Billions spent and hypothetical returns: the AI boom explained with six charts

1. AI has sent stocks soaring

The S&P 500, which tracks nan 500 biggest US companies, has been connected a tear complete nan past 5 years – rising by astir 80%. That jump has been driven by large tech stocks pinch a liking successful nan AI boom, nan “magnificent seven” of Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla. 

The investor attraction connected exertion is unprecedented, says Jim Bianco of nan US institution Bianco Research, which recovered that 41 AI-related stocks now relationship for astir half nan S&P 500’s marketplace value.

Neil Wilson, an expert astatine nan finance level Saxo UK, says nan imaginable of a 1970s-style ostentation shock, lofty tech valuations successful wide and a imaginable frost successful nan backstage in installments marketplace do not bode good for stocks.

“The full marketplace has go 1 elephantine AI edifice,” he says. “The threat is simply a repetition of nan dotcom bubble – a immense crash, and years of mislaid returns. By immoderate measures valuations aren’t arsenic stretched arsenic past but this looks for illustration an incredibly vulnerable market.”

2. Expenditure is increasing astatine a staggering rate

Spending connected AI – from datacentres to chips – is racing ahead, from $765bn this twelvemonth to $1.6tn successful 2031, according to Goldman Sachs. The finance slope acknowledges location could beryllium problems pinch this standard of commitment. What if nan datacentres are delayed?

“At nan standard of superior being committed, moreover humble delays successful execution induce existent scrutiny astir nan request assumptions utilized to underwrite these investments,” opportunity Goldman analysts, though they adhd that if nan spending plans spell up without hitches, it could unleash a caller activity of AI demand. Nonetheless, nan expenditure shows really overmuch world financial resource, and anticipation for a return, is being committed to AI.

3. Firms and consumers are adopting AI astatine pace

Despite mixed reports connected nan benefits, nan immense mostly of companies are starting to usage AI – up from 33% successful 2023 to astir 80% now, according to nan consultancy group McKinsey. Usage among nan wide nationalist is besides high, pinch OpenAI’s ChatGPT now reaching 1bn monthly progressive users, according to information from Sensor Tower – a grounds for immoderate app.

The mobility now for AI developers is really to make money from this immense nationalist and backstage customer base. Companies request to beryllium capable to show that AI improves outcomes and reduces their costs capable to warrant nan bill. That intends utilizing it to build full workflows – business motto for carrying retired an full task from opening to end. There is simply a agelong measurement to spell connected that. 

4. Claude is snapping astatine ChatGPT’s heels

Anthropic began to summation crushed connected OpenAI precocious past year, erstwhile its Claude Code instrumentality went viral among mostly San Francisco-area package developers, earlier spreading much widely. Claude Code represented a displacement successful really ample connection models – nan halfway exertion down chatbots – are used, ushering successful a modulation towards autonomous AI agents that transportation retired tasks without quality intervention, enabling moreover nan non-tech-savvy to create package and do a wide scope of tasks.

OpenAI still has nan acold larger wide personification base, but information from nan net study institution Kentik – which tracks usage crossed a number of US net work providers – shows that Anthropic is quickly catching up. Claude’s personification postulation grew importantly faster than that of ChatGPT and Google’s Gemini betwixt January and April, spiking aft nan Pentagon declared it a proviso concatenation risk successful March. At this complaint of growth, Kentik projects that it could overtake ChatGPT by summertime – 1 much logic why Anthropic might spot an easier way to an IPO than its rival.

5. AI is getting much costly to usage

Every clip an AI chatbot aliases supplier issues a response, it is measured successful “tokens” – building blocks of connection that tin beryllium words, punctuation marks aliases syllables. (For example, OpenAI says the building “You miss 100% of nan shots you don’t take” is worthy 11 tokens.) It besides uses tokens to measurement inputs, specified arsenic nan punctual you type into ChatGPT.

The costs of these alteration per model; OpenAI prices it astatine $5 a cardinal input tokens for GPT-5.5, and $30 a cardinal output tokens (ie nan consequence fixed to your prompt).

The problem for subscribers is that token costs are going up massively, moreover arsenic companies everyplace are encouraging labor to “tokenmaxx”, that is, really spell difficult connected utilizing AI. The problem for AI companies is that they still aren’t charging enough.

The inherent committedness successful AI usage is that nan money a institution spends connected utilizing these devices is much than paid backmost successful improved productivity – a measurement of economical efficiency, wherever improved productivity intends you get much output from each worker. If this trade-off isn’t happening, past nan assumptions underpinning AI valuations – and policies – is undermined.

“The costs are getting wholly retired of control,” says Liam Betsworth, laminitis of nan British AI startup Pendra. Software developers successful his circle are utilizing agents to code, he said, starting pinch nan cheapest subscription, and very quickly moving connected to nan astir costly package. They aren’t unsocial – news tract Axios precocious reported connected an unnamed institution that spent $500m successful a period connected licences for Claude Code. 

6. Datacentre building mightiness not support gait pinch demand

Datacentre building represents nan cardinal tense strategy of AI products truthful increasing improvement and usage of AI devices must beryllium matched by much capacity – different location will beryllium a compute crunch, which intends rising costs for AI companies and users.

The sector’s standard of ambition for datacentres is immense and seemingly improbable. Bloomberg estimates that 23GW of capacity was nether building globally successful 2025 (capacity is measured successful electrical power, because that is nan constraint connected really overmuch computing a tract tin perform).

The US spot institution JLL predicts that 100GW will beryllium added betwixt 2026 and 2030 – a doubling of what they estimate arsenic existent capacity- balanced to 1,200 datacentres. JLL says its estimate takes into relationship speculative projects that ne'er break ground.

Where nan money – and power proviso – will travel from to fulfil this forecast is an unfastened question. Cecilia Rikap, an subordinate professor astatine University College London, says galore projects astir nan world remainder connected governmental commitments to grow nan grid and present nan power; but governments mightiness not person nan wherewithal to deliver. 

She asks: “Has nan authorities calculated whether specified an description is feasible? Do they person nan money to do it? Have they taken into relationship nan associated biology damage?”

7. What AI models tin do is expanding quickly

The abilities of AI models person improved by leaps and bounds since 2023, according to METR, a investigation organisation that measures AI capabilities. 

METR’s measurements are based connected whether AI models tin transportation retired a coding task, quantified by nan magnitude of clip it would return a quality to do so. By this metric, AI models are doubling successful capacity each 4 months. For instance, Anthropic’s Claude Mythos exemplary is calculated to scope a 50% occurrence complaint connected tasks that would return a quality master betwixt 8 hours and 2 days. 

However, location is nary commensurate effect connected jobs – truthful far. A March study from Anthropic contained investigation showing that, successful theory, AI could execute a big of jobs from computing to ineligible work, but has yet to do truthful successful immoderate awesome force. 

Bouke Klein Teeselink, an world astatine King’s College London and an master connected nan effect of AI connected work, says location are bottlenecks successful adopting AI successful nan workforce. For instance, really overmuch of a main executive aliases elder manager’s occupation tin beryllium safely outsourced to a bot? Can legally delicate tasks beryllium done by thing different than a human? Nonetheless, he says, alteration is coming.

“We are very overmuch astatine nan early stages of nan AI gyration still. There are galore group doing tasks that could beryllium done by an AI. The magnitude of alteration we are going to spot will beryllium huge.”

8. Datacentres are propping up US GDP

Despite nan simplification successful US authorities employment nether Donald Trump’s management and wide layoffs crossed a wide swath of industries, US GDP has continued to turn – 2.1% successful 2025 and 1.6% successful Q1 2026, according to nan US Bureau of Economic Analysis. A Harvard economist, however, calculates that without nan datacentre boom, these figures could beryllium acold smaller – that is, that “investment successful accusation processing instrumentality & software” accounted for 92% of nan US’s GDP maturation successful nan first half of 2025. 

This intends that datacentres – and nan AI roar – transportation a disproportionate stock of US growth, and a ample portion of why nan world’s largest economy, contempt important headwinds, still looks healthy. Any dent successful this expenditure could person economic, and frankincense political, consequences.