Sixty-nine percent of Americans now want to take half of the AI industry and put it in a public fund.
That number comes from Verasight, a national survey of 1,690 adults fielded in June. Futurism covered it last week under the headline “A Majority of Americans Now Support Seizing Wealth From AI Industry,” and misspelled the pollster’s name in the second paragraph, which tells you roughly how carefully any of us are reading these numbers before we react to them. The details are worth more than the headline. When the same policy was explicitly attached to Bernie Sanders, support only fell to 64 percent, which means the idea is now more popular than its most famous advocate. Eighty-nine percent want AI companies to publish the results of their internal safety testing. Eighty-one percent want the federal government to have the authority to block a model it deems unsafe. And only 30 percent trust that same government over the labs to actually make that call.
Read those last two together. The public wants a referee that it also does not trust. That is not incoherence. That is what it looks like when people have decided that everyone involved is acting in bad faith and they just want somebody, anybody, to put a hand on the wheel.
I have been shipping production software on the Microsoft stack for almost thirty years, and for the last three of those I have been building agentic systems for a living. MCP servers, RAG pipelines, vector search, orchestration frameworks, observability for all of it. I use this technology every working day. I bill for it. It is not a hypothetical to me and it is not a religion to me. So I want to make an argument that I think most of the people cheering for that 69 percent and most of the people horrified by it are both missing.
The AI companies are not the AI.
Those are two different things and we have to stop letting them be discussed as one. One of them is a set of weights that already exists, already sits on disk, already runs on hardware you can buy, and cannot be recalled. The other one is a balance sheet. You can kill the balance sheet without touching the weights. In fact, you don’t even have to kill it. You just have to stop paying for it.
What they actually did
Let’s be honest about why 69 percent of the country wants to take their stock, because “people are scared of new technology” is the story the industry tells itself and it is a lie. People are not reacting to the technology. They are reacting to a specific list of things that specific companies did on purpose.
They invented a race. The single most effective piece of lobbying in the last decade has been the claim that the United States is in a neck-and-neck sprint with China toward artificial general intelligence, and that any regulation is a handicap in that sprint. There is a good academic literature arguing that this framing was largely manufactured. One survey of the doctrine puts it plainly: the narrative of a US-China AI race for global dominance began as, and in significant regards remains, a fiction, promoted almost exclusively in the West, and used to justify progress at all costs, including policies that benefit AI companies over everyone else. The authors note the obvious tell, which is that the people pushing the race hardest are the people who directly profit from the conclusion. Another analysis, published in a volume on AI governance, traces the race narrative from 2017 to 2025 and finds limited evidence for a concentrated Chinese state effort to win the AGI sprint, while Chinese officials were publicly calling for cooperation and safety. A race needs two runners. If only one guy is running and screaming that he’s in a race, he’s not racing. He’s fundraising.
They bought the referees. As of the end of June, the two largest AI super PACs had already dropped at least $44 million into forty House and Senate candidates, out of more than $200 million raised. Leading the Future, the big one, is funded by Andreessen Horowitz and OpenAI’s Greg Brockman, with Palantir’s Joe Lonsdale and Perplexity along for the ride. Public Citizen counts $294 million in combined spending from crypto, AI, Big Tech, and online betting in the 2026 midterms, part of a record-breaking $517 million in total corporate election spending. The Campaign Legal Center has filed an FEC complaint alleging two super PACs in the Leading the Future network laundered their spending through shell LLCs to hide where the money went. The goal of all this is not mysterious and it is not hidden. It is federal preemption of state AI law. They want one national rulebook, written with them in the room, replacing the thirty-eight states that passed over a hundred AI laws in 2025.
And they are winning that fight without even needing the PACs. In December, the President signed an executive order creating a DOJ AI Litigation Task Force whose sole job is to sue states over their AI laws, and conditioning $42 billion in broadband funding on states repealing regulations the administration finds onerous. Attorney General Bondi stood the task force up on January 10. Congress rejected a state AI moratorium twice. The executive branch did it anyway, by memo.
They socialized the power bill. Good Jobs First, which has been tracking corporate subsidies for twenty-seven years, calls data centers one of the most subsidized industries they have ever seen. Virginia’s exemption alone runs about $1.6 billion a year. Indiana handed Amazon an estimated $8.2 billion package. Meanwhile a guy in Manassas opened a $281 electric bill in January after paying about a hundred the month before. Maryland has filed a complaint at FERC saying its residents are eating roughly $2 billion in grid upgrade costs for data centers across the state line in Virginia. West Virginia ratepayers are looking at $440 million of the same. And here is the part that should end the argument: in a survey of data center operators, only about three percent ranked tax credits as the most important factor in choosing a site. States have been paying trillion-dollar companies billions of dollars to do something they were going to do anyway. Seventy-one percent of Americans now oppose a data center in their community, which is worse polling than nuclear power plants get.
They pirated the training data. I have to name my own house here, because I make my living on Claude and I will not pretend otherwise. Anthropic downloaded more than seven million books from LibGen and Pirate Library Mirror and settled with authors for $1.5 billion, roughly $3,000 per title across about 500,000 works. It is the largest copyright settlement in American history. Judge Alsup’s ruling drew the line exactly where it belonged: training on books you bought is transformative fair use, and downloading them off a pirate site is just theft with extra steps. Meta’s internal documents in the Kadrey case tell the same story in uglier language, one employee writing that LibGen was essential to hit state-of-the-art numbers. These companies built a business on other people’s work and only started writing checks when a judge made them.
They fired people for a capability that didn’t exist yet. This is the one that actually earns the 69 percent. Gartner surveyed 350 executives at companies with over a billion in revenue and found that roughly eighty percent admitted to cutting staff in order to fund AI and automation investments. Read that again, because the causality is backwards from the story we were told. They didn’t lay people off because AI replaced them. They laid people off to pay for the AI. And then it didn’t work. Orgvue found 32 percent of organizations that made AI-linked layoffs had to rehire. A Careerminds survey of 600 HR professionals found two in three employers who cut jobs for AI were already rehiring, often within months. Robert Half puts it at 29 percent. Forrester estimates 55 percent of employers regret the cuts. Gartner projects half of the companies that cut customer service for AI will be rehiring for the same roles by 2027, under new titles. IBM automated 94 percent of routine HR requests, discovered the remaining six percent were the ones that mattered, and announced it would triple entry-level hiring. Ford is rehiring engineers to fix quality problems the automation created. Klarna quietly hired its support reps back. McDonald’s turned the drive-thru AI off after it put hundreds of dollars of chicken nuggets on somebody’s order.
So a hundred thousand people lost their jobs this year to a technology that, by the industry’s own consultants’ numbers, could not do those jobs. That is not disruption. That is a management fad with a body count. And the same Gartner survey found the companies actually seeing gains were the ones using AI for what it called people amplification, meaning they gave the tools to their employees instead of firing them. The thing works. The layoffs were never about whether it works.
The race is fake, and here is the receipt
Now watch what happens when you follow the China argument all the way down.
The labs say we must not be regulated, because China. Fine. What is China actually doing? China is giving the models away. GLM-5.2 from Z.ai, DeepSeek V4, Qwen 3.6, Kimi K2.7 from Moonshot, MiniMax M3. MIT and Apache licenses. Open weights. Download them, run them on your own metal, fine-tune them, ship them, never call home. Chinese labs hold four of the top five slots in open-weight AI. GLM-5.2 scores 51 on the Artificial Analysis Intelligence Index, fifth overall, ahead of models people pay a lot of money for. The open-weight scores on SWE-bench-style suites now sit within single digits of the Western frontier. Zhipu trained a frontier model entirely on Huawei Ascend chips without a single Nvidia part, which is what happens when you put an engineer under a constraint.
And American companies are buying it. DoorDash, Airbnb, and Siemens are adopting Chinese models, according to the Financial Times. On OpenRouter, DeepSeek and Z.ai have overtaken Claude and ChatGPT in usage. DoorDash’s cofounder said publicly that they’re routing lower-level work to Moonshot and saving real money. Lindy dumped Anthropic entirely for DeepSeek V4. One organization reportedly burned half a billion dollars on Claude in a single month. Firms deepest into AI are spending about $7,500 per employee per month. Of course they’re switching. “Enterprises are starting to realize, hey, we don’t need the best model, we can use the faster, cheaper models,” as Featherless AI’s CEO put it.
So the actual competitive situation is this. The Chinese labs commoditized the thing the American labs are trying to sell as a subscription. That is not a race. That is a price war, and our side is losing it while asking Washington for protection. The China argument was never about national security. It was a moat. And the moat has been drained by the very people it was supposed to keep out.
It gets better. Cohere’s CEO told the FT that the Mythos export restrictions were “the most tangible event” pushing customers away, because it exposed the risk of depending on a single American vendor who can have his product yanked by the Commerce Department on a Tuesday. We built the case for buying Chinese ourselves.
The burn
Here is where I have to be more careful than the people I’m arguing with, because the sloppy version of this argument is wrong and the precise version is devastating.
OpenAI burned $3.7 billion in the first quarter of 2026 on $5.7 billion of revenue. That’s about 65 cents of every dollar, and the Q1 operating margin was negative 122 percent. Internal projections put 2026 losses around $14 billion, some estimates say $17 billion, on cumulative cash burn of roughly $115 billion through 2029. The company has signed something north of $1.4 trillion in multi-year compute commitments against a $25 billion run rate. HSBC thinks it needs over $200 billion more capital through 2030 and won’t be profitable by then anyway. Gross margin fell to 33 percent in 2025 against an internal target of 46. Inference costs quadrupled in a year. Sacra estimates $27 billion of burn this year and $63 billion next.
Now the honest part. That is not the whole industry. Anthropic hit a $47 billion annualized run rate in May, closed a $65 billion round at a $965 billion valuation, and expects to be profitable in the current quarter if it hits its revenue target. Its burn is projected to fall to roughly a third of revenue this year and around nine percent next year. Same sector, completely different animal. One of these companies is a business. The other one is a religion with a data center.
I am not saying this because Anthropic pays my bills, indirectly. I’m saying it because it is the entire point. “The AI industry” is not a thing you can nationalize, because it is not one thing. It’s a handful of companies with wildly different unit economics, and the burn is concentrated in precisely the companies doing the loudest lobbying. Greg Brockman funds Leading the Future. Sam Altman’s CFO floated a federal backstop. The correlation between “burning the most cash” and “screaming the loudest about China” is not a coincidence. It’s the business model.
And under all of it is the circular financing. Nvidia invests in OpenAI, OpenAI buys Nvidia chips. Microsoft invests, OpenAI spends it on Azure, and OpenAI is something like 45 percent of Microsoft’s cloud backlog. Amazon puts billions into Anthropic, Anthropic commits over $100 billion to AWS. Analysts have mapped more than $800 billion of these arrangements. AMD handed OpenAI and Meta warrants for 160 million shares at a penny each in exchange for purchase commitments. Defenders call it vendor financing and point out that it built the railroads. Critics call it Lucent and Nortel, who lent money to their own customers, booked the loans as revenue, and took the sector down with them when the customers folded. Jensen Huang said in March that his $30 billion in OpenAI “might be the last time,” and that the $100 billion is “not in the cards.” When the guy selling the shovels starts hedging, pay attention.
Why the sovereign wealth fund is the worst available answer
Sanders’ American AI Sovereign Wealth Fund Act would impose a one-time 50 percent tax, payable in stock, on any AI company with more than $200 million in annual gross receipts. The shares go into a Treasury trust run by a seven-member Independent Commission for Democratic AI, nominated by the President and confirmed by the Senate. He estimates $7 trillion, with a five percent annual draw funding roughly $1,000 checks to every man, woman, and child in America.
I want to take this seriously, because the moral premise is not stupid. The models were trained on the collective output of humanity, a fair amount of it pirated, and the buildout is running on public infrastructure and public subsidies. If you take the commons and turn it into a private toll road, the public has a claim. That argument is sound. Sanders is not wrong about the diagnosis.
He is catastrophically wrong about the treatment, and here is why.
If the government owns half of OpenAI, the government owns half of a company that is going to lose $14 billion this year and needs $200 billion more to survive. Congratulations, taxpayer. You are now the largest shareholder in the burn. And the second that fund is on the books, every incentive in Washington inverts. The Treasury now has a fiduciary interest in the share price. The preemption fight the industry has been spending hundreds of millions to win becomes a federal budget item. Every safety rule is now a hit to the sovereign wealth fund’s NAV. Every state law is now an attack on your neighbor’s thousand-dollar check. The Guardian made this argument in June and it is the correct one: hand the government a giant equity stake and you have given AI companies more influence over the government, not less. You have not seized the industry. You have merged with it.
And notice what it actually delivers. Sarah Friar sat on a Wall Street Journal stage in November and mused that a government “backstop, the guarantee that allows the financing to happen” would drop their cost of capital and raise their loan-to-value. The reaction was so violent she walked it back within hours and Altman posted that OpenAI does not want government guarantees and that “taxpayers should not bail out companies that make bad business decisions or otherwise lose in the market.”
The sovereign wealth fund is the backstop. It’s the same policy with a red rosette on it. The industry asked for taxpayers to co-sign the loan and got laughed out of the room, and now the left is proposing to co-sign the loan and call it socialism. Friar should send Bernie a thank-you note.
David Sacks accidentally wrote the correct policy
When Friar floated the backstop, the White House AI czar posted eleven words that are, unintentionally, the best AI policy anyone has proposed:
“There will be no federal bailout for AI. The U.S. has at least 5 major frontier model companies. If one fails, others will take its place.”
He’s right. And he didn’t follow his own logic one inch further than he had to, so I’ll do it for him.
If a frontier lab can fail without taking the technology with it, then the failure is not a crisis. It’s a market clearing. And if that’s true, then every single thing the industry is asking for — the preemption, the subsidies, the export protection, the manufactured urgency — is unnecessary by the administration’s own stated reasoning. You cannot simultaneously argue that the sector is so resilient that any given company’s death is a non-event, and so fragile that a Colorado disclosure law will hand the future to Beijing. Pick one.
So here is my proposal, and it is going to be unsatisfying to everyone, which is usually how I know I’m onto something.
Do nothing. Specifically, stop doing things for them.
Don’t seize anything. Don’t nationalize anything. Don’t create a commission. Just withdraw the public subsidy from a private bet and let arithmetic finish the job. Let the data center tax exemptions expire, the way Illinois, Arizona, Ohio, New Jersey, and North Carolina are already doing, and the way Virginia did when it put a penny-per-kilowatt-hour tax on data center power. Make them pay for their own grid, with legal force and not a nonbinding White House pledge that Harvard’s electricity law director called “a show designed to sweep this issue under the rug.” Let the states regulate, because they’re the only level of government currently doing it and Congress has declined twice. Enforce copyright, which is already producing licensing markets. And when a company that is burning 65 cents on the dollar with no path to profitability can’t raise its next round, let it not raise its next round.
That’s it. That’s the policy. It requires no new law, no new agency, and no constitutional fight. It requires the government to stop being the marginal buyer, the co-signer, the litigator, and the hype man for a private company’s balance sheet.
What survives
Here’s the part the doomers and the boosters both get wrong, and it’s the part I actually know something about, because it’s my job.
If OpenAI evaporated tonight, my systems would still run tomorrow.
I’d change a base URL. Maybe a day of work. The MCP servers I’ve built don’t care what’s behind the endpoint — that’s the entire point of the protocol. The RAG pipelines don’t care. The vector indexes in SQL Server don’t care. The orchestration doesn’t care. The weights for GLM-5.2 and DeepSeek V4 and Qwen and Kimi are already downloaded, already mirrored a thousand times over, already permissively licensed. You cannot uninvent a file. You cannot repossess a tensor. The transformer paper is nine years old and it’s free.
This is what everyone screaming about the AI bubble refuses to understand, and what everyone defending the burn is desperate for you not to notice: the technology and the capital structure are separable. When the dot-com bubble burst, the fiber stayed in the ground. Pets.com died. TCP/IP did not. The people who lost money were investors, which is the job description. The web kept working. It got better, actually, because the noise cleared out and the survivors had to build something people wanted to pay for.
The same thing is true here and it’s true right now, today, without any legislation. A small business running Qwen on a single GPU to answer their email is not participating in the burn. My clients running agentic pipelines on open weights are not raising anyone’s electric bill. The productivity gain is real — Gartner found it, in the companies that used the tools to amplify their people instead of replacing them. That gain does not require a $965 billion valuation or a trillion dollars of compute commitments or a manufactured war with China or a single dollar of Virginia’s tax base. It requires a model, a machine, and somebody who knows what they’re doing.
The valuations and the layoffs and the lobbying and the ratepayer theft are not the price of the technology. They’re the price of a handful of men trying to own it. Those are separable too.
So let them run out of money. Not because I want the technology to fail — I’ve bet my career on it and I’d bet it again. Because it’s the only outcome that punishes the balance sheet without touching the weights. The models will still be there. The GPUs will get sold at auction to people with actual businesses. The engineers will go build things that have customers. And the 69 percent who want to seize half the industry will discover they didn’t need to seize anything, because the thing they were angry at was never the technology. It was the bill.
Stop paying it.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment. He has spent nearly thirty years shipping production software on the Microsoft stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy, and Allegiant Air. He was employee number six at Break.com and a Los Angeles Business Journal CTO of the Year finalist.
His current work is production agentic AI: MCP servers, the Microsoft Agent Framework, RAG pipelines, SQL Server 2025 vector search, and the PLogger observability framework. He is the .NET AI guy that ships, which means he has an unusually low tolerance for people who talk about this technology without ever having deployed it, and an unusually high tolerance for the technology itself.
He writes about the AI industry from inside it. He is also the author of the novel The Archive of Lost Suns, trains Brazilian jiu-jitsu under Sergio Penha in Las Vegas, rides motorcycles, and co-hosts Edge Grip Podcast.