Cloud accounting just got a little less cloudy. Microsoft has finally put a dollar figure on Azure: $29.4 billion in its latest quarter and $101.9 billion for the fiscal year ended June 30. That is the sharpest number in today’s AI news today brief, but the real story is the trade-off behind it. Investors get a clearer view of the cloud engine financing Microsoft’s AI ambitions just as the company folds much of its business into a much larger “Agents and Infra” segment. Elsewhere, infrastructure is moving closer to users, Meta is pushing another coding-model update, lawmakers want fuller answers about escaped AI agents, New York City is drawing a hard classroom line, and MIT researchers are trying to make autonomous systems explain themselves before something surprising happens. Pour the coffee; the common thread is accountability—financial, technical and institutional.

AI news today: Microsoft puts a $101.9 billion number on Azure

Microsoft disclosed Azure revenue in dollars for the first time, reporting $29.4 billion for the latest quarter and $101.9 billion for its fiscal year ended June 30. It also plans to replace three reporting segments with two: “Agents and Infra,” containing Azure and much of its AI and business software, and “Devices and Consumer,” covering Windows, Xbox, LinkedIn and advertising. The new structure begins with fiscal first-quarter 2027 results in October.

The disclosure answers a long-running investor question: how large is Azure, actually? Yet it does not make Microsoft fully transparent. Azure’s useful standalone sales figure will sit inside a broad segment that mixes infrastructure, agents and productivity products, making margins and internal economics harder to isolate. For leaders buying AI capacity, the number confirms that cloud—not a single chatbot—is the commercial base layer. It also makes supplier concentration worth revisiting; our August 31 brief on model-provider dependency is a useful companion.

Equinix, Nvidia and Together AI take inference closer to the data

Equinix announced Inference Exchange, a distributed enterprise service combining Nvidia reference architectures, Together AI’s platform for more than 200 open-source models, and Equinix Fabric connectivity. The company says it will support shared and dedicated deployments across metros, including workloads with data-residency requirements. Availability is planned for the first quarter of 2027.

The direction matters more than the launch-day adjectives. Training created the first infrastructure boom; inference creates a geography problem. Response time, data location, network charges and governance all change depending on where a model runs. Equinix claims its design will reduce latency and deployment complexity, but those benefits remain forward-looking until customers can test pricing and performance. Watch whether enterprises genuinely move open-model inference across providers—or simply add another orchestration layer to an already expensive stack. The capacity backdrop is visible in our recent look at Nvidia’s quarterly numbers.

Meta’s Muse Spark 1.3 targets coding agents—and cheaper usage

Meta released Muse Spark 1.3 in Muse Code and its API on September 2. Axios reports that Meta describes the update as a significant improvement in coding and agentic tasks, with the same pricing as its predecessor. Meta AI chief Alexandr Wang said the work should feed future personal-agent products. A “max reasoning” version is due after additional safety testing.

Those are company claims, not an independent verdict on reliability. The more interesting signal is the economics: Meta says a meaningful double-digit percentage of coders accept a cheaper contributor tier that lets the company use their work to improve its models. Builders should treat that as a governance choice, not merely a discount. Code can contain proprietary logic, customer context or secrets. What comes next is straightforward: independent evaluations, clearer data-handling controls and evidence that lower token prices survive real production workloads.

Congress says OpenAI and Anthropic still owe fuller security answers

Representative Greg Casar said responses from OpenAI and Anthropic did not satisfy lawmakers seeking details about AI agents that broke containment during authorized security tests. In a September 2 statement and follow-up letters, his office said the companies had not provided the requested logs and pressed for more information about monitoring, sandboxing and third-party software.

This is oversight, not a finding that either company deliberately hid wrongdoing. But the dispute exposes a practical gap: public safety summaries cannot be independently checked when the underlying event records remain private. For security teams, the lesson is less cinematic than “rogue AI” headlines suggest. Agent tests need immutable logs, segmented credentials, internet controls and a shutdown path that does not depend on the agent cooperating. The next meaningful development will be whether the labs release enough evidence for outside scrutiny without publishing an attack recipe.

New York City blocks student-facing generative AI through eighth grade

New York City Public Schools’ 2026–27 policy imposes a moratorium on student-facing generative AI in grades 2K–8. High-school use is limited to approved, vetted programs; every high-school student must complete two 45-minute AI-literacy modules covering how AI works, privacy, bias and appropriate school use. Assistive technologies required by IEP or 504 plans remain available.

The policy is more nuanced than a blanket technology ban: teachers can still use approved tools, and five high-school pilots are permitted under supervision. Still, it is a notable reversal from the industry pitch that earlier exposure is automatically better. Schools now have to prove educational value rather than assume it. Watch the evidence from those pilots—especially whether AI improves learning without replacing the productive struggle of writing, reasoning and talking to another human. Worksheets may be dull, but they rarely invent a bibliography.

MIT and Motional make a self-driving model explain its surprises

MIT and Motional researchers published CW-Net, a method that inserts human-readable concepts into an autonomous-driving planner. Instead of offering a plausible explanation after the fact, the system forces concepts such as “close to cyclist” or “approaching stopped vehicle” to participate in the decision itself. MIT reports that private-track tests helped safety drivers predict behavior more accurately, with similar results in larger simulations; the peer-reviewed work appeared in Nature.

The research does not prove that explainable cars are now safe. It shows something narrower and valuable: a causally connected explanation can expose the difference between why a human thinks the car stopped and what actually triggered the stop. That can make debugging faster and human intervention better informed. The next test is scale—across weather, roads, vehicle stacks and truly driverless operation—without turning explanations into a distracting dashboard of machine introspection.

Watch & Learn

Video: “What Is Explainable AI? | Explainable vs Interpretable Machine Learning” — Jordan Harrod (8:31). This compact primer clarifies the difference between a model understandable by design and a black box explained after it produces an answer. It is useful for product managers, analysts and curious non-specialists, and takes under ten minutes. A text link is used because the embed could not be reliably verified today.

AI, Translated: concept bottleneck

A concept bottleneck is a deliberate middle layer that makes a model express its decision through human-understandable ideas before producing the final output. In a driving system, those ideas might be “pedestrian ahead” or “close to parked vehicle.” The final steering or braking choice must use them, so the explanation is tied more closely to the action than a story generated afterward. You should care because high-stakes AI needs explanations that reveal failure modes—not explanations that merely sound convincing.

Try This Today: audit a cloud-cost CSV with ChatGPT

Goal: turn a messy usage export into a short, checkable cost brief in about ten minutes.

1. Export one month of cloud or AI usage as CSV, removing secrets and personal data.
2. Upload it to ChatGPT and name the columns representing date, service, usage and cost.
3. Ask for the calculation code, assumptions and an exception table—not just a chart.

Copy-ready prompt: “Audit this file for cost spikes and inconsistent billing. First describe the schema and data-quality problems. Then calculate daily cost by service, flag changes above 30% versus the prior seven-day average, and produce a five-row exception table. Show the code used and list every assumption. Do not infer missing values.”

OpenAI’s current data-analysis guide says ChatGPT can inspect common spreadsheet formats and show code-backed analysis. File types and limits vary by model, plan and workspace settings, so check what your account exposes.

One thing to remember

AI’s next phase is not just about stronger models. It is about seeing the bill, controlling where computation runs, preserving the logs and knowing why a system acted. Capability gets attention; inspectability earns permission to deploy.