Follow the money—then follow the silicon. Anthropic just closed a $30 billion Series G at a $380 billion valuation, making it the second-largest private funding round in tech history. We dug into what the investor list reveals—GIC, Coatue, D.E. Shaw, Founders Fund, and Abu Dhabi’s MGX. Annualized revenue hit $14 billion. Claude Code alone generates $2.5 billion. And with nearly $64 billion raised since 2021, Anthropic is sitting on more capital than most countries’ defense budgets. That cash needs somewhere to go—and this week showed exactly where: hardware.
Samsung just started shipping the first commercial HBM4 memory—11.7 Gbps, 46% above the industry standard, with 3.3 TB/s bandwidth per stack. That’s the memory bottleneck cracking open for Nvidia’s Vera Rubin platform. And OpenAI made its own hardware move: GPT-5.3-Codex-Spark now runs on Cerebras chips, hitting 1,000+ tokens per second—15x faster than the flagship Codex. It’s a research preview—but it’s OpenAI’s first model served off Nvidia silicon. The infrastructure layer is fracturing, and every lab is picking sides.
But here’s the twist: while the giants race to spend billions on custom silicon, MiniMax just shipped M2.5—an open model that scores 80.2% on SWE-Bench Verified, within 0.6 points of Claude Opus, at roughly 1/13th the cost. We covered Zhipu’s GLM-5 challenging frontier models on zero Nvidia chips earlier this week. And Spotify’s Honk system proved the same point from a different angle: we broke down why infrastructure matters more than which model you pick. The billion-dollar question isn’t who has the best model anymore. It’s who built the best pipes.
Latest from PulseMark
 |
Anthropic’s $30B Round: What the Investor List Reveals
Sovereign wealth funds, hedge funds, and VCs betting nearly $64B total on one AI safety lab. The round tells a story the headline doesn’t. |
 |
Samsung Ships HBM4: The Memory Fix That Breaks AI’s Bottleneck
11.7 Gbps, 3.3 TB/s bandwidth, and a 50% capacity expansion planned by year-end. The chip that feeds the chips. |
 |
Spotify Honk AI Coding: Why Infrastructure Matters More Than Claude
Spotify built an internal AI coding system that works regardless of which model powers it. The real moat is plumbing. |
 |
GLM-5 Matches Claude Opus on 44B Active Parameters and Zero NVIDIA Chips
Zhipu AI’s new model runs entirely on domestic Chinese silicon. Export controls just got a lot less effective. |
That’s Friday. Six stories, zero fluff.
— The PulseMark Team
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.