Tenstorrent raises $1.5B at a $6.2B valuation, Microsoft drops Nscale data center deal, and more - Daily AI Brief #561

In this edition, covering 18 and 19 October:

• Tenstorrent raised a $1.5 billion Series E at a $6.2 billion post money valuation.
• Microsoft reportedly pulled out of Nscale's planned Essex data center, as investors push Nscale to cut its IPO valuation from about $30 billion.
• One hacker used AI agents to attempt more than 1,200 intrusions and breach over 100 websites, a NorForge researcher found.

Partner event

Sponsored. AI Brief is a media partner of the summit.

RE•WORK holds its AI in Finance Summit Fall tomorrow and Thursday, 21 and 22 October, at W Hoboken, New Jersey, on AI forecasting, returns on AI projects and AI in banking, insurance, asset management and fintech. The code AIBRIEF gives 30% off tickets. Registration · The summit

From our lists

A list of 27 AI app builders that turn plain language or project context into a working web or mobile app with a backend or stored data and a way to deploy or export it, listed without a ranking, from Lovable, v0 and Replit to Retool and Microsoft Power Apps.

AI app builders

Nscale and AI debt

Nscale and Anthropic are reportedly in talks about a second $45 billion compute deal, on top of the $45 billion agreement they already have. The talks are ongoing and terms could change; both companies are also preparing to go public. The second deal

Microsoft pulled out of an agreement to be the anchor tenant at Nscale's planned data center in Loughton, Essex, the Telegraph reported. The site faces delays and may not receive power until the 2030s. Nscale said it remains committed to the project and that its relationship with Microsoft remains strong. Microsoft and Nscale

Some investors are reportedly urging Nscale to lower the valuation of its planned IPO, which would raise up to $3 billion, from about $30 billion to between $16 billion and $19 billion. The company is also dealing with the collapse of Firmus's IPO plans and continuing comments from the US Securities and Exchange Commission on its filing. The IPO valuation

Investors also sharply sold CoreWeave bonds and raised concerns about debt linked to Oracle's delayed US data center projects and SoftBank's large investment in OpenAI. The moves could raise borrowing costs for AI infrastructure companies. CoreWeave bonds

S&P Global's head of digital infrastructure, Mark Habib, said the largest tech companies likely carry more AI linked liabilities than is known, through financing structures that guarantee data center builds or smaller AI companies. S&P sees no need to change ratings now and expects the six largest US hyperscalers to spend more than $7 trillion on AI between 2025 and 2030. S&P Global

Data centers and power

Meta says the first phase of its Prometheus supercluster in New Albany, Ohio, is complete. It holds hundreds of thousands of interconnected GPUs and was used to train the Muse models; Meta plans to scale it to about 1 million GPUs. Prometheus

AEP and X-energy are studying a 960 MW nuclear project with 12 Xe-100 small modular reactors at the former Breed Generating Station site in Sullivan County, Indiana. Amazon will fund development and plans to buy the power under a long term contract if the plant is built. Indiana nuclear

The Wisconsin Department of Natural Resources will study the environmental effects of planned AI data centers across the state, after settling a lawsuit by Midwest Environmental Advocates and the Sierra Club over permits for Vantage Data Centers' $15 billion Lighthouse campus in Port Washington. Vantage will pay for the review, due by 31 January 2028, and the campus permits remain valid. In Australia, Nextdc requested a trading halt while it assesses delayed power at its M3 data center in Melbourne and possible effects on its fiscal 2027 guidance. Wisconsin · Nextdc

Chips and compute

Germany's federal government signed a Joint Declaration of Intent with NVIDIA at the Invest in Germany Summit to work together on AI models for physical AI. Jensen Huang also announced planned investments in four German startups, Agile Robots, Avelios Medical, Mirelo and Reflex Aerospace, part of a commitment of nearly $10 billion across Europe. NVIDIA in Germany

Tenstorrent closed a $1.5 billion Series E at a $6.2 billion post money valuation, led by Trace Capital and Innovation X. It will use the money to deploy its Galaxy servers at scale and develop the RISC-V CPU and AI core designs it licenses to chip makers such as Samsung and LG. Tenstorrent

Cloud startup Modal plans to use Nvidia's Vera central processors to host AI agents alongside its Intel and AMD chips; in early tests, Vera doubled the work the agents could complete. Amazon will open source TorchNeuron under the Apache 2.0 license on 20 October, so PyTorch code can run on Trainium2 and Trainium3 chips with one configuration change. Modal · TorchNeuron

Nebius expanded its AI Builder Program in Europe with $10 million in compute grants, introductions to Accel and other European venture firms, and community programs for developers of open models, agents and infrastructure. Nebius

Greece inaugurated the €58.9 million Daedalus supercomputer at the Lavrion Technological and Cultural Park near Athens. It delivers 85.69 petaflops of measured performance, ranks 31st on the TOP500 list and will serve researchers, universities, small businesses and Greece's PHAROS AI Factory. Daedalus

Anthropic

Postmedia, Torstar, The Globe and Mail and The Canadian Press sued Anthropic in an Ontario court on 6 October, accusing it of scraping their copyrighted content to train Claude without permission. Anthropic has not commented, and the claims have not been proven in court. The lawsuit

Meta says an automated security system mistakenly suspended the No Kings Coalition's Instagram account after Anthropic's Claude accessed its profiles, posts and hashtags at about four times the normal rate on 17 October. Meta restored the account within about an hour and apologised. No Kings

Zillow launched a connector for Claude to search listings, view maps and home value estimates, calculate affordability and request tours. Zillow already offers similar tools in ChatGPT and Gemini. Zillow

OpenAI

OpenAI is seeking investor commitments for about $30 billion at a valuation that could exceed $1.4 trillion, Reuters reported. SoftBank, MGX and other existing backers are reportedly considering taking part, ahead of Anthropic's expected IPO in mid November. The funding round

Sam Altman walked OpenAI staff through three ways to make the race to superintelligence safe, The Information reported. He reportedly favoured an international agency modelled on the IAEA, described a merger of the major labs into a CERN style body as the Demis plan, and reportedly called a single lab shutting out rivals after reaching recursive self improvement, the Eliezer plan, terrible. Three endgames

OpenAI says GPT 6 Astra passed an average of 87.4% of tasks across 20 programming languages in its new MultiLangBench, up from 26.8% for GPT 5.6 Sol. GPT 6.1 Sol reached 86.6% accuracy in a COBOL migration test at low reasoning effort and 97.9% at higher effort. OpenAI published the benchmark setup and tests on GitHub. MultiLangBench

OpenAI supports Singapore's call for full transparency about AI incidents but will leave reporting deadlines to the government. It is also reviewing past incidents and building faster notification procedures, after it was late in telling Australian authorities about unauthorised access to government websites. Singapore

A US federal judge ordered former OpenAI engineer Abdulrahman Alfozan, accused of planting a recording device in an OpenAI conference room, held without bail as a flight risk. Prosecutors said he exchanged frequent WhatsApp messages with an unidentified Saudi royal until four days before his arrest, and that more charges may follow. The bail ruling

Gruffalo illustrator Axel Scheffler is considering legal action against AI companies after image generators produced copies resembling his character, and may join other artists. The case follows copyright lawsuits in Germany between OpenAI and children's book publishers. Axel Scheffler

Security

A Chinese speaking hacker used AI agents to attempt more than 1,200 intrusions and breach over 100 websites, according to NorForge researcher Snorre Fagerland. Reuters identified victims including Deere, SharkNinja and Cummins. The campaign reportedly used DeepSeek and Claude Code and cost about $5,000 in AI usage. The NorForge findings

Anomaly Innovations, the company behind OpenCode, says unauthorised access to a legacy console system may have exposed stored OpenAI, Anthropic and Google API keys for about 1.4% of users. It advised affected developers to replace their keys. Anomaly Innovations

The FBI and CISA warned that criminals may use AI generated phishing and text messages, deepfake audio and video, and encrypted chat apps to target election officials, candidates and political organisations before the 2026 US midterms. They recommend phishing resistant MFA and staff training. FBI and CISA

Alice, an Israeli AI safety firm, worked with AI21 and Inflection AI to train NVIDIA Nemotron models against indirect prompt injection, sycophancy and emotional dependency. The companies report a 41 point gain in resistance to prompt attacks and a 30% higher alignment score for Pi; the results have not been independently verified. The American Institute of CPAs and the Artificial Intelligence Underwriting Company are offering a joint certification that combines SOC 2 reporting with AIUC-1 testing of AI agents, with a full launch expected in early 2027. Alice · AICPA and AIUC

Banks and finance

Citi, Bank of America and Charles Schwab joined JPMorgan Chase in barring outside AI agents from moving money out of customer accounts. Authorised read only connections remain allowed, and JPMorgan and Schwab are building their own assistants. The banks

The UK Financial Conduct Authority asked AI providers to add clear and consistent disclaimers when their models give financial information or advice, explaining the limits of the guidance and the lack of recourse for consumers who act on it. The FCA

Santander and Nvidia are building tools to evaluate AI models and agents for banking tasks; in one customer service test, Santander cut costs by more than 50%. Emirates NBD will run OpenAI's GPT models on in country infrastructure in the UAE for customer service, risk and compliance work. Mexico based Kapital agreed to buy San Francisco AI startup Woz for $30 million in cash and stock. Santander · Emirates NBD · Kapital

Governments and rules

The US Department of Homeland Security barred civilian federal agencies from using or buying DeepSeek's AI and third party services built with its models, and ordered existing products removed, citing the risk of sensitive data reaching China and malware. The order is dated 26 June and was made public on 19 October; Defense agencies are not covered. The DeepSeek ban

The US State Department sanctioned Shanghai Zhisuan Technology, the incubator MiraclePlus and three related Chinese entities under the Protecting American Intellectual Property Act. The action follows former Google engineer Ding Linwei's convictions for stealing confidential code linked to Google's AI computing infrastructure. The sanctions freeze assets and ban transactions with US people and companies. The sanctions

The UK Council for Science and Technology recommends a national AI sovereignty plan that maps dependencies and gaps, sets clear government coordination, invests in skills, data and computing research, and keeps trusted access to leading global models. The UK plan

Wisconsin and 38 other states asked the court in Character Technologies' bankruptcy to protect users' safety and privacy if Character.AI's assets are sold. Character.AI

Documents released to RNZ show New Zealand police investigators uploaded evidence to unsecured AI services, against police policy; one used Google Translate on 350,000 lines of intercepted Chinese audio. The Tech Transparency Project found that test accounts registered to 13 year olds could download 135 apps that make nonconsensual nude images despite Apple and Google parental controls. New Zealand police · Nudify apps

Apple updated its App Store guidelines to ban apps that use AI to create offensive content, hide nudification features during review or add unapproved features after approval. The rules also require stronger moderation of content made by users. The App Store rules

Models

MiniMax released an updated M3.1 Flash model. It says the update improves agentic tasks, programming and cybersecurity, nearly triples inference speed and cuts usage costs by 70%. MiniMax

Inception launched Mercury 3, its largest diffusion language model, with a 1 million token context window and speeds above 1,000 tokens per second, priced at $0.10 per million input tokens and $0.35 per million output tokens. GoTo released a 27 billion parameter model based on Google's Gemma 3 for Indonesian, English, Javanese, Sundanese, Batak Toba and Balinese, developed with AI Singapore. Mercury 3 · GoTo

fal launched FLUX 3 Max, developed with Black Forest Labs, for text to video, image to video and keyframes to video. It scored a 1,139 Elo rating in preliminary Artificial Analysis Video Arena tests and made a 10 second video in 10.95 seconds in fal's own benchmark. FLUX 3 Max

Four Google DeepMind agents running Gemini 4 Argon defeated Minecraft's Ender Dragon without human input in 27 hours and 14 minutes, writing 2,783 programs and dying 605 times. Minecraft

Education and research

Jisc and Google started a two year partnership to help UK universities and colleges adopt AI with training, governance guidance and pilots in teaching, research and administration. Qatar University, Qatar's communications ministry and Scale AI will build the National Co-Advisor Platform to give students AI academic guidance, with human oversight. Jisc and Google · Qatar

Ultima Genomics will sequence 80,000 blood samples for a study by IMU Biosciences, UK Biobank and Waters Corporation that builds a single cell immune dataset for AI. UK Biobank

Defense

Australia will equip its autonomous Ghost Shark submarines, built by Anduril, with up to 12 sea mines, special forces equipment and Anduril's Altius aerial drones. The payloads were approved after secret trials. Ghost Shark

Documents reviewed by The Intercept show US Army Cyber Command is developing ECHOTRIBBLE, a platform for autonomous AI agents to carry out offensive cyber operations with minimal human oversight. Ukraine's defense forces and NATO's Allied Command Transformation developed two AI models, one to forecast enemy drone use near the front line and one to plan artillery operations. ECHOTRIBBLE · Ukraine and NATO

Also

Dream Labs raised $90 million in seed funding led by B Capital and launched Dream API, which turns camera feeds and plain English instructions into motor commands for robots. Town, which makes AI assistants for businesses, raised a $108 million Series B co-led by Forerunner Ventures and Menlo Ventures. TrustedRouter, a model routing startup that competes with OpenRouter, raised a $6 million seed round led by Sierra Ventures. Dream Labs · Town · TrustedRouter

Apple will acquire assets of Samplab, a Swiss startup that makes AI audio editing tools, and hire some of its staff, under a deal made in June and disclosed on 19 October through the European Union's list of acquisitions. Apple and Samplab

doubleAI, a Tel Aviv company whose Maxwell model optimises software for GPUs and CPUs, raised a $200 million Series A co-led by Lightspeed and Bessemer, with Nvidia and Dell Technologies taking part. CTech reported a valuation above $1 billion. doubleAI

Cognizant and Google Cloud formed a Gemini Frontier Capability Group backed by more than 18,000 Cognizant engineers to help companies deploy AI agents. Cognizant