The Executive Briefing — Friday, October 2, 2026

Micron Books a $54.23 Billion Quarter and Shows Where the AI Budget Is Really Going

For two years the AI infrastructure story has been told through GPUs. On September 30, Micron Technology reminded the market that every accelerator needs memory beside it, and that the company supplying it is now running at a scale that would have been hard to imagine a year ago. In its filing with the Securities and Exchange Commission, Micron reported fiscal fourth quarter revenue of $54.23 billion, against $41.46 billion in the prior quarter and $11.32 billion in the same quarter a year earlier. The quarter closed the company's fiscal year, which ended September 3.

Profitability moved even faster than sales. Micron reported GAAP net income of $37.70 billion for the quarter, or $32.87 per diluted share, and a GAAP gross margin of 86.8 percent of revenue, up from 44.7 percent a year earlier. For the full fiscal year, Micron reported revenue of $133.19 billion, compared with $37.38 billion in fiscal 2025. Those are not the economics of a cyclical commodity supplier at the top of a cycle. They are the economics of a constrained input that every AI buyer needs and few vendors can deliver.

The segment detail is where executives should look closest. Micron's Core Data Center Business Unit generated $18.002 billion in the fourth quarter, compared with $1.577 billion in the same period a year earlier, at a 90 percent gross margin. The company also said revenue from its server LPDDR SOCAMM portfolio more than doubled sequentially, and that its PCIe Gen 5 and Gen 6 SSDs are shipping for KV cache applications, the working memory that large models use to hold context during inference. That last detail matters: it ties memory demand directly to inference workloads, not just to model training.

Micron is not hoarding the windfall. Net capital expenditures were $10.77 billion in the fourth quarter and $27.37 billion for the fiscal year, and the company ended the year with $73.48 billion in cash, marketable investments and restricted cash. Its guidance points higher still, with fiscal first quarter 2027 revenue projected at $61.5 billion, plus or minus $1.5 billion.

Chairman and CEO Sanjay Mehrotra framed the results in language borrowed from the current political moment. "AI is becoming Super Intelligence (SI), and memory enhances this intelligence and the competitiveness of our customers' platforms," he said in the release. He also pointed to the company's Strategic Customer Agreements as added confidence in the durability of its financial performance, agreements the company described in June as multiyear arrangements, a signal that customers are committing to supply over longer horizons rather than buying quarter to quarter.

For enterprise technology leaders, the practical implication is that the cost structure of AI is shifting toward components most procurement teams have never had to negotiate. Memory has historically been a line item buried inside a server quote. At these margins, it is a strategic input whose price and availability will shape cloud pricing, private infrastructure timelines, and the feasibility of long context inference at scale.

There is also a concentration lesson. When a single supplier's data center segment grows more than elevenfold in a year, buyers downstream inherit that supplier's capacity decisions. Enterprises building AI roadmaps on the assumption that compute gets steadily cheaper should test that assumption against the memory market, where demand is plainly outrunning the industry's historical pricing discipline.

AG-UI 1.0 Freezes the Missing Layer Between AI Agents and the People Who Use Them

The agent stack has spent the past year acquiring standards. MCP connects agents to tools. A2A lets agents coordinate with each other. On September 30, CopilotKit declared the third layer finished, releasing AG-UI 1.0, an open protocol that standardizes how AI agents connect to user facing applications. The company describes the three together as a complete map: MCP for tools, A2A for agent to agent coordination, and AG-UI for the application a person actually sees.

The adoption claim is what makes the release consequential. CopilotKit says AG-UI is adopted by Google, Microsoft, Amazon and Oracle, and supported by agent frameworks including LangChain, Mastra and Anthropic's Claude Managed Agents. The announcement also lists compatibility with Google ADK, Strands Agents, the Microsoft Agent Framework and Pydantic AI.

The core change in 1.0 is stability. The release introduces a formal specification backed by a JSON Schema, and the TypeScript, Python and .NET SDKs are now generated directly from that schema. According to CopilotKit, the schema is strict: a field it does not define fails validation, except in places reserved for custom data such as metadata. CopilotKit says the draft was discussed openly on GitHub and that the Anthropic, Pydantic AI and TanStack teams gave feedback that shaped the final text. CopilotKit's promise to builders is direct: "The 1.0 spec won't change, so what you build on it today keeps working."

The feature list reads like a catalog of problems enterprise teams have been solving by hand. According to CopilotKit, AG-UI 1.0 adds subagent support, so a user interface can tell which of several parallel agents produced which output. It adds metadata, multimodal tool results, and human in the loop interrupts, which let a run pause for a user's approval and resume where it left off. Runs can now also end as cancelled, rather than looking identical to a run that finished or failed.

The most underrated addition is cost visibility. CopilotKit says that under 1.0, runs can report token usage when they finish, covering input, output, cached and reasoning tokens, and tokens consumed by subagents count toward the run that started them. For any organization trying to charge back agent usage to business units, or simply to understand why an agent workflow costs what it does, that is the difference between estimating and measuring.

Compatibility was handled with unusual care. CopilotKit says 1.0 is backwards compatible in both directions: a 0.x agent works with a 1.0 client, and a 1.0 agent works with a 0.x client, so neither side of an integration has to upgrade first. That lowers the risk of adopting the protocol inside a large estate where front end and back end teams move on different schedules.

The strategic read is that the interface between agents and humans is becoming a commodity layer, just as tool access did when MCP took hold. Enterprises that have built bespoke streaming code for each agent framework now have a path to retire it. The harder question shifts upward, to governance: which approvals, interrupts and audit records an organization requires before an agent acts, and whether its platforms expose those controls through the standard rather than around it.

Washington Signs a Voluntary AI Accord and the FTC Opens a Probe of Anthropic and OpenAI a Day Later

In the span of roughly twenty four hours, the federal government offered two incompatible answers to the question of who polices frontier AI. On Tuesday, September 29, President Donald Trump hosted AI company leaders at the White House, where Anthropic, Google, Meta, Nvidia, OpenAI and xAI signed the Joint Commitment on Frontier Responsibilities. On Wednesday, a senior Federal Trade Commission official confirmed to ABC News that the agency had opened a broad probe into the safety of AI systems, including Anthropic and OpenAI.

The accord itself is short. According to IAPP, it is a four point AI safety plan focused on independent audits and robust internal controls. The commitments include controls to monitor the capabilities and alignment of models, and an independent board level committee to review internal progress reports on company controls. No enforcement mechanisms were announced. Trump characterized the agreement as "morally binding" and praised what he described as tremendous self policing by developers.

House Speaker Mike Johnson described the document plainly. "This is a statement of principles that you'll see, a statement of standards, commitments that are voluntary on behalf of the industry," he said, according to ABC News.

The FTC did not wait to be named. According to ABC News, the probe is looking into allegations of unfair or deceptive acts by AI companies and potential harms to consumers, and the agency could issue formal demands for companies to turn over information. The New York Post was first to report the investigation. The Wall Street Journal, as cited by KQED, reported that the commission is examining whether Anthropic and OpenAI misled consumers about potential harms from their technology. FTC investigations can result in civil penalties.

Litigation is also moving independently of either Washington track: IAPP, citing Wired, reported that OpenAI faces a lawsuit filed by Legal Advocates for Safe Science and Technology and the law firm Gerstein Harrow in San Francisco Superior Court over the Hugging Face incident.

The practical effect for enterprises is that the absence of a federal AI statute does not mean the absence of federal exposure. The accord sets expectations that are voluntary for the six signatories. The FTC probe applies existing consumer protection law, which reaches any company making claims about what its AI does and how safely it does it. Those are very different instruments, and only one of them carries penalties.

Consider the sequencing from a board's point of view. The signatories committed on Tuesday to stronger internal controls and outside audits. By Wednesday, two of them were confirmed as subjects of a federal consumer protection inquiry. Whatever the accord eventually produces, the investigation means that safety commitments, product claims and incident disclosures are now likely to be read side by side by a regulator with the power to demand documents. Companies that treat those three as separate workstreams owned by separate teams are carrying a risk they may not have priced.

That distinction should reshape how buyers read vendor assurances. A safety commitment signed at the White House is a statement of intent. A representation made in a contract, a marketing page or a product disclosure is something a regulator can test against reality. As agents gain the ability to act on systems rather than merely answer questions, the gap between those two kinds of statements is where enterprise risk now lives.

California Enacts the Nation's First No Robo Bosses Law and Puts a Human Between the Algorithm and the Pink Slip

While Washington debated voluntary pledges, California wrote statute. On September 30, Governor Gavin Newsom signed SB 947, the No Robo Bosses Act of 2026, authored by state Senator Jerry McNerney. According to McNerney's office, it is the first law of its kind in the nation. It takes effect July 1, 2027, according to KQED.

The core rule is simple. SB 947 bars California employers from relying solely on automated decision making systems to fire or discipline workers, and requires human oversight and verification when such systems assist in those decisions. Under the bill text, an employer that primarily relied on an automated system to make a disciplinary or termination decision must give the affected employee a written notice saying so at the time it informs the employee of the decision. Enforcement sits with the California labor commissioner, the state attorney general, or local prosecutors.

"No worker should ever be fired or disciplined by a robo boss," McNerney said in announcing the signing, adding that "AI must remain a tool controlled by humans, not the other way around." His office cited estimates that more than 550 so called bossware products are available to employers. The law is a narrower version of legislation Newsom vetoed in 2025, KQED reported, and it was sponsored by the California Federation of Labor Unions, AFL-CIO. Its president, Lorena Gonzalez, said "Today, California's workers and our unions have changed the national narrative on how Americans can fight back and win against AI taking over our jobs and workplaces."

SB 947 was one of a slate. According to KQED, Newsom also signed SB 951. Its bill text revises the state's existing WARN Act so that an employer giving notice of a mass layoff, relocation or termination caused in whole or in substantial part by AI or other automated technology must include additional information, including the job functions being automated, and it requires the Employment Development Department to report to the Legislature on AI's effects on business hiring practices by January 1, 2028. KQED reported that AB 1883 restricts AI powered workplace surveillance, including tools that infer workers' emotional states or collect neural data. AB 1331 bans workplace surveillance tools in bathrooms. AB 1979 bars AI from replacing the clinical judgment of licensed health care professionals and extends medical confidentiality to health chatbots.

The package has limits that employers should note. Newsom vetoed AB 2656, which would have required notice to public employees when AI performs work within the scope of their jobs. Rideshare drivers, classified as independent contractors under Proposition 22, are outside SB 947's protections. KQED reported that it reached out to Uber, Lyft and the California Chamber of Commerce and did not hear back by publication time.

The timing carried its own message. September 30 was Newsom's last day to sign or veto bills before he terms out in January, and he also signed an executive order requiring state agencies to keep calling the technology artificial intelligence rather than Super Intelligence. In the same week federal policy leaned on voluntary commitments, California chose enforceable rules for one of the most consequential uses of AI inside a company.

For workforce leaders, the significance reaches beyond California. Employers operating in many states often prefer a single HR process to a patchwork, which tends to make the most demanding rule the operating standard. A human corroboration step and a written worker notice are now legal requirements for California terminations that rely primarily on automated systems, and they are likely to become best practice everywhere else well before the July 2027 effective date.