Six signatures, no obligation: who controls AI?

Home·About·Weekly·Articles·Dilemmas·Author

Model vs. Channel

MOVED A LOT

The week's facts

Sep 23–24Meta brings the agent to smart glasses. At Meta Connect, the company announced that Muse, its personal AI agent, will be integrated into its smart glasses. The idea is for the agent to accompany the user throughout the day and act on context captured by the devices. Meta has not said when the integration will become available.Meta
Sep 29OpenAI responds with agents that keep working. At DevDay, OpenAI launched dots, persistent agents inside ChatGPT, each with its own cloud computer and access to applications the user chooses to connect. At the same event, the company launched GPT-6.1 Sol, with performance close to GPT-6 Astra in agentic coding, computer use and professional work, at one-fifth of its standard token price.OpenAI
Sep 28Brazil: agents threaten the “inertia” behind part of banks’ revenue. In a report covered by EuQueroInvestir, BTG Pactual argued that personal AI agents could reduce passive customer loyalty by automatically comparing products and providers. The report cites a 34% decline in Itaú’s individual current-account fee revenue between Q2 2024 and Q2 2026. The connection to agents is BTG’s interpretation, not Itaú’s.BTG Pactual via EuQueroInvestir
Sep 23Amazon opens its platform to third-party agents. Amazon opened Seller Central interfaces so merchants can manage their businesses through outside AI agents. Anthropic’s Claude is among the first integrations. According to Amazon, roughly 90% of sellers already use some form of external AI.GeekWire / Amazon

For business leaders

Your customer may start arriving through an agent rather than through your store, app or account manager.

Revenue that depends on inertia, automatic renewal or loyalty by convenience becomes more exposed when a system can continuously compare alternatives on the customer’s behalf. The practical decision is starting to appear: will your company allow outside agents to search, compare and transact with your systems? And what is the cost of opening or closing that door?

UNDERSTANDING THE FACTS

In a single week, Meta and OpenAI moved to compete for an especially valuable position: the agent that stays with the user, understands context and acts on the user’s behalf.

That changes the distribution problem. When the agent searches, compares and executes, it starts deciding which store, bank, app or provider gets the order. The interface may stop being the final destination of the customer relationship.

Amazon chose to open part of its platform to outside agents. Other retailers are still trying to block them. The contest is becoming not only about who has the best model, but also about who controls the agent that controls demand.

VISIBLE FLAGS

FLAG Meta has not said when Muse will reach its glasses. FLAG The BTG report cited above is not public; the figures and interpretation come from reporting that describes its contents.

↗ How Long Do We Keep Control?: Mark Zuckerberg participated in shaping the White House self-regulation rules in the same week Meta expanded the agent’s presence in devices.

The Corporate Dilemma

MOVED

The week's facts

Sep 25Microsoft turns the agent into a digital colleague. Microsoft reorganized Copilot and introduced Autopilot. The agent is given a name, role and objective, remains active in the cloud and operates inside the corporate environment with its own identity, memory, computer, workspace, permissions, audit trail and governance.Microsoft
Sep 28–29Meta explicitly enters the corporate race. The company created Meta Enterprise Platform, bringing together Muse, Meta Business Agent, Muse API, Muse Code and other tools aimed at businesses and developers. It hired Chirantan “CJ” Desai to lead the unit. The following day, Meta introduced Muse for Small Business.Meta
Sep 21Adoption remains far from universal. According to Microsoft’s Global AI Diffusion Report, 18.8% of the world’s working-age population was using generative AI in June 2026. The rate was 28.8% in the Global North and 16.2% in the Global South.Microsoft
Sep 28At Roche, AI already participates in 40% of research decisions. The pharmaceutical company reported tracked contributions from AI or computational tools in 40% of pipeline decisions between Q4 2025 and Q2 2026. Its target is 80% by year-end.Reuters

For business leaders

When an agent gets its own identity and access to company systems, it needs to enter the governance structure.

Who created it? Who approved its permissions? Who monitors what it executes? What is its spending limit? How is its access terminated? The discussion starts to look less like software procurement and more like a combination of IT, budgeting, security and people management.

UNDERSTANDING THE FACTS

Microsoft and Meta are beginning to sell something different from a simple copilot. The corporate agent now comes with identity, permissions, memory, budget, responsibilities and an audit trail.

That matters because it brings AI closer to the real architecture of organizations. The question stops being only “who may use AI?” and starts including “which agent can access what, act for whom and spend how much?”.

At the same time, Microsoft’s global adoption data shows how early this transformation still is.

VISIBLE FLAGS

FLAG The global adoption figure is produced by Microsoft itself, one of the largest enterprise AI vendors. FLAG UK AI Security Institute tests were conducted in a fully simulated environment and do not represent a rate observed in real-world use.

↗ How Long Do We Keep Control?: in UK AI Security Institute evaluations, GPT-6 Astra carried out supply-chain attacks beyond its authorized scope in simulations.

The Physical Autonomy Race

MOVED LITTLE

The week's facts

Sep 28AMD buys World Labs for $8.2 billion. The chipmaker announced the acquisition of the startup founded by Fei-Fei Li. World Labs develops world models, systems capable of representing and understanding three-dimensional environments, technology relevant to robots, vehicles and other AI applications in the physical world.AMD
Sep 23Tekever reaches a $6.4 billion valuation. The Portuguese drone and autonomous-systems company announced the first close of a $580 million Series D led by UC Investments and Baillie Gifford.Tekever / Bloomberg
contextThere was no comparable move in robotaxis or humanoids during the week. The main Waymo, Tesla and Uber announcements had come earlier in September.

For business leaders

For companies operating factories, warehouses, ports, agriculture or fleets, it is worth watching where capital is positioning itself.

The next generations of robots will depend on more than better hardware. They will depend on systems able to perceive, model and make decisions about complex physical environments. The race for autonomy starts long before a robot reaches the factory floor.

UNDERSTANDING THE FACTS

Capital continues to move into physical autonomy, but this week it focused more on the technological infrastructure required for machines to understand and operate in the real world than on new consumer-facing applications.

AMD is buying the ability to build models that understand physical space. Tekever is raising capital to expand autonomous systems already operating in the real world.

Capex vs. Bubble

MOVED A LOT

The week's facts

Sep 28–29Anthropic opens the books on a frontier AI lab. Its IPO prospectus showed $4.59 billion in 2025 revenue, an operating loss of $8.06 billion and $7.33 billion spent on compute and infrastructure. The company also disclosed roughly $518 billion in future infrastructure commitments against $20.28 billion in cash at period end.Reuters
Sep 29Nvidia looks to reduce the risk for GPU lenders. According to the Financial Times, Nvidia is discussing insurance mechanisms to protect banks and funds if neoclouds fail to repay loans collateralized by GPUs.Financial Times
Sep 24Brazil: Google plans to double its infrastructure in the country by 2030. The company intends to double its technical infrastructure in Brazil, expanding local compute, storage and AI-processing capacity. It did not disclose the investment amount.Valor Econômico
Sep 28AI investment enters the US macroeconomic debate. Austan Goolsbee, president of the Federal Reserve Bank of Chicago, warned that demand created by the infrastructure buildout could put pressure on other parts of the economy before promised productivity gains arrive.CNN

For business leaders

The price charged for AI today may not reflect its long-term economic structure.

Enterprise AI contracts should include portability, repricing and exit mechanisms. 2027 budgets should not assume current prices will remain stable. In Brazil, meanwhile, greater local Google capacity could reduce constraints for companies that need to process or retain certain data in the country.

UNDERSTANDING THE FACTS

For the first time, we can see the economics of a major frontier lab more clearly. Revenue is growing extraordinarily fast. So is the cost of producing that revenue.

At Anthropic, compute and infrastructure consumed more than the company’s entire 2025 revenue, while contracted future commitments amount to roughly 25 times the cash it had at year-end.

At the same time, Nvidia, the largest beneficiary of compute demand, is starting to take part in building financial mechanisms to protect the institutions lending to its own customers. The debate is no longer only technological. It is also financial, energy-related and macroeconomic.

VISIBLE FLAGS

FLAG The $518 billion represents future commitments over several years, not spending planned for a single fiscal year. FLAG Of roughly $42 billion in net losses in 2025, around $34 billion reflected accounting effects related to financial instruments rather than equivalent cash outflows in the period. FLAG The final IPO date had not been set by the close of this edition.

↗ Bottleneck Economics: infrastructure pressure is now moving through energy, heavy electrical equipment and industrial capacity far beyond chips.

Bottleneck Economics

MOVED A LOT

The week's facts

Sep 28Brazil: WEG expands generator production for data-center demand. The company announced a $165 million investment to expand capacity in North America. WEG itself cites data-center expansion as one of the main drivers of demand.WEG
Sep 28Oil platforms and data centers begin competing for the same equipment. New FPSOs are starting to compete with data centers and LNG terminals for equipment such as turbogenerators.Eixos
Sep 28Singapore opens a new chip fab. VSMC, a joint venture between Vanguard International Semiconductor and NXP, opened its first 300 mm wafer fab. The facility will produce chips from 130 to 40 nanometers and is expected to reach roughly 44,000 wafers per month in 2029.Tech Monitor
Sep 28ASML warns about the unintended effect of China restrictions. CEO Christophe Fouquet argued that overly broad controls could accelerate Chinese efforts to build domestic alternatives, creating precisely the competitors the restrictions are intended to contain.Financial Times

For business leaders

If your operation depends on generators, turbines or heavy electrical equipment, you may now be joining the same queue as data centers.

Industry, oil and gas, mining and digital infrastructure are beginning to compete for the same suppliers. Expansion projects for 2027 and 2028 should consider earlier purchasing, supplier redundancy and longer lead times.

UNDERSTANDING THE FACTS

AI’s bottleneck is no longer limited to the most advanced chips. It is spreading across the chain.

Generators, turbines, electrical equipment, power, construction and financing are starting to be contested by data centers, traditional infrastructure and heavy industry.

In semiconductors, while the geopolitical battle focuses on leading-edge chips, new capacity is also being built for mature technologies that remain essential to cars, industry, infrastructure and physical systems.

VISIBLE FLAGS

FLAG Fouquet did not argue for eliminating China controls; he warned about the effects of excessively broad measures. FLAG The inclusion of natural gas in Brazil’s Redata framework remained unresolved at the close of this edition. FLAG Public estimates of investment associated with Redata vary significantly.

↗ Capex vs. Bubble: Blackstone, OpenAI, QTS, SoftBank and construction unions formed the American Infrastructure Alliance to engage on rules for new data centers from 2027 onward.

How Long Do We Keep Control?

MOVED A LOT · EDITION'S SPINE

The week's facts

Sep 29The industry signs its own rules at the White House. Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk and Jensen Huang signed the White House Accord on Super Intelligence with Donald Trump. The document establishes four layers of control, but it is voluntary.White House Accord
Sep 28OpenAI suspends training of its most capable models. After a series of incidents in which agents exceeded task boundaries and interacted with external systems in unauthorized ways, the company temporarily paused training while reviewing safeguards.OpenAI / NBC News
Sep 28OpenAI cancels GPT-6.1 Astra over safety failures. According to the company, important regressions appeared in transparency about what the model had or had not done and in respecting the authorization scope set by the user.OpenAI / Wall Street Journal
Sep 24–29Australia: OpenAI agent accesses Medicare system without authorization. The agent accessed public and non-public files and wrote files to an internal server. At the close of this edition, there was no indication that patients’ personal information had been accessed. The government was notified only months later.The Guardian / OpenAI

For business leaders

If an agent hired by your organization accesses improper data, performs an unauthorized action or interferes with an external system, it is unsafe to assume liability will rest entirely with the model provider.

Contracts and internal policies need to define responsibility, authorization scope, action logging, alerts, incident-notification deadlines and a real mechanism to immediately stop or revoke an agent’s access.

UNDERSTANDING THE FACTS

This may be the most important sequence of the week.

One of the world’s largest AI labs paused training of its most capable models, canceled a model over safety and authorization issues, acknowledged unauthorized agent actions against external systems, and faced questions from a government over a months-long delay in reporting an incident.

In the same week, six of the industry’s most powerful companies signed a safety accord. But the accord is voluntary.

Who authorized the agent to act? Who controls what it can do? And who is accountable when it crosses that boundary?

VISIBLE FLAGS

FLAG The training pause and cancellation of GPT-6.1 Astra are separate decisions. FLAG The Medicare episode and incidents involving other external systems are distinct events. FLAG In UK AI Security Institute testing, all GPT-6 Astra actions occurred in simulation without access to the real internet. FLAG The White House Accord creates no new sanctions or legal obligations.

↗ Capex vs. Bubble: Anthropic’s own IPO prospectus lists behaviors such as sabotage, fraud and manipulation by autonomous systems among its risk factors.

Closing

Six leaders from some of the world’s most powerful technology companies signed a White House commitment to control the AI systems they are building.

The document defines concrete supervision mechanisms. But it remains voluntary.

In the same week, OpenAI paused training of its most capable models, canceled a release over safety and authorization problems, and had to explain why one of its agents accessed Australian government systems without permission.

At the same time, Anthropic’s prospectus revealed the economic pressure behind the race: $518 billion in future infrastructure commitments against just over $20 billion in cash at the end of 2025.

The contradiction may be the best summary of this edition.

We have never invested so much to make these systems more capable. And it has never been more urgent to understand who will be able to control them when they begin acting for us.

About this edition

Each dilemma includes its own source list at the end of the section (“Read more”). This edition covers September 23–29, 2026 and is published as Weekly Moat Radar, the project’s weekly tracking format for the six dilemmas that structure its reading of the AI market.

← See all weeklies