San Francisco SoMa. Friday morning, February 27, 2026.

Every surface in this city has been sold to AI. Digital billboards boast about language-model benchmarks, and bus stops are covered in AI startup slogans. Taxi ad panels, subway walls, billboards. Every vertical surface within sight promises a faster, smarter future.

Figure 01 / 03Source Telegram #81

The area around Howard Street, where Anthropic's headquarters is located, is the hottest-pulsing part of that heart. Salesforce Tower rises within a single block, and Anthropic took over the old Slack headquarters in its entirety. OpenAI, Uber, Adobe, and Block are spread across a radius of a few hundred meters. Real estate people began calling the area “AI Alley” last year. In the middle of that street, people reached for the cheapest tool of all.

Hundreds of people gathered one by one on the pavement outside Anthropic's headquarters and took out chalk. They were probably ordinary hands. Hands that worried about next month's bills, lay awake at night trying not to be pushed aside by algorithms, and sometimes questioned their own usefulness. What those hands truly feared was not a machine uprising. It was the fear that, before a cold and perfect intelligence, human warmth and imperfection might become mere excuses.

“GOD LOVES ANTHROPIC.”

“THANK YOU FOR SAYING NO.”

“KEEP YOUR SOUL.”

Roy Bahat of Bloomberg Beta was passing by and took out his camera. The video spread instantly. People cried when they saw the writing.

On January 9, at the Department of War—the building where the Trump administration had restored the Pentagon's old name—Secretary Pete Hegseth issued a memo. AI companies must remove restrictions on the use of their technology. Changing the name from “Department of Defense” to “Department of War” is not a simple renaming. What “defense” asks for and what “war” asks for are completely different pressures, even when the substance is the same.

When Anthropic signed a $200 million contract with the Pentagon last July, it attached two conditions. Claude would not be used for fully autonomous lethal weapons. It would not be used for mass domestic surveillance of American citizens. Claude was the first AI model deployed on the U.S. military's classified networks, and these conditions were written into the contract.

Hegseth demanded that the conditions be withdrawn. AI models, he said, must be available for “all lawful uses.” If Anthropic did not comply: cancellation of the contract, invocation of the Korean War-era Defense Production Act, and designation as a “supply chain risk” company.

On Tuesday, February 24, Amodei met Hegseth at the Pentagon. An ultimatum was issued. Deadline: 5:01 p.m. Eastern Time on Friday, February 27. There is something oddly bureaucratic about the number one minute. These are people who do not forget administrative precision, even in an ultimatum.

On Thursday, February 26, Dario Amodei issued a statement.

“We cannot in good conscience accede to their request.”

The heaviest word in this sentence is not “conscience” but “accede.” He did not oppose them. He declined to comply. The two are different. A protest is directed at the other person, but a refusal is an act of drawing a line for oneself.

That night, people probably took their chalk with them to bed. The next morning, before the deadline arrived, words appeared on the pavement.

The retaliation was immediate.

Trump wrote on Truth Social: “Immediately cease all use of Anthropic technology by all federal agencies. We don't need it, we don't want it, and we will never do business with them again.” Hegseth designated Anthropic a national security supply chain risk company. No business with military contractors. An end to its use across federal agencies. For six months.

Think about where the phrase “supply chain risk” was originally used. It was language applied to Chinese semiconductors and Russian software. The moment a state puts that frame around one of its own companies, the company becomes an internal enemy. That is how language legitimizes the target of violence.

That same day, Sam Altman said that “OpenAI has the same red lines.” At the same time, OpenAI announced a contract for classified networks with the Pentagon. If it were truly the same line, one side would not have been forced to cross it. Elon Musk's xAI filled the place Anthropic was pushed out of. The state directly designed a structure in which the side willing to accept lower standards gets the contract. This is not competition. It is capture.

There have been similar moments in history. IBM supplied punch cards to the Nazi regime. History only recorded later what those machines were used to classify. In 2018, thousands of Google employees signed a petition and resigned in opposition to the military AI project Maven. It was a rebellion from the bottom up.

Anthropic's refusal is different in character. It was a decision made first by management. The CEO himself used the word “conscience.” Socrates did not refuse the hemlock. Galileo knelt before the Inquisition but did not give up heliocentrism. An AI company has now joined the lineage of those who paid a price for their conscience.

The phrases people left in chalk were not protest slogans. They were closer to prayers, or offerings. It follows the same grammar as laying flowers and candles in front of a church, a courthouse, or a hospital. In front of the modern spires made of glass and concrete, humans of flesh and bone knelt down. In an age when people trust neither the government, the media, nor religion, when it seems that no one is caring for our souls, people are looking for moral authority in front of an AI company.

As AI enters the realm of the transcendent, we have begun, without realizing it, to borrow the language of theology. In an age that needs salvation, people look for saviors anywhere they can. That language simply happened to erupt in front of a glass building.

Inside that glass building, trillions of parameters were probably turning coldly and silently. Outside, chalk scraped against the pavement and crumbled with a soft hiss. The silence of a machine calculating eternity, and the friction of humans scraping away at a moment.

The AI safety movement has begun to take on a religious grammar. There is doctrine, a martyr's narrative, and heresy. Anthropic is becoming the cathedral of this secular church. Chalk is an offering. This is its most powerful asset and its most dangerous liability. Once people begin projecting their souls onto you, the weight of that expectation becomes real. The distance between a place where people kneel and pray and a place where they carry torches to execute heretics is much shorter than we think. For now, people are writing prayers in chalk, hoping for salvation. But the moment that hope turns into betrayal, that white powder can become something else at any time.

What did Dario Amodei feel? What he is making is the product of mathematics and statistics, but the world is already asking him to bear the responsibility of God. Telling a company that makes AI to protect its soul is a strange request. The soul belongs to humans, after all. But was the person who wrote those words trying to save Anthropic, or trying to save himself? Perhaps that boundary had already grown blurry even to him.

Chalk disappears when it rains. It is a message written in the weakest, most vulnerable way. That vulnerability becomes proof of its sincerity. The white powder, blurred and carried away by rainwater, will seep into the cracks in the concrete. The most earnest sentences are not recorded and preserved. They melt into the world.

The morning after the chalk is erased, the people who walk over it will not know. They will not know that the ground beneath their feet was once a canvas for prayer. If this scene appears in textbooks ten years from now, what question will students face?

“What had humanity lost that made it kneel at the door of a technology company, and did that company ultimately bear the trust placed in it?”

The chalk has already been erased, and in this city where billions of dollars' worth of billboards illuminate the future of AI all night, only the question that will last the longest has seeped beneath the concrete, still not dry.