AI by the numbers: what it is worth, what it runs into, and why "which shares should I buy" is the wrong question

Talk about artificial intelligence almost always collapses into two positions: everyone is about to be replaced, or it is another bubble. Both are about feelings rather than money. So we looked at the numbers — what the companies are worth, what they earn, what the growth actually runs into, and what any of it means for an ordinary person with a little in savings and a lot of questions.

What it is worth in money

At the close on 8 September 2026 Nvidia was worth $5.45 trillion, with the share at $225.73 — first among all companies in the world, and the highest market capitalisation a public company has ever reached. On that same day, incidentally, the stock fell 2%. Over five years it has gained something like 1,300%.

Your ad could be hereAdvertise here

Private companies have caught up with public ones. In its spring 2026 round Anthropic was valued at about $965 billion, ahead of OpenAI's $852 billion for the first time. Anthropic filed confidentially for an IPO with the SEC on 1 June 2026; OpenAI has no confirmed filing.

Revenue and losses are more interesting than valuations. Anthropic's annualised revenue run-rate was around $47 billion by May 2026; OpenAI's was about $25 billion in February. Their operating economics differ in kind, not degree: Anthropic has moved into profit, while OpenAI's operating margin is estimated at minus 122% — more than a dollar of loss for every dollar earned.

That is the number that matters to anyone thinking about investing: two companies in one industry, valued comparably, running opposite economics.

What growth actually runs into: electricity

Here is the part the AI coverage rarely shows. Models are run in data centres, data centres eat electricity, and electricity cannot be printed.

Morgan Stanley puts US data-centre demand at 74GW by 2028, with roughly 49GW of that unserved — the grid will not deliver it. Global electricity demand is set to rise by more than a trillion kilowatt-hours a year by 2030, and nearly 20% of the entire increase comes from this one sector. Utility spending is up 30% since 2019, and the shortfall risk window is 2027–2028.

The money going in matches: hyperscalers will spend more than a trillion dollars across 2025–2026, and global energy investment ran at $1.5 trillion in 2025.

Britain: a grid queue up to fifteen years long

The same story in miniature, and the numbers are sobering. The country has around 2.4GW of data centres today, 1.7GW of it within twenty miles of London. The government reckons at least 6GW of AI capacity is needed by 2030 — the output of about four large nuclear reactors.

Waiting in the grid connection queue are 140 projects totalling almost 50GW. Fifty gigawatts is the whole of Britain's peak demand. Half the projects are already funded. Some face waits of up to fifteen years.

And the most telling pair of figures: in 2025 the country approved close to £10 billion of data centres and built less than £1 billion of them. That tenfold gap is the real state of the industry once you take out the press releases. It is precisely why the government created five AI Growth Zones with accelerated planning and discounted power.

Science fiction that turned into a Tuesday

It is worth stopping occasionally to notice what has already happened. Live speech translation is not a prototype, it is a phone feature. Transcribing a doctor's appointment, drafting a contract, reading a thousand-row spreadsheet in a minute — routine now, and five years ago you paid a specialist for it. A model that writes working code from a description of the task. This is not the future; it is the day before yesterday's news, and we have simply stopped being surprised.

Which is where the catch sits: the technology became ordinary faster than people learned its limits.

Why some people use it and others have not understood the question

The divide does not run by age or education. It runs on one thing: whether a person tried asking in their own words.

Those who use it discovered something simple — you do not have to speak in commands. You can say "I have a letter from the tax office, explain what they want from me and what happens if I ignore it", and get an answer. Those who do not use it mostly tried once, typed three words as if into a search box, got a generic reply and concluded it was a toy.

What AI genuinely does well: explains complicated things plainly, translates, condenses and structures long texts, drafts, finds holes in an argument, works through options, takes a document apart clause by clause.

What it does not do, and where it goes wrong: it does not know what happened after its training unless it looks online; it will invent details confidently if nobody checks — dates, figures, the names of laws and court cases most of all; it does not replace a doctor, a solicitor or an accountant, it prepares you for the conversation with them; and it knows nothing about your situation until you describe it.

The practical rule is short: anything with a cost attached to being wrong, check against the primary source. A model's answer is a good draft, not a document.

How politicians know about shares — and what is being done about it

The question comes up constantly. The answer is duller than the conspiracy and more interesting than it sounds.

In the United States, members of Congress trading shares is not a secret: the STOCK Act obliges them to disclose purchases and sales, and public trackers make those trades visible to anyone. The openness is exactly what fuels the argument, because it keeps turning out that a legislator bought into an industry they regulate — and sometimes missed the disclosure deadline entirely.

On 22 July 2026 the House of Representatives passed a bill barring members, their spouses and their children from buying new stock, by 232 votes to 198. Existing holdings may be kept; selling requires 7 to 14 days of public notice. The penalty is $2,000 or 10% of the transaction, whichever is greater, plus disgorgement of any profit. Broad-based funds are exempt. The Senate's position is unclear: a tougher version requiring full divestment has sat in committee for over a year.

So "how do they know" is a question about conflicts of interest that their own legislature is arguing over, not about secret sources. And note the timing: disclosed trades reach the public with a lag of up to 45 days, so copying them after the fact means being late by design.

"Which shares should I buy before they go up?"

The honest answer: we do not give investment advice and we are not financial advisers. Not out of caution — because any answer to that question without knowing your situation is bad advice.

What we can say is what the figures above suggest, read soberly.

• The growth is already in the price. A company up 1,300% in five years is not being bought at a pre-boom price. The question is not whether the sector grows, but whether that growth is already in today's quote.
• A sector is not a company. Of the two leading AI companies, one is profitable and the other loses more than a dollar per dollar earned. A bet "on AI" does not distinguish between them.
• The bottleneck is physical. Fifteen years in a connection queue, and £10 billion approved against under £1 billion built, is not market sentiment — it is transformers, cable and lead times.
• Private companies cannot be bought on an exchange. Neither Anthropic nor OpenAI is listed. Anything offering you a private "stake" in them deserves extreme suspicion; this is classic ground for fraud.
• The British frame. Here investing lives inside ISAs and pension schemes with their own limits and tax rules, and investment firms must appear on the FCA register. Checking that registration is part of the decision, not a formality.

One last thing. If someone answers "what should I invest in" quickly, confidently and for free, that is the most reliable reason not to believe them.