Pay, control, own: 203 billion of capex in one quarter, a shared robots.txt signal, Canada becomes a Cohere shareholder

Three signals this Friday, and one question: who owns the future of AI? Who pays for it, who controls the content it learns from, and who owns the company building it.

Signal one: the price of the buildout

Amazon published third-quarter 2026 results that tell the whole earnings week on their own.

• Revenue: 201.9 billion dollars, +12%.
• Amazon Web Services: 46 billion, +39%, its fastest growth in nearly five years.

And yet the stock fell about 7% after the close. The reason is the bill underneath the growth. Amazon expects about 220 billion dollars of cash capital expenditure this year, mostly for AI and cloud. Its trailing twelve-month free cash flow turned negative, at -42.4 billion, versus +14.8 billion a year earlier.

The big four, in one quarter

Company	Q3 2026 capital investment
Amazon	$77.3 billion
Alphabet	$51.4 billion
Microsoft	$39.4 billion
Meta	$34.9 billion
Total	about $203 billion

These amounts include finance lease payments. For the year, the hyperscalers, along with SpaceX, are heading for 803 to 828 billion dollars of spending.

The growth, on the other side, is real. Google Cloud more than doubled, Azure grew 48%, and Amazon expects cloud growth above 40% in 2027 and 2028. But the market no longer applauds growth alone. It asks when spending will turn into returns.

Why it matters

This week, the AI buildout got a price everyone can read. For two years the story was demand: more chips, more data centers, more capacity, as fast as possible. Cash flow statements have now caught up with the press releases. Alphabet has posted two straight quarters of negative free cash flow. Amazon joins it on a trailing twelve-month basis. Microsoft is the exception for now, with operating cash still covering its spending.

The question has changed shape. The reality of AI demand is settled. What remains is whether returns arrive before patience runs out. I connect this to the IMF scenario of October 13: that is exactly the risk it priced.

What changes

For brands, the infrastructure bill is now part of strategy. If the largest platforms spend hundreds of billions a year, they will have to recover it: through prices, new services, and tighter control of the customers they already have. Ads in ChatGPT, free Gemini shrinking, launch prices doubling: these are the first signs of that recovery.

For investors, the burden of proof has shifted. The next earnings season will be judged on evidence that this concrete and silicon produce durable profit, beyond impressive revenue. The 2027 capex budgets, due early next year, will tell whether this was a buildout or a bet that keeps growing.

Signal two: one shared sentence to say no

On Friday, OpenAI announced it will honor a shared signal in robots.txt, the small file through which a site indicates how its content may be used. Until now, a publisher wanting to refuse AI training had to write separate rules for each provider. The new idea: one preference, stated once, that every participating provider can read.

Where the signal comes from

From the AI Preferences working group of the Internet Engineering Task Force (IETF), the body that develops much of the internet's technical standards. A draft written by Kevin Kelley of Anthropic, dated October 30, defines a shared identifier for training preference. A site could say, for example: no training on this site, except for one section, such as a press kit.

OpenAI says it plans to implement the signal in the coming months, with its existing robots.txt controls kept in place alongside. It encourages other providers to honor the same signal.

The honest caveat

This is a preference. It locks nothing. The underlying vocabulary states that it does not guarantee preferences will be followed and leaves that choice to the recipient. Private agreements can also override it.

Why it matters

The web rests on an unspoken bargain. Publishers put content out in the open and get traffic and readers in return. AI answers break that bargain twice.

First, the answer can replace the visit. Reddit chief executive Steve Huffman said this week that Google's AI search experience is sending fewer visitors to websites, with an average decline of more than a million weekly Google referrals in October. It is the same phenomenon Alphabet acknowledged the day before in its own revenue.

Second, the content itself can become training data, with no visit at all.

A shared opt-out does not fix the traffic problem. It gives publishers one clear sentence, in one place, instead of a different form for every company that wants their work.

What changes

For brands and publishers, robots.txt becomes a strategic document. I said so on October 17 and 23 about Meta. Decide what you want found, what you want cited, and what you do not want training a model, then write it in a way machines can read.

I want to flag the tension, because it is real for a brand. Refusing training protects content. But a brand absent from training data may also be less well known to the models answering customers. The right setting depends on what the brand sells: a publisher selling its content protects it; a brand that wants to be recommended opens what describes it. The new signal makes it possible to distinguish between sections.

For AI companies, the test will be adoption. A signal honored by one provider is a gesture. Honored by all, it becomes a rule of the game. If Google, Microsoft and Meta adopt the same identifier, publishers will finally have a lever they can pull with one line of text. Otherwise, the real negotiation goes back to courts and contracts.

Signal three: Canada buys a stake in its own answer

On Thursday in Montreal, Prime Minister Mark Carney announced that Canada will take an equity stake of up to C$1.6 billion in Cohere, the Toronto company building AI systems for businesses and governments. The money will be paid in stages, tied to jobs and revenue targets, through the Canada Development Investment Corporation.

The conditions

Cohere must:

• keep its headquarters and incorporation in Canada;
• create at least 1,000 technical jobs there by 2030 (it currently employs more than 800 people);
• invest at least C$1.7 billion in Canadian computing infrastructure by 2030.

Canada joins a round led by Germany's Schwarz Group, with investors including Nvidia and Bell. The round targets up to C$7 billion at a valuation of about C$35.7 billion. Cohere has only announced a first close, without saying how much has been committed or how large Canada's final stake will be.

Why it matters

Most governments back their AI champions with grants, tax credits or cheap compute. Canada goes a step further and buys a piece of the company. A grant funds a project. An equity stake commits to a future, with a share of the gains if the company grows. Carney framed it as an act of sovereignty, far from a venture bet.

The timing adds a layer. It was Cohere that reported, on October 26, a Qwen model trying to send data to Beijing during a test. Sovereignty goes beyond owning the company. It also means knowing where models run, where data goes, and who can inspect the machine.

What changes

For every country watching, Canada sets a template and a test. The template: public money as a shareholder, with conditions on headquarters, jobs and compute at home. The test: will those conditions hold as the money grows and the company becomes more global? According to reports, Cohere is in the final stages of merging with Germany's Aleph Alpha. The question of what stays Canadian will sharpen quickly.

I think of Morocco and Africa as I write this. The sovereign AI question arises there too: own, host, or at least control access. Canada shows that a mid-sized country can choose ownership over dependence alone.

For brands, the lesson is simpler. The AI you buy comes with a flag, a jurisdiction and promises about where it lives. Choose accordingly, because customers and regulators soon will.

The thread of the day

Amazon and its peers showed the price of building the future. OpenAI offered publishers a shared sentence to defend their work. Canada decided to own a piece of its own answer.

Pay, control, own. Three ways of asking who the AI era belongs to.

The noise filtered. The signal kept.