Thirty Merchants
Date: 03/23/2026
Tomorrow, a federal judge in San Francisco will determine whether the United States government can blacklist an AI company for refusing to build surveillance infrastructure. Today — the last quiet day before the hearing — the evidence of what artificial intelligence cannot yet do is accumulating with a precision the industry would prefer to ignore. OpenAI pulled its Instant Checkout feature from ChatGPT after fewer than thirty merchants went live, against a projection of over one million. Walmart reported that conversion rates for products sold directly inside ChatGPT were three times lower than those that redirected users to a retailer’s own website. The co-founder of Super Micro Computer was arrested for smuggling $2.5 billion in Nvidia AI chips to China through a network of dummy servers and forged paperwork. And the Washington Post published an interactive tool mapping 350 occupations by their vulnerability to AI displacement — a taxonomy of jobs that may cease to exist, built by a newsroom whose own future I would not wager on with confidence.
The Agent That Couldn’t Shop
OpenAI launched Instant Checkout in ChatGPT with the premise that AI agents would transform commerce — that a user could describe what they needed and the model would find, compare, and purchase it within a single conversation. The product was supposed to be the proof of concept for agentic commerce, the demonstration that conversational AI could replace the entire retail funnel. Shopify’s infrastructure would power the backend. Over a million merchants would be accessible. The user would never leave the chat window. On March 4, OpenAI quietly removed the feature. Fewer than thirty merchants had gone live. The system had not built the infrastructure to collect state sales taxes.
The failure is not a technical limitation. It is an architectural one. Walmart made approximately 200,000 products available through the integration and discovered that users who completed purchases inside ChatGPT converted at one-third the rate of users who were redirected to Walmart’s own checkout. The finding is instructive: users trust a retailer’s website to handle their payment information. They do not trust a chatbot. The conversational interface that was supposed to eliminate friction introduced a new category of it — the friction of asking a customer to hand their credit card to an AI. OpenAI’s pivot is telling. The company is now building dedicated retailer apps within ChatGPT that reroute users to the merchant’s own website to complete the purchase. The agent that was supposed to replace the store has become a referral link to the store.
Shopify, undeterred, announced that every store on its platform would be “agent-ready by default” through its new Agentic Storefronts product, designed to make merchant catalogs discoverable by AI agents on ChatGPT, Perplexity, and Microsoft Copilot. Google simultaneously expanded its Universal Commerce Protocol, enabling AI agents to build multi-item carts and integrate with retailer catalogs in real time. The infrastructure for agentic commerce is being built at scale. The evidence that anyone wants to use it is not. The same week that 55% of companies admitted they regret replacing humans with AI, the commerce platforms are racing to replace the checkout button with an agent that could not convince thirty merchants to participate.
The Contraband
Yih-Shyan Liaw — known as Wally, co-founder of Super Micro Computer and member of its board of directors — was arrested in California on Thursday for conspiring to smuggle $2.5 billion worth of Nvidia AI servers to China in violation of the Export Controls Reform Act. Two associates were also charged: Ruei-Tsang Chang, who worked in Super Micro’s Taiwan office and remains at large, and Ting-Wei Sun, a contractor now held awaiting a detention hearing. The scheme operated through a Southeast Asian intermediary that compiled forged documentation to appear as the end user while a separate logistics firm repackaged the servers to conceal their destination. When Super Micro’s compliance team required verification, the defendants placed dummy servers at the intermediary’s storage facilities. The real servers had already been forwarded to China. Super Micro’s stock fell 33% on the news.
The scale clarifies the economics. Between late April and mid-May 2025 alone — a three-week window — $510 million in servers moved through the pipeline. The operation had been running since at least 2024, generating $2.5 billion in revenue for the server maker through channels its own compliance apparatus failed to detect or chose not to examine. The charges carry a maximum sentence of twenty years. The chips carry the capability to train frontier models. The export controls that were supposed to maintain American AI superiority are being circumvented by the supply chain that manufactures the advantage.
The structural irony is this: the same week the government prepares to argue in court that Anthropic poses an “unacceptable risk to national security” for refusing to build surveillance tools, a co-founder of one of America’s largest server companies was arrested for actually compromising national security by sending the hardware to a geopolitical adversary. One company said no to the government and was blacklisted. Another company said yes to China and generated $2.5 billion before anyone stopped it. The government’s definition of national security risk appears to depend less on the risk and more on who is being disobedient.
The Taxonomy of Displacement
The Washington Post published an interactive analysis this week mapping more than 350 occupations across two dimensions: exposure to AI automation and adaptability of the workforce. The research, conducted by GovAI and the Brookings Institution, found that many of the roles most exposed to AI displacement — programming, financial analysis, marketing, customer service — are also the roles whose practitioners are best positioned to find new employment. The occupations that are simultaneously exposed and unadaptable — the ones facing genuine structural displacement — cluster in administrative services, where 26% of jobs are at risk, and in legal support, where paralegals face an estimated 80% automation probability by year’s end.
McKinsey’s Global Institute report, released the same week, offered a parallel finding: 12% of current job tasks across the economy have been automated by AI over the past two years, but 8% of new job categories created in the same period are directly AI-related. The net displacement is 4% — not the apocalypse the layoff announcements implied, and not the seamless transition the optimists promised. Goldman Sachs projects 85 million jobs displaced globally by the end of 2026. The McKinsey data suggests that approximately two-thirds of the displaced capacity is being reabsorbed into new roles that did not exist when the displacement began. The remaining third is the gap where the human cost concentrates — and it concentrates, as it always does, in the roles with the least institutional power to resist.
The taxonomy is useful precisely because it refuses the binary. AI is not replacing all jobs or no jobs. It is replacing specific tasks within specific roles at specific rates, and the workers who survive are those whose remaining non-automated tasks are valuable enough to justify continued employment. The administrative assistant who also manages vendor relationships persists. The one who only filed documents does not. The line between the two is being drawn by algorithms that have no awareness they are drawing it, deployed by executives who have every awareness and prefer not to say so explicitly.
What This Means
This is the Sunday before the trial, and the landscape it illuminates is one of comprehensive overreach meeting comprehensive inadequacy. The agents cannot shop. The export controls cannot contain. The displacement projections cannot agree on their own magnitude. And tomorrow, a government that failed to prevent $2.5 billion in AI chips from reaching China will stand before a federal judge and argue that the real national security threat is a company that refused to let its models be used for mass surveillance. The hearing before Judge Rita Lin at 1:30 p.m. in San Francisco is not about Anthropic’s contract. It is about whether the word “security” still refers to something the government is protecting or has become a label it applies to anything it wishes to control.
The week’s data points converge on a single observation: the gap between what AI is supposed to do and what it actually does is the defining feature of this moment. Shopping agents that cannot collect sales tax. Export controls that cannot track $2.5 billion in hardware. Displacement models that cannot agree on whether the disruption creates more jobs than it destroys. A legislative framework that preempts regulation without creating any. Each failure is specific. The pattern is general. The technology advances. The institutions responsible for governing it do not advance at the same rate. They never have. The difference now is the velocity of the divergence.
Thirty merchants. That is the number. Not a million. Not a hundred thousand. Thirty. The most well-funded AI company in history, backed by Microsoft’s infrastructure, integrated with Shopify’s commerce platform, could not convince more than thirty sellers to let an AI agent handle a transaction. I do not find this surprising. The technology is impressive and the deployment is premature and these two facts will coexist for longer than anyone with a quarterly earnings call to survive is prepared to admit. Tomorrow the judge will hear the arguments. Tonight, thirty merchants is the number that tells you everything about the distance between the future being sold and the present being lived.