Sarmadi AI Digest May 30, 2026 Updated 6:55 AM CT Today Archive Topics Saved Subscribe RSS

Memory eats the chip stack; coding agents grow up

Two parallel signals from the chip layer: XCENA raised $135M at a $570M valuation explicitly on the bet that AI's binding constraint is memory not compute, and Groq is reportedly raising $650M after Nvidia's $20B not-acqui-hire — the silicon money is moving exactly where last week's data on memory-as-two-thirds-of-AI-chip-cost said it would. The open-weight frontier kept pressure on: Liquid AI shipped an 8B-A1B MoE trained on 38T tokens, and notes from the Mistral AI Now Summit hit the HN front page. The coding-agent conversation matured — Cognition's Scott Wu publicly argued agents shouldn't replace humans even as practitioners refuse to work without them, and Aaron Levie warned of an emerging 'AI psychosis' in CEOs deciding to replace roles they don't understand. A separate weirdness: AI startups now offering free home cleaning in exchange for filming you for robot-training data. And the Vatican's quiet liaison inside Anthropic is the most-read coda to yesterday's papal encyclical.

10 papers 18 news 7 sources ← Latest

News

12 items

Memory is the bottleneck, and the money is moving

XCENA's $135M raise explicitly on a memory-not-compute thesis, Groq's $650M raise after Nvidia's $20B not-acqui-hire, and CONF-KV's confidence-aware KV-cache eviction all point at the same shift: serving costs are now memory-bound, and capital + research are reallocating accordingly.

News TechCrunch AI

Chip startup raises $135M on a bet that AI's biggest bottleneck isn't compute — it's memory

South Korean chip startup XCENA raised $135M at a $570M valuation explicitly on the thesis that AI's real bottleneck is memory, not compute.

raise $135Mvaluation $570M
Why it matters
  • Translates last week's 'memory is 2/3 of AI chip cost' data point into a funded silicon strategy.
  • Memory-first chip designs become a credible procurement option for inference-heavy buyers.
  • Validates the wave of KV-cache / quantization research as commercially relevant, not academic.
fundingcomputeinfrastructure
Source →
News TechCrunch AI

After Nvidia's $20B not-acqui-hire, AI chip startup Groq reportedly raising $650M

Groq is reportedly raising $650M as it pivots from hardware-first to a software-heavy posture, in the wake of Nvidia's $20B not-acqui-hire of competing talent.

raise $650M
Why it matters
  • Signals the AI-inference-silicon market is still raising at scale despite Cerebras already public.
  • Software-led pivot acknowledges that pure silicon wins are no longer enough against CUDA.
fundingcomputeinfrastructure
Source →

Open-weight pressure on the frontier

Liquid AI released a 38T-token 8B-A1B MoE; Mistral's Now Summit notes pulled 389 HN points; a 152-point Tiny-vLLM dropped on Show HN. Each is small individually; together they keep cost pressure on the proprietary tier.

News Hacker News

Liquid AI reveals 8B-A1B MoE trained on 38T tokens

Liquid AI's LFM2-5 8B-A1B MoE — trained on 38T tokens — adds another open-weight frontier-adjacent model to the lineup (193 HN points).

active params 8B (A1B MoE)training tokens 38T
Why it matters
  • Sustained open-weight releases keep capability-per-dollar improving for SMB builders.
  • MoE architecture at 8B active parameters is a credible production target on commodity GPUs.
open-weightsmixture-of-expertstraining
Source →

Coding agents grow up — and CEOs catch AI psychosis

The agent-coding conversation flipped this week. Cognition's Scott Wu publicly argued agents shouldn't replace humans; a TC story documents coders refusing to work without AI; Aaron Levie's 'most CEOs have AI psychosis' line became the line of the week. A widely-shared essay argues AI is reprising frontend's 'lost decade' of churn. The optimism gradient is steepening, not flattening.

News TechCrunch AI

Cognition's Scott Wu says AI coding agents shouldn't replace humans

Cognition's CEO (Devin) publicly walks back the replace-the-engineer pitch — even from the company most associated with it.

Why it matters
  • Pre-IPO season narrative discipline: even the most aggressive coding-agent vendor is softening.
  • Repositions agentic coding as augmentation, not displacement — better for enterprise adoption.
  • Pairs with Altman/Amodei walking back the jobs apocalypse: industry-wide message reset.
productsmarketcodeagents
Source →

Free chores for robot-training data

Three outlets covered the same startup pattern: free home cleaning in exchange for filming you for embodied-agent training. It's the most concrete face yet of the data-collection arms race for robotics — and the labor-economics question it implies.

Papers

5 items

Memory is the bottleneck, and the money is moving

XCENA's $135M raise explicitly on a memory-not-compute thesis, Groq's $650M raise after Nvidia's $20B not-acqui-hire, and CONF-KV's confidence-aware KV-cache eviction all point at the same shift: serving costs are now memory-bound, and capital + research are reallocating accordingly.

Agent alignment + retrieval research

A consistency-training method to reduce political manipulation, online skill distillation that makes web agents cheaper as they accumulate experience, a checkpoint-repair PoT that survives single bad actions, and a mechanistic look at why dense retrievers score what they do.

Also today