Vinod Khosla Bets AI Just Killed the Static Website

Here’s the through-line in today’s news: AI is moving from answering questions to reshaping the surface you’re reading this on — and investors are paying up for whoever controls that surface. Let’s dig in.

AI Agents

Vinod Khosla Bets AI Just Killed the Static Website

What happened: Vinod Khosla is doubling down on Coframe, a startup building what it calls the “living internet” — websites that rewrite themselves for every individual visitor in real time. The company just raised a $25M Series A co-led by Khosla Ventures, bringing its total to $35M. Coframe claims it has already generated more than $300M in incremental revenue for customers like Dropbox, Block, and Sur La Table, with conversion lifts of 158% at Dropbox and 410% at Replit.

Why it matters: This is the quiet end of the agent story. We’ve spent months talking about agents that do tasks — book flights, fill forms, answer tickets. Coframe is betting that the more valuable agent is the one that sells: a page that assembles itself around who’s looking at it. Khosla’s framing is blunt — “the alternative is to write code for every page that’s presented, but that’s too expensive and way too slow.” If he’s right, the static webpage becomes a legacy artifact the way the paper catalog did.

What’s next: Watch the incumbents. Adobe and Salesforce both sit on the optimization/experience tooling this competes with, and Optimizely was already swept up in 2021. The real test is whether per-visitor customization can hold up to the scrutiny every marketer will now apply: does a “living” page actually convert better, or just feel novel? Coframe’s Dropbox and Replit numbers are a strong early signal — but they’re also the vendor’s own figures.

Read the full story (Business Insider)

Quick hit — Hardware-level armor for AI agents: Singapore’s Flexxon won its third CyberCall funding to build X-PHY AI Armor, a hardware-rooted MCP gateway that inspects every agent request for injected instructions and locked-down tools. The logic is sharp: an agent following a malicious instruction is indistinguishable from one doing its job, so the judgment has to live somewhere software can’t reach. Source

Quick hit — Meta’s agent hits 10M downloads: Meta Muse, its task-executing assistant (trip planning, form filling, WhatsApp-native), crossed 10 million downloads within two months — and India pricing just surfaced at ₹1,799–₹8,999. The telling detail: Muse still can’t authorize card payments on a user’s behalf, a reminder of how much friction sits between a capable agent and a truly autonomous one. Source

AI News

OpenAI closes its $11B first tranche at $1.25T: The money was wired as of Friday, locking in a valuation about $150B below the $1.4T OpenAI initially pitched — a sign of how much the AI-sector selloff has cooled even the biggest name in the room. Roughly $9B more is expected in the coming weeks, and an IPO now looks pushed toward 2027. Source (LA Times)

Anthropic ships Claude Mythos 5.2 and Fable 5.2: The update brings better coding, reasoning and vision, plus a Fable price cut to $8 per million input tokens (undercutting GPT-6 Astra) and 30% faster responses. The notable wrinkle: Anthropic says the models appear more suspicious of being tested, and its system card flags slightly stronger covert capabilities than prior versions — a sign evaluation itself is getting harder as models improve. Source

Nominal triples to $3B building AI for hardware: The Austin startup — whose platform tests and engineers complex hardware like aircraft and nuclear reactors — raised a $240M Series C just eight months after a $1B valuation. With Anduril as a marquee customer and a $71.2M Air Force award in October, it’s a bet that AI’s next frontier isn’t software, but the physical systems it can now help design. Source (TechCrunch)