AI was sold as a way to get more done for less. In 2026 AI is costing far more than promised, and the bill lands in three places: in company budgets and the jobs cut to balance them, in the lives of the workers who make AI safe, and in the energy and water it consumes.
A small version of that cost is visible on LinkedIn. Open the platform on any morning and AI content is already waiting. Staying visible in the daily flood of information eats hours, and the pressure to post more and keep your audience engaged has become a trap that puts your own voice at risk. So instead of posting less but better, many of us reach for AI as a quick fix.
The result is a feed of thought leadership and motivational posts we have read a hundred times. That feed has become a factory floor where AI writes, humans click Like and whatever signal there was disappears into the hum. This is what you get when volume beats value. Is it simply how things are now? Do we keep up or push back?
The same logic plays out on a far larger scale. In 2026 more than 166,000 tech workers have been laid off, and 56% of companies name AI as the reason.
Oracle put that link in writing. Its annual SEC filing connects its workforce reductions to AI: "The adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce." That sentence is a legal statement to regulators. The effect shows on the payroll, which carried about 21,000 fewer people in fiscal year 2026, nearly 13% of the workforce.
Meta explicitly tied its May 2026 cut of about 8,000 positions, 10% of its global workforce, to the funding of its AI buildout. At SAP the CFO Dominik Asam told Handelsblatt that annual job cuts of one to two percent of its global workforce are now the company's permanent rhythm. He compared those cuts to brushing your teeth, something done every day without a second thought.
Yet the savings come with a price. Uber's CTO Praveen Neppalli Naga said the company burned through its entire 2026 AI coding budget in four months. Because the bills had grown too high, Microsoft reportedly told engineers in a major division to stop using an AI coding assistant. And Bryan Catanzaro, Nvidia's vice president of applied deep learning, told Axios, "For my team, the cost of compute is far beyond the costs of the employees."
Across the industry, spending runs far ahead of income. Big Tech has announced roughly USD 725 billion in capital expenditure this year, about 77% more than in 2025. Gartner projects that spending on AI agent software alone will reach USD 207 billion in 2026, up 139%. Sequoia Capital partner David Cahn estimates that AI companies need about USD 600 billion in annual revenue to justify current infrastructure spending. They are nowhere near that figure. As of mid-2026, the gap is widening, not closing.
Yet the economic case for replacing people with AI is weaker than the layoffs suggest, and two findings show why. A 2024 MIT study found that AI automation is cost-effective in only about 23% of roles where computer vision is central to the work, which leaves humans cheaper in the other 77%. And Goldman Sachs chief economist Jan Hatzius told the Atlantic Council that AI investment contributed "basically zero" to US economic growth in 2025.
As a result, some companies already regret the cuts, and analysts expect more to reverse course. A 2025 Orgvue survey of more than 1,000 business leaders found that 39% had laid off workers because of AI, and that 55% of those leaders now admit it was the wrong decision. Gartner predicts that by 2027 half of the companies that cut customer service roles because of AI will rehire staff for similar work, often under different job titles.
Companies regret the cuts because what AI cannot do turned out to matter most. AI cannot hold a client relationship together over years, weigh an ambiguous situation or answer for what went wrong. These abilities are not soft skills but the core of most high-value work, and they do not come in a subscription. I suspect this also explains why return-to-office mandates and AI expansion are arriving together. Companies that overshot on automation are rebuilding the human infrastructure they quietly dismantled, while trying to justify the AI spend to investors.
Employees pay twice, first when the job goes and again when they return to find the terms have changed.
Another group pays without ever appearing in these figures. Behind every AI product stands a workforce that never reaches the front page, in Nairobi, Manila, Bogotá, Accra and Bulgaria. These workers look at the worst the internet produces, from child sexual abuse material to live-streamed violence and graphic torture, and they label and filter that content so that AI does not hand it back to someone on the other side of the world. They sign NDAs and may not talk about their work or their earnings with anyone, so they carry it alone.
The harm is documented. In 2025 the human rights organisation Equidem interviewed 113 data workers in Colombia, Kenya and the Philippines and recorded psychological harm that included PTSD and substance dependence. One former content moderator read up to 700 sexually explicit and violent texts a day. The psychological cost was permanent.
In Nairobi this harm has reached the courts. More than 140 former Facebook content moderators were diagnosed with PTSD, and their lawsuit against Meta and Samasource Kenya alleges that exposure to necrophilia, child sexual abuse and terrorism caused severe trauma. Separately, workers in Nairobi who labelled data for OpenAI earned between 1.32 and 2 dollars an hour in 2021 and 2022, TIME reported. Kenya's labour law does not cover psychological injury from digital content, so it requires no therapy, sets no limit on exposure and offers no hazard pay.
"They call it Artificial Intelligence," a former supervisor told reporters. "There is nothing artificial about the trauma these young Kenyans are carrying home every night." The AI that writes LinkedIn carousels was made safe by someone in Nairobi who is not safe themselves.
The cost does not stop with people. A June 2026 report from the United Nations University projects that by 2030 data centres powering AI will draw 945 terawatt-hours of electricity, nearly triple what Pakistan, Bangladesh and Nigeria use together in a year. Those three countries are home to more than 650 million people. The same data centres will have a water footprint equal to the basic annual domestic water needs of all 1.3 billion people in Sub-Saharan Africa and a land footprint above 14,500 square kilometres, about twice the Jakarta metropolitan area.
At the level of one model, researchers at UC Riverside estimated that training GPT-3 alone used 700,000 litres of freshwater for on-site cooling, enough to manufacture 370 BMW cars. At the level of the industry, AI data centres are now estimated to cause 2.5 to 3.7% of global greenhouse gas emissions, more than aviation's 2%, and those emissions are growing by 15% a year.
Every ChatGPT prompt has a water bill and every AI-generated LinkedIn post a carbon cost. The figures for individual prompts and posts are not public, and the companies that hold them are not required to share them.
The story of AI and work in 2026 is that of a very expensive bet. The companies that placed it underestimated the cost of compute, the stubbornness of human judgment and the damage to people and ecosystems in places like Nairobi that barely reach the front page.
I keep asking whether this race is worth it, and whether a slower way of building an audience, a business and a future might be more sustainable and more enjoyable. I expect the race to go on for a while. But sooner or later we will have to face what AI has cost us.
If we do not step back soon, the consequences will not stay with the workers in Nairobi or the ecosystems around data centres but will reach all of us, in the jobs that quietly disappear, in the water we share and in the trust we place in content no human hand wrote.
