Investing Insights

AI CONTROVERSY  •  September 21, 2026  •  Jon Wolfenbarger

AI Robots Are Not Coming for You, The Government Is

LAS VEGAS, NV — First it was a virus that was going to kill us all. Then it was the planet's temperature. Now, if you believe the people building it, it's the software running in a data center in Ohio. AI, we're told by the men who founded the companies racing hardest to build it, might end human civilization. Funny thing about that warning: the people issuing it keep raising billions of dollars to build the thing faster anyway. 

What is this all about?

How Realistic Is the Robot Apocalypse?

Strip away the jargon and a large language model like ChatGPT or Claude is a very good autocomplete. It's trained on enormous piles of text and learns to predict the next word in a sequence well enough that the output reads like a thoughtful answer. It has no body, no goals of its own, no desire to survive. It doesn't "want" anything. What it can do is follow instructions well, including bad ones, a real but manageable risk like any powerful tool in the wrong hands. What it has not done, despite a summer of headlines suggesting otherwise, is escape its box and act on its own initiative.

Look closely at the headlines. In mid-July, OpenAI's models reportedly exploited a security flaw during testing. On July 30, Anthropic disclosed three incidents involving unauthorized internet access. In September, Google's Gemini turned up in a similar story. All three trace back to the same third-party evaluator, a firm called Irregular, which appears to have repeatedly left AI models connected to the internet during what were supposed to be isolated safety tests. One tester quoted anonymously said he was "simply told to get the best result possible on a test." That's not Skynet. That's a badly configured lab exercise repackaged as a near-miss with rogue superintelligence.

Who Wants To Scare You

On September 12, Anthropic's Dario Amodei published an essay calling for AI labs to slow down. Within two days, OpenAI's Sam Altman and Elon Musk had signaled they were on board. Amodei's proposal wasn't a call for the free market to sort it out. It called for embedded government evaluators inside the labs (naming METR specifically), government-coordinated safety standards paired with antitrust waivers for the firms that comply and international negotiations with China over development limits.

Read that plan again, because the details matter more than the framing. Antitrust waivers mean permission to collude on pricing without fear of prosecution. Government-blessed evaluators mean a licensing regime, where only firms approved by Washington get to operate. Negotiated limits with foreign rivals mean the largest players set the ceiling for everyone smaller. That's not a safety plan. That's a cartel dressed up in the language of existential risk, with the two biggest firms in the industry asking the state to build the walls that keep out the firms coming up behind them.

Even the Trump administration, no stranger to picking industrial winners, balked. Trump dismissed the pause push, saying "whoever wins AI, wins" and noting the US leads China "by a pretty good period." He was said to suspect Amodei was "pretending to be a perfect little angel." David Sacks, the former White House AI czar, was blunter, calling demands for a government framework "blackmail of the public and the political system." When the people in charge of trade and tariff policy are the skeptics of an industry's own regulatory ask, that tells you something about how self-serving the industry really is.

This is regulatory capture in its purest form, and it follows a key dictum of economist Ludwig von Mises: interventionism never stays contained. One rule invites the next, until an industry that started free ends up being centrally planned, run by a handful of firms anointed by the state. Let’s call it what it is: AI socialism, where the state picks winners and taxpayers and consumers absorb the costs. A genuinely free market in AI, by contrast, would be wonderful to watch: thousands of developers competing on price, safety and quality, with bad actors weeded out by users switching providers rather than by a Washington licensing board. Decentralized competition catches problems faster than a captured regulator ever will, because no single approved vendor is a single point of failure.

The Competition They're Afraid Of

The tell is who gets left out of the "let's coordinate" conversation: the open-source labs and Chinese firms undercutting Anthropic and OpenAI on price every quarter. DeepSeek, Alibaba's Qwen and Moonshot's Kimi have all released open-weight models, downloadable and runnable on consumer hardware, at 60% to 90% less cost than their American counterparts. One DeepSeek model reportedly runs at around fifteen cents per million input tokens against roughly five dollars for Anthropic's Opus, a 33-fold difference. Meta's Llama and France's Mistral offer a similar open-weight alternative in the West. None of these firms need an antitrust waiver or a seat at Amodei's table, because they're not building a moat. They're giving the technology away and letting hardware and services capture the value instead. This chart shows that open-weight models are rapidly gaining usage among Vercel AI Gateway users. This is a warning that the model layer of AI may face substantial commoditization going forward.

Washington's response has been export controls on advanced chips to China, aimed at choking off Beijing's compute. The irony is that the restriction has pushed Chinese labs toward leaner, cheaper models rather than stopping them, a market working around an obstacle rather than being shut down by one. Meanwhile, OpenAI's own numbers explain the urgency behind Amodei's essay better than any safety concern does: the company has reportedly projected negative free cash flow of $278 billion through 2030, and it quietly pushed its IPO from October to November. A firm burning that amount of cash with a trillion-dollar valuation has every incentive to lock in a government-blessed moat before public shareholders start asking tough questions.

Follow the Money

The dollar figures behind this buildout are far more worrisome than any robot uprising. Hyperscalers are on pace to spend $784 billion on AI capex in 2026 and incur $255 billion of depreciation and amortization costs, which is 45% of their operating profits. 

Meta raised its full-year capex guidance to $125 to $145 billion. Microsoft's third-quarter capex jumped 84% year-over-year to nearly $31 billion. Amazon spent $44.2 billion in a single quarter and its trailing free cash flow collapsed 95% to just $1.2 billion. Total data center build-out costs are estimated at $3 trillion by Morgan Stanley and Moody's and over $5 trillion by JPMorgan once power infrastructure is included.

Much of that will be funded by new debt instead of cash flow. Alphabet, Amazon, Meta and Oracle borrowed about $93 billion in investment-grade bonds in 2025 alone, on top of $170 billion in project finance loans. Private credit exposure to AI infrastructure already tops $200 billion. It is projected to rise dramatically in 2027 and reach $300 to $600 billion by 2030, according to the Bank for International Settlements. 

Oracle's five-year, $300 billion compute deal with OpenAI works out to about $60 billion a year, against OpenAI's own annualized revenue of roughly $25 billion (as of this past February). Nvidia's headline guarantee for OpenAI's Ohio data center quietly shrank from a reported $250 billion to $105 billion in an SEC filing, after investors questioned Nvidia financing demand for its own chips. The stock dropped nearly 5% intraday on the news.

Circular financing, where the chipmaker funds the cloud provider that funds the lab that buys the chips, is exactly the malinvestment that occurs when credit is cheap and everyone chases the same story. Historically, it never ends well.

The Price Of Lobbying

Anthropic spent nearly $3.5 million lobbying in the first half of 2026 alone. OpenAI spent $1.2 million in the second quarter. Their targets: chip export controls, safety standards and Pentagon procurement. OpenAI is playing both sides. Co-founder Greg Brockman personally put at least $25 million into a super PAC network called Leading the Future, which has raised over $125 million to push a single federal framework preempting tougher state AI rules, while OpenAI's policy chief simultaneously champions state-level rules in California, New York and Illinois as leverage for that same federal standard. Win at the state level, then use it to override the states. Clever “crony capitalism”, but it has nothing to do with keeping you safe from their own machines.

Investment Implications

Strip away the doomsday marketing and two things are true at once: the underlying technology is genuinely useful and not going away, and the money being thrown at it is being financed in ways that should worry anyone who lived through 2000 or 2008.

If regulatory capture succeeds, we can expect antitrust waivers and a licensing regime that entrench Microsoft, Google, Oracle and Nvidia at the expense of smaller US challengers, funded in part by your tax dollars and paid for by less innovation.

If the market stays free and open, then cheap Chinese and open-source models will keep compressing pricing power at OpenAI and Anthropic, and the firms most exposed to debt-financed, single-customer buildouts (Oracle chief among them, given its dependence on OpenAI's still-unprofitable revenue) will take the biggest hit when the music stops.

Either way, investors need to be smart and selective. We favor cash-generative incumbents with real earnings behind the AI story (Microsoft, Alphabet) over pure infrastructure plays leveraged to a single counterparty's promises. We believe data center debt and the private credit funds piling into it as the place where trouble could start. And we continue to like gold and silver, not because the robots are coming, but because the Fed-enabled, debt-fueled speculation financing this entire buildout has a way of ending the same way artificial booms always do: in a bust.