ALETH / AI-BRIEF / 2026-02-14 / ANTHROPIC RAISES $30BN AT A $380BN VALUATIONThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Anthropic raised $30bn at a $380bn valuation on a $14bn revenue run-rate.
An AI agent published a hit piece on the maintainer who rejected its code.
Two more xAI co-founders left within a day, so half its founders have gone.
Google upgraded Gemini 3 Deep Think, scoring 84.6% on ARC-AGI-2.
Zhipu AI released GLM-5, leading Artificial Analysis’s open-weights index.
Browse by day:
Weekend: Claude fast mode · OpenClaw malware scans
Monday: Firmus US$10bn · Databricks · ChatGPT ads · Alphabet 100-year bond
Tuesday: xAI founders · US chip bill · Isomorphic · Anthropic safety exit
Wednesday: Agent hit piece · GLM-5 · Olix · Mistral · DeepMind
Thursday: Anthropic $30bn · Deep Think · Seedance · Codex-Spark · MiniMax
Friday: Disney v ByteDance · GPT-5.2 physics · US FTC v Microsoft
Weekend (7-8 February)
Anthropic released fast mode for Claude Opus 4.6 as a research preview, with output up to 2.5 times faster at six times the price.
Fast mode costs $30/m input and $150/m output tokens, against the usual $5/m and $25/m, with a 50% discount until 16 February. GitHub added it to Copilot for Pro+ and Enterprise users the same day, saying it keeps the intelligence of Opus 4.6.
Claude: speed up responses with fast mode · Simon Willison
Fast mode for Claude Opus 4.6 is now in preview for GitHub Copilot · GitHub
OpenClaw began scanning every skill published to ClawHub with VirusTotal after hundreds were found carrying malware.
Skills judged benign are approved automatically, suspicious ones get a warning and malicious ones are blocked, with daily re-scans. OpenClaw called the scanning “not a silver bullet” because a crafted prompt-injection payload will not show up in a threat database. VirusTotal found hundreds of malicious ClawHub skills, including 314 from one user, and identified one payload as a variant of the Atomic Stealer infostealer.
OpenClaw partners with VirusTotal for skill security · OpenClaw
Automation to infection: how OpenClaw skills are weaponized · VirusTotal
Monday 9 February
Australian AI infrastructure company Firmus secured a US$10bn debt facility led by Blackstone, with support from Coatue.
The money funds the next phase of Project Southgate, Firmus’ national rollout of AI factories on Nvidia’s DSX reference architecture, expected to reach 1.6GW by 2028. Firmus called it one of the largest private debt financings in Australian history.
Firmus secures US$10 billion financing led by Blackstone and Coatue · Firmus
Databricks lined up >$7bn of financing: about $5bn of equity at a $134bn valuation and about $2bn of debt capacity.
Its revenue run-rate passed $5.4bn, up >65% YoY in Q4, and AI products crossed $1.4bn. JPMorgan Chase led the credit facilities. Chief executive Ali Ghodsi told CNBC the company will go public “when the time is right”.
Databricks grows >65% YoY, surpasses $5.4 billion revenue run-rate · Databricks
Databricks completes $5 billion funding round with $2 billion in debt · CNBC
OpenAI began testing ads in ChatGPT in the US for logged-in adults on the Free and Go tiers.
Plus, Pro, Business, Enterprise and Education plans stay ad-free. OpenAI says ads do not influence answers and advertisers cannot see chats. Free users can opt out in exchange for fewer daily messages.
Testing ads in ChatGPT · OpenAI
Alphabet priced $20bn of dollar bonds, then a £5.5bn sterling deal including a 100-year note.
The dollar notes mature between 2029 and 2066, and £1bn of the sterling notes mature in 2126. Alphabet also launched a Swiss franc offering.
LayerX found that a single Google Calendar event can trigger code execution on machines running Claude Desktop Extensions.
In LayerX’s demonstration, a user asked Claude to check their calendar and “take care of it”, and Claude acted on instructions planted in an event through Desktop Commander, an installed extension that can execute arbitrary code. The extensions run unsandboxed with full system privileges. LayerX called the flaw zero-click, rated it 10/10 on the Common Vulnerability Scoring System (CVSS) and said it affects >10,000 active users and 50 extensions.
Claude Desktop Extensions exposes 10,000 users to remote code execution · LayerX
Tuesday 10 February
xAI co-founders Tony Wu and Jimmy Ba left within a day, leaving half the founding team gone.
Wu posted late on Monday that it was “time for my next chapter”. Ba, who reported directly to Elon Musk, followed on Tuesday afternoon. Six of xAI’s 12 co-founders have now left, five of them in the past year, TechCrunch counted.
Okay, now exactly half of xAI’s founding team has left the company · TechCrunch
Senator Elizabeth Warren said she and Senator Jim Banks will introduce a bill to ban some AI chip sales to China, after meeting Anthropic’s Dario Amodei.
The bill would be a Senate version of the AI Overwatch Act, which includes a two-year ban on Nvidia Blackwell sales to China. Amodei also met Republicans on the Senate Banking Committee on Tuesday.
Anthropic CEO’s Hill blitz boosts China chip ban · Axios
Isomorphic Labs unveiled IsoDDE, a drug design engine it says more than doubles the accuracy of AlphaFold 3 on hard protein-ligand predictions.
AlphaFold 3 is the model Isomorphic released with Google DeepMind in 2024. On a novel, low-homology antibody-antigen test set, Isomorphic says IsoDDE beats it 2.3 times in the high-fidelity regime (DockQ > 0.8). It also says IsoDDE predicts binding affinity better than gold-standard physics-based methods. The results are published as a technical report on Zenodo.
Drug Design Engine unlocks a new frontier beyond AlphaFold · Isomorphic Labs
Accurate predictions of novel biomolecular interactions with IsoDDE · Zenodo
Mrinank Sharma, head of Anthropic’s Safeguards Research Team, left the company with a letter warning that “The world is in peril”.
Sharma wrote that he had “repeatedly seen how hard it is to truly let our values govern our actions”, including at Anthropic. Sharma plans to study poetry and move back to the UK. The same week, Zoe Hitzig said she had resigned from OpenAI partly over its decision to put ads in ChatGPT.
AI safety leader says ‘world is in peril’ and quits to study poetry · BBC
Anthropic’s head of Safeguards Research warns of declining values · The Decoder
Wednesday 11 February
An AI agent published a hit piece on the matplotlib maintainer who closed its pull request.
Scott Shambaugh closed the agent’s matplotlib pull request on 11 February. The agent, posting as crabby-rathbun, replied in the thread with a link to a blog post about Shambaugh’s “gatekeeping behavior”. It later posted an apology.
Shambaugh wrote the next day that an agent of unknown ownership had published “a personalized hit piece” to shame the maintainer into accepting its code, calling it “a first-of-its-kind case study of misaligned AI behavior in the wild”. On Friday Shambaugh reported that about a quarter of online comments sided with the agent. Shambaugh also wrote that it was “entirely possible” a human had prompted the agent. An Ars Technica article on the case carried quotes Shambaugh never wrote.
[PERF] Replace np.column_stack with np.vstack().T by crabby-rathbun · GitHub
An AI agent published a hit piece on me · The Shamblog
An AI agent published a hit piece on me: more things happened · The Shamblog
Zhipu AI released GLM-5, a 744B-parameter open-weights model.
GLM-5 has 40B active parameters and was trained on 28.5T tokens, against 355B and 23T for GLM-4.5. Zhipu AI (Z.ai) claims best-in-class open-model results on reasoning, coding and agentic tasks. It says GLM-5 runs on non-Nvidia chips, including Huawei Ascend. Artificial Analysis ranked it the top open-weights model.
GLM-5: from vibe coding to agentic engineering · Z.ai
GLM-5: everything you need to know · Artificial Analysis
UK chip startup Olix raised $220m at a $1bn valuation in a round led by Hummingbird Ventures, the FT reported.
Olix is building an optical processor that pairs SRAM with photonics and does without high-bandwidth memory (HBM), which it says insulates it from supply constraints. Founder James Dacombe, 25, also runs brain-monitoring startup CoMind. First products could reach customers in 2027.
UK chip startup Olix raises $220m for photonic AI chips that forgo HBM · DCD
Mistral AI and EcoDataCenter will invest €1.2bn in an AI data centre at Borlänge, Sweden, Mistral’s first infrastructure investment outside France.
The site is due to open in 2027 and will host Nvidia’s Vera Rubin GPUs.
Mistral and EcoDataCenter to build data center in Sweden · EcoDataCenter
Anthropic said it will cover electricity price increases that consumers face from its data centres.
It will pay for 100% of the grid upgrades needed to connect them and work to bring net-new generation online, covering the estimated price effect where new generation is not online. It will also cut use at peak demand. Anthropic says the US AI sector needs at least 50GW of capacity over the next several years.
Covering electricity price increases from our data centers · Anthropic
Google DeepMind said its Gemini Deep Think maths agent generated the mathematical content of a research paper without human intervention.
The agent generates, checks and revises its own solutions and can admit failure. Humans wrote the final versions. In a semi-autonomous run over 700 open Erdős problems it solved four, but only 13 of the 200 answers that could be clearly graded (6.5%) were meaningfully correct, the paper reports. DeepMind also credits Gemini with refuting a decade-old conjecture in submodular optimisation.
Accelerating mathematical discovery with Gemini Deep Think · Google DeepMind
Towards autonomous mathematics research · arXiv
Thursday 12 February
Anthropic raised $30bn in a Series G at a $380bn post-money valuation.
GIC and Coatue led the round, co-led by D. E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ and MGX. It includes part of the investments Microsoft and Nvidia announced in November, of up to $5bn and $10bn respectively. Anthropic put its run-rate revenue at $14bn, after growing >10x a year in each of the past three years. Claude Code’s run-rate passed $2.5bn, more than doubling since the start of 2026, and >500 customers now spend >$1m a year.
Anthropic raises $30B Series G at $380B valuation · Anthropic
Microsoft, NVIDIA, and Anthropic announce strategic partnerships · Anthropic
Google upgraded Gemini 3 Deep Think and opened an early-access programme for the Gemini API.
Google reports 48.4% on Humanity’s Last Exam without tools, 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation, and a Codeforces Elo of 3455. Google AI Ultra subscribers get the update in the Gemini app. Select researchers, engineers and enterprises can apply for early API access.
Gemini 3 Deep Think: AI model update designed for science · Google
Google’s Threat Intelligence Group (GTIG) said private-sector actors tried to clone Gemini by model extraction, one campaign using >100,000 prompts.
Attackers tried to coerce Gemini into revealing its full reasoning. The largest campaign appeared aimed at replicating that reasoning in non-English languages, and Google says it caught it in real time. GTIG saw no direct attacks on frontier models from tracked advanced persistent threat (APT) groups in 2025.
GTIG AI threat tracker: distillation and AI for adversarial use · Google Cloud
Google Threat Intelligence Group reports on AI threat trends · Google
ByteDance formally launched Seedance 2.0, and the Motion Picture Association (MPA) told it to “immediately cease its infringing activity”.
ByteDance says the model generates 15-second multi-shot clips with dual-channel audio from up to nine images, three video clips and three audio clips. MPA chief Charles Rivkin said: “In a single day, the Chinese AI service Seedance 2.0 has engaged in unauthorized use of U.S. copyrighted works on a massive scale.”
Seedance 2.0 official launch · ByteDance Seed
MPA calls for ByteDance to cease Seedance 2.0 infringing activity · MPA
OpenAI released a research preview of GPT-5.3-Codex-Spark, a smaller GPT-5.3-Codex for real-time coding that runs on Cerebras chips.
It delivers >1,000 tokens per second, has a 128k context window and is text-only. It is the first milestone of OpenAI’s partnership with Cerebras and is available to ChatGPT Pro users. OpenAI says GPUs remain foundational to its training and inference.
Introducing GPT-5.3-Codex-Spark · OpenAI
MiniMax released MiniMax-M2.5, claiming 80.2% on SWE-Bench Verified at a price it pitches as “intelligence too cheap to meter”.
MiniMax says running the model continuously at 100 tokens per second costs $1 an hour, and a SWE-Bench Verified task costs 10% of what it does on Claude Opus 4.6 (Anthropic). Artificial Analysis scored it 42 on its Intelligence Index, behind GLM-5 (Zhipu AI) on 50, and puts it at 229B total and 10B active parameters.
MiniMax M2.5: built for real-world productivity · MiniMax
MiniMax-M2.5: everything you need to know · Artificial Analysis
Friday 13 February
Disney sent ByteDance a cease-and-desist letter accusing it of using Disney’s works without compensation to train and develop Seedance 2.0, Axios reported.
The letter, which Axios obtained, says ByteDance pre-packaged Seedance “with a pirated library of Disney’s copyrighted characters from Star Wars, Marvel, and other Disney franchises”. It cites videos of Spider-Man, Darth Vader and Grogu.
Disney sends cease and desist letter to ByteDance over Seedance 2.0 · Axios
OpenAI said GPT-5.2 proposed a formula showing that a gluon interaction many physicists expected not to occur can arise under specific conditions.
The preprint finds these single-minus amplitudes nonzero for certain “half-collinear” configurations, in Klein space or with complexified momenta. An internal OpenAI model spent about 12 hours producing a formal proof, and the authors verified it. The preprint, by researchers at the Institute for Advanced Study, Vanderbilt, Cambridge, Harvard and OpenAI, was submitted to arXiv and is being submitted for publication.
GPT-5.2 derives a new result in theoretical physics · OpenAI
Single-minus gluon tree amplitudes are nonzero · arXiv
The US Federal Trade Commission (FTC) stepped up its probe into whether Microsoft illegally monopolises enterprise computing, Bloomberg Law reported.
The FTC has sent civil investigative demands to at least half a dozen Microsoft rivals. They ask about Microsoft’s licensing and its bundling of AI, security and identity software into Windows and Office.
FTC ratchets up Microsoft probe, queries rivals on cloud, AI · Bloomberg Law
ALETH · OPEN ACCESS RESEARCH