This week, AI moved deeper into the infrastructure phase.

OpenAI launched GPT-5.5, its newest model for complex work across coding, research, productivity tasks, tool use, and multi-step execution. The rollout began for ChatGPT Plus, Pro, Business, and Enterprise users, along with Codex integrations.

Google announced its eighth-generation TPUs, including TPU 8t for model training and TPU 8i for inference. The split shows how AI workloads are separating between training frontier models and running fast, cost-effective agents at scale.

Anthropic expanded its Amazon relationship, securing up to 5 gigawatts of compute capacity to train and run Claude models.

The legal spotlight also intensified. Florida Attorney General James Uthmeier launched a criminal investigation into OpenAI and ChatGPT tied to the 2025 Florida State University shooting, after prosecutors reviewed chat logs between ChatGPT and the accused shooter.

The big story: AI companies are no longer just competing on model intelligence. They are competing on compute, agents, enterprise workflows, chips, cloud capacity, distribution, and regulatory trust.

OpenAI introduced GPT-5.5 this week, positioning it as a stronger model for coding, debugging, online research, productivity tasks, planning, tool use, and multi-step work.

The model is being rolled out across ChatGPT Plus, Pro, Business, and Enterprise tiers, along with Codex integrations. GPT-5.5 Pro is available for higher-tier users.

Why it matters:
OpenAI is pushing ChatGPT and Codex further away from simple chatbot use and closer to delegated work. The product direction is clear: AI that can research, reason, use tools, complete tasks, and operate across business systems.

What to watch:
Codex adoption, GPT-5.5 usage inside enterprise workflows, and whether companies begin replacing one-off prompts with repeatable AI work systems.

Florida Attorney General James Uthmeier announced that the Office of Statewide Prosecution has launched a criminal investigation into OpenAI and ChatGPT. The investigation follows an initial review of chat logs between ChatGPT and Phoenix Ikner, the accused gunman in the 2025 Florida State University shooting.

Reuters reported that the shooting killed two people at Florida State University. The investigation is focused on OpenAI and its AI app ChatGPT, and it represents one of the most serious legal escalations yet involving a major AI chatbot company.

This probe is separate from an earlier Florida investigation into OpenAI and ChatGPT. On April 9, Reuters reported that Florida’s attorney general had opened a probe into OpenAI ahead of a potential IPO, focused on broader concerns around ChatGPT.

Why it matters:
This could become a major regulatory test for AI companies. The question is not only whether a chatbot gave harmful information. The bigger question is how much responsibility AI providers have when users ask dangerous questions, and what obligations companies have to detect, restrict, escalate, or report high-risk interactions.

What to watch:
Whether Florida seeks criminal liability, whether other states follow, whether OpenAI changes its safety and escalation policies, and whether this becomes a larger overhang for OpenAI’s future IPO plans.

Google announced two specialized eighth-generation TPUs: TPU 8t and TPU 8i.

TPU 8t is built for training, while TPU 8i is aimed at inference. TechCrunch reported that Google Cloud is splitting its eighth-generation custom AI chips into a training chip and an inference chip, directly positioning the hardware against Nvidia’s dominance in AI infrastructure.

Google described the chips as built for the “agentic era,” where AI systems need both massive model training and fast, efficient inference to support agents that can reason and act repeatedly.

Why it matters:
The cost of inference is becoming one of the biggest constraints in AI. If agents are going to run constantly in the background, companies need faster and cheaper infrastructure to support them.

What to watch:
Google’s ability to use custom silicon to improve margins, compete with Nvidia, and make Gemini-powered enterprise agents more cost-effective.

Anthropic expanded its Amazon relationship, securing up to 5 gigawatts of compute capacity to train and power Claude models. The agreement includes AWS Trainium chips and a major expansion of international inference capacity across Asia and Europe.

The key signal is not just the size of the deal. It is what the deal says about the AI market: model demand is now directly tied to data center capacity, chips, power, cooling, and cloud infrastructure.

Why it matters:
AI is now a compute war. Model quality still matters, but frontier labs need massive, reliable, affordable infrastructure to keep scaling.

What to watch:
AWS Trainium adoption, Claude enterprise growth, and whether Anthropic can keep scaling without becoming too dependent on Amazon.

The market is starting to separate AI winners into two groups: companies that benefit from AI infrastructure spending, and software companies that may be disrupted by AI-native workflows.

Chip, cloud, networking, data center, and custom silicon companies continue to benefit from visible AI spending. Traditional software companies face more investor scrutiny because AI could either increase their value or make parts of their existing software easier to replace.

Why it matters:
Investors are asking a harder question: does AI make a company more valuable, or does it make the company’s existing software less defensible?

What to watch:
Semiconductors, networking, cloud infrastructure, custom silicon, and data center names may continue getting cleaner AI demand signals than traditional SaaS companies.

OpenAI’s GPT-5.5 launch intensified competition with Anthropic, especially around coding, enterprise productivity, and agent-like work. The Verge reported that the model improves coding, debugging, online research, and productivity tasks across tools.

Axios reported that GPT-5.5, codenamed “Spud,” was trained using Nvidia GPUs and is focused on coding, office tasks, computer use, and early-stage scientific research.

Google Cloud Next 2026 made custom silicon a major theme. TechCrunch reported that Google’s TPU 8t is geared toward training, while TPU 8i is aimed at inference.

Reuters reported that Florida launched a criminal probe into OpenAI and ChatGPT over the Florida State University shooting, making AI safety and liability a much larger public issue this week.

Florida’s attorney general said prosecutors reviewed chat logs between ChatGPT and the accused FSU shooter before launching the criminal investigation.

Reuters also reported earlier this month that Florida opened a separate probe into OpenAI ahead of a potential IPO.

Use this stack to turn research into strategy.

Best for: founders, newsletter writers, market researchers, investors, creators, and operators.

How to use it:
Use Gemini for brainstorming, research expansion, scenario planning, and first-draft thinking. Use NotebookLM as the grounded source hub where you upload articles, transcripts, PDFs, docs, slides, and source material.

Workflow:

Prompt to use:

Use this stack to build a long-term research vault.

Best for: creators, consultants, analysts, investors, newsletter writers, students, and knowledge workers.

How to use it:
Use NotebookLM to synthesize source material. Use Obsidian as the permanent knowledge base where your ideas, frameworks, tags, and evergreen notes live.

Workflow:

Prompt to use:

Use this stack to turn research into deliverables.

Best for: teams creating strategy memos, client reports, SOPs, sales enablement docs, investor updates, internal training, and executive briefs.

How to use it:
Use Google Docs, Slides, Sheets, and Drive as the workspace. Use NotebookLM to analyze and synthesize the source material, then turn the output into Docs, Slides, Sheets, or team-ready briefs.

Workflow:

Prompt to use:

This week’s AI stock watchlist is focused on the broader AI infrastructure buildout: chips, custom silicon, cloud capacity, networking, and data centers.

Nvidia remains the center of the AI infrastructure trade because GPUs, CUDA, networking, and full-stack AI systems are still core to training and inference.

Catalyst: Hyperscaler AI spending, model training demand, inference growth, and the next wave of AI data center buildouts.

Current price: about $202.68.

Broadcom is one of the key custom silicon names in the AI market.

Catalyst: Hyperscalers want custom AI accelerators, networking silicon, and ASICs to improve performance and reduce dependence on general-purpose GPUs.

Current price: about $413.92.

AMD is the main challenger GPU name.

Catalyst: Demand for alternative AI accelerators as cloud providers and AI labs look for more supply, pricing leverage, and less vendor concentration.

Current price: about $345.88.

Oracle is becoming a larger AI cloud infrastructure story.

Catalyst: Demand for GPU cloud capacity, large AI infrastructure contracts, and enterprise workloads moving into Oracle Cloud Infrastructure.

Current price: about $172.66.

TSMC is the manufacturing backbone of the AI chip economy.

Catalyst: Demand for advanced-node chips used in GPUs, AI accelerators, networking chips, and high-performance computing.

Current price: about $397.37.

CoreWeave is one of the purest public AI cloud infrastructure plays.

Catalyst: AI labs and hyperscalers continue renting large amounts of cloud capacity instead of waiting to build everything themselves.

Current price: about $114.64.

Arista is the AI networking watchlist name.

Catalyst: AI data centers need high-bandwidth networking, switching, and data movement at massive scale.

Current price: about $175.80.