AI Sector Daily Digest — July 7, 2026

The five

1. Synopsys redirects engineering away from fab-control software

• Synopsys plans to stop offering new versions of its EES and FDC manufacturing-control tools while honoring support obligations, as it shifts resources toward higher-margin AI design products. 1
• The affected tools help semiconductor fabs monitor equipment and detect anomalies; Reuters says more than 10 chipmakers, including Samsung, SK Hynix, Kioxia, and Qorvo, were notified in April and May. 1
• Samsung says it has alternatives and expects no production hit, but some sources warned that weaker maintenance could hurt yields at certain fabs. 1

2. Samsung's AI-memory boom runs into investor nerves

• Samsung estimated second-quarter operating profit of 89.4 trillion won, about $58.44 billion, a 19-fold jump from a year earlier; revenue is expected to rise 129% to 171 trillion won. 2
• The stock still fell as much as 10.1%, while SK Hynix dropped as much as 10.6%, as investors questioned whether AI data-center spending can keep memory prices rising. 2
• The read-through is mixed: AI demand is still producing extraordinary chip earnings, but markets are now testing how much of that growth depends on nonstop hyperscaler capex. 2

3. Illinois signs a frontier-model safety law with annual audits

• Gov. JB Pritzker signed Senate Bill 315, the Artificial Intelligence Safety Measures Act, adding transparency and accountability requirements for the largest AI model developers. 3
• The law applies to large models tied to more than $500 million in annual revenue and massive compute use; developers must publish safety frameworks and report serious incidents within 72 hours, or 24 hours for imminent death or serious-injury risks. 3
• Illinois also added mandatory annual third-party audits and civil penalties of up to $1 million for a first offense and $3 million for later violations; the law takes effect Jan. 1, 2028. 3

4. Bespoke Labs raises $40 million for agent-training environments

• Bespoke Labs raised $40 million across a Series A led by Wing VC and a seed round led by 8VC, with participation from Mayfield, The House Fund, and angel investors from Anthropic, OpenAI, Meta, and other AI companies. 4
• The company builds realistic environments for agents to practice long-horizon work, including codebases, microservices, logs, tickets, email, and Slack-like workflows. 4
• The funding is a sign that agent infrastructure is moving beyond model wrappers toward training, evaluation, and reinforcement-learning environments that can make autonomous workflows more reliable. 4

5. EvoAgentBench tests whether agents really improve from experience

• A new arXiv preprint introduced EvoAgentBench, a benchmark for agent self-evolution across web research, algorithmic reasoning, software engineering, and knowledge-work tasks. 5
• The benchmark uses a 528-task training split and 267-task test split to measure whether reusable abilities from prior traces transfer to related but unseen tasks. 5
• The result is a useful caution for agent builders: curated ability content transfers across model families, but no automatic self-evolution method in the study sustained positive gains in every setting. 5