The Rogue Swarm and the Federal Backstop

How Safety Warnings Could Become a Backstop for Big Tech

Construction underway on a Project Stargate AI infrastructure site in Abilene, Texas, in April 2025. Photo credit: Daniel Cole

There is a particular kind of tweet that only the CEO of a frontier lab can send at midnight on a Saturday. Dario Amodei’s was one of them: a warning that within six to twelve months a “rogue swarm” could seize control of the internet and inflict hundreds of billions in damage. The message was delivered with the grave, unsentimental cadence of a man who “has seen things that you people wouldn’t believe.” Within minutes, Sam Altman and Elon Musk had joined the chorus—they had also “seen things.” By Sunday morning, the alarm was general.

Consider the timing and the cycle. Within the same week, an ex‑OpenAI, ex‑Anthropic researcher, Jacob Coxon, posted that both companies were ‘racing straight to self-improving superintelligence and gambling with our lives.’ The alarmist thread drew more than 120 million views in its first day and prompted more than 20 politicians to respond with calls for AI regulation. Around the same time, Anthropic’s own economists released a scenarios report which featured roughly 15% annual GDP growth, 11.9% economy‑wide mass unemployment, and labour’s share of income falling from about 60% to 45.2%—as though the bleak forecast of technological unemployment required a formal scenarios report to be taken seriously. And within days of Amodei’s warning, Semafor reported that a previously stalled bipartisan AI safety bill had suddenly emerged as the most viable legislative option before 2027. Meanwhile, Bernie Sanders, who had announced a superintelligence‑ban bill on 3 September, was preparing to introduce it formally and convene expert briefings on AI’s “extraordinary dangers.”

Let’s be serious and ask the only question that truly matters here: what does this sudden AI-panic wave actually produce? One thing it certainly does not and will not produce, is regulation in any binding sense. Instead, it creates the terrain or precondition for something entirely different: a federal backstop for AI capital expenditure (CapEx) that OpenAI and other labs have sought from the beginning. In this framework, Amodei’s swarm is the alibi. It gives Altman the perfect reason to postpone the trillion-dollar IPO—“an ill‑advised moment to go public’, he commented—while handing Congress the emotional warrant to authorise emergency federal funding to secure AI (and, it seems, the world) against catastrophe. The existential threat is real enough to justify the money and vague enough to justify anything else.

This is the Machiavellian techno-financial machinery in its purest form. The point to always keep in mind, amid all the noise, is that the AI buildout runs on mountains of debt: capital investments by tech giants on data centres and AI infrastructure will exceed a trillion dollars in 2027; off-balance-sheet commitments are at $3.1 trillion with $1.3 trillion added in a single quarter; AI bond issuance has hit $266 billion this year and a further $400 billion are expected next year. Several of the largest hyperscalers are now free‑cash‑flow negative, as AI data‑centre and chip spending outstrips operating cash generation and pushes more financing off their balance sheets. In short, private balance sheets can no longer absorb the debt bingeing on their own. So, the conditions for the next round of credit creation must be conjured politically: first the panic, then the guarantee, then the issuance. The safety discourse is the lubricant—the initial ideological form taken by this issuance. The sequence is straightforward: private credit finances the buildout; mounting fragility demands a public guarantee (through procurement, energy subsidies, liability protection, debt guarantees, or emergency appropriations), and the guarantee requires an emergency vocabulary. AI safety supplies that vocabulary.

The Rehearsal

As many will remember, in April 2026, Anthropic announced its Mythos model had “found thousands of high-severity vulnerabilities, including some in every major operating system and web browser.” The White House restricted access to roughly 40 organisations, including Amazon, Microsoft, JPMorgan. The rest of the world’s banks, hospitals and governments were left waiting outside that circle. They had to confront the implications indirectly: through regulatory warnings, emergency briefings, and a rapidly expanding language of systemic risk.

The IMF called it a potential “macro-financial shock,” warning that ‘correlated failures’ could “disrupt financial intermediation, payments and confidence at the systemic level.” The Bank of England’s governor, Andrew Bailey, requested that Anthropic brief the Financial Stability Board—finance ministry officials and central bankers from the G20. Meanwhile, the ECB gave 110 banks until 31 October to submit “comprehensive action plans” against AI-enabled cyber threats, raising its systemic risk assessment to “severe.” And Klaus Schwab, of course, had already said it plainly back in 2020: compared to a major cyberattack, “the COVID-19 crisis would be seen as a small disturbance.”

It is hard not to conclude that the rehearsal is now done and the script is largely written. After all, SIFMA’s Quantum Dawn exercises have simulated large-scale cyber incidents on the US financial system since 2011, rehearsing exactly the kind of systemic crisis that would justify emergency authority. If an AI-driven financial emergency ever materialises—whether real, allowed, or manufactured—the institutional response architecture is already pre-positioned, and the pattern of response has been trained. The public won’t ask whether it is real. More likely, they’ll beg Anthropic to save them.

Oracle and the Tell

If you want to see what the market believes, ignore the commentary and watch the credit. On 9 September—the same 24-hour window in which the safety panic crested—Oracle reported results celebrated as a triumph. The stock jumped 8% in after-hours trading. The triumph consisted of one thing: more capital expenditure and more debt. The celebration was about the borrowing itself, as though the size of the bet were the proof of its wisdom.

Then look at what the credit market did. Oracle’s credit default swaps—the price of insuring against its default—had already climbed to record highs in the weeks before earnings and remained elevated even as the stock jumped. Within twenty-four hours the equity had given back the entire gain and more. That divergence is the whole story in miniature. While the equity market is trading narrative, the credit market is pricing risk. When the two disagree this violently about the same company over the same 24‑hour window, someone is mispricing risk—and it is rarely those who buy insurance. The AI buildout is ultimately financed by debt, leases, private credit, and the expectation that future cash flows will validate present borrowing. When the equity story and the credit price diverge, the key question is no longer whether investors believe in AI. Rather, it is who will absorb the loss when that belief stops producing cash.

This is why the safety pantomime—as if reckless AI billionaires had suddenly developed a conscience—is functional. The issue is not that the executives are insincere or that the danger is unreal. It is that the language of safety converts private exposure into public obligation. Private balance sheets are approaching the limit of what they can absorb. When that limit is reached, the state is drafted in—not to bail out an entire failing sector, which now would be politically impossible, but to secure it as strategic, which is patriotic. What else is the AI Safety Bill? It converts a solvency problem into a civilisational emergency, and an emergency into a federal appropriation. The whole point of it all is the guarantee on CapEx, while the “rogue swarm” is the packaging.

The Only Way Out Is... Through

The logic is almost banal in its brutality. A system built on credit cannot survive prolonged deleveraging; it requires continuous credit creation to keep the machinery running. When private demand for that credit weakens—when free cash flow turns negative, CDS spike, and the equity pop lasts less than a day—the conditions for the next round must be manufactured elsewhere. Sometimes through monetary policy. Increasingly, through an external narrative that justifies fiscal policy: an existential threat that only state money can credibly backstop. The September AI panic is nothing but the credit machine’s current operating procedure.

The backstop, then, is the first move: socialise the risk while private balance sheets can still absorb it. The rehearsal is the second move: prepare the emergency architecture for the moment they can’t. Both are forms of the same conjuring, aimed at the same end: more credit, politically guaranteed.

This is why the timing of Amodei’s tweet is important, and why the Coxon account, the sudden thawing of a bipartisan bill, the 9/11‑style rhetoric of existential threat, and the convenient postponement of OpenAI’s IPO all belong to the same picture. A picture in which the incentives converge: the labs need protection from the consequences of their leverage; politicians need an emergency through which authority and spending can be expanded; the media need a catastrophe legible as a story; and the credit system needs the next tranche. The panic writes itself, because everyone benefits from it.

And then there is Trump’s response. On Monday 14 September, true to his style, he posted on Truth Social that Amodei was “pretending to be a ‘perfect little angel’.” that his administration had already ‘stopped AI “people” from doing bad, or potentially bad, “things”’ (Amodei included) and that ‘the only control or “guardrails” that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT’. He called the safety debate a ‘SICK conspiracy’ benefiting only China and then declared, ‘WHOEVER WINS AI, WINS!’

If Trump is right that Amodei’s essay is a competitive moat play, he is wrong about the conclusion. The correct inference is not that regulation is unnecessary; it is that the insincerity is structural, operating on both sides of the aisle—and on both sides of the panic. What Trump’s post offers is not an alternative to the backstop but an alternative form of it: the replacement of legislation with executive discretion. No new rules, no Safety Bill, no congressional oversight—just the guarantee that whoever holds the pen will keep the credit machine running. The backstop is the President’s word, and the market’s confidence that the word will hold.

Which leaves the question I want to close on, via Bridgewater’s Greg Jensen: AI today is “where Covid was in February 2020”: “Until the AI starts killing people, unfortunately, history would suggest we’re not going to do anything, but we are going to face that. That’s going to happen, and it’d be much better if we started dealing with it before then.” The trigger need not be financial to serve the same function: what matters is that the response architecture is already pre-positioned, ready to convert whatever emergency arises into politically guaranteed credit.

What we are watching, then, is not simply a bubble being inflated but a bubble being prepared for rescue. The political groundwork is being laid in advance so that when the credit seizes, the backstop will already be justified. The rogue swarm is an application for federal insurance, filed in advance in the language of salvation.

Postscript

The Fed raised rates by 25 basis points this week, pushing the target range to 3.75—4.00%, even as Brent crude held above $108 and diesel sat at record highs. The trap tightens from both directions: an energy shock the Fed cannot control and a credit machine it cannot afford to puncture—conditions that make a politically guaranteed AI backstop more, not less, likely. The first stage of the conjuring—socialise the risk—is now being tested in real time. The system’s answer to the contradiction remains what it has always been: more credit, conjured politically when private credit begins to fail.