Afternoon Brief, October 11th 2026
A contractor pleaded guilty in the Super Micro chip-diversion case. Asana found a large cost saving in how its browser kept its history. Britain is funding help for businesses adopting robots. Amazon's now include a Gemma-based family. Updated 3:03 p.m. Pacific. This edition now includes five additional stories from the parallel run: Tenstorrent acquisition talks, Nvidia's reported d-Matrix investment, Anthropic discount caps, Grok Bot email and Stuut's raise.
A guilty plea advances the Super Micro chip-diversion case
Ting-Wei “Willy” Sun pleaded guilty Thursday in Manhattan to four counts involving the illegal export of AI servers to China, Reuters reported Friday from a court filing. The counts cover conspiracies to violate , smuggle goods, defraud the United States and obstruct justice. Prosecutors charged Sun and two others in March over an alleged scheme involving about $2.5 billion in technology.
The indictment describes dummy servers and false statements to inspectors used to conceal equipment already diverted to China. Those details remain allegations about the scheme; Sun's plea does not establish the guilt of other defendants. Super Micro told Reuters it was not named as a defendant and referred to earlier statements that the case had not affected operations. The company terminated Sun and cut ties with the other defendants earlier this year. SourcesB
Asana cuts browser-agent costs by preserving the history it was throwing away
Asana's October 8th engineering account reports that changing a browser agent's history management and switching models reduced estimated model cost from at least $36.21 to $0.47 per run. The original workflow repeatedly removed screenshots and trimmed text, breaking the unchanged input prefix that needs. The revised version retained more history and removed screenshots in batches.
Most of the saving preceded the model switch: optimizing the original model reduced cost to $1.24. Moving the optimized workflow to GPT-6.1 Sol produced the remaining reduction. The study used a fixed book-catalog task, with three runs per condition, and some baseline runs stopped at their step limit unfinished. These are internal test results, not a general price cut or a production-wide savings audit.
Britain launches robotics hubs to help buyers get machines into use
The UK government announced eight Robotics Adoption Hubs Friday, backed by £40 million. They will offer advice, demonstrations and help connecting businesses and public services with suppliers, integrators and finance providers. The network covers surgical robotics, agriculture, manufacturing and other applications; Lincoln's HARVEST hub will provide testbeds for farming and food production.
The program funds the work between seeing a demonstration and making a purchase: assessing costs, selecting equipment and planning deployment. Buyers get somewhere to test whether a machine fits their operation. Faster recovery from surgery and higher agricultural productivity are the government's intended benefits, not measured outcomes of these new hubs. SourcesA
Strands Decider adds Gemma models and revises its Qwen training data
Amazon's Strands Labs released four Gemma-based decision models and a revised Qwen-based model on October 9th, according to its changelog. The Gemma versions broaden the hardware choices for a system that scores supplied answers without generating prose. The repository lists about 23 GB of weight memory for the 12B version and about 51 GB for the 26B-A4B version.
The revised Qwen release removes training rows from datasets whose terms restrict commercial use and adds code and written-policy tasks. Its evaluation notes also say the data-building code is not yet on the main branch, so the training mix cannot be reconstructed from that branch alone. Buyers should separate downloadable from a fully reproducible training pipeline. Performance figures remain the developers' measurements, and the Gemma support requires installation from the repository rather than the published package. SourcesAAA
Vultr joins a shared computing pool aimed at shorter AI workloads
Vultr said Wednesday it is joining National Compute as a founding consortium member and contributing capacity to its Grid Exchange. The proposed pool connects providers and chip architectures so research teams can reserve computing resources without relying entirely on long-term contracts. The consortium targets 2 of pooled capacity by 2030.
That is a target, not delivered capacity. Vultr's announcement confirms a participating supplier but does not establish how much a customer can reserve today, at what price, or with what service guarantee. For smaller labs, the useful test will be whether a short experiment can secure suitable machines when needed. SourcesA
Anthropic recruits Google's former construction chief for regional capacity delivery
John M. Wilson, Google's former global design and construction chief, has joined Anthropic to oversee capacity delivery in Europe, the Middle East and Africa, TechRadar reported Friday. The report also identifies former Google specialists moving into Anthropic's electrical engineering, design and operations roles.
The appointments add experience in getting facilities built and operated. They do not establish new power supply or available computing capacity. Wilson's assignment shows Anthropic staffing the delivery work behind its infrastructure commitments; the amount of capacity his team will bring online was not disclosed in the report. SourcesB
DuplexAgent repairs the handoffs between live conversation and background work
An October 8th introduces DuplexAgent, which keeps a voice conversation responsive while other agents handle reasoning and coding. Its coordination software separates jobs such as accepting work, cancellation and result delivery into editable modules. Simulated conversations produce failure traces that guide targeted repairs, and an archive records what earlier repairs changed.
The authors report stronger spoken-knowledge and executable-tool scores than the delegated systems they compared against, while preserving interruption response. The abstract supplies no numerical effect size. This is an experimental approach to a recognizable failure: an assistant can complete a background task and still mishandle a user's interruption or deliver the wrong result to the conversation. The results have not been independently reproduced here. SourcesA
Intel and Qualcomm are both reported to be in early talks to buy Tenstorrent
Jim Keller's AI chip startup is running two processes at once: it has hired investment banks to explore a sale, and it is sounding out investors for a new round. Bloomberg reported that a sale could value Tenstorrent at $5 billion or more B. Tom's Hardware separately put a Qualcomm deal at $8 billion to $10 billion C. The two figures come from different outlets and differ by up to twice.
Nothing has been signed. Neither Intel nor Qualcomm has commented, Tenstorrent has not confirmed the talks, and people familiar with them said they could still fall apart. Intel shares were reported up as much as 8% intraday, though the report does not say which session that was C. SourcesC
Nvidia is reported to be investing in d-Matrix, an inference chip rival
The Information reported on Saturday, citing three people with knowledge of the matter, that Nvidia plans to invest in d-Matrix, which sells processors that compete with Nvidia's own. Neither company has confirmed it, and the size and timing are undisclosed C.
The odd part is the plumbing. d-Matrix's next-generation Raptor processors are slated to connect to Nvidia's platform through , announced September 10th, and its racks are designed to run next to Nvidia systems. Reported funding history conflicts: Data Center Dynamics has a $275 million at a $2 billion valuation in November 2025, while Reuters cites $450 million at the same valuation C. The Information said in July that d-Matrix was seeking money at $5 billion C. SourcesC
Anthropic reportedly ends enterprise discounts the moment a customer reaches its contracted cap
Large customers who exhaust the allotment in their contract now pay list price for further use, or must renegotiate, The Information reported, citing managers at three software companies that buy from both Anthropic and OpenAI. Anthropic did not comment B. The discount at stake is reportedly about 15% off list C.
The comparison in the report is with OpenAI, which reportedly lets discounted customers keep their rate through the rest of the month plus a further month before list prices apply. Amazon, Microsoft and Google typically keep discounts on usage above the committed amount C. Customers who plan to overshoot a commitment now have a reason to negotiate a larger one up front. The three-company sample is small, and the primary article is paywalled. SourcesBCC
xAI's Grok Bot can now hold its own email inbox
Rollout began Friday. A user claims an address ending in @mail.grokbot.com by asking the bot or tagging @bot on X, and can request a custom name. xAI says the bot can use the inbox to sign up for services, contact businesses and schedule time with people B.
The coverage describes no usage limits or safeguards, so what stops the bot from registering for things its owner did not intend is unanswered. 9to5Mac calls it the bot's second significant upgrade this week, after one that added help from Anthropic's Claude, which the article links to but does not explain. SourcesB
Stuut raises $52.5 million to send AI agents after unpaid invoices
The New York company's Series B was led by Insight Partners, with Andreessen Horowitz and Microsoft's M12 participating, announced October 7th. Total funding is now $93 million, ten months after its Series A C. Stuut says more than 150 customers use it and over $3 billion has moved through the platform ?. Both figures are the company's own and unverified.
The product handles order-to-cash: invoicing, chasing late payers and reconciling what arrives. Stuut puts global unpaid receivables near $16 trillion, a number that is hard to reconcile with the $7.48 trillion in US trade receivables the Federal Reserve reported for the end of the first quarter ?. SourcesC
Sources
- B Reuters via AOL: Super Micro contractor's guilty plea.
- A Asana: Browser-agent cache study.
- A OpenAI: Asana customer study.
- A UK government: Robotics Adoption Hubs.
- A Strands Decider changelog, repository and evaluation record.
- A Vultr: National Compute membership.
- B TechRadar: Anthropic's infrastructure hiring.
- A DuplexAgent-RSI preprint.
- B The Information via Dealroom: Anthropic moves to cut off discounts when customers hit their cap. Paywalled; facts limited to the published summary.
- B The Information: Anthropic and OpenAI fighting over enterprise spending. Not inspected directly.
- C Newsquawk: Anthropic ends enterprise discounts at contracted limits.
- C Startup Fortune: Intel and Qualcomm chase Tenstorrent. Bloomberg and Tom's Hardware figures relayed second-hand.
- C TBreak: Nvidia reportedly plans to invest in d-Matrix. The Information report relayed second-hand.
- B 9to5Mac: Grok Bot gets its own email address.
- C HackerNoon: Stuut raises $52.5M. Company press release.
Research and vendor results were not independently reproduced. The guilty plea is reported from Reuters' account of the court record; other defendants' liability is not settled here. Wilson's original LinkedIn post could not be opened. The search retained unresolved leads for review, so coverage is incomplete. No new Chinese flagship weight release was verified; Strands extends models derived from Qwen and Gemma. Paywalled items from the parallel run are reported from published summaries and tiered on the originating outlet.