In 2015 OpenAI launched as a nonprofit with a charter promise to make sure artificial general intelligence benefits all of humanity. A decade later it is a company investors have reportedly valued at half a trillion dollars, signing commitments to spend hundreds of billions more on data centres. Somewhere between that founding language and the balance sheet, the AGI debate quietly became an economics debate.
Which is probably the more productive place for it. Arguments about whether the next model “counts” as AGI tend to stall, because nobody can agree on the test. Arguments about money resolve faster: capital has to be raised, chips bought, power generated, and every one of those bills comes due long before the productivity does.
The whole bet rests on one unusual claim
OpenAI’s thesis isn’t really that people will pay $20 a month for a chatbot. It’s that intelligence becomes a cheap, general-purpose input to almost every kind of work, and that a small cost per task multiplied by billions of tasks is a very large number. Sam Altman has floated figures on the order of $100 trillion in economic value. Outside analysts have gone further; one widely shared Open Philanthropy analysis sketched scenarios topping $9 quadrillion if growth compounds hard enough.
Take those as directional. The mechanism is what matters. Make a cognitive input 100 times cheaper and the number of tasks worth applying it to grows faster than the price falls. Electricity did this. Cloud computing did this. It’s also why the terminology fights matter more than they look: the arguments over what AGI actually means are partly arguments about which economic assumptions you’re willing to accept.
Costs land years before returns do
Stargate, announced in January 2025 with SoftBank, Oracle and MGX, commits $100 billion up front and up to $500 billion across four years. The first site in Abilene, Texas runs to hundreds of thousands of Nvidia chips and its own gas-fired generation. Add Microsoft’s roughly $13 billion of cumulative investment and OpenAI’s reported losses in the billions against about $3.7 billion of 2024 revenue, and the shape of the wager is clear: huge fixed costs, deferred payoff, no guarantee the demand curve arrives on schedule.
Infrastructure builds have a habit of running in that order. Britain’s railway mania laid track that reshaped the economy and ruined plenty of the people who paid for it. The fibre glut of 1999 to 2001 built the internet’s backbone and bankrupted a generation of telecoms firms. The assets outlived the balance sheets. Whether OpenAI becomes the Cisco of this cycle or its Global Crossing is a separate question from whether AGI works.
Productivity appears in narrow slices first
Today’s measurable gains cluster in work that is text-heavy, checkable, and already happening on a keyboard.
Coding
Sundar Pichai said in late 2024 that more than a quarter of new code at Google is generated by AI and then reviewed by engineers. Microsoft has made similar claims. Vendor numbers deserve scepticism, but the direction is consistent across labs, and it’s why coding agents were the first product category to win real enterprise budgets.
Support and back office
Klarna’s AI assistant handled the equivalent of 700 full-time agents’ work at launch, a claim the company later softened when it started rehiring humans for complex cases. Both halves of that story are instructive. The volume work really does compress, and the messy edge cases really don’t.
Science
Protein structure prediction turned years of lab crystallography into hours, and drug discovery is the prize everybody is chasing. This is where AGI’s economic impact could eventually dwarf everything else, and also where the timelines are longest and least legible from the outside.
The pilot-to-P&L gap
MIT’s NANDA group reported in August 2025 that around 95% of enterprise generative AI pilots produced no measurable effect on profit and loss. That isn’t proof the technology doesn’t work. It’s proof that integration, process redesign and data quality are where the money leaks out. The distance between a demo and a live workflow is where most of the real economic impact will be won or lost.
Labour effects cut along tasks, not job titles
Research from Stanford’s Brynjolfsson, Chandar and Chen found a roughly 13% relative employment decline among 22 to 25 year olds in the occupations most exposed to AI, even where total employment in those occupations held up. Firms stop hiring juniors for tasks the model now handles, while senior staff with context and judgement stay valuable.
That detail matters more than headline job-loss numbers. If entry-level roles are where people learn the trade, substituting them out becomes a pipeline problem five or ten years down the line. Wages are likely to move before employment does: compression at the bottom of knowledge work, premiums for people who can direct, verify and take responsibility for machine output.
Output moves before wages do
GDP is a poor guide to who ends up better off. If a lot of AI’s value arrives as consumer surplus, in free answers and cheaper software and faster service, the statistics will miss it, the same way they missed Wikipedia and GPS. Meanwhile the returns concentrate with whoever owns compute, weights and distribution. The clearest read on how fast capability and cost curves are actually moving is in these charts on the current state of AI, and the pattern is steep capability gains paired with stubborn inference costs away from the frontier.
OpenAI’s own structure shapes the distribution too. The 2025 reorganisation turned the capped-profit arm into a public benefit corporation, gave the nonprofit foundation a large equity stake, and left Microsoft with roughly 27%. The original profit cap, 100x for early backers, was a deliberate limit on how much private capital could capture. Lifting it turns a philosophical question about who benefits into a contractual one, negotiated in term sheets.
Regulation is now a cost line
Governance stopped being abstract for investors. More than a thousand employees at frontier labs, OpenAI among them, signed a statement asking for protections and a slowdown mechanism, part of the push for an AI brake pedal coming from inside the industry. The EU AI Act’s obligations for general-purpose models began applying in August 2025. Licensing regimes, liability rules and compliance overhead all feed into discount rates, and one adverse ruling in a major market can reprice a five-year capital plan overnight.
What to price in now
If you’re running a business, planning a career, or allocating capital, a handful of assumptions are worth building into your models.
- Assume inference gets cheaper, not free. Unit costs fall fast, but usage grows faster. Budgets tend to rise, not shrink.
- Fund the redesign, not the licence. The 95% pilot failure rate is mostly a process problem. The teams that win rewire workflows before they buy seats.
- Watch your own junior hiring. It’s the earliest and most honest signal of task-level substitution, visible in your headcount plan long before it shows up in national statistics.
- Track the capex-to-revenue ratio at the frontier labs. If spending keeps climbing while revenue growth flattens, the correction will be sharp and it will hit the entire supply chain.
- Follow the technical signals, not the press releases. Benchmark saturation, agent reliability on multi-step tasks, and cost per unit of useful work tell you more than any launch event. The AGI technology trends worth tracking make a reasonable starting map.
The honest answer to “what will AGI do to the economy” is that it depends on which bottleneck breaks first. If reliability improves faster than costs fall, gains stay concentrated in software and back office. If costs collapse before reliability does, you get a lot of cheap mediocre output and a muddled productivity story. And if both move at once, the trillion-dollar figures stop looking like marketing. OpenAI’s spending only makes sense in the third scenario, which is precisely why the company keeps arguing for it.