The memory crunch is here. A new class-action antitrust suit accuses Samsung, SK Hynix, and Micron of driving DRAM prices up roughly 700% over four years — filed June 25 by 14 consumers and three small businesses. Analysts see RAM climbing up to 50% more, with no relief until 2028. The capital response is national: South Korea just lined up $380B in private investment across chips and data centers, with Samsung earmarking 350+ trillion won for AI. The shortage is becoming policy.
The shortage is acute enough to bend supply chains. Apple’s CXMT bid has it seeking US approval to buy from China’s largest DRAM maker — still on the Pentagon’s 1260H blacklist — even as memory prices quadrupled over three quarters and pushed the MacBook Pro 16-inch from $2,499 to $2,999. Apple shares fell more than 6%, erasing about $265 billion. TSMC went the other way: its Winbond wafer-stacking deal builds 3D DRAM to route around Korean HBM suppliers. Everyone’s racing to own the memory they can’t buy.
The squeeze isn’t only physical silicon. Anthropic says Alibaba’s distillation campaign ran about 25,000 fake accounts to query Claude 28.8 million times between April 22 and June 5, targeting its Mythos Preview model — the largest such attack Anthropic has disclosed, detailed in a June 10 letter to two senators. Earlier February attacks tied to DeepSeek topped 150,000 exchanges. Memory, compute, and model IP are now the same fight: every layer of the AI stack is a contested resource someone is trying to corner.
The Pulse
Ford has been rehiring veteran quality engineers after AI fell short
Turns out you still need the people who know what “good” actually looks like.
General Intuition raises $320M to train AI agents on video-game footage
If Fortnite can teach a robot to walk, the sim-to-real thesis just got real.
Patronus AI lands $50M to build digital worlds that stress-test AI agents
Before your agents break in production, someone has to break them on purpose.
Apple’s Vision Pro chief is reportedly leaving for OpenAI’s hardware team
OpenAI keeps raiding Apple’s bench — the device war is really a talent war.
Latest from PulseMark
 |
GLM-5.2 API Tutorial: Build With Z.ai’s 1M-Token Model
While frontier memory and model access get harder to buy, an open-weight alternative ships a 1M-token window and 128K max output at $1.40 per million input tokens. We walk through auth, tool use, streaming, structured output, and MCP support with copy-paste code. |
That’s the signal.
— The PulseMark Team
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.