Saturday, April 18, 2026

Start of archive·May 16

Top story

The NumbersTechCrunch Startups

AI chip startup Cerebras files for IPO

A major AI chip startup is pursuing public markets on the back of landmark enterprise deals with two of the world's largest AI buyers, signaling validation of alternative chip architectures and a potential shift in compute procurement patterns away from NVIDIA-dominated supply chains.

Cerebras filing for IPO

The briefs

Recursive Superintelligence's massive pre-product funding round signals investor conviction that self-improving AI systems are an imminent technical frontier—and that former DeepMind/OpenAI researchers can execute it. This validates a thesis that's been theoretical for years.

Recursive Superintelligence raised $500M+ at $4B valuation

Google's Gemma 4 represents a meaningful shift in on-device AI efficiency—a 2.3B model matching 70B-scale performance on minimal hardware removes a key barrier to local-first AI adoption and challenges the cloud-dependency thesis.

Gemma 4: 2.3B parameters

AI infrastructure is compute-constrained not just by GPUs, but by memory bottlenecks. A multi-year DRAM shortage directly impacts data center buildout timelines and raises the cost of training/inference infrastructure across the industry.

Global DRAM demand will only be 60% met by end of 2027

Meta is making a strategic choice to reallocate human capital toward AI compute infrastructure. This signals how seriously the company is betting on AI dominance and willing to sacrifice near-term headcount to fund long-term compute capacity—a model other tech leaders may follow.

8,000 jobs to be cut on May 20 (first wave)

DeepSeek's first external funding round marks a strategic inflection point for one of China's most closely watched AI labs, suggesting competitive pressures from both domestic rivals and US tech giants are forcing a recalibration of its previously independent model.

DeepSeek seeking outside funding for the first time

As AI agents move into production, context management becomes a bottleneck. Cloudflare's new Agent Memory feature solves the cost/recall tradeoff by storing conversation history separately, letting builders scale multi-turn interactions without exploding inference costs.

Cloudflare launches Agent Memory feature

Anthropic's market valuation is inflating dramatically on investor appetite alone, signaling extreme capital concentration in frontier AI labs and setting a new ceiling for AI startup valuations even without a financing event.

Anthropic received expressions of interest from investors at $800B valuation

Google's Gemma 4 demonstrates a major capability leap in efficient, on-device LLMs—enabling privacy-preserving AI at smartphone scale without internet, which could reshape how enterprises and consumers deploy AI infrastructure.

Gemma 4: 2.3B parameters

Bipartisan legislation tightening export controls on semiconductor manufacturing equipment directly constrains AI infrastructure buildout by limiting access to advanced chipmaking capabilities—a critical chokepoint for both domestic and international AI companies competing on compute.

Bipartisan bill introduced to tighten controls on sensitive chipmaking equipment