AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
Google's $30B GPU Bet & AI-Designed Vaccines
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey, welcome to AI Mornings with Andreas Vig. It's Monday, June 8th. Let's start with a massive infrastructure story. Google just signed what might be the largest compute deal in AI history. The company will pay SpaceX 920 million US dollars per month from October this year through June 2029. That's roughly 30 billion US dollars total for access to about 110,000 Nvidia GPUs at SpaceX's XAI data centers in Memphis. The deal was disclosed in an SEC filing on Friday and positions Google to handle surging demand for Gemini Enterprise Services. What makes this particularly interesting is the timing. SpaceX is expected to begin trading on the NASDAQ this week at an estimated US$15 trillion valuation. That would be the largest IPO ever. And uh Google isn't the only one writing big checks. Anthropic separately agreed in May to pay SpaceX $1.25 billion US dollars monthly for compute access. The AI arms race has officially become an infrastructure arms race. Now to a breakthrough in medical AI. Researchers at the University of Cambridge have completed the first human trial of a vaccine whose core component was designed entirely by artificial intelligence. They call it a super antigen vaccine and it targets the entire SARB-Co coronavirus family. The idea is to protect against current variants and future strains that don't even exist yet. The FACE 1 safety trial involved 39 participants. A larger study with about 200 people is now underway to measure immune responses. The Cambridge team is already applying the same AI-driven approach to influenza, bird flu, and Ebola-related viruses. This could fundamentally change how we prepare for pandemics designing vaccines for threats before they emerge. In the Model Wars, DeepSeek is making headlines again. The Chinese lab's V4 Pro model just beat OpenAI's GPT 5.5 Pro in a head-to-head precision benchmark. DeepSeq scored 38 points to OpenAI's 33. The test evaluated both models on fresh tasks, including Python log redaction, vendor communication, meeting summarization, and JSON conversion. DeepSeq won three tasks and Tide 1. The key difference. DeepSeek followed instructions more literally and produced more reliable code under constraints. OpenAI's model drifted from the prompt, adding unprompted details and breaking schema requirements. This continues DeepSeek's pattern of competing hard on both price and performance against Western frontier models. Microsoft is sending a message to the AI market about pricing. The company's decision to shift GitHub Copilot from a flat monthly fee to per token billing is being called the token pocalypse. And it's a sign of things to come. As AI companies like Anthropic head toward public markets, they face awkward questions about profitability. The heavily subsidized ecosystem we've gotten used to is starting to pass real costs to customers. The question analysts are asking: can AI labs collapse their costs fast enough to meet what customers are willing to pay? Or are we going to see widespread usage caps and restrictions as the new normal? Uber, interestingly, already hit this wall. They burned through their annual AI budget in four months and had to impose spending caps on employees. Alright, a few more things worth knowing about today. Meta's AI systems keep having security problems. Attackers exploited Meta's AI-powered customer support agent to hijack prominent Instagram accounts. The technique was simple. They asked the AI for a password reset and email change, and the AI sent verification codes to the attackers' email addresses. Compromised accounts included ones associated with Barack Obama, Sephora, and a Space Force official. Meta says the exploit is now fixed and they're securing affected accounts, but it's a reminder that basic AI agents handling sensitive tasks can be vulnerable to simple social engineering. Small businesses are going all in on AI agents. A growing number of entrepreneurs are using platforms like OpenClaw to automate customer service, bookkeeping, marketing, scheduling, and research. Some report dramatic productivity gains and reduced staffing needs, but users are also dealing with reliability issues, hallucinations, security vulnerabilities, and prompt injection risks. It's a preview of how white-collar work might evolve powerful tools that still need human oversight. In Science News, a Columbia University team has posted the first use of precision-based editing on human embryos. They made single-letter DNA changes related to cholesterol regulation and fetal hemoglobin. The embryos showed mosaicism, meaning some cells were edited and others weren't. The lead researcher emphasizes this technology isn't ready for clinical use, but bioethicists worry the preprint could trigger premature commercial attempts anyway. Amazon introduced a new conversational warehouse robot called Proteus that follows natural language commands. The company also announced a 10 billion euro investment in European infrastructure, continuing its push to integrate AI-powered automation into logistics. And finally, a quick note on a minor disruption. Notion temporarily disabled all Anthropic models yesterday due to elevated errors on Claude. Twelve hours later, they restored access after Anthropic confirmed a brief infrastructure issue had been resolved. That's all for today. See you tomorrow.