Does the AI Discount Really Exist? | Investing.com
Accenture had the worst trading day in its history last week. On Thursday, shares fell as much as 20% intraday and closed down nearly 18%, the steepest one-day drop the company has ever recorded. The headline numbers from the quarter itself were not the problem: earnings per-share beat estimates, free cash flow came in at $3.6 billion, and operating margins expanded. The damage came from the Outlook, bookings fell, guidance for the current quarter landed below what Wall Street expected, and management trimmed its full-year growth forecast, with roughly $400 million of the quarter’s softness tied to the war in the Middle East.
That single session has compounded a brutal run. Accenture’s stock is now down roughly 57% over the past 12 months. Stretch the lens to five years and the picture barely improves: total return over that period sits at around -50%, meaning $1,000 invested in the stock five years ago would be worth less than $500 today, dividends included. Bloomberg Intelligence’s read on the sell-off was blunt: AI is disrupting demand across consulting and managed services. The same fear has been hammering the entire IT-services, cybersecurity, and enterprise-software complex for months.
The implicit assumption behind that fear is simple: AI is cheap, it will keep getting cheaper, and cheap intelligence eventually eats expensive human labor. That assumption deserves more scrutiny than it’s getting.
The starting premise, that AI is inherently cheap, doesn’t hold up well against what training and running these models actually consumes.
On the training side, an analysis published by Knowledge at Wharton, the research publication of the University of Pennsylvania’s Wharton School, found that training GPT-3 alone required roughly 1,287 megawatt-hours of electricity and emitted about 502 metric tons of CO2, the equivalent of 112 gasoline-powered cars running for a year, and that was a 2020-era model, orders of magnitude smaller than what OpenAI and Anthropic train today. The same analysis notes that inference, the everyday business of answering queries, can account for up to 60% of AI’s total energy footprint, and that a single ChatGPT query uses roughly 100 times the energy of a typical Google search.
Disclosure here is the real story, and it’s thin. OpenAI has said an average ChatGPT query uses about 0.34 watt-hours, and Google has published a more detailed figure of roughly 0.24 watt-hours for a Gemini text prompt. Anthropic has disclosed neither a per-query energy figure nor any Scope 1, 2, or 3 emissions data.
Independent estimates put a typical Claude Sonnet query somewhere between 0.8 and 5.5 watt-hours, already several times OpenAI’s disclosed number, before factoring in that reasoning models change the math entirely. AI o3 and DeepSeek-R1 have been benchmarked consuming over 33 watt-hours for a single long prompt, more than 70 times the energy of a lightweight model handling the same task.
Scale that up and the numbers stop being abstract. U.S. data centers consumed about 224 terawatt-hours of electricity in 2025, more than 5% of the country’s total electricity use, up from just 1.9% in 2018. MIT’s Energy Initiative puts U.S. data centers above 4% of national electricity consumption in 2023, projecting that share could reach 9% by 2030, with a single hyperscale data center drawing as much power as 50,000 homes. Water tells a similar story: a Washington Post and University of California study found that generating a 100-word email with GPT-4 consumes roughly 519 milliliters of water for cooling, about a bottle’s worth, and as much electricity as running 14 LED bulbs for an hour.
If AI-specific U.S. electricity demand reaches the projected 300 terawatt-hours annually by 2028, cooling alone could consume some 720 billion gallons of water, equivalent to more than a million Olympic-size swimming pools, or the annual water use of 18.5 million American households. When Microsoft trained GPT-4 in Iowa, its data centers there consumed around 11.5 million gallons, 6% of the local district’s wáter, in a single month, during a three-month training run, in a state two years into a drought.
None of this means AI is unaffordable. It means the "AI is cheap" premise rests on numbers the companies selling it mostly haven’t shown anyone.
The gap between today’s AI demand and the grid’s capacity to supply it is being closed through three mechanisms, and none of them is simply "the market sorting itself out."
1. Outside companies financing the build-out. In March 2026, the White House brought together Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI to sign a Ratepayer Protection Pledge, under which the companies committed to build, finance, or acquire their own power generation and cover the infrastructure upgrades tied to their data centers, regardless of how much electricity they end up using. On paper, that shields households. In practice, the pledge is voluntary, and a December 2025 report from Climate Power found residential electricity bills had already risen 13% across the U.S. in 2025, with states hosting dense data center clusters seeing sharper spikes, with Illinois up 16%, Virginia up 13%, Ohio up 12% year-over-year. The companies most exposed to this dynamic, Microsoft, Google, Amazon, Meta, aren’t AI labs themselves, but they capture much of the indirect upside from cloud and chip demand while bearing the political cost of the energy buildout.
2. Tax relief aimed squarely at data centers. A Good Jobs First database tracks 251 separate instances of data centers receiving tax credits or subsidies across 16 states between 2020 and 2026, led by Washington with more than 120 awards and Texas with roughly 86. Indiana’s package tied to Amazon Data Services is the largest on record at an estimated $8.2 billion. Virginia alone forfeited $1.6 billion in tax revenue to data center exemptions in fiscal 2025, up 118% from the year before, in a state that now hosts more than 600 data centers, over 10% of estimated global hyperscale capacity. Several states are now reversing course, Illinois will stop processing new applications to its data center incentive program as of July 1, 2026, and at least a dozen states have moratorium bills in this year’s legislative sesión, which itself tells you how large these giveaways had become.
3. Custom-built infrastructure designed around a handful of corporate tenants, not around what a region’s residents or businesses need. New transmission lines, substations, and generation capacity get sited and sized for one hyperscaler’s roadmap.
What’s notably absent from most coverage of this build-out is the question of whether the public actually needs this capacity, as opposed to whether it should come from gas, nuclear, or renewables. The renewable-versus-non-renewable framing dominates the debate, whether households should be subsidizing infrastructure built almost entirely for a handful of private companies, justified, when challenged, by appeals to national security and the race against China, barely comes up.
Set the energy question aside and look purely at the financing, and the picture gets stranger, not clearer.
OpenAI’s own disclosed numbers put 2025 revenue at roughly $13 billion, up from about $3.7 billion in 2024 and $1 billion in 2023, genuinely extraordinary growth. But in the first quarter of 2026 alone, OpenAI reported $5.7 billion in revenue against a $9.3 billion operating loss and a net loss north of $21 billion, with stock-based compensation of $2.3 billion. Internal projections reportedly show $14 billion in losses for 2026 and cumulative losses of $115 billion through 2029 before the company expects to turn a profit. Against that backdrop, OpenAI has committed to roughly $1.4 trillion in infrastructure spending and filed confidentially for an IPO on June 8, 2026, targeting a valuation above $1 trillion.
Anthropic’s trajectory is the more interesting comparison, because it’s the company often cited as proof the AI business model can work. Anthropic’s run-rate revenue crossed $47 billion by mid-May 2026, up from roughly $1 billion at the start of 2025, a climb with almost no precedent in software history. The company raised $65 billion in a Series H round at a $965 billion post-money valuation, and has now taken in over $126 billion across 26 funding rounds in total. By 2030, Anthropic’s training costs are projected at around $30 billion a year, versus roughly $125 billion for OpenAI over the same period, a fourfold gap. Even on the more favorable side of that comparison, Anthropic is still burning enormous amounts of capital relative to what it earns, it has simply found a more efficient ratio than its main rival, not a profitable one.
What this financing reveals isn’t just the scale of the bet, it’s the tone investors are being asked to accept around it. In a November 2025 appearance on the BG2 podcast, investor Brad Gerstner pressed Sam Altman on how a company generating roughly $13 billion in revenue could justify $1.4 trillion in spending commitments. Altman disputed the revenue figure, then cut the exchange short: "If you want to sell your shares, I’ll find you a buyer."
That response is worth sitting with. Faced with a question about the long-term viability of the business model, the CEO didn’t explain how the spending would eventually pay for itself, but he pivoted straight to share demand and liquidity. That’s the language of a speculative asset being marked up, not of a company defending a business plan. It raises a fair question about what an OpenAI IPO, or Anthropic’s, which filed confidentially for its own listing on June 1, 2026 at a $965 billion valuation, actually represents: a chance for outside investors to buy into proven economics, or a chance for insiders and early backers to cash out while hype is at its peak.
Move from infrastructure financing to the unit economics of the product itself, and transparency gets worse, not better. Neither OpenAI nor Anthropic publishes what a given prompt, coding task, or agentic project actually costs them to run, what compute capacity a token corresponds to, or how token usage maps onto something concrete like lines of code produced. Developers using tools like Claude Code work this out empirically, through trial and error, because the companies themselves haven’t said.
Layer the reliability problem on top of that opacity and the pricing question gets worse. By the end of 2025, roughly 85% of developers were using AI tools regularly, but errors and hallucinations compound over the course of an autonomous agent’s run, getting baked into the final output by the time the task finishes. Independent testing has found that 29% to 45% of AI-generated code contains security vulnerabilities, and close to 20% of recommended software packages don’t actually exist. A separate empirical study found nearly a third of AI-generated coding projects required manual debugging afterward, with the majority of failures traced to basic errors in code logic and structure rather than missing dependencies. Even on benchmarks designed to test reliability directly, the strongest publicly tested configuration, Claude Opus 4.5 with web search enabled, still produced unsupported claims roughly 30% of the time in difficult, multi-turn settings. In law, where precision matters most, Stanford’s RegLab measured hallucination rates of 69% to 88% on specific legal queries, and U.S. courts issued $145,000 in sanctions in the first quarter of 2026 alone against attorneys who filed AI-fabricated citations, the highest quarterly total on record.
The consulting industry itself has already been burned by this, and the case is barely a week old. KPMG published a flagship report on agentic AI back in October 2025, titled "Total Experience: Redefining Excellence in the Age of Agentic AI." It took until June 2026 for anyone to actually check the sourcing. An investigation by the AI-detection firm GPTZero found that of the report’s 45 citations, only five pointed to real, intact sources, and 40 of the 45 citation titles turned out to be fabricated, a pattern GPTZero termed "vibe citing," the citation cousin of vibe coding, where a model stitches together fragments of real sources and invents the rest.
UBS, the U.K.’s National Health Service, Swiss Federal Railways, and Transport for London all told the Financial Times that the report’s claims about their own AI usage were either untrue or misleading. The report even cited internal "KPMG research" claiming 55% of CEOs rank AI as their top investment priority, while the firm’s own 2025 CEO Outlook, published that same month, put the real figure at 71%. KPMG pulled the report once the findings surfaced. One of the same Big Four firms whose business model AI is supposedly about to disrupt just spent eight months distributing a report about AI’s promise that AI itself had quietly filled with fabrications, a fairly direct demonstration of why paying for AI output doesn’t guarantee getting something usable back.
Put bluntly: you pay for tokens upfront, with no refund mechanism, and what comes back may or may not be usable. That’s not the service model of a mature, fungible commodity, it’s closer to a gamble priced like a utility.
And the price itself is propped up by something temporary. Current AI pricing is subsidized by hundreds of billions of dollars in venture and sovereign capital, not by revenue covering cost, closer to an extended free trial than a stable market price. If that capital eventually demands a return, prices will need to rise toward what the infrastructure actually costs to run, which, as the energy data above suggests, may be considerably more than what users pay today. If AI becomes genuinely expensive and remains unreliable, the case for ripping out consulting, IT services, and software vendors in favor of it gets a lot weaker.
None of this means Accenture’s stock fell purely because investors got the AI story wrong. The company had a genuinely weak quarter on its own terms: bookings fell, guidance missed, and a war in the Middle East cut measurably into revenue. Software and consulting firms can underperform for reasons that have nothing to do with artificial intelligence.
But the AI-substitution narrative driving so much of the recent sell-off across consulting, cybersecurity, and enterprise software rests on an assumption that hasn’t been tested: that AI is, and will remain, cheaper than the labor it’s supposed to replace. The evidence above suggests the opposite is at least as plausible. The energy and water costs are larger and less disclosed than the marketing implies. A meaningful share of the infrastructure paying for today’s low prices is being financed by taxpayers and ratepayers rather than by AI companies’ own revenue. The two best-funded labs in the industry are burning tens of billions of dollars a year against revenue that, even at its most impressive, is a fraction of their spending commitments. And the product itself remains expensive to run and unreliable enough that paying customers sometimes get nothing usable for their money.
If AI’s current cheapness turns out to be a function of subsidy rather than genuine efficiency, the market may be making two mistakes at once: underestimating the businesses it’s selling off, and overestimating the technology it’s buying into instead.