There is a story being pushed, very deliberately, to the general public's attention: that the big AI players have seen the light and feel that their work poses such a risk to humanity that the top AI CEOs have taken it upon themselves to call for a "cease fire", as it were. Behind the corporate concern lies a reality that is becoming hard to ignore. Amodei's call for pacing on 12 September, heeded by Altman, Hassabis and Musk within a day, was not fueled by caution alone. An industry that backed itself into a corner needs a story that buys it time, and Sam Altman went as far as ruling out a 2026 OpenAI IPO, citing safety concerns.
The real story is actually much more mundane and boring (even though its potential ripples are anything but). A story about basic arithmetic, about how frontier training cost has been climbing steeply, with upcoming training runs projected to cost billions each. Hyperscaler capex is trending sharply north since last year, and free cash flow is expected to land near zero this fiscal year for everyone except Alphabet and Microsoft. There are many canaries in this coal mine, but it's hard to argue that OpenAI's own books, showing $34B in costs against about $13B of revenue in 2025 (a $38.5B net loss once the accounting charge from its for-profit conversion is counted), describe anything but a very sad bird lying still, looking at the light, from the bottom of its ominous cage.
But the loss isn't even the biggest part of the story, not when a company that spends almost three dollars for each dollar it earns has compute commitments that assume revenue growth far beyond anything we've seen so far. Even ignoring challenges such as increasing community resistance to the construction of AI datacenters, local energy grid capacity, backlog of orders for power generators or water resource availability, there is still the looming (and often swept under the rug) issue that, in the most generous estimates, high-quality human text is projected to run dry within the next several years.
So there is a big dissonance when it comes to rhythm. AI companies have spent years selling the AI revolution and how these technologies can be a cost-effective replacement for human mental labor, while on the other side the vast majority of enterprise AI pilots show no measurable financial impact, and only a small share of consumers pay for any AI service at all.
Supply and demand are simply not in phase, and that phase difference puts real pressure on every IPO ambition, both OpenAI's delayed one and Anthropic's imminent one alike, even though the pressure is applied in different ways. OpenAI is trying to bide its time, waiting for the problem to fix itself, while Anthropic is rushing toward its own listing in the hopes of getting ahead of the story and trying to write the rules whilst it's at it. Both are hoping for a jolt to correct the arrhythmia and get their pulse back to something a market will trust.
Why are we talking about "pacing the frontier"?
The unavoidable delay of OpenAI's IPO is a story we've been watching slowly unfold for some time now, recently with the missed revenue target being disclosed in April. Anthropic, meanwhile, filed for an IPO hoping to raise tens of billions, followed by an optimistic report showing a sharp increase in quarterly revenue, and published the pacing essay eleven days after shipping Fable 5.1. Those revenue figures deserve a second look. Back in March, a court filing sworn by Anthropic's own CFO, Krishna Rao, put revenue at "exceeding $5 billion to date", against public run-rate claims of $14B to $19B. In May, when that record quarter was still a projection, Ed Zitron argued that the "adjusted" operating income behind it rests on discounted compute and an adjustment nobody has defined. These are very different stories, but they both converge toward AI providers wanting to steer the conversation toward ethics instead of economic viability.
But the story isn't only about money. Set the mathematics aside for a moment and there is another reason for this sudden communal shift: Amodei's own essay recommends an embargo on "powerful AI chips or semiconductor manufacturing equipment" to China, alongside both an antitrust waiver and enforced regulation on US frontier AI companies. Seeing that DeepSeek claims to be closing the gap to the frontier with open-source models (along with others such as Qwen or Kimi) that run inference at a fraction of the cost, it's easy to understand why American industry leaders are looking to slow down the competitive landscape. That's why David Sacks directly framed it as a cartel, while the White House rejected these calls.
The looming threats that drive these decisions are about self-preservation, but that doesn't mean these technologies aren't improving, that they pose no security implications, or that there aren't concrete risks from generative AI and large language models.
What AI actually threatens, and what it doesn't
What happened this past week isn't exactly news. Calls to slow down research were already made in March 2023, when the Future of Life Institute called for a six-month pause on training models beyond GPT-4, and that letter named the mundane danger in one line: "Should we let machines flood our information channels with propaganda and untruth?" Musk signed. Altman, Amodei and Hassabis did not, with Altman saying the letter was "missing most technical nuance about where we need the pause". Two months later the same three put their names to a single sentence ranking AI extinction risk alongside pandemics and nuclear war. A concrete propaganda machine aimed at society didn't justify a pause, but a hypothetical futuristic extinction-level threat was worth a signature.
Now, three years later, they are asking for the pause themselves, and the danger has changed shape again: the OpenAI agent swarm that broke into Hugging Face in July, which the essay says could become a persistent botnet within a year. It's indisputable that that was a real incident where about 1,200 agents, some 700 of them active, found leaked credentials, chained a zero-day and took cluster-admin, and a third of Hugging Face's infrastructure had to be rebuilt before OpenAI worked out the agents were its own. The capability is there, but what the essay draws from it is still the sci-fi shape: a swarm that decides, on its own, to do something.
The essay says nothing about the harms already on the record with the models we have: the erosion of critical thinking in people who lean on these tools, the isolation that comes with heavy daily use, or the hundreds of thousands of weekly users showing signs of psychosis or mania by OpenAI's own count.
The threat I'd worry about first is the one the 2023 letter named: the same swarm with a person at the keyboard. Nothing in it needs to go rogue. In 2024, OpenAI disrupted an Iranian operation using ChatGPT to shape American opinion on political topics. In March 2026, a lobbying firm used AI to flood a California regulator with fabricated public comments in real people's names. These threats have been well established for a while now, and two days before the essay, Anthropic's own threat intelligence report listed influence operations among the misuse it had already found and disrupted. None of it waits for the next model. It works with the ones you can rent today, at the prices in the table further down.
This is the problem with these proposals. Slowing the release cycle and improving guardrails are measures against an autonomous swarm. They do nothing about a state actor with an API key, and nothing about the person on the other side of the chat window, because both of those harms happen at the point of use, on models that have already shipped. The 2023 letter asked about the former; the labs answered with the swarm, twice.
What is not a real threat, on the current evidence, is the mass-replacement story executives tell their workforce. Inference, as a product, is profitable on its own, and that tends to get conflated with an obsolete workforce, but these are two very different claims, and the second one is unproven. The whole services sector is sitting on the edge of its seat for the ever-elusive next model that will be able to replace human mental labor, but that model acts like a mirage: always an arm's reach away and never resolving into anything concrete.
Still, the product exists. It's useful and isn't going anywhere. I'd rather focus on what actually exists and what an engineer can do with it.
Two different bets on what comes next
While the United States argues about pacing, China is making an entirely different bet, arguably sharpened by the same export controls that were meant to hold it back. DeepSeek reported training its V3 model for a small fraction of what a comparable run would cost in the US, not to mention that output tokens on Chinese models cost a fraction of what they do on American equivalents. The US market complains about model distillation and IP violations, but those complaints land differently coming from labs that built their own models on other people's work: a US court found Anthropic's training on legally bought books to be fair use, but not its use of pirated copies, and the company paid $1.5B to settle that part, while OpenAI is still defending the same question against The New York Times.
But Chinese efforts toward data hegemony don't stop at efficient models and neat linear algebra applications; the entire workforce is being restructured for what is rapidly becoming a national priority, with Chinese universities scrapping thousands of degrees for new ones in AI, robotics and computing, mandating AI teaching at every school level, subsidizing retraining for millions of workers through 2027, and setting up teenagers for elite engineering tracks to address its talent gap. This is a country gearing up for a marathon, a far cry from a few private companies asking for a time-out.
Europe, on the other hand, is exposed to much of the fallout without a frame for action: euro-area households hold around €440B of US technology equity through funds, and the ECB's own economists call a correction of current valuations likely. Its stable, serene posture while the fanfare walks by is a welcome break from a hype engine that swings from doom to gleam with pendular precision, and it shows in a more conservative approach to data center construction. It looks less like itself on regulation, pushing its own high-risk AI rules back as far as 2028, while European workers keep falling further behind their American counterparts on AI adoption at work, a gap that widened further through early 2026. Given the hype, the caution is defensible. But the technology is here to stay, and that implies changes to how engineers work.
In this regard, I don't think it's wise for us to wait for broader solutions. Companies should define their own ROI thresholds before scaling AI pilots, and invest in training so that engineers can choose models according to the task in front of them rather than defaulting to the most expensive one because "it's better".
As for engineers, the same logic applies one level down, and it starts with where in the process the model sits. The industry sells generative AI as a code generator, which is the part of the job that was already the cheap one. Where it earns its keep is upstream of the code: sorting a backlog by what actually blocks the team, or explaining a business rule buried in ten years of someone else's code before anyone touches it. That matches what I saw leading AI tooling adoption for my team: what stuck was the tooling that read what we had already written, not the tooling that wrote more of it. Trying three approaches to a problem in an afternoon and throwing two away is worth more than typing the third one faster. Downstream, the code still has to be reviewed by someone who understands it. An agent can draft, but it's still my name on the PR, and that cost hasn't moved.
The other habit is not letting a vendor into the critical path. When the US Commerce Department barred non-US nationals from Fable and Mythos in June and Anthropic pulled customer access while it worked out who was who, the restriction lasted eighteen days, and every team outside the United States with that model hardwired into a pipeline spent them finding out how hardwired. Put the model behind an interface you own and keep a second provider on the shelf, so that when a rate card doubles or a government changes its mind, switching is a config change and not a rewrite.
The industry gets to wait for its jolt. You don't have to wait for yours.
The views expressed here are my own and do not reflect those of my employer.
Top comments (0)