The Full Picture of Anthropic's Compute Deal: $1.25 Billion a Month to xAI
The price of the deal in which Elon Musk's xAI supplies compute resources to Anthropic has come to light. According to TechCrunch, Anthropic pays xAI $1.25 billion per month. The contract works out to $15 billion a year — an unprecedented scale for a compute deal between AI companies.
The deal pairs an odd couple: xAI, founded by Musk after he left OpenAI, and Anthropic, which itself was built as an OpenAI rival. What stands out is that the importance of compute in AI development has created a structure where cooperation now transcends competition.
xAI's $6.4 Billion Loss Revealed in SpaceX's IPO Filing
SpaceX's IPO filing has disclosed xAI's 2025 financials for the first time. According to TechCrunch, xAI recorded a $6.4 billion loss in 2025. With a massive Grok expansion plan underway, spending is expected to climb even further.
The numbers lay bare the scale of Musk's AI ambitions, but burning through billions with no visible path to profitability is also raising questions about sustainability. Selling compute — the Anthropic deal being one example — is seen as a way for xAI to help offset the red ink.
NVIDIA Commits $90 Billion in AI Deals
According to Semafor, NVIDIA has committed $90 billion to AI-related deals. It signals that the company is accelerating its strategy of pairing startup investments with dealmaking, leveraging its position at the heart of AI infrastructure.
According to TechCrunch, NVIDIA posted record quarterly revenue, while also disclosing that its equity stakes in startups have reached $43 billion. AI demand remains strong, but the company has flagged slowing revenue growth, and investors are reacting cautiously.
Cohere Releases Command A+ — Its First MoE Model
Cohere has released a new open-weights model, Command A+. According to an announcement on Reddit by Cohere co-founder Nick Frosst, it is the company's first MoE (Mixture of Experts) model, with 25 billion active parameters and 218 billion total parameters.
It's available under the Apache 2.0 license and optimized to run efficiently on just one or two GPUs. The model specializes in multilingual support and agentic use cases, with a design that prioritizes enterprise practicality. It has also been ranking high on Hugging Face's trending charts.
HalBench: Testing Four Models for Sycophancy and Hallucination
HalBench, published on Reddit, is an open benchmark that measures LLM sycophancy and hallucination. It put four frontier models (Sonnet 4.6, Grok 4.3, GPT 5.4, and Gemini 3.1 Pro) to the test using 3,200 false-premise prompts.
Sonnet 4.6 posted the highest score (0.565), followed by Grok 4.3 (0.498), GPT 5.4 (0.381), and Gemini 3.1 Pro (0.339). A higher score means the model was able to push back on a false premise.
Particularly interesting is how differently each model fails: GPT tended to comply outright, while Gemini tended to generate the content first and then attach a disclaimer. All four models also struggled to detect "false attributes of real entities," which turned out to be the weakest area across the board. The dataset and code are fully public.
Cloudflare's CEO Tells the WSJ How the Company Picks Employees to Replace with AI
Cloudflare's CEO wrote a Wall Street Journal op-ed explaining how the company decides which employees to replace with AI. It has picked up 16 points on Hacker News and sparked debate.
It has been rare for a tech CEO to publicly discuss the criteria for AI-driven headcount cuts, but such statements are multiplying as AI adoption accelerates. With observers also pointing out that replacing New Zealand's civil servants with AI carries hidden costs, debate over AI's impact on jobs is heating up across countries.
---