Bloomberg’s AI Today newsletter on 23 October 2026 put the week in one line: “AI leaders want their models everywhere and under their control. They may not get both.” Its opening sentence says the race is fought on two fronts: a push to make the technology cheaper and more open, and a drive to keep it tightly controlled within national borders. We read only that opening; the rest of the newsletter is behind a paywall.
Our own reporting from 16 to 24 October shows both fronts moving in the same nine days. By our count, four open AI launches went out (three open-weight models and one open-source inference framework) and four new controls were set or made public. This piece lines them up side by side, with each figure kept as the piece it comes from reported it.
Business Pill · OPEN WEIGHTS
A short Business Engineer explainer on open weights: renting a model versus running it yourself. It teaches the general idea only and says nothing about any company or government in this piece.
The key insight: As we read it, the open launches came with terms and the controls came with gaps: MiniMax’s licence asks commercial products generating more than $20 million in yearly revenue for prior written authorization, and China’s draft rule puts open-sourcing outside the controls or on a fast track while it controls local deployment of a restricted model abroad.
Open weights means the trained model file can be downloaded and run on your own hardware, instead of only being rented through a company’s paid service; open-source code, like DeepSeek’s DeepInfer, is the software that runs a model, published for anyone to use. A fence here means a rule or order that limits who can buy, fund, deploy or export an AI system.
Move and date
Open or fence
What our piece reported
China’s draft export catalogue revision, on 16 October
Fence
The draft adds 23 AI technologies, covering large models, agents, computing clusters and information processing, with a temporary pre-export filing step (export filing piece)
US Homeland Security order on DeepSeek, on 19 October
Fence
Civilian federal agencies told to exclude DeepSeek from procurements and to remove DeepSeek products from their IT systems (DeepSeek ban piece)
DeepSeek open-sources DeepInfer, on 20 October
Open
The inference framework it says powers V4.1-Flash in production, under the Apache License 2.0; the README lists Huawei’s Ascend 950 NPUs alongside Nvidia’s SM100 GPUs (DeepInfer piece)
US Treasury proposed rule, on 22 October
Fence
Would bar US investors from backing companies in China and other countries of concern that develop any AI system; comments due by 7 December (Treasury piece)
MiniMax publishes M3.1 Flash weights, on 23 October
Open
428B parameters in total, 23B active per token; free for non-commercial use, and commercial products generating more than $20 million in yearly revenue need MiniMax’s prior written authorization (MiniMax piece)
Moonshot launches Kimi K3.1 Pro, on 23 October
Open
Described as an open-weight model; Moonshot says the weights will be released by 4 November, with an API price of $2.00 and $10.00 per million input and output tokens (Kimi piece)
Sweden’s Government Offices block two US assistants, on 23 October
Fence
ChatGPT and Claude blocked on staff work computers from Monday, with approved services such as Microsoft Copilot instead, Sveriges Radio’s Ekot reported (Sweden piece)
Abacus.AI releases Smaug-Nemotron-3.5, on 24 October
Open
An open-weight model built by fine-tuning Nvidia’s Nemotron 3.5 Super, which Bindu Reddy called “the best US open-source model in the agentic category” (Abacus piece)
Two further moves sit between the columns. On 20 October China’s Commerce Ministry answered the Homeland Security order and said China will “take all necessary measures to resolutely safeguard its legitimate rights and interests in response to any action that substantively harms China’s interests,” without naming a measure. On 22 October, Chinese state media said DeepSeek and MiniMax had completed registration under the new filing rule and had gone ahead with open-sourcing and overseas licensing.
Tokens processed on OpenRouter between 5 and 11 October 2026: 172 trillion in total, 86.74 trillion on Chinese models and 18.97 trillion on US models, per OpenRouter figures cited by Trevor Noren of SageRoad Research. ‘Non-Chinese’ (85.26 trillion, which includes the 18.97 trillion on US models) and the shares are our arithmetic. OpenRouter counts only traffic routed through it.Tokens processed on OpenRouter between 5 and 11 October 2026: 86.74 trillion on Chinese models and 18.97 trillion on US models, out of 172 trillion, per OpenRouter figures cited by Trevor Noren of SageRoad Research. Shares are our arithmetic; OpenRouter counts only traffic routed through it.
Open, With Licence Terms
The open launches came with licence terms. DeepSeek chose the Apache License 2.0 for DeepInfer. MiniMax’s licence file grants use free of charge “for non-commercial purposes”, requires commercial products to display “Built with MiniMax M3.1-Flash”, and sets the $20 million revenue line for written approval.
China’s draft rule draws its own line. As our export filing piece reported, open-sourcing and online model services sit outside the controls or on a fast track, according to the Ministry of Commerce, while local deployment of a restricted model abroad is controlled.
What the Fences Cover
The Homeland Security order also reaches “third-party products and services developed using DeepSeek implementations,” according to the notice filed for public inspection on 19 October. Treasury’s fact sheet says its proposal “would eliminate the existing notification requirement for these transactions,” so many such deals that until now only had to be reported would be barred under the proposal.
Sweden’s block points the other way: a European government limiting two US assistants. The Government Offices’ digitalisation director wrote that using AI services without agreements can have “konsekvenser för vår informationssäkerhet”, consequences for our information security (our translation).
Update (24 Oct): a commercial fence, too. The fences above are set by governments. Software vendors can set their own. The Information reported on 22 October that Salesforce customers are bracing for new fees to use Claude in its apps. In a post on X, the outlet said customers using Claude and other external AI tools to automate work “could soon face new charges,” and that some businesses “are already reconsidering their AI usage and may move data out of Salesforce if the fees are too high.” The article is paywalled, and we read only its headline and the outlet’s post.
The billing mechanism was described in September by Salesforce Ben, a trade site. It reported that agents registered with Salesforce get their own identity and permissions, and that each successful call they make through MCP or the API counts as a “Headless Platform Interaction” paid in Flex Credits. Salesforce had not disclosed the rate, and customers are to get 30 days’ notice before metering starts, according to the site. A Salesforce spokesperson told it that “the cost per interaction does not vary,” while the number of charged interactions depends on how the outside agent behaves.
If the fees land, a customer running Claude on its Salesforce data would pay twice, once to the model provider and once to the system of record, Salesforce Ben noted. That is the same question this piece tracks for open models: who controls access once the model itself is cheap.
Where the Usage Went
Trevor Noren of SageRoad Research, citing OpenRouter, wrote that between 5 and 11 October models on the platform processed 172 trillion tokens, with Chinese models accounting for 86.74 trillion and US models for 18.97 trillion. He called it the 24th consecutive week in which Chinese models outpaced US ones. By our arithmetic, that is about 50% for Chinese models and about 11% for US models.
OpenRouter’s rankings page, read by FourWeekMBA on 24 October, listed the five most-used models by tokens for the current week as all from Chinese developers, with Anthropic’s Claude Opus 5.5 sixth at 8.04 trillion tokens. OpenRouter counts only traffic routed through OpenRouter, so it is one market, not the whole one.
Connecting the Coverage
Our DeepSeek ban piece reported the Homeland Security order on 19 October, and our Beijing response piece reported the Commerce Ministry’s answer a day later. Our DeepInfer piece covered the framework release the same week, and our export filing piece covered the state-media account of DeepSeek and MiniMax clearing the new rule.
On the US side, our Treasury piece covered the proposed investment ban and its 7 December comment deadline. On the open side, our MiniMax piece covered the M3.1 Flash licence, our Kimi piece the K3.1 Pro price, and our Abacus piece the US open-weight model built on Nvidia’s base. Our China fleet piece carries the OpenRouter figures, and our Sweden piece the government block.
The Structural Read
As we read it, the four controls drew their lines around who pays, who buys and who deploys: the Treasury proposal targets US investment, the Homeland Security order targets federal procurement, China’s rule targets local deployment abroad, and Sweden’s block targets staff computers.
As we read it, the usage figures cover one platform only: Noren’s OpenRouter numbers put Chinese models at 86.74 trillion of 172 trillion tokens in the week of 5 to 11 October, and OpenRouter counts only traffic routed through it.
As we read it, the control aimed at US assistants came from a European government: Sweden’s Government Offices blocked ChatGPT and Claude, and its digitalisation director tied the block to using AI services without agreements.
Bloomberg AI Today, 23 October 2026
“AI leaders want their models everywhere and under their control. They may not get both.”
Three Implications
FOR BUILDERS USING CHINESE OPEN WEIGHTS MiniMax’s licence requires prior written authorization above $20 million in yearly revenue, and the Homeland Security order reaches third-party products developed using DeepSeek implementations in civilian federal agencies.
FOR INVESTORS Treasury’s proposal would bar US investors from backing companies in countries of concern that develop any AI system; comments are due by 7 December.
FOR THE OPEN-WEIGHT RACE Moonshot says Kimi K3.1 Pro’s weights will be released by 4 November, and Abacus.AI released a US open-weight model built on Nvidia’s Nemotron 3.5 Super.
The framework puts its rule as “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, this week’s releases opened model weights and serving code, while licences and export filings kept conditions on commercial use and on deployment. The framework was published in March 2026 and does not discuss these releases.
What Is Not Established
We have not read the full Bloomberg newsletter, so we do not know which release its headline refers to or what else it argues. The four-and-four count is ours, over a window we chose. The sources do not show whether any of the fences has changed how much these models are used. Kimi K3.1 Pro’s weights were not yet published when our piece ran, and the Treasury rule is a proposal.
In nine days, four open AI launches (three open-weight models and one open-source framework) and four new controls, by our count. Bloomberg’s line is that AI leaders want their models everywhere and under their control. Our reporting shows both happening at once, with licence terms and filing rules setting conditions on the open side.
A note on sourcing. Figures from our reporting; original sources are linked in each piece. This analysis adds no new reporting: it sets side by side moves our pieces reported between 16 and 24 October 2026, keeping each source’s attribution and hedges. We read only the opening of the Bloomberg newsletter, which is behind a paywall. The open-or-fence labels and the count are ours.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
