Mistral previews a trillion-parameter open model, SpaceX seeks $40bn for Nvidia chips, and 3 more AI moves
Five verified developments from October 6 to 7, 2026: Mistral opens a public preview of Mistral Large 4, a trillion-parameter open-weight model trained in Europe; SpaceX seeks $40 billion of debt to buy Nvidia chips; OpenAI publishes 372 mathematics results from an unreleased internal model; Boston Dynamics names Amazon's Rohit Prasad as chief executive; and Finland orders Google to halt work at two data centre sites.
In brief
Between October 6 at 12:00 UTC and October 7 at 12:00 UTC, Mistral opened a public preview of a trillion-parameter open-weight model trained in Europe; SpaceX began raising $40 billion to buy Nvidia chips; OpenAI published 372 mathematical results from a model nobody outside the company can run; Boston Dynamics handed its top job to the Amazon executive who built Alexa; and Finland ordered Google to stop clearing forest at two data centre sites until an environmental assessment catches up with the work.
- Launches: Mistral released Mistral Large 4 as a public preview, a model with 1 trillion parameters, 49 billion of them active on any given token, and said the weights will follow by the end of October. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres, and the independent evaluator Artificial Analysis scored it 38 on its Intelligence Index — the highest for a model from outside the United States and China, and level with OpenAI's GPT-6 Luna. Price separates it from its Chinese peers: $1.13 per index task at list prices, more than four times what GLM-5.3-Flash and DeepSeek V4.1 Flash cost for comparable scores. 12
- Funding: SpaceX is seeking $40 billion to buy Nvidia chips, in a financing led by the asset manager Apollo Global Management that the Financial Times reported on Tuesday. About $10 billion of that is bank loans and about $30 billion is investment-grade debt, and the bond fund Pimco is among a small group of lenders in talks to provide it, with the transaction expected to close in 2027. Bloomberg confirmed the talks from its own sources and reported that they are at an early stage and could end without a deal. 345
- Research: OpenAI published 372 mathematical results on Tuesday evening, each resolving or substantially advancing an open problem in mathematics or theoretical computer science, in a GitHub repository that also carries computer-checked versions of many of the proofs. A spokesperson told Scientific American that nearly all of the results came from a single prompt handed to a single AI agent, and that the average result consumed the equivalent of about three hours of ChatGPT Pro thinking. The model behind them is internal and unreleased, which is the objection mathematicians and OpenAI's own advisory group have raised. 67
- Hires: Boston Dynamics appointed Rohit Prasad, who spent 12 years at Amazon and led both the Alexa assistant and the Nova foundation-model family, as chief executive from October 7. He is the third chief executive of the Hyundai-controlled robotics company and succeeds Robert Playter, who stepped down in February. 8910
- Regulation: Finland's Licensing and Supervision Agency ordered Google's subsidiary Tuike Finland Oy to suspend preparatory building work at its planned data centres in Muhos and Kajaani by October 23, pending environmental impact assessments. Work at one of the sites went ahead without the assessment the law requires, and more than 300 hectares of forest had already been cleared. The two sites are part of the €13 billion Google committed to Finland on September 9, its largest single investment in Europe. 1112
Five developments
1. Europe's largest open model arrives as a preview
Mistral put Mistral Large 4 — "le Chonk" in the company's own styling — into public preview on its API and in Mistral Studio on Tuesday, and says the weights will be released by the end of October. Until then, the only way to run it is through Mistral. 1
The model is the largest Mistral has built: 1 trillion parameters in total, with 49 billion active for any given token, which is the sparse mixture-of-experts design most labs now use to keep the cost of running a huge model near the cost of running a small one. It reads text and images and writes text, and it has a 512,000-token context window. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in data centres it runs itself in Europe, serves the preview from the same infrastructure, and says a significant share of the training data was multilingual, spanning more than 160 languages including every official language of the European Union. 12
Artificial Analysis, which runs models through a fixed set of evaluations and publishes the scores and the cost of each one, put Mistral Large 4 at 38 on its Intelligence Index, comparable to OpenAI's GPT-6 Luna at maximum reasoning effort (38) and DeepSeek V4.1 Flash at maximum effort (39). That makes it the most intelligent model from outside the United States and China, ahead of South Korea and the United Arab Emirates, and the evaluator describes France as having regained that position. 2
The price is where the model separates from the Chinese open-weight models it is being compared with. At list prices of $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14, Mistral Large 4 costs $1.13 per Intelligence Index task. GLM-5.3-Flash costs $0.25 and DeepSeek V4.1 Flash at maximum effort $0.27 for comparable scores, so Mistral's model costs more than four times as much per task. A 50% launch discount running for two weeks brings it to $0.57 and still leaves it above both. 2
Mistral's own benchmark table makes a different case, and cybersecurity carries most of it. On the Artificial Analysis Cyber Index, which tests whether a model can find and fix flaws in real software, Mistral Large 4 scores 50, level with GLM-5.3-Flash and behind Xiaomi's MiMo-V2.6-Pro at 56; its strongest result on the index's parts is 82% on CyberGym-E2E-AA, a task that asks a model to reproduce a real vulnerability and then patch it. Mistral points out that several leading closed models, including Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra, score near zero on that test because they decline the task, and argues that provider-level refusals block legitimate vulnerability research and incident response. The company also reports 93% on Cybench, a set of 40 security-competition exercises. 12
On ordinary engineering work the picture is mixed, and the numbers come with their own baselines. Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, for a combined coding-agent index score of 49.8% that the company places ahead of DeepSeek V4 Pro 0813 and Alibaba's Qwen3.8 Max. On AutomationBench, which runs 657 business workflows across apps such as Gmail, Google Sheets, Slack and Salesforce, it scores 59.9%. In a blind human evaluation that Mistral commissioned from Surge AI, professional annotators rated outputs on a five-point scale without knowing which model produced them; Mistral Large 4 placed second of five at 3.74, ahead of Kimi K3 at 3.59 and GLM-5.3 at 3.60 and behind Claude Opus 5 at 4.22. 1
What a reader can act on today is narrow. The weights are unpublished, so nothing here can be run, audited or self-hosted yet, and every figure beyond the two Artificial Analysis index scores is the vendor's own measurement. Mistral says it is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who receive the same model with reduced moderation and expanded cyber capabilities — which means the version being tested is not the version most customers will get. The company also says Mistral Large 4 will be the base for a new generation of specialised models and that it uses the same training and reinforcement-learning environment Mistral sells to enterprises through Mistral Forge. 1
2. SpaceX borrows against chips
SpaceX is seeking $40 billion to buy Nvidia chips, according to a Financial Times report on Tuesday that Reuters relayed the same evening. About $10 billion of the total is bank loans and about $30 billion is investment-grade debt, Apollo Global Management is expected to lead the deal and help place the debt with a broad range of investors, the bond fund Pimco is among a small group of lenders in talks to provide financing, and the transaction is expected to close in 2027. 34
Bloomberg reported the same talks from its own sources, naming Pimco as one of the investment managers looking at the deal, describing the discussions as at an early stage and saying they could end without a deal being completed. Bloomberg's report calls the proposed financing one of the biggest ever debt financings for the AI buildout. SpaceX, Apollo and Nvidia did not immediately respond to Reuters requests for comment; Pimco declined to comment. SpaceX shares fell 1% in extended trading after the report and Nvidia's rose 0.5%. 45
The size has a published baseline to sit against. Morgan Stanley estimates that AI infrastructure will need $1.5 trillion of external financing by 2028, even as lenders and investors grow more cautious about funding the industry's expansion — which is the tension the SpaceX deal sits inside: the amounts keep rising while the appetite is being tested. 4
Three facts from the past month set up what the borrowing is for. Musk took SpaceX public in June in a record $86 billion initial public offering, and said last month that xAI's Colossus 2 data centre could more than double the number of Nvidia chips it uses by December, with the company planning to build its data centres on Nvidia hardware exclusively. Nvidia in August partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on financing platforms intended to mobilise more than $500 billion for AI infrastructure projects, so the lenders being approached are also the ones already organised around this trade. 4
The detail that decides who can buy the debt is the credit rating. The Financial Times reports that insurance and pension funds would be able to hold SpaceX's debt because of its BBB rating, the second-lowest tier of investment grade. That is the mechanism by which a company buying chips turns into a fixed-income asset held by retirement money — and it is also why the rating, and any change to it, matters more to the deal's reach than the headline number does. 3
Everything above rests on people familiar with private talks, relayed by one newspaper and confirmed in outline by a second outlet. Neither SpaceX nor Apollo nor Nvidia has confirmed the amount or the structure, and Bloomberg's own reporting says the talks may end without a deal. A reader weighing the figure should treat it as a reported intention with a 2027 closing date, not as committed capital. 45
3. OpenAI publishes 372 mathematics results and withholds the model
OpenAI released 372 mathematical results on Tuesday evening, publishing them in a GitHub repository rather than as papers. Scientific American reports that each result resolves or makes substantial progress on a major open question in mathematics or theoretical computer science. The repository carries formalisations of many of the proofs in Lean, a programming language in which a computer checks a proof's logic, which the report treats as making those results all but certain to be correct. 67
Among the claims are a solution to the four-dimensional Kakeya conjecture, improvements on some of the world's most important computer algorithms, and progress toward the Riemann hypothesis. OpenAI says the average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking, and that the repository includes ten summaries of the model's reasoning, estimates of compute spent expressed in the same ChatGPT Pro terms, and statistics on how many problems were attempted. 67
That compute figure is the claim with the largest consequences. A month ago OpenAI's model produced a solution to the Navier-Stokes problem through what Scientific American describes as a 10,000-strong agentic swarm that cost millions of dollars in computing power. A spokesperson told the magazine that nearly all of the new results came from a single prompt handed to a single agent, though some may have taken multiple attempts. If that holds, the cost of getting a frontier model to do research-grade mathematics has fallen by orders of magnitude — and the same tool becomes available to anyone who can rent a model. 7
The model that produced the results is internal and has not been released. That is where the scepticism attaches. Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, told Scientific American: "Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified," adding, "We should ask for receipts." 7
OpenAI assembled an independent advisory group on September 21, after its handling of the Navier-Stokes release drew criticism, and asked it for recommendations on publishing AI-generated results responsibly. The Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study recommended that a company releasing such results make public the model, the exact prompt and the compute time behind each one. OpenAI published an average compute figure and additional statistics, and no prompts; a spokesperson told Scientific American that the team takes the guidelines seriously and is doing its best to comply, and that the company is not bound by the recommendations. 7
Mathematicians disagree about whether releasing the results this way helps or harms the field, and both positions are on the record. Daniel Litt of the University of Toronto told Scientific American that publishing is the right call: "If we want to know the answers to these math questions, I see no reason why we should ask the company to keep them secret from us." Terence Tao has criticised the pace of AI-generated results from frontier labs. A spokesperson also said that many of the newly released results are not yet understood by OpenAI's own mathematicians. 7
OpenAI says it will fund a series of workshops, conferences and special programmes on understanding major results produced by AI, that it is consulting the advisory group to develop best practice, and that it is working to release the model that produced the results as quickly and responsibly as it can. For a researcher, the practical position today is that 372 results now exist in a public repository, some of them machine-verified, and the means of reproducing them does not. 67
4. Boston Dynamics hands the humanoid to an AI executive
Boston Dynamics appointed Rohit Prasad as chief executive on Tuesday, effective October 7, ending a search that opened when Robert Playter stepped down in February. Prasad becomes the third chief executive of the robotics company, which is majority-owned by Hyundai Motor Group. 810
Prasad spent 12 years at Amazon, most recently as senior vice president and head scientist for Alexa and artificial general intelligence. He was involved in building Alexa from its earliest days into a consumer AI service used in hundreds of millions of households, and more recently led the development of the Amazon Nova family of foundation models used by enterprises. Before Amazon he spent nearly 14 years at Raytheon BBN Technologies leading machine-learning research for US government and commercial customers. He has lived in the Boston area for more than 25 years. 8
The company he takes over sells three machines. Spot is a quadruped that walks industrial sites and inspects equipment; Stretch moves boxes in logistics and retail; Atlas is an electric humanoid still in development. Hyundai, which acquired majority ownership in 2021, said earlier this year that it plans to deploy Atlas at its plant in Georgia starting in 2028 and to build the capacity to produce as many as 30,000 humanoids a year — the two figures against which Prasad's tenure will be measured. 9
Both sides framed the hire as a bet that model-making skill transfers to machines that move. Prasad said: "Boston Dynamics is uniquely positioned to advance Physical AI through its world-class robotics expertise," and that combining advanced AI with robotics would "shape the next era of Physical AI". Jaehoon Chang, vice chair of Hyundai Motor Group and chair of the Boston Dynamics board, said the company's robotics, Prasad's "proven ability to turn AI into products at global scale", and Hyundai's manufacturing, logistics and mobility operations "give us a powerful foundation to build and scale Physical AI". Prasad is expected to join the board subject to the relevant approval process. 8
The appointment is the clearest signal yet of where the humanoid race is being decided. Boston Dynamics has led on the hardware — Atlas is famous for movements most machines cannot reproduce — and has lagged on the software that decides what a robot should do next. Hiring the person who ran Amazon's assistant and its enterprise model family, rather than a manufacturing or automotive executive, is a statement that the next bottleneck is the model, and that Hyundai's plant in Georgia is the place it will be tested. 89
5. Finland stops Google's forest clearing
Finland's Licensing and Supervision Agency, known by its Finnish initials LVV, issued a notice to Google's subsidiary Tuike Finland Oy on Tuesday ordering it to suspend preparatory work at the data centre sites it is developing in Muhos and Kajaani. The work must stop by October 23, until the environmental impact assessments required for the sites have been completed. 1112
The suspension covers tree removal, topsoil removal, excavation, quarrying, crushing, the transfer of soil masses, ditching, and the construction of roads and storage areas. Planning, measurements, soil surveys and other measures with minor impacts that can be easily restored are exempt. The BBC reports that the orders relate to preparatory work already carried out, and that more than 300 hectares of forest have been cleared across the two sites; one site's work went ahead without the assessment the process requires. 1112
Tommi Muilu, head of the environmental department at the licensing authority, explained the order in the agency's statement: "In our opinion, the measures in question will change the environment and cause impacts, the identification and assessment of which are among the key objectives of the EIA procedure." Tuike has been asked to provide an explanation and set out how it intends to proceed by October 14, and the agency said it could begin enforcement proceedings if the company fails to comply. The Finnish Association for Nature Conservation, which had called for work to stop while the requirements were examined, said through its chair Hanna Halmeenpaa that "nature sites which should be preserved" had been felled. 1112
Google did not dispute the finding. "We understand the concerns and have fallen short of our own high standards in this instance," a spokesperson told CNBC, adding that the company would follow the agency's guidance. The same spokesperson said Google "acted in good faith under the Forestry Act, and ran nature surveys and put in place measures to protect high-value areas", and that it is advancing landscaping and biodiversity plans that include planting trees across 130 hectares at Muhos. 1112
The sites sit inside the largest single investment Google has made in Europe. On September 9 the company committed at least €13 billion ($15 billion) to Finland through 2027 and 2028, covering new data centres in Kajaani, Muhos and Vaala alongside its existing facility in Hamina. Google says the construction phase will support more than 37,000 jobs and contribute an average of €3.6 billion a year to the Finnish economy, and it has signed a 22-year power purchase agreement with the Finnish utility Fortum covering up to half the capacity of the Loviisa nuclear power plant. 12
Finland's attraction is the same one that made the halt possible. Cool air lowers the energy needed to cool server halls and the grid is largely low-carbon, which has drawn a queue of projects with potential capacities in the hundreds of megawatts and earned the country the label "Texas of Europe" in industry coverage. Land and power have been the binding constraints on AI data centres for two years; this order introduces a third, an environmental assessment that a national agency can enforce on weeks' notice, and it lands on a company that had already started moving earth. 11
The read-through
Four of this issue's five are companies and governments deciding who holds an asset the operation depends on. Mistral trained and now serves its largest model on hardware it runs itself in Europe, and will hand the finished weights to customers rather than renting access indefinitely. SpaceX is borrowing $40 billion so that the chips feeding its data centres sit on its own balance sheet rather than a supplier's. OpenAI published the output of a model it keeps, and the loudest objection is that nobody else can run the thing that produced the results. Finland has shown that a national agency can stop a hyperscaler's construction with a notice, which is the state holding the land. That reading is this issue's own inference from five separate announcements; none of the parties framed its move as part of a pattern. 14611
The fifth runs the other way, and the difference is worth naming. Boston Dynamics is buying capability rather than building it: Hyundai owns the robot maker, and the company has just hired an executive from a foundation-model lab's parent to supply the intelligence that decides what the machines do. Mistral and OpenAI both treat the model as the thing to keep; Boston Dynamics treats it as the thing to buy in. 8
Cost appears in all five, which is not a theme so much as the arithmetic of the same year. Mistral's model costs more than four times as much per task as its closest open-weight peers, and its pitch is control rather than cheapness. SpaceX is financing chips with investment-grade debt that pension and insurance money can hold. OpenAI's claim is that a research result that recently cost millions of dollars in compute now costs hours of a consumer subscription. The Finnish halt costs a construction schedule, and Hyundai's plan to produce as many as 30,000 humanoids a year is a bet that its unit economics work. 2369
The evidence behind the five is uneven, and where it is thin matters. Mistral's benchmark table is the company's own measurement, with the two Artificial Analysis index scores the only independently produced numbers in the item; the weights are unpublished, so none of it can be checked by running the model. The SpaceX figures come from people familiar with private financing talks, relayed by one newspaper and confirmed only in outline by a second outlet, with every company approached declining to comment or failing to respond. OpenAI's results are published by the company that made the model, in a repository it controls, and the scepticism about them is documented rather than resolved. Boston Dynamics' appointment is first-party, with the manufacturing context from Bloomberg. Finland's order rests on the licensing agency's own statement as carried by two outlets, and Google's response came to both. 257911
Watch next
- Whether Mistral's weights arrive by the end of October, and what the price is afterwards. Mistral has promised the weights, architecture detail and post-training methodology for the end of the month, and the preview is running on a 50% discount that expires after two weeks. A published set of weights with an independent score would settle whether this is a European frontier model or a preview that stays one. 12
- Whether the SpaceX financing closes, and what happens to its rating. The deal is reported with a 2027 closing date and depends on investment-grade debt that insurers and pension funds can hold. A completed financing, the talks ending, or any change to the BBB rating would each say something different about how much AI spending the debt markets will absorb. 34
- Whether OpenAI releases the model behind the 372 results. The advisory group's recommendations ask for the model, the exact prompt and the compute time behind each result; the company has published an average compute figure and no prompts. Replication by mathematicians who did not build the model is what would turn the release from a claim into a result. 67
- Whether Atlas reaches Hyundai's Georgia plant in 2028. Hyundai has said it plans to deploy the humanoid there starting in 2028 and to build capacity for as many as 30,000 robots a year. Prasad's tenure will be measured against those two dates, and against whether Atlas leaves the demonstration floor. 89
- Whether the Finnish assessment clears the sites or enforcement follows. Tuike's explanation is due on October 14 and the suspension takes effect by October 23. A completed assessment, a revised construction schedule, or the enforcement proceedings the agency has threatened would each settle what the halt costs Google's €13 billion plan and whether other projects in the country face the same test. 1112
References
- 1Mistral — Introducing Mistral Large 4
mistral.ai
- 2
- 3
- 4
- 5
- 6
- 7
- 8Boston Dynamics — Appoints Rohit Prasad as Chief Executive Officer
bostondynamics.com
- 9
- 10
- 11
- 12
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.