The neutral parts of your stack now have owners.
Tuesday and Wednesday put three announcements on the table. A model hub is being bought, one cloud started reading its rivals' catalogues, and renting chips in Singapore now needs paperwork from Washington. Not one of them is about a smarter model. All three are about who holds the keys to the floors we build on.
Advertisement
Lowest error rate on real dictation. Ahead of Google, OpenAI
Speech models love clean lab audio. Your real prompts don't sound like that. So Wispr Flow tested its in-house speech model, Canto, on 10 hours of real dictations from 2,300+ speakers.
The result: the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. It also tied for the lowest error rate on LibriSpeech, a standard public benchmark.
For you, that means fewer fixes and longer, richer prompts. Speak your context, constraints and examples into ChatGPT or Claude, and Wispr Flow turns it into clean text in any app. Works on Mac, Windows, iPhone and Android.
1. Nvidia is buying the shelf your open models sit on
Nvidia's quarterly release on Tuesday is where the deal turned up. Between a $150 billion increase to its buyback authorisation and the Vera Rubin NVL72 racks starting to ship, one line says it has agreed to acquire Hugging Face, in its own words to make intelligence “more open, capable and accessible”.
We took the revenue number out of that release on 18 November. This line was in the same document.
The hub had a loud week before the deal surfaced. On 16 November California sued OpenAI over a hack at Hugging Face it blames on a rogue agent, and on 17 November Sarvam opened a token factory in India that its own coverage presents as the local answer to OpenRouter and Hugging Face.
I do not expect downloads to stop or licences to change next quarter. My worry is narrower: the hub is where a lot of teams leave their model provenance, and provenance you hold no copy of is provenance you rent. I would spend an afternoon this month listing which production models came from there, which weights and model cards we actually keep ourselves, and what mirroring the rest would cost. Not because Nvidia is the villain here. Because that answer should not take an afternoon to find.
Our research · Apollo says Udemy looks like Accenture. Practitioners name Deloitte.
We tested three sources of competitors on 29 companies: look-alikes, the competitor list on Crunchbase and the pairs practitioners name in their articles. Each answers a different question, and none is a competitor list on its own.
Advertisement
Your Next Side Hustle is Already on This List
Ready to stop scrolling job boards and start building something of your own?
The Hustle's free "Side Hustle Ideas Database" uncovers 100 unconventional, actually-actionable ways to earn on the side.
Inside you'll find:
A curated list of 100 real side hustle ideas spanning e-commerce, services, content, and weird corners of the internet most people never think to look
Opportunities at every level, whether you have 5 hours a week or 50, zero capital or a small runway
Offbeat, under-the-radar ideas you won't find in generic "how to make money" listicles
Subscribe for free today and get instant access. Your next income stream is one click away.
2. Google's data cloud started reading its rivals' catalogues
Google Cloud put out a batch of data announcements on 17 November. Four of them read as one plan:
- Iceberg catalogue federation with Databricks Unity, Snowflake Horizon and AWS Glue, plus zero-copy integrations with SAP, Workday and Salesforce.
- AlloyDB can query Iceberg tables in cloud storage and join them with live operational data in a single PostgreSQL query.
- Data Agent Kit is generally available, wired into VS Code, Claude Code and Antigravity.
- Cross-cloud interconnect and caching for AWS and Azure are generally available, and Spanner Omni is generally available across on-premises, edge and multi-cloud.
The plumbing under this has been moving for a week. Fivetran published a path off Snowflake's native tables into Apache Iceberg on 13 November, and Commvault and Databricks set out governed data access for AI applications on 12 November.
The open-table story we have all been telling has a hole in it. We picked Iceberg so that no single vendor would hold the data, and that part worked. What nobody wrote down is that whoever reads every catalogue becomes the place the questions get asked, and the questions are where the agents and the bills live. So my expectation for next year is not a storage fight, it is an argument inside data teams about which vendor's agent is allowed to ask. I treat the catalogue as something we govern and the engine as something we swap.
Playbook · Follow a list of companies without reading the news: add the outside layer
We tested our own news collection as a monitor for 25 companies and it failed: it found a real October event for 14, one short of the bar we set beforehand. Here is the test, to run on your own list.
3. Renting chips abroad now needs a permit from Washington
The United States has widened chip export controls to Indonesia, Malaysia, Singapore, Thailand and Brazil, so advanced chips going to any firm in those countries that is not US-headquartered need a licence, the Straits Times reports.
Data centres in third countries also need a US licence before supplying significant AI computing power to Chinese-owned or Chinese-headquartered firms.
Singapore runs an estimated 1.6 GW of data centre capacity across more than 70 facilities. Prime Minister Lawrence Wong said his government could facilitate US checks on companies and would take “clear and firm” action against firms that circumvent export controls. Malaysia took the largest share of the region's build-out and carries the most exposure to the rules, and analysts quoted in the piece expect AI capital spending across ASEAN to slow through 2027.
Two days earlier Washington moved to limit remote access to advanced chips, reported the same day as a tightening of controls on remote access to AI technology after the Nvidia affair.
Region usually turns up in procurement as a latency and privacy question. It is a licensing question now, and the licence is held by a government most of us pay no taxes to. The firms I expect to get hurt are not the ones with Chinese owners. They are mid-sized companies whose cloud region happens to sit in Johor or Singapore, who learn what their provider promised about jurisdiction at the renewal. I would ask our provider in writing which licence covers the capacity we sit on, and who applies for it when it lapses.
Today's context on 7wData
- Ramp raises $1.85 billion at a $60 billion valuation Ramp, the finance operations company, announced a $60 billion valuation, and Tech Funding News reports the round as $1.85 billion in primary financing led by Dragoneer Investment Group.
- Dynatrace completes its purchase of Arize, the AI observability company On 1 October 2026 Dynatrace completed its acquisition of Arize, the AI observability and evaluation company, and says Arize AX and the open source Phoenix stay available on their own while the two are integrated over time.
Advertisement
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
Also today
- California and Oregon filed separate lawsuits against OpenAI over AI safety and ChatGPT's effect on children. (3 min read)
- OpenAI said NATO's communications and cyber agency is using its Daybreak security models, and that ENISA used its models to find four software flaws, with fixes published for all four. (3 min read)
- Anthropic was labelled high risk as investors' concerns rise before its listing. (6 min read)
The ArgumentThe open parts have landlords too
Open, in this business, has never meant unowned. It means somebody else is paying the hosting bill and has not asked you for anything yet. This week that arrangement moved on three layers at once: the hub the models sit on, the catalogue the definitions sit in, the building the chips sit in. No open specification was broken anywhere. The list of people you have to ask got longer.
Which part of your stack would you have to rebuild if its owner changed the terms, and how long would that take?
My call: by 30 June 2027, at least two US-listed companies will name a chip or compute export licence in an earnings call or an annual filing as a reason an AI workload moved or stalled.
If you have ever actually moved a model registry or a catalogue to another vendor and know what it cost you, hit reply and tell me.
Yves
Tomorrow, the Scoreboard: I grade my own calls of the past week, the misses included.
Sponsor this newsletter in 2027: 8 placements on dates you choose, from EUR 2,400, until 30 November. Book: 7wdata.be/sponsor-2027
Ins7ghts Daily is published by 7wData.







