The Meltdown — March 8-14
Refusing the Pentagon turned into a lawsuit, Meta admitted its model isn't good enough, and Europe started writing checks that rival Silicon Valley.
01 — Anthropic is suing the Pentagon
Anthropic filed lawsuits in California and DC after the Department of Defense designated it a "supply chain risk," a label normally reserved for companies like Huawei and Kaspersky, killing hundreds of millions in federal contracts. Last week we covered Anthropic quietly reopening talks with the Pentagon, but that door is firmly shut now. This reads like them deciding that getting punished for having principles is better spent setting legal precedent than negotiating from weakness. The same week they launched a $100M Claude Partner Network for enterprise customers, a deliberate pivot: if the government won't buy, corporate will.
02 — Meta's next model isn't ready, and they know it
Meta pushed its next-gen "Avocado" model from March to May or later after internal benchmarks showed it trailing Gemini 3.0, OpenAI, and Anthropic on reasoning, coding, and writing. This is a company spending $115-135B on AI capex in 2026, and they're reportedly discussing temporarily licensing Google's Gemini to fill the gap. Worth sitting with that for a moment: you don't spend $130B and then rent from a competitor unless something has gone genuinely sideways in your training pipeline.
03 — Yann LeCun raised a billion dollars to bet against LLMs
AMI Labs, LeCun's new Paris-based venture, closed a $1.03B seed round at a $3.5B pre-money valuation, the largest seed in European history, with Nvidia, Temasek, and Bezos Expeditions in. The thesis is "world models" over language models, machines that build internal simulations of reality rather than predicting the next token. LeCun has been saying for years that autoregressive LLMs are a dead end, and now he has the funding to prove it. The skeptical take: four-month-old companies don't normally command $3.5B valuations, so investors are clearly paying for the name as much as the thesis.
04 — Atlassian cut 1,600 jobs and called it an AI strategy
Atlassian CEO Mike Cannon-Brookes said AI is changing the "skills mix" at the company, and 10% of the workforce is out, costing $225-236M in restructuring charges. This follows Block cutting 4,000 people with similar AI justifications last month, and Atlassian's stock has lost over half its value this year. We're probably watching "AI-washing" become the default cover story for layoffs that would've happened anyway. Some of these cuts are real capability shifts, but when every CEO uses the same language, the specificity disappears and it just becomes an excuse investors won't question.
05 — ByteDance quietly built a supercomputer in Malaysia
ByteDance deployed roughly 36,000 Nvidia B200 chips through a Southeast Asian cloud partner called Aolani Cloud, about $2.5B worth of hardware across ~500 Blackwell systems in Malaysia. U.S. export restrictions ban direct chip sales to Chinese companies, but third-country cloud deployments technically comply with current rules, and Nvidia confirmed it has "no objections." Everyone in the industry knows this is happening, Washington included. Hard to see how the rules survive contact with this playbook, because every Chinese AI lab with enough cash will copy it.
06 — OpenAI bought its way into agent security
OpenAI acquired Promptfoo, an open-source AI security platform used by 350K+ developers and over 25% of the Fortune 500, for what was previously an ~$86M valuation. It'll plug into OpenAI Frontier, their enterprise agent platform, and Promptfoo stays open source. That price feels like a bargain for a tool already embedded in most major companies' AI stacks, and as agents get more autonomy, security testing becomes the bottleneck. OpenAI is buying the plumbing before everyone else realizes they need it.
07 — Washington State wants AI chatbot guardrails
Washington passed HB 2225, requiring AI chatbots to remind users they're talking to a machine (hourly for minors, every three hours for adults), detect suicidal ideation, and take reasonable measures to prevent chatbots from simulating romantic relationships with minors. There are now 78 chatbot-specific bills alive in 27 states. The hourly reminder feels silly, but the minor protections and suicidal ideation detection are serious policy. State-by-state regulation creates a patchwork that's expensive to comply with and easy to ignore, which probably accelerates the push for a federal framework, exactly what the big AI companies want since they'll have more influence over one bill than twenty-seven.