News · · 10 min
Period: September 11 — morning of September 12.
Today, three lines are most important: AI infrastructure is becoming critical physical infrastructure; enterprise agents are moving from individual tasks to working for weeks; autonomy is drastically reducing the cost of AI abuse.
What happened. The UAE is revising plans for an AI campus with up to 5 GW of power following March attacks on the region's technological infrastructure. Instead of a single 10-square-mile mega-campus in Abu Dhabi, a distributed network of facilities across the country is being considered, including blast-resistant structures, underground components, and additional physical protection systems.
The first phase of Stargate UAE, costing about $30 billion, was supposed to bring approximately 200 MW online as early as 2026. The project involves G42 and US technology companies, including OpenAI, Oracle, and SoftBank.
Why it matters. An AI datacenter is becoming an asset of the same class as a power plant or a telecommunications hub.
Until now, AI cluster architecture has been primarily optimized for:
power → cooling → networking → GPU utilization.
Now, the following is being added:
physical resilience → geographic redundancy → blast risk → continuity of operations.
Distributed architecture is typically worse for tightly coupled training clusters due to networking latency and bandwidth, but better for inference resilience and data storage.
What the data shows. 5 GW is the planned scale of the entire project, not already built capacity. The specific new topology has not yet been approved; Reuters describes options being discussed by authorities and project participants.
What you need to understand now: AI cluster, scale-up/scale-out networking, geographic redundancy, disaster recovery, sovereign compute.
What this could change. For most companies — observe for now. But for critical AI products, a new vendor risk question arises: in which physical region is inference actually performed, and what happens if an entire data center or region is lost?
Date: September 11. Source: Reuters.
Results were released after market close on September 10, so their economic impact is already part of yesterday's news cycle.
Oracle reported that quarterly Cloud Infrastructure revenue grew 121% year-over-year to $7.4 billion. During the quarter, the company added another 850 MW of data center capacity. The total backlog/RPO reached $664 billion.
Total cloud revenue grew 62% to $11.6 billion.
Why it matters. AI is changing the economics of a classic enterprise software vendor.
Oracle has historically been primarily a database/software company. Now, one of the fastest-growing parts of the business is literally:
land → electricity → GPU → networking → datacenter → AI cloud.
This is capital of a significantly heavier type than typical SaaS.
Earlier Oracle disclosures show the other side of the economics: in FY2026, free cash flow was −$23.7 billion, as the company financed AI infrastructure. At the same time, some large customers are either prepaying for GPUs or supplying hardware to Oracle themselves — such prepaid/customer-supplied hardware commitments have accumulated to about $75 billion.
What you need to understand now: IaaS, CAPEX, free cash flow, RPO, GPU capacity.
What this could change. For business, this explains why AI API price cannot be viewed as typical SaaS margin. Behind every token, a very capital-intensive supply chain is gradually emerging.
Date: results published September 10 after market close; analysis — September 11. Primary source: Oracle Q1 FY27 results
This is the main architectural release of the day.
What happened. Salesforce introduced a new generation of Agentforce, where an agent can receive a goal not for a single conversation, but for days and weeks.
The first such agent is Hunter for outbound sales. Example task:
save at-risk deals by the end of the quarter.
The agent builds a plan, researches accounts, conducts outreach, tracks changes, adjusts next steps, and involves a human when approval is required.
Underlying this, Salesforce highlights three separate runtime functions:
In parallel, Multi-Agent Orchestration became GA: specialized agents can hand off work to each other.
Why it matters. This is a qualitatively different software abstraction.
Not:
request → agent → answer.
But:
business goal → persistent plan → events → actions → memory → coordination → approvals → result after several weeks.
This is already closer to a workflow engine or a virtual employee than to a chatbot.
The supply-chain agent Marshall uses deterministic execution and leaves an audit record of every action. Salesforce is also promoting Agent Script — an open-source language that allows combining AI reasoning with ordinary rigid rules.
This aligns with the conclusion of several recent issues:
The LLM selects and interprets; deterministic software constrains and executes where guarantees are needed.
This is especially logical for supply chain, payments, CRM updates, and other actions with external consequences.
What the data shows. Salesforce reports 7 billion Agentic Work Units already, including 3.2 billion in the second quarter alone. This is Salesforce's own usage telemetry, not an independent effectiveness assessment.
The company's customer case studies:
These numbers cannot be compared with each other: the denominator and task complexity are different.
What you need to understand now: long-horizon agent, durable execution, persistent memory, multi-agent orchestration, deterministic execution.
What this could change. If a company is planning enterprise agents, it should stop modeling them only as conversational UX. It will be necessary to design state machines, recovery, goal expiration, permissions, and the human as a participant in a long-running workflow.
Date: September 11. Primary source: Salesforce; the official newsroom confirms the September 11 release.
Instead of a generic agent builder, Salesforce released ready-made roles:
Why it matters. Enterprise AI is starting to move from selling capability:
"here is a model that can reason"
to selling a job-to-be-done:
"here is an agent that manages the outbound pipeline."
If the model becomes a commodity, economic value shifts to:
enterprise context + permissions + integrations + workflow + auditability + domain-specific evaluation.
That is, the moat of an enterprise-AI product is increasingly less necessarily in the model itself.
What you need to understand now: job-specific agent, workflow, system of record, enterprise context, vendor lock-in.
What this could change. SaaS companies should analyze the threat not as "Salesforce added AI," but as Salesforce attempting to automate an entire category of work that was previously performed manually within applications.
A fresh threat-intelligence report from Anthropic contains a much more important signal than individual jailbreaks.
The company observed operations where one person with Claude simultaneously managed dozens of targets, and some multi-agent frameworks autonomously performed reconnaissance, exploitation, and data theft for hours or days. There were also scheduled agents that autonomously updated stolen access tokens and downloaded data from cloud storage.
Anthropic formulates the economic effect quite precisely: autonomy primarily reduces the cost of an attack, not necessarily increasing the maximum severity of a single attack.
That is:
the same potential payoff
÷
less human labor
= it becomes profitable to attack targets that were previously not worthwhile.
A separate European case received additional confirmation on September 11.
According to Anthropic, one French-speaking operator used Claude to create infrastructure and attack approximately 40+ European political, media, and related SaaS organizations. Compromise occurred in at least 14 tracked targets; in one case, about 140 thousand records were extracted via a vulnerable campaign-management platform, including sensitive information about political views.
This is more important than the question "can Claude write malware."
New capability:
one operator → AI engineering → automation harness → dozens of parallel targets.
What you need to understand now: agentic cyber operations, credential harvesting, attack automation, scheduled agent, defense in depth.
What this could change. A company's security assumptions must account for not only a more powerful attacker, but also a much cheaper attacker. Previously economically uninteresting SaaS, APIs, and employee accounts become more attractive if reconnaissance and exploitation can be automated.
Date of new publication/analysis: September 11. Primary source: Anthropic — September Threat Intelligence Report
One of the most noticeable GitHub signals on September 11 is i-have-adhd: a small skill whose goal is to prevent the coding agent from overwhelming the user with long, noisy output.
Third-party historical GitHub Trending snapshots recorded approximately +3.8 thousand stars in a day. obra/superpowers, a framework from a set of reusable engineering skills, simultaneously gained about +700 more. These values should be considered approximate, as they are third-party snapshots, not official GitHub historical statistics.
The project itself is less important than the category.
First, developers optimized:
the model.
Then:
the agent framework.
Now, more and more activity is around:
skills → behavioral constraints → engineering methodology → communication discipline.
That is, the next optimization of coding agents often requires not greater intelligence, but better model behavior within a human team.
What you need to understand now: Agent Skills, harness, context engineering, developer workflow.
What this could change. Internal benchmarks for coding agents should be evaluated not only by task pass rate, but also by the cost of human supervision: how much time a developer spends reading, correcting, clarifying, and reviewing.
The agent stack is beginning to transition from the mode of "automating a single action" to the mode of "delegating long-term responsibility."
Salesforce is building agents that work for days and weeks; security operations observed by Anthropic show that the same persistence and concurrency radically change the economics of abuse. And from below, Oracle and Stargate show the physical price of all this autonomy: more long-lived agents means more continuous inference — which means more GPUs, energy, networking, and infrastructure.
If an agent works for several days, where is its authoritative state stored, and how do we understand after a restart that it is still pursuing a relevant goal?
Which actions of a long-running agent require deterministic rules or human approval, even if the model's reasoning seems convincing?
Do we calculate the economics of an agent based on the price of a single model call or the full cost of a long-running workflow — including retries, memory, tools, and supervision?
Does cheap automation of attacks change our security threat model for less critical APIs, internal tools, and employee accounts?
When choosing agent tooling, do we measure human-attention cost alongside task completion rate?