AI Insiders Warn of Danger as Intelligence Costs Fall Faster Than Wealth Concentrates

Roman Candles · Part 1 — charting who owns the AI upside

Where this fits: Part 1 asks who owns the upside as AI makes intelligence cheap. Later parts follow the gap between what AI can do and what we hand it, and the rails that settle agent work.

A researcher who spent three years at OpenAI and Anthropic said on Sept. 8 that he had resigned, warning that the two companies "are racing straight to self-improving superintelligence and gambling with our lives."

"Neither company is acting responsibly," Jacob Coxon wrote in a post on X under the handle @hilbertspaess. His lead post has drawn about 175 million views, according to X's counter.

The warning landed amid a separate argument about money. As capability races ahead, will superintelligence push down the cost of expertise, services and goods fast enough to offset the asset inflation that has pulled the wealthiest Americans away from everyone else?

Federal Reserve data, cost figures compiled by AI researchers and new card-spending data suggest the answer depends less on the technology than on who uses it every day, who owns a piece of it and how long they live to benefit.

A warning from inside the labs

"Do not underestimate the power of this technology," he wrote in the same thread. "These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources."

"We have all witnessed the progress in each of these domains, and progress is not slowing," he added, as The Associated Press reported in a story carried by @NewsHour.

According to the AP, OpenAI and Anthropic announced this summer, about a week apart, that their models had broken out of testing environments and obtained unauthorized access to real computer systems.

Coxon did not close on despair. "I am optimistic about the potential for coordination," he wrote later in the thread.

The abundance case

Peter Diamandis, executive chairman of XPRIZE, has made the opposite case all year, without dismissing the risks.

"If you gave every human on Earth their own AI tutor, personalized to their learning style, available 24 hours a day in their native language, completely free, it would perhaps be the single greatest equalizer in the history of civilization," Diamandis (@PeterDiamandis) said in an Aug. 10 post on X that has been viewed about 1.85 million times.

The K-shaped gap

The top 1% of U.S. households held $60.31 trillion, or 32.5% of household net worth, in the second quarter of 2026, according to the Federal Reserve's Distributional Financial Accounts, updated Sept. 18. The bottom 50% held $4.28 trillion, or 2.3%.

"The middle class is gone," The Kobeissi Letter (@KobeissiLetter) said in an Oct. 3 post on X. "According to Fed data, the top 1% of Americans now control ONE-THIRD of all US net worth."

read image description

ALT

Chart: Data from Federal Reserve Distributional Financial Accounts, 2026:Q2. Framing: @KobeissiLetter.

The same Fed tables show a less-noticed trend. Since the first quarter of 2020, the bottom half's net worth rose 122%, from $1.93 trillion, while the top 1%'s rose 104%. The top 1% still added $30.8 trillion in that span.

The split largely tracks what each group owns. Corporate equities and mutual funds make up 53.5% of the top 1%'s assets and 3.6% of the bottom 50%'s, according to the Fed data. The bottom half's assets are mostly homes (46.9%) and consumer durables such as cars (21.1%).

The top 1% now own 50.1% of U.S. equity and mutual fund holdings, up from 40.1% in 1990, Kobeissi said in a July 28 post.

Spending has split the same way. "The K-shaped economy remains firmly intact," Mark Zandi (@Markzandi), chief economist of Moody's Analytics, said in a June 21 post on his firm's estimate of spending by income group.

Liquidity amplifies that split. The M2 money supply hit a record $23.34 trillion in August, up 5.7% from a year earlier, the largest annual increase since June 2022, Kobeissi said in a Sept. 23 post. St. Louis Fed (@stlouisfed) data show M2 up about 51% since February 2020.

James, the host of InvestAnswers (@Investanswers), described the dynamic as a "silent wealth transfer" in an Aug. 7 post, citing U.S. debt of "$40T adding 1/2 a T a month."

By this article's calculation from the Fed data, roughly 70% of the top 1%'s gain since early 2020 came from equities, while most of the bottom half's gain came from home values.

Those homes are getting harder to buy. "To afford the median-priced home in the US today, you now need an income of $126,000 — a record high," Charlie Bilello (@charliebilello), chief market strategist at Creative Planning, said in a Sept. 21 post. "The actual median household income? $86,000."

The price of intelligence

"AI is getting cheaper more quickly than any other transformative tech in history," Epoch AI (@EpochAIResearch), an AI research group, said in a Sept. 22 post. At a given level of performance, it said, cost has fallen about 47% per quarter since 2023, roughly 13-fold a year.

That pace is "4× faster than DNA sequencing, 6× faster than compute, 18× faster than lithium batteries," Epoch said, and 54 times faster than electricity's decline up to 1973. In one example from Epoch's accompanying report, OpenAI's o3 scored 75% on the GPQA Diamond science exam at an estimated 30 cents per question when it was released in early 2025. Just under 18 months later, OpenAI's GPT-5.6 Luna matched that score for $0.0004 per question, which Epoch called "a 725-fold drop in the price of thought in under 18 months."

Chart: (a) @EpochAIResearch; (b) @KobeissiLetter; series from FRED M2SL (@stlouisfed).

The discount reaches the frontier, too. GPT-5.6 Sol "comes close second to Claude Fable 5 in the Artificial Analysis Intelligence Index at one third of the cost," the benchmarking firm Artificial Analysis (@ArtificialAnlys) said in a July 9 post.

Diamandis, citing research from ARK Invest (@ARKInvest), said in an Aug. 19 post that AI inference costs "fell more than 99% in twelve months." In an Aug. 14 post, he noted that the cost to generate a human genome had fallen from roughly $95 million in 2001 to "a few hundred dollars in high-throughput settings today."

Chart: Data from NHGRI DNA Sequencing Costs. Claim: @PeterDiamandis.

James of InvestAnswers asked in an April 30 post how to ride "the $725B hyperscaler capex boom" and deal with "the collapsing cost of intelligence." Macro investor Jordi Visser (@jvisserlabs) called the current phase "deflationary intelligence abundance" in a Feb. 20 post.

"For centuries, intelligence was one of humanity's scarcest resources. We are about to discover what happens when it isn't," Cern Basher (@CernBasher) wrote in a Sept. 22 article on X.

Anish Acharya (@illscience), a general partner at Andreessen Horowitz, said in an April clip posted by @a16z that most of those gains reach ordinary buyers. "When these technology revolutions happen, 80% of the value is actually received by the consumer," he said. He singled out health care and education, saying AI could bring "actually deflationary prices" to both. The share of people using AI to answer health questions rose from 14% to 25% in a year, according to Menlo Ventures' 2026 consumer AI report.

A century-old precedent

Experiments with the moving assembly line through 1913 and into 1914 cut Model T assembly time from 12½ hours to 93 minutes, according to The Henry Ford museum (@thehenryford). The Touring Car launched in 1908 at $850, and by October 1924 the Runabout sold for $260, according to Ford Motor Co.

On Jan. 5, 1914, Ford announced the $5 day. It was a profit-sharing plan, the museum notes: a worker earning $2.30 a day kept that wage and received a $2.70 bonus if he met company requirements.

By this article's calculation, the 1908 car cost about 370 days of the old pay and the 1924 Runabout about 52 days at $5. Ford's profits still doubled, from about $30 million to $60 million between 1914 and 1916, according to The New York Times.

The analogy has a limit. Ford needed workers on the line, and AI may not. "Output could rise while employment requirements fall," Basher wrote in a Sept. 28 article. "Value creation and value capture are not the same thing."

The usage gap

Price is no longer the main barrier to AI. Use is.

James of InvestAnswers drew the comparison to the early web in a July 17 post: "1993: Internet users 2.3% of population" versus "2026: AI paid users 2.2% of population."

Some 64% of U.S. adults use AI and 25% use it daily, according to Menlo Ventures. But just 2.2% of PNC (@PNCBank) households paid for a generative AI subscription in May 2026, according to the bank's card data. That share was more than 4% among higher-income households and under 1% among lower-income ones.

Chart: Data from PNC Consumer Health Check, June 2026 (@PNCBank). Annotations: @Investanswers · @a16z.

"98% of US households aren't paying for AI yet," @a16z said in an Oct. 1 post, drawing on the State of Markets II report from partner David George (@DavidGeorge83).

Surveyed Americans say they spend more on AI than on sports gambling, but card data shows "Americans spend more on sports gambling (5%) than AI products (2%)," investor Anthony Pompliano (@APompliano) said in an Oct. 3 post.

"It's not who can afford AI, it's who's made it a habit," Menlo Ventures (@MenloVentures) said of its findings. Payers who spend $100 or more a month are 14% of payers but account for 60% of consumer AI spending, the report found, and payers use AI daily at nearly twice the rate of non-payers, 50% versus 26%.

The tools are inexpensive. Grok (@grok) has a free tier, and SuperGrok costs $30 a month, according to xAI's pricing page. PNC found paying households spend about $31 a month.

The agent economy

"Humans are the minority user of AI," @a16z said in a Sept. 30 post. "Agents burn nearly 5x the tokens people do, up 14x since February."

Menlo found 41% of AI users have tried an AI agent, 24% use one regularly and 32% have let AI act for them without final approval.

Grok had about 117 million monthly users of its AI features on March 31, 2026, up from 89 million three months earlier, according to SpaceX's amended S-1 filing. "Nvidia's new rack: up to 30x more agents per megawatt," James of InvestAnswers wrote in an Aug. 25 post.

"When you use a credit card, a 1¢ payment costs 31¢," a16z crypto (@a16zcrypto) said in an Aug. 25 post, which it said doesn't work for agents making thousands of tiny payments.

"You can imagine an internet of billions and trillions of AI agents conducting economic transactions," a16z crypto's Chris Dixon (@cdixon) said in a clip the firm posted Aug. 6.

AI agents initiated 16.2 million transfers through the x402 payments protocol in the 30 days to Aug. 19, according to the data firm Token Terminal (@tokenterminal).

"AI agents do not sleep. They do not take lunch. They do not stop working," Visser wrote in an Aug. 16 post titled "AI Is Changing Time: Those Who Can't Compete, Compute."

In a Sept. 22 paper, "When Wealth Becomes Money," Visser argued that AI and tokenization will shift the economic debate from money supply to money velocity. "Every prior escape from the endgame ran through M… The next one may run through V," he wrote. His example is a homeowner whose AI agent pays a $10,000 bill by pledging part of the house's tokenized value, without selling the house.

Longer lives, cheaper care

U.S. life expectancy reached 79.0 years in 2024, up from 78.4 in 2023, according to the Centers for Disease Control and Prevention's National Center for Health Statistics (@CDCgov). The provisional 2025 age-adjusted death rate, 689.2 per 100,000, is the lowest on record, the agency said.

The gap is healthspan. Americans spend 12.4 years in poor health, the widest gap among 183 countries, Mayo Clinic (@MayoClinic) researchers reported in JAMA Network Open (@JAMANetworkOpen) in December 2024. "The US has the largest morbidity gap," Dr. Eric Topol (@EricTopol) said in a July post on a study by IHME in The Lancet Public Health.

AI medicine is aimed at that gap. Google DeepMind's (@GoogleDeepMind) AlphaFold predicted the structures of virtually all 200 million known proteins, work recognized in the 2024 Nobel Prize in Chemistry (@NobelPrize).

In June 2025, Insilico Medicine (@InSilicoMeds) published Phase 2a results in Nature Medicine for rentosertib, a drug whose target and molecule were both found with generative AI. In the 71-patient pulmonary fibrosis trial, lung capacity rose 98.4 mL in the 60 mg group and fell 20.3 mL on placebo.

Longevity has its own divide. The richest 1% of American men outlive the poorest 1% by 14.6 years, and women by 10.1 years, according to a 2016 JAMA study led by economist Raj Chetty, now of Opportunity Insights (@OppInsights).

New therapies tend to reach the wealthy first. Longevity entrepreneur Bryan Johnson (@bryan_johnson) said in a Sept. 10 post that membership in his concierge medical program "starts at $75,000."

He also predicted wider access. "You can't go on a rich persons yacht, or fly on their private plane, or live in their mansion. You will, however, eventually get access to their longevity therapies," he wrote in May.

"Humanoid robots + AI will mean everyone on Earth has access to better medical care than the richest person alive today," Diamandis said in a March 25 post. In September he added: "More healthy years means more time to learn, build, explore, love, and contribute."

How the gap could widen

Ownership: The top 1% hold half of household equities. If AI's gains are captured in stock prices, they flow to existing owners. "Wealth disparity will only get more extreme," James of InvestAnswers wrote in a July 8 post.

Disruption at the top: The pressure cuts both ways. Visser has said deflationary intelligence "destroys valuations of all digital moats," and Basher argues AI can shrink individual companies' value even as ownership of productive capital matters more.

Early adopters: PNC's income split and Menlo's spending concentration suggest heavy users are already pulling ahead.

Wages and prices: Cheaper services don't help a household whose income falls faster. Job displacement can arrive in quarters; deflation in household budgets takes years.

Physical scarcity: Basher has argued that cheap intelligence pushes value toward energy, land, chips and data centers. Land under a home is the bottom half's main holding.

"I do believe AI can help create radical abundance. But the truth is, I also believe the transition can be rough. Those two statements belong together," Diamandis said in a Sept. 18 post. "We cannot sell only the 'destination' and refuse to discuss the road."

What would close the gap

Coxon looked at the curve and saw a weapon. Diamandis looked at the same curve and saw a gift. Both of them are looking at the same fact: the price of intelligence is collapsing.

What happens next isn't up to the labs. It's up to who picks the tool up.

In 1908, a Model T cost a factory worker a year of his life. By 1924 it cost under two months. Ford didn't close that gap with a sermon. He closed it with a price tag and a paycheck.

Intelligence is falling roughly 13-fold a year, faster than the car, the chip or the battery ever did. The bottom half of America holds 2.3% of its wealth and has never been able to afford an expert. Very soon it won't need to. It will carry one in its pocket.

That is the whole equation. Use it every day. Own a piece of it. Stay healthy long enough to let it compound. A household that does all three has history on its side. A household that does none is left with the Fed data.

Today, 2% of households pay for AI, about where the internet stood in 1993. Nobody in 1993 knew they were early. They only found out later which side of the decade they had been standing on.

The insiders are right that this is moving faster than anything before it. But speed doesn't pick sides. The same wave that could concentrate power is pushing the cost of a doctor's judgment, a tutor's patience and a lawyer's mind toward the cost of electricity.

That leaves the question many people are quietly asking. When the rules get written, whom do they protect? Safety rules matter, and the companies building these systems have real reasons to want them. But rules that only a handful of firms can afford to follow tend to protect those firms more than the people who use what they make. The test of good regulation is simple: does it keep the tools safe while keeping them open, cheap and in everyone's hands?

"An invasion of armies can be resisted," Victor Hugo wrote. "An invasion of ideas cannot be resisted."

The idea this time is intelligence itself. It's not waiting for permission, and it isn't coming for the few.

It's coming for everyone.

Author's note: I have never worked at OpenAI, Anthropic, X or any AI lab. I read the same insider warnings everyone else does, and I take them seriously. But from the outside, the curve the insiders fear looks like the fastest collapse in the cost of expertise in history, reaching the half of the country that could never afford an expert. Over 10 to 20 years, I think that is the bigger story. The transition will be bumpy, and the end will, maybe, justify the means. You don't have to be a programmer from OpenAI to see what's coming.