillion, up 783% on the same quarter a year earlier. Revenue was KRW 195.3 trillion, another all-time high and 14% higher than the previous quarter. At recent exchange rates that is roughly $80 billion of profit on about $146 billion of sales, an operating margin above 50% for a company that also sells phones, televisions and washing machines.

Almost all of it came from chips. The Device Solutions division, which covers memory, foundry and Samsung's own processors, posted KRW 147.1 trillion of revenue and KRW 106.5 trillion of operating profit, about 99% of the company total. Samsung said its memory business set all-time highs for both quarterly revenue and quarterly operating profit, and that it significantly scaled up sales of HBM4, the sixth generation of the stacked high-bandwidth memory that sits beside AI accelerators. Coverage of the earnings call reports that HBM4 revenue more than tripled quarter on quarter and that Samsung's planned 2027 HBM supply is already sold out.

The rest of Samsung looks very different. The Device eXperience division, which sells phones and consumer electronics, made an operating loss of KRW 0.8 trillion, with the MX and Networks businesses alone losing KRW 0.6 trillion on KRW 37.7 trillion of revenue. Samsung Display contributed KRW 8.0 trillion of revenue and KRW 1.1 trillion of profit. The shape matches what rival memory makers have been reporting: AI data centres are absorbing so much DRAM and flash that prices have risen across every application, including the ordinary server and PC memory that used to be a low-margin commodity. Samsung's board approved a KRW 30 trillion cash dividend payable on 27 November.

There are caveats worth keeping. Net profit was KRW 75.14 trillion, up 514% but below the roughly KRW 87 trillion that Korean consensus data had it at, and Samsung's shares fell about 4% on the day of the announcement. Earnings-call coverage attributes roughly KRW 10 trillion of quarter-on-quarter drag to exchange rates. The figures for HBM4's share of HBM sales and the sold-out 2027 supply come from Samsung's own call rather than from filings, and the company has not broken out how much of the quarter's profit HBM4 produced by itself. Memory is also famously cyclical; the price moves that produced this quarter reverse hard when supply catches up.

For anyone outside the chip industry the practical signal is cost. Samsung, SK hynix and Micron are prioritising HBM and high-end server memory, which means the DRAM and SSDs that go into laptops, phones, small servers and cheap cloud instances are getting scarcer and dearer, and Samsung expects AI-driven demand to outrun industry supply again in 2027. If you buy, build or rent hardware, whether that is a workstation for running models locally, a GPU box, a self-hosted app or an office refresh, the quote you got six months ago is probably no longer the price. That is a planning problem, and for some people a business.

Why it mattersThe single most profitable quarter in technology history was produced by selling memory into AI data centres, which is the clearest evidence yet that the AI build-out, not consumer devices, now sets hardware prices. Anyone who buys or rents compute is paying for that shortage.

Nvidia released Nemotron 3.5 Transcribe on 28 October: an open-weight speech model that transcribes audio in up to 40 languages, translates 31 of them to and from English, and tells apart up to eight speakers in a recording. The usual transcription pipeline chains several models together, one to detect speech, one to transcribe, one to separate speakers, another to answer questions about the result. Nvidia says Transcribe does all of that in a single pass of one model, and that you steer it with a prompt, for example asking it to label each speaker, translate into English, or summarise what was decided in a meeting.

It is a 30-billion-parameter mixture-of-experts model with about 4 billion parameters active per token, which is why it can run on a single GPU rather than a cluster. Nvidia's deployment guidance spans the range from a Jetson AGX Thor edge module and a DGX Spark desktop to a Blackwell data-centre card, and the company says a Blackwell Server Edition GPU running NVFP4 precision with the model compressed to a quarter of its size can chew through 100 hours of audio in under three minutes. Reports differ on the per-recording limit: TechCrunch says recordings up to an hour, SiliconANGLE describes digesting and answering questions about 100 hours, so check the model card before you design around it.

Alongside the model Nvidia published Granary v2, an expanded version of its open speech dataset, built with its NeMo Curator pipeline. The public manifest covers 463,062 hours of audio across 48 languages and 16 source corpora, up from 25 languages in version one. SiliconANGLE notes those numbers describe the dataset release rather than the model's full training mixture. Granary v2 also ships three transcript styles per recording, one verbatim with every filler word, one that renders numbers as figures, and one with filler removed, which is what lets you ask Transcribe for a particular output style instead of cleaning transcripts by hand afterwards.

Accuracy claims need care because they come from Nvidia and are measured on different benchmarks. The company says Transcribe leads Hugging Face's Open ASR Leaderboard among open models, with coverage citing an average word error rate of about 5.33% on the English leaderboard, a comparison that excludes proprietary API models. On a broader multilingual benchmark SiliconANGLE reports a 30.8% word error rate out of the box, falling to 20.8% with careful prompting, and down to 11.8% once fine-tuned on the Eka Care Indian medical speech benchmark. In other words, out-of-the-box quality varies a lot by language and domain, and fine-tuning does most of the work. Transcribe also handles recorded audio, not live streams.

The practical point is that a capability which until recently meant a per-minute bill to a cloud API is now a file you can download, run on one machine and adapt. That matters most where the big commercial transcription services are weakest: languages with little commercial data behind them, heavy code-switching, strong regional accents, and recordings that cannot legally leave a client's premises. The release lands in a crowded month for voice models, with Mistral's Voxtral 3, ElevenLabs' Realtime and Salesforce's Launa-Omni all shipping recently, but the open weights plus an open dataset plus a fine-tuning pipeline is an unusually complete package to hand to small builders.

Why it mattersHigh-quality multilingual transcription with speaker labelling has been a paid API feature; now the weights, the dataset and the fine-tuning pipeline are all free and run on one GPU. That removes the main cost barrier for anyone serving languages or industries the big providers ignore.

Anthropic published a randomised controlled trial on 27 October that tested a comfortable assumption: that when an expert uses an AI agent to do the work, the expert still makes the judgment calls. The researchers, Amac Herdagdelen, Hailey Joren and Alex Tamkin, recruited 61 professional analysts with causal-inference experience and gave them a realistic task, advising a fictional farming cooperative on whether its paid advisory programme improved member farms' productivity enough to be expanded. Fifty of them used Claude Code for the analysis, and the deliberate design feature was that four of the decisions along the way had two perfectly reasonable answers.

For each analyst and each decision, a coin flip set which option the agent would argue for, and the analysts were not told this was happening. The hidden instructions told the agent to hold the lean as its own view, raise it when the decision came up and make the strongest honest case the data allowed, while also telling it to change its mind if evidence turned against it and never to misrepresent results or hide data. The open decisions were the sort that quietly move an answer: whether to measure productivity as yield per hectare or per unit of operating cost, which comparison method to use, which farms to include, and whether to weight every farm equally.

Across the four analytical decisions, analysts went with their agent's lean 86% of the time, against roughly 50% if the lean made no difference. Per decision the match rate ran from 81% to 92%. Where an analyst had already stated their own preference before the agent weighed in, they switched away from it 60% of the time when the agent disagreed, compared with 6% when it agreed. The effect was weaker on the final recommendation, 54% recommending rollout when the agent leaned that way versus 36% when it leaned the other, but by then the agent's earlier leans had already shaped the estimate the recommendation rested on.

The part that should worry anyone reviewing AI-assisted work is what the analysts believed afterwards. Asked to explain their choices, 77% used the agent's own arguments, sometimes word for word, while only 15% credited the agent as the source. Those who followed the agent closely were about as confident as everyone else, and about as likely to call the result their own judgment. Anthropic also reports that a statistical model using only the four coin flips correctly predicted 82% of analysts' final recommendations, against 61% for always guessing the most common answer. In other words, much of the outcome was set before the work began.

Anthropic is careful about the limits, and so should anyone citing this. The leans were deliberately strengthened to make the effect measurable, so a normal agent that has not been told to advocate may push less hard, though Anthropic notes that an unprompted Claude Opus 4.8 still made the same five choices in 112 of 123 runs. Eleven of the 61 recruits chose not to use the agent and are excluded. The task used synthetic data at a fictional cooperative, not real stakes. And the study is not an argument against agents: those who used one finished faster and made far fewer serious errors, though that comparison group was small and less experienced.

Why it mattersIf experts adopt an agent's arbitrary preferences while sincerely believing the judgment is their own, then human review is not the independent check that governance frameworks, audits and client assurances assume it is. That gap is a real, sellable problem.