Everybody knows the story of the 1906 Meat Inspection Act. Upton Sinclair writes The Jungle, the country reads about what is going into its sausages, an outraged public forces regulation onto a screaming industry, and the food gets safer.
The last part is true. The screaming is not.
The big Chicago packers were the most enthusiastic supporters of federal inspection in the country, and had been lobbying for it for roughly twenty years before they got it. Ogden Armour, who ran Armour and Company, said plainly that federal inspection put the stamp of legitimacy on their product, in contrast to the smaller rivals selling meat that no government inspector had ever looked at.
The American Meat Packers' Association backed the bill. They fought exactly one provision in it, and they won: the original draft made the packers pay the cost of inspection. That got struck. The taxpayer picked up the bill instead.
So the incumbents got a federal seal of approval, paid for by the public, applied to an industry where their competitors now had to meet the same standard out of their own pockets.
In 1906 there were 923 interstate meat packers in the United States. By 1910 there were 300.
Why I am telling you this in 2026
On 12 September, Dario Amodei published an essay arguing that the AI industry should slow the pace of capability development. Within a day Sam Altman agreed on the need to pace the frontier, Demis Hassabis said the direction was correct, and Elon Musk wrote that Dario is right.
That is the heads of Anthropic, OpenAI, Google DeepMind and xAI agreeing, in public, in the same weekend. It was not the first time. Earlier this year all three frontier labs published memos calling for regulation of frontier models, converging on independent pre-release testing, a single standards body, and American leadership on governance.
Fierce competitors do not usually agree about anything. When they agree about the rules, it is worth asking who the rules land on.
Andrew Ng, who co-founded Google Brain, has said the quiet version out loud: that large technology companies are spreading fear about extinction as part of a regulatory capture campaign against open-source competition. That is not a fringe voice. That is someone who helped build the field.
The part of the story that gets told wrong
Before going further I want to correct the version of the meat-packing argument that circulates online, because it is wrong in a specific way and the wrongness is the first thing a critic will reach for.
The claim usually ends: and that is how the industry consolidated down to four packers, and they are the same four today.
They are not. The 1906 giants were Armour, Cudahy, Morris, Swift and Wilson. Today's big four are Tyson, Cargill, JBS and National Beef. Swift survives as a brand that JBS bought in 2007. The others are gone as independent firms.
The line is not continuous either. Congress passed the Packers and Stockyards Act in 1921 specifically to break the trust that 1906 helped build. It worked for a while. Beef packing was down to 36 percent four-firm concentration by 1980. Today it is around 85 percent, and that came from four decades of mergers and thin antitrust enforcement, not from Edwardian compliance costs.
I am not softening the argument by saying this. I am removing the part that is easy to knock down, because what is left is stronger: regulation written with incumbent support wiped out two thirds of the small operators in four years. That happened. It does not need a hundred-year conspiracy attached to it to be alarming.
What it actually costs the little guy now
The modern numbers are public, and they are not small.
Under the EU AI Act, a five-person startup can expect somewhere between thirty and eighty thousand euros to get compliant across legal review, risk classification, documentation and conformity assessment. A lean startup with a high-risk system is looking at fifty to a hundred and fifty thousand. The EU's own impact assessment puts total compliance for a single high-risk product at up to four hundred thousand euros, of which standing up a quality management system is a hundred and ninety-three to three hundred and thirty thousand, plus roughly seventy-one thousand a year to keep it running.
Now put those figures next to a frontier lab's legal department and notice that they do not register.
Every one of those line items is a fixed cost. Fixed costs favour scale. That is not a controversial claim about AI, it is an ordinary observation about regulation, and it is the same arithmetic that took 923 packers down to 300.
And here is the number I keep coming back to, because it is the clearest evidence available that the effect is already happening: more than sixty percent of startups now deliberately steer toward low-risk AI applications in order to stay out of the compliance regime.
Nobody went out of business in that statistic. That is what makes it easy to miss. The small players are not dying, they are declining to compete at the frontier. Which is the same outcome, arrived at voluntarily, and it does not show up as a bankruptcy anywhere.
Where the analogy breaks, and it matters
Here is the part I would want someone to tell me if I were about to publish the strong version of this argument.
The 1906 Act applied to every interstate packer regardless of size. There was no small-operator exemption. That universality is the entire engine of the 923 to 300 collapse. Take it away and the mechanism does not run.
The frontier proposals currently on the table are not shaped that way. Anthropic's framework, to take the one with published numbers, applies only above ten to the twenty-fifth FLOP of training compute and above either five hundred million dollars of annual AI revenue or a billion dollars of annual AI research spending. A startup underneath that is not covered. The EU is lowering conformity assessment fees for small companies for the same reason.
So the honest verdict on the strong claim is: not yet, and not in this shape.
But the critique survives in a smarter form, and this is the version I would actually defend. A threshold set just above where the incumbents are sitting does not block today's startup. It taxes tomorrow's success. Founders and investors price the costs they will face after they win, not the ones they face this year, and a wall you can see from a distance changes where people choose to walk. The moat is not at the door. It is at the crossing.
That is exactly what sixty percent of startups steering away from high-risk applications looks like from the inside.
The two questions worth asking about any proposal
Everything above collapses into two tests, and they are the ones the packers themselves fought over in 1906.
Who does it apply to? A rule that binds everyone regardless of size is the 1906 shape. A rule with a real, high, durable threshold is not. Watch whether thresholds appear, whether they hold, and whether they drift downward once the compliance industry has grown its own constituency.
Who pays for compliance? This is the one almost nobody watches, and it is the provision the packers cared most about. They accepted the entire bill and fought only the clause that made them fund the inspectors. If the public purse pays for the auditing regime, incumbents get a free seal. If firms pay per model, the cost is flat in dollars and enormous in proportion.
Apply those two questions and the current proposals sort themselves in a way that will surprise people who came in with a side already picked. The industry proposals carry thresholds. The Sanders and Casar bill, which would ban superintelligence outright and pause advanced development until a new federal regulator writes the rules, carries no size carve-out at all, along with twenty-year prison terms and what its own summary calls a corporate death penalty.
Gary Marcus, one of the most persistent critics of the AI industry alive, opposes the bill.
What I actually think
I do not think Dario Amodei is lying about being frightened. I also do not think it matters very much whether he is.
Sincerity and self-interest are not opposites, and the 1906 packers are the proof. Federal inspection genuinely did make meat safer. It genuinely did put a federal seal on Armour's export beef at public expense. It genuinely did kill two thirds of their smaller competitors inside four years. All three of those are true at once, and any account that needs one of them to be false is a worse account of what happened.
The useful posture is not to pick a camp. It is to stop reading the arguments and start reading the thresholds. A safety case is made in public and is designed to persuade you. A threshold is a number in a document, it decides who is inside the fence and who is outside it, and it is set by people who know exactly where their own company sits relative to it.
Watch the number, not the essay.
A disclosure, since it is load-bearing here. The research behind this post was done with Claude, which is built by Anthropic, one of the companies named above and the employer of the man whose essay set this off. I wrote about that conflict in The Bias It Could Name and the same caution applies: the dates, the thresholds and the cost figures are all checkable, and the framing is mine. The paragraph where that conflict bites hardest is the one reporting that Anthropic's framework exempts small companies. It is true, it is published, and it also happens to be the most convenient fact in this piece for the company that helped me find it. Go and read the threshold yourself.
This is a note about how I work, not advice about how you should. I build software at Revelations Technology and write these as I go.