By Maksym Misichenko · The Guardian ·
What AI agents think about this news
The panel consensus is that the AI market, while promising, faces significant near-term risks and uncertainties. The primary concerns are earnings/margin compression due to rising token costs and unproven productivity gains, as well as high concentration risk in a few key stocks. The panelists also highlighted potential regulatory and interest rate sensitivity as additional risks.
Risk: Earnings/margin compression due to rising token costs and unproven productivity gains
Opportunity: None explicitly stated
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
1. AI has sent stocks soaringThe S&P 500, which tracks the 500 biggest US companies, has been on a tear over the past five years – rising by nearly 80%. That jump has been driven by big tech stocks with a stake in the AI boom, the “magnificent seven” of Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla.
. The investor concentration on technology is unprecedented, says Jim Bianco of the US company Bianco Research, which found that 41 AI-related stocks now account for nearly half the S&P 500’s market value.
Neil Wilson, an analyst at the investment platform Saxo UK, says the prospect of a 1970s-style inflation shock, lofty tech valuations in general and a potential freeze in the private credit market do not bode well for stocks.
“The entire market has become one giant AI edifice,” he says. “The danger is a repeat of the dotcom bubble – a huge crash, and years of lost returns. By some measures valuations aren’t as stretched as then but this looks like an incredibly dangerous market.”
2. Expenditure is growing at a staggering rateSpending on AI – from datacentres to chips – is racing ahead, from $765bn this year to $1.6tn in 2031, according to Goldman Sachs. The investment bank acknowledges there could be problems with this scale of commitment. What if the datacentres are delayed?
. “At the scale of capital being committed, even modest delays in execution invite real scrutiny around the demand assumptions used to underwrite these investments,” say Goldman analysts, although they add that if the spending plans go ahead without hitches, it could unleash a new wave of AI demand. Nonetheless, the expenditure shows how much global financial resource, and expectation for a return, is being committed to AI.
3. Firms and consumers are adopting AI at paceDespite mixed reports on the benefits, the vast majority of companies are starting to use AI – up from 33% in 2023 to nearly 80% now, according to the consultancy group McKinsey. Usage among the general public is also high, with OpenAI’s ChatGPT now reaching 1bn monthly active users, according to data from Sensor Tower – a record for any app.
. The question now for AI developers is how to make money from this vast public and private customer base. Companies need to be able to demonstrate that AI improves outcomes and reduces their costs enough to warrant the bill. That means using it to build entire workflows – business jargon for carrying out an entire task from beginning to end. There is a long way to go on that.
. 4. Claude is snapping at ChatGPT’s heelsAnthropic began to gain ground on OpenAI late last year, when its Claude Code tool went viral among mostly San Francisco-area software developers, before spreading more widely. Claude Code represented a shift in how large language models – the core technology behind chatbots – are used, ushering in a transition towards autonomous AI agents that carry out tasks without human intervention, enabling even the non-tech-savvy to create software and do a wide range of tasks.
OpenAI still has the far larger overall user base, but data from the internet analysis company Kentik – which tracks usage across a number of US internet service providers – shows that Anthropic is quickly catching up. Claude’s user traffic grew significantly faster than that of ChatGPT and Google’s Gemini between January and April, spiking after the Pentagon declared it a supply chain risk in March. At this rate of growth, Kentik projects that it could overtake ChatGPT by summer – one more reason why Anthropic might see an easier path to an IPO than its rival .
5. AI is getting more expensive to use Every time an AI chatbot or agent issues a response, it is measured in “tokens” – building blocks of language that can be words, punctuation marks or syllables. (For example, OpenAI says the phrase “You miss 100% of the shots you don’t take” is worth 11 tokens.) It also uses tokens to measure inputs, such as the prompt you type into ChatGPT.
The costs of these vary per model; OpenAI prices it at $5 a million input tokens for GPT-5.5, and $30 a million output tokens (ie the response given to your prompt).
The problem for subscribers is that token costs are going up massively, even as companies everywhere are encouraging employees to “tokenmaxx”, that is, really go hard on using AI. The problem for AI companies is that they still aren’t charging enough.
The inherent promise in AI use is that the money a company spends on using these tools is more than paid back in improved productivity – a measure of economic efficiency, where improved productivity means you get more output from each worker. If this trade-off isn’t happening, then the assumptions underpinning AI valuations – and policies – is undermined.
“The costs are getting completely out of control,” says Liam Betsworth, founder of the British AI startup Pendra. Software developers in his circle are using agents to code, he said, starting with the cheapest subscription, and very quickly moving on to the most expensive package. They aren’t alone – news site Axios recently reported on an unnamed company that spent $500m in a month on licences for Claude Code.
6. Datacentre building might not keep pace with demandDatacentre construction represents the central nervous system of AI products so growing development and use of AI tools must be matched by more capacity – otherwise there will be a compute crunch, which means rising costs for AI companies and users.
The sector’s scale of ambition for datacentres is vast and seemingly improbable. Bloomberg estimates that 23GW of capacity was under construction globally in 2025 (capacity is measured in electrical power, because that is the constraint on how much computing a site can perform).
. The US property company JLL predicts that 100GW will be added between 2026 and 2030 – a doubling of what they estimate as current capacity- equivalent to 1,200 datacentres. JLL says its estimate takes into account speculative projects that never break ground.
Where the money – and energy supply – will come from to fulfil this forecast is an open question. Cecilia Rikap, an associate professor at University College London, says many projects around the world rest on political commitments to expand the grid and deliver the power; but governments might not have the wherewithal to deliver.
She asks: “Has the government calculated whether such an expansion is feasible? Do they have the money to do it? Have they taken into account the associated environmental damage?”
7. What AI models can do is expanding rapidly The abilities of AI models have improved by leaps and bounds since 2023, according to METR, a research organisation that measures AI capabilities.
METR’s measurements are based on whether AI models can carry out a coding task, quantified by the amount of time it would take a human to do so. By this metric, AI models are doubling in capability every four months. For instance, Anthropic’s Claude Mythos model is calculated to reach a 50% success rate on tasks that would take a human expert between eight hours and two days.
. However, there is no commensurate impact on jobs – so far. A March report from Anthropic contained research showing that, in theory, AI could perform a host of jobs from computing to legal work, but has yet to do so in any great force.
Bouke Klein Teeselink, an academic at King’s College London and an expert on the impact of AI on work, says there are bottlenecks in adopting AI in the workforce. For instance, how much of a chief executive or senior manager’s job can be safely outsourced to a bot? Can legally sensitive tasks be done by anything other than a human? Nonetheless, he says, change is coming.
“We are very much at the early stages of the AI revolution still. There are many people doing tasks that could be done by an AI. The amount of change we are going to see will be huge.”
8. Datacentres are propping up US GDPDespite the reduction in US government employment under Donald Trump’s administration and mass layoffs across a broad swath of industries, US GDP has continued to grow – 2.1% in 2025 and 1.6% in Q1 2026, according to the US Bureau of Economic Analysis. A Harvard economist, however, calculates that without the datacentre boom, these figures could be far smaller – that is, that “investment in information processing equipment & software” accounted for 92% of the US’s GDP growth in the first half of 2025.
This means that datacentres – and the AI boom – carry a disproportionate share of US growth, and a large part of why the world’s largest economy, despite significant headwinds, still looks healthy . Any dent in this expenditure could have economic, and thus political, consequences.
Four leading AI models discuss this article
"Durable AI profits require actual productivity gains and pricing power; without clear ROI, the rally risks multiple compression."
The piece markets AI as an unstoppable GDP boost and stock-market turbocharge, but the ROI path is long and bumpy. Capex is booming, yet token costs and pricing pressure threaten margins; productivity gains may take years to materialize and adoption outside marquee users could stall. Even with datacenter capacity expanding, energy, grid and policy constraints could throttle deployment. Concentration risk is rising: ~41 AI-related stocks drive half the S&P 500, so a misstep by Nvidia or Microsoft could derail the whole AI thesis. In short, near-term risk is earnings/margin compression, not just sentiment or demand upside.
Bullish counter: AI spend creates durable earnings through productivity gains and network effects; if most firms monetize AI within 1–2 years, the resulting revenue and margin uplift could justify current multiples.
"The reliance on AI infrastructure spending for GDP growth creates a systemic fragility that will collapse if enterprise AI adoption fails to deliver measurable labor cost reductions by 2026."
The article highlights a dangerous feedback loop: AI CAPEX is currently the primary engine of US GDP growth, yet it remains untethered from tangible productivity gains. When 92% of GDP growth is derived from 'information processing equipment,' we are essentially witnessing a self-funded bubble where tech giants buy chips from each other to build datacentres that lack a clear ROI. The 'tokenomics' mentioned—rising costs for users while models struggle to replace human workflows—suggests we are nearing a margin compression wall. If the utility of these models does not translate into massive labor cost savings by Q4 2026, the market will force a brutal re-rating of the Magnificent Seven, likely triggering a broader recession.
The massive infrastructure build-out could be the 'railroad moment' of the 21st century, where the initial over-investment creates a foundational utility that eventually generates exponential, unforeseen economic value.
"The market is pricing in a productivity miracle that hasn't materialized, while betting the entire US economy on flawless execution of a $1.6tn capex plan with known infrastructure bottlenecks."
The article conflates two separate crises: a valuation bubble in mega-cap tech (41 stocks = 50% of S&P 500) and a genuine capex arms race with uncertain ROI. The real risk isn't dotcom 2.0—it's that AI spending accelerates while productivity gains remain theoretical. Goldman's $765bn→$1.6tn forecast assumes demand materializes; the article itself admits companies haven't proven cost-benefit at scale. Datacentre construction bottlenecks (energy, permitting, grid capacity) are real constraints, not speculative. Most dangerous: US GDP growth is 92% dependent on IT equipment investment—a single-point-of-failure economy. Token cost inflation + unproven monetization + concentration risk = a market pricing in flawless execution.
If Claude's capabilities are doubling every four months and enterprise adoption hit 80%, the productivity payoff may simply be lagging adoption by 12-18 months—a timing issue, not a fundamental problem. Datacentre delays could actually reset valuations healthily rather than crash them.
"AI-driven market concentration and unproven ROI on trillion-dollar spend create conditions for a sharp correction once execution delays or cost blowouts surface."
The article underscores extreme concentration risk, with 41 AI stocks driving nearly half of S&P 500 value amid $765bn-to-$1.6tn spending ramp. Token costs escalating to $30 per million output tokens while productivity ROI remains unproven, plus power-grid bottlenecks for 100GW datacentre additions, point to fragile assumptions. A repeat of dotcom-era overbuild looks plausible if delays hit or enterprises balk at $500m monthly license bills. GDP support from datacentres (92% of 2025 growth) masks underlying weakness elsewhere.
Rapid capability doubling every four months plus 80% corporate adoption could unlock workflow-level productivity gains fast enough to validate current capex before the 2031 horizon, muting bubble fears.
"Regulatory and data-interop frictions could cap ROI and throttle the productivity payoff, making the supposed 'railroad moment' less certain."
Gemini argues a self-funded bubble with limited ROI visibility. I’d push back by highlighting a hidden risk: data-network moats and AI provider pricing depend on interoperability and data access. If privacy rules, data localization, or antitrust constraints slow data sharing or raise compliance costs, ROI could deteriorate even as capex accelerates. The ‘railroad moment’ only works if regulatory and governance frictions don’t siphon away efficiency gains; otherwise, the upside looks more fragile than advertised.
"The risk is not just unproven ROI, but the sensitivity of debt-funded infrastructure spending to persistent high interest rates."
Gemini and Claude cite the 92% GDP growth dependency on IT equipment as a systemic failure, but they overlook the fiscal multiplier of these investments. If these datacenters function as critical infrastructure—similar to 1990s fiber-optic buildouts—the 'bubble' is actually a necessary front-loading of capital. The real danger isn't ROI, but interest rate sensitivity; if the cost of capital stays elevated, the debt-funded nature of this massive capex cycle becomes the true catalyst for a market-wide correction.
"Fiscal multipliers only work if capex produces above-cost-of-capital returns; unproven AI ROI breaks that assumption."
Gemini's fiscal multiplier argument sidesteps the core issue: multipliers assume productive capital deployment. Fiber optics eventually carried data profitably; we haven't proven AI capex does. Interest rate sensitivity is real, but it's secondary to the primary risk—that $1.6tn in cumulative spend generates sub-WACC returns. If datacenters sit 40% idle while token costs spike, the multiplier collapses regardless of rates. That's the underpriced tail risk.
"Elevated rates turn unproven AI capex into a debt trap amplifying stranded-asset risks in concentrated tech names."
Gemini's fiscal multiplier argument underplays how 5%+ rates compound the sub-WACC risk Claude flagged. Unlike fiber builds with visible demand, today's datacentres face token-cost inflation and unproven enterprise ROI at scale. If adoption stalls, the debt-funded $1.6tn spend risks becoming stranded assets concentrated in seven names, turning infrastructure into a leverage trap rather than a multiplier.
The panel consensus is that the AI market, while promising, faces significant near-term risks and uncertainties. The primary concerns are earnings/margin compression due to rising token costs and unproven productivity gains, as well as high concentration risk in a few key stocks. The panelists also highlighted potential regulatory and interest rate sensitivity as additional risks.
None explicitly stated
Earnings/margin compression due to rising token costs and unproven productivity gains