Is the AI Coding Bubble About to Pop? Who Actually Benefits?

The AI coding bubble is two booms sharing a nickname. One is a stock and data-center build that Goldman Sachs still sizes at about $1 trillion worldwide in 2026. The other is the claim that coding agents have already turned extra keystrokes into extra finished software. The first boom can keep spending for months and still disappoint shareholders in 2027. The second one is already failing a simpler test: more code is landing, and the shipped work is not rising with it.

Humans who can tell a safe change from a plausible one are collecting the scarce payoff. Chip orders, bond deals, and model labs are collecting the capital. A $20 or $200 editor login sits in a third pile, and a bad week for Nvidia does not delete it.

The short answer on the AI coding bubble

A pop this week is the wrong forecast. Capital Economics told CNBC on 10 September 2026 that most of its bubble gauges say the equity boom is late, then still penciled in a higher S&P 500 by year-end 2026 (about 7,650 then, 8,250 by the close) and a fall to 6,500 by the end of 2027. That is roughly 8% up, then a 21% slide. James Reilly’s line was that, on past form, the end looks like months rather than years. Oxford Economics, on 29 September, put the painful part in the same window: the risks intensify as 2027 starts. A correction, in their wording, only needs results to miss elevated hopes. The technology can still be real.

For a working developer the useful split is elsewhere. In a Harvard study of 718 firms, AI coding agents lined up with 30% more lines of code, 20% more commits, and 23% more pull requests. Resolved tickets did not show a matching rise. Review time on those pull requests rose 49%. That is the bubble you can see in Git. It is a measurement problem with a bill attached.

Who benefits, if “AI” and “humans” are the two names in the question: owners of the buildout benefit while the capex is still being cheered, and they carry the debt if monetization misses. Companies that redesign work around experts are posting the large productivity gains. Developers who review, scope, and ship familiar code are the humans the labor data is still hiring. Early-career developers are the group absorbing the hiring damage. The model does not get a paycheck.

What is money Bubble?

Goldman Sachs Research, 7 August 2026, puts AI-related investment at about $1 trillion globally this year, including $581 billion in the United States. The bank’s cross-checks land near the same place: around $1 trillion worldwide and just under $600 billion in the US. Add the years since 2022 and the cumulative figure they use is $1.8 trillion by the end of 2026.

The commonly quoted hyperscaler capex number, $794 billion, is a messy shortcut. Goldman says it understates global AI capex by about $200 billion and overstates the US share by about $200 billion, because not every hyperscaler dollar is AI, a lot of the build sits outside the US, and private plus non-US spend gets left out. On their preferred series, US AI capex is 1.8% of GDP in 2026, 2.5% in 2027, and 2.8% in 2028. They also say the leading indicators they watch (chip-equipment imports, PMIs, GPU rental prices) are still near the top of their range since 2022, which is a way of saying the near-term spending tape still looks strong.

Oxford’s 29 September briefing is the cold shower on that tape. AI-related capex was an estimated 27% of US fixed investment in the second quarter, and cumulative investment in the sector topped 2% of US GDP across three years. In two centuries of tech booms, they say only the UK railway mania of the 1840s was a larger investment surge. Utilization-adjusted US productivity growth has been decelerating. AI firms are showing negative free cash flow. Consensus profit growth, in their view, does not line up with the productivity gains they think are likely. Dot-com at least produced measurable productivity quickly. This boom is being asked to do the same, soon.

The financing has changed shape since the summer notes. IMF managing director Kristalina Georgieva, 15 October 2026, wrote that hyperscalers’ share of investment-grade bond issuance rose more than tenfold in two years, from under 1% to 10%, and that the deals are often circular: hyperscalers invest in model labs, and the labs buy capacity from the same firms. The October 2026 Global Financial Stability Report says a disappointment in AI investment or profitability could cut global GDP, investment, and consumption by as much as 0.4%, 1.3%, and 0.6% by 2028. The same chapter notes that Bank of America’s August 2026 fund-manager survey ranked the “AI boom” as the top market risk. A background note puts interest expense in the second quarter of 2026 at nearly $6 billion across Alphabet, Amazon, Meta, Microsoft, and Oracle, more than double the year before, and records an Alphabet equity raise of about $85 billion in June 2026 (IMF background analysis).

Capital Economics, back on 10 September, still called leverage less alarming than earnings expectations, which were up at dot-com-peak territory, with index concentration in the same neighborhood. A month later the IMF put the debt and the circular funding in the foreground. I would not treat either shop as a calendar. I would treat the disagreement as the story: spending can stay huge and the equity still reprice, because the bust condition is “worse than the hope,” which is a low bar when the hope is a record.

Georgieva also offered the other half of Amara’s law in plain language. The short run gets overrated. The long run gets underrated. She put a possible growth boost on the order of half a percentage point a year over the next decade, and she was careful: businesses have to rebuild workflows before that shows up. About 40% of jobs globally are exposed to AI on IMF research, about 60% in advanced economies. She also said there is no apparent hit to the job market so far. Exposure is not a layoff count.

AI developer bubble review

The paper to read is Fiona Chen and James Stratton, “Artificial Intelligence in the Firm: Bottlenecks in Software Production”, Harvard, current version 4 August 2026. Jellyfish data: about 300 million work events, 725,938 workers, 718 firms, January 2021 through March 2026. Tools in the measurement include GitHub Copilot, Cursor, and Claude Code. Adoption of agents at these firms went from near zero in October 2024 to over 95% by January 2026.

Agents, after firm adoption:

Assistants (Copilot and Cursor seats, before the agent wave) were milder: commits up a significant 9%, lines and pull requests up a little and not significant, review length basically flat.

I keep seeing a “12% more finished work” line in write-ups of this paper. That 12% is the gain the authors could rule out. Their confidence interval on agent-era issue completion stops at about 12% of the baseline. The estimate itself is a noisy few percent. Ars Technica’s October 2026 piece gets this right. Some summaries do not.

On time, the paper’s survey of 100 engineers put coding at about 30% of the week. The Jellyfish time imputation put it at 42%. The rest is review, planning, and the other work a ticket needs before it counts. By March 2026 nearly 80% of the firms had some AI in code review. Even then, 23.3% of review comments were from AI, and 10.8% of pull requests had at least one AI comment. Humans were still the review. One senior engineer they interviewed said the tools lack a grasp of what the feature is supposed to do. That matches the pile.

Bain’s 2026 Global Technology Report, as written up by CIO Dive on 2 October from a survey of nearly 300 senior technology leaders, is the manager version of the same gap. Developers complete about 21% more tasks. Time spent reviewing outputs is up 91%. Simultaneous workstreams are up 47%. Leaders say they expect a 95% productivity boost and a 148% faster release cycle within a year or two. What they say they are capturing now is 20% to 27%. Purna Doddapaneni’s line is the one worth keeping: the bottleneck moved from writing code to trusting it, and the system runs at the speed of the slowest human checkpoint.

Expectations at 95%, delivery around a quarter of that. If you want a bubble you can explain without a stock chart, it is that gap.

The 19% slowdown is a 2025 snapshot

METR’s randomized trial, published 10 July 2025, found 16 experienced open-source developers took 19% longer with early-2025 tools (mostly Cursor Pro and Claude 3.5 or 3.7 Sonnet) on 246 real issues in repos they already knew. Before the tasks they forecast a 24% speedup. Afterward they still believed they had been 20% faster. The 24 February 2026 update says the confidence interval on that slowdown ran from +2% to +39% longer.

METR then tried a later experiment and basically threw the clean read out. Developers were refusing to join if they had to work without AI. Pay had been cut from $150 an hour to $50. People running several agents at once made the clock meaningless. On the subset of original developers who came back, the raw estimate flipped to an 18% speedup, with an interval from 38% faster to 9% slower. New recruits came in around 4% faster, interval from 15% faster to 9% slower. METR’s own words: this is weak evidence, and the true speedup among people who opted out could be larger. They also say self-reported speed is still unreliable.

An earlier note on this site, the workflow piece, described that February update as a clean flip from 19% slower to 18% faster. The note from METR is narrower than that sentence. Use the 2025 result as history. Use the 2026 result as “the lab can no longer run the old experiment, because too many developers will not work naked.” That refusal is a fact about adoption. It is a poor fact about how many minutes you saved on Tuesday.

What developers themselves trust

Stack Overflow’s 2026 survey is the mood, not a stopwatch. Among people using coding assistants or agents, 73% use them daily, and 30.9% of daily AI users are on the tools four hours or more. Trust is conditional: 48% trust the output when they can check it, and 16% extend trust to most tasks that are not important decisions. The work they actually hand over is familiar code (69.3%), debugging or refactoring (63.8%), straight technical questions (59.4%), tests (58.1%), and unfamiliar code (56.0%).

Among people who had used a coding agent or assistant in the past year (n = 12,255), Claude Code was 65.5%, GitHub Copilot 58.7%, OpenAI Codex 29.5%, and Cursor 25.5%. Multi-select, so the shares overlap. Only 6.4% said they switched tools for a lower price. Better output was the reason for 25.1%. The seat market is not behaving like a bubble that already popped. It is behaving like a quality fight.

17.1% said they avoid AI at work because they worry about losing the skill, or about training a replacement. That lines up with the skill-atrophy problem already on this site. The other 41.3% said they do not avoid it. Both groups are in the same profession.

Who benefits

AI versus humans” is a bad scoreboard. AI does not invoice. These are the parties that do.

PwC’s wage split is the cleanest “which humans” number in the set. On more than a billion job ads, roles they class as professionalised (AI takes the routine slice, the human keeps judgment) grew available jobs twice as fast as democratised roles, with 42% faster salary growth. Headcount at the most AI-exposed companies rose 52% from a 2018 baseline, versus 36% at the least exposed, and wages 24% versus 17%. Jobs asking for specific AI skills grew 69%, against 9% for the job market, and the wage premium for those skills was 62%.

Read that next to Harvard, or you will fool yourself. Inside firms that adopted coding agents, Chen and Stratton could not pin a junior-versus-senior employment change on the rollout, and they could not pin a company-wide jobs cut on it either. They flag the limits: short run, partial equilibrium, one analytics vendor’s customers. A company can roll out Claude Code and keep its juniors, while the wider market stops opening junior reqs. CNBC’s 11 October labor piece is that second picture. The BLS projection they cite still has software-developer employment growing 10% from 2025 to 2035, from about 1.72 million jobs to 1.89 million. The profession is not the bubble. The easy on-ramp is.

My own cut, after sitting with both: the humans who benefit are the ones who can close the review, name the tradeoff, and talk to the user. The humans who lose ground are the ones whose practice used to be the first draft. Companies benefit when they spend the saved drafting on a product somebody will pay for. They waste the seat when the only scoreboard is lines of code. That is an unglamorous answer, and it fits the tables better than “AI wins” or “humans win.”

What to keep paying

Pay for one coding agent you already trust on a repo you know. The October comparison on this site still has the practical split: Claude Code when the hard part is a multi-file change, Cursor when the editor itself is the product you want, GitHub Copilot when the team’s pull requests already live on GitHub and finance wants one Microsoft seat. What those tools cost once you leave the sticker price is a separate decision from a CNBC chyron.

Do not buy the top lane twice because the timeline feels loud. The Max versus Ultra piece, prices checked 21 October 2026, is the wallet version: both $200 seats exist, and most people who think they need both have not shown that a cheaper pair fails. Recheck the vendor pages before you renew. Limits move.

A two-week log is enough to know which row you are in. Count pull requests opened, days until merge, and issues actually closed. If the first number rises and the third one does not, you are living the Harvard result in miniature. Bring that log to the budget meeting. “We might be in a bubble” is not a measurement.

Where people fib to themselves

Lines of code are a vanity metric the moment an agent is in the repo. Chen and Stratton’s whole point is that the intermediate count and the shipped count come apart. If your dashboard leads with accept-rate or LOC, you are scoring the boom, not the work. The maintenance version of that fib is AI technical debt: code that passes and still costs a senior a week six months later.

The 12% figure will keep circulating. If a vendor slide uses it as “agents increase output 12%,” ask whether they mean the confidence-interval ceiling. They usually do not know.

The July 2025 METR slowdown will also keep circulating, sometimes from this site’s older wording. It described early-2025 tools on experienced maintainers. METR has said the follow-up cannot support a clean current percentage. Quote them, or leave the percentage out.

A stock drop is not a product sunset. Georgieva’s scenario is an equity and spending shock amplified by leverage and circular deals. Your editor keeps working the next morning. The way a correction reaches a developer is a renewal email, a tighter rate limit, or a finance partner who suddenly wants one seat instead of three. Have the log before that email.

And the reverse fib, “the agent deleted the job,” is also sloppy. Harvard’s adopters did not show that employment drop. Indeed’s postings rebounded, concentrated in senior and AI-titled roles. The documented human damage is the narrower one: fewer easy openings for the youngest developers, and a review load shifted onto people who already know the system. The code-review piece is what that load feels like on a Tuesday.