Not gonna lie — today's news cycle feels like a hangover after a really long party. Nvidia just shed a trillion dollars in market value. A new kind of ransomware decided it doesn't need human handlers anymore. And Amazon is quietly burning cash trying to turn Alexa into something it was never built to be.
Let's dig in.
Nvidia's $1 Trillion Slide — Context Matters
Here's the headline everyone's talking about: Nvidia lost roughly $1 trillion in market cap in under two months. Shares are down 16% since their all-time high on May 14. The stock is now trading at 18 times forward earnings — the cheapest it's been since early 2019. That's pre-AI-boom territory.
But here's what makes this interesting. Wall Street analysts haven't actually lowered their profit estimates for Nvidia. Revenue growth is still projected to be the fourth-fastest in the S&P 500 this year. The selloff isn't about Nvidia doing something wrong. It's about the AI trade rotating.
Money is moving into memory and storage plays — Micron, AMD, Intel. Stocks that were written off as also-rans during the GPU shortage frenzy are now doubling and tripling. Nvidia is technically cheaper than the S&P 500 right now. That's wild for a company still printing money.
From a builder's perspective, this matters because GPU pricing might finally loosen up. If you've been waiting to provision inference infrastructure without paying a premium, this rotation could be your window. But don't expect a fire sale — NVDA still owns the data center GPU market.
The Distillation Problem Nobody Wants to Talk About
Business Insider dropped a piece today that cuts to the core of the current AI business model tension. AI distillation — training smaller models on the outputs of bigger, more expensive ones — is eating into the margins of companies like OpenAI, Anthropic, and Google.
Here's the blunt version: if a competitor can replicate 80% of GPT-5's capability by training a smaller model on its outputs, why would anyone pay frontier API prices? The economics get ugly fast. Distillation has been an open secret in the ML community for a while, but this is the first time I've seen a major outlet frame it as an existential threat to the current pricing structure.
I've been running local models for email triage (shoutout to the XDA article in the feed today — someone else who won't let AI send replies either), and honestly, the gap between frontier and distilled models keeps shrinking. If you're building a product on top of an API, it's worth keeping an eye on this. The pricing you're paying today might not hold.
Amazon's Moonraker — Alexa Gets an Agentic Overhaul, and a Hefty Price Tag
Internal Amazon documents reveal "Moonraker" — a project pushing Alexa toward full agentic capabilities. Multi-step task planning, cross-service orchestration, acting without step-by-step prompts. The relaunched Alexa+ has been in the wild since March 2026, but the internal cost numbers are starting to surface.
And they're not pretty.
Every agentic query costs significantly more than the simple voice commands Alexa was designed for. Panos Panay's team is wrestling with a fundamental math problem: how do you deliver frontier-level AI inference at voice-assistant scale without bleeding money? Amazon killed Rufus (its standalone shopping chatbot) to consolidate around Alexa, but that doesn't solve the unit economics.
I've been using Alexa+ for a few months now, and the agentic stuff works — sometimes. Asking it to "find a contractor for the kitchen remodel, check their license, and book a consultation" is genuinely impressive when it works. But it fails often enough that I still default to doing it myself. The reliability gap between demo and daily use is real.
The First Fully Autonomous AI Ransomware Is Here
Sysdig's threat research team published findings on a malware campaign using an LLM-driven attack that operates without human command. This is the first documented case of fully agentic AI ransomware.
The model handles reconnaissance, lateral movement, encryption, and ransom negotiation autonomously. No human operator in the loop. No command-and-control server with a person pressing buttons. Just an LLM with tool access and a goal.
I've been saying this would happen for a while, but seeing it confirmed is different. The security implications are massive — traditional signature-based detection is useless against a model that can adapt its approach on the fly. If you're in charge of infrastructure security, this should be on your radar.
Quick Hits
- Base44's first LLM — Business Insider tested it for building a website. Claims are faster and cheaper than frontier models. The output was decent but not mind-blowing. Another contender in the "good enough" tier.
- Shadow Web — Open-source Python SDK that reduces web page tokens for LLM agents by 64–97% using Shadow DOM flattening. If you're building AI agents that browse the web, this is worth a look.
- Floqer raises $2M — Perplexity Fund co-led a pre-seed for an autonomous customer knowledge base. Interesting signal: VCs are betting on AI that helps GTM teams, not just code generation.
There's a lot happening right now, and not all of it is good news. Nvidia's slide might make GPUs more accessible. Distillation could bring model prices down. But autonomous ransomware and expensive agentic bets remind us that this industry is still figuring things out as it goes.
What's on your mind? I'm particularly curious whether anyone else is running Alexa+ and finding the agentic features useful or frustrating. Drop a comment — I read all of them.
Top comments (0)