Five stories from the neural fringe this week. Microsoft had to rent servers from Amazon to keep GitHub online because AI coding agents flooded the platform so badly that availability fell to 88 percent. OpenAI’s first publicly confirmed audited financials show the company spent thirty-four billion dollars in 2025 to earn thirteen billion and posted a net loss of thirty-eight and a half billion. Anthropic published a paper calling for a coordinated global pause on AI development three days after secretly filing for a one trillion dollar IPO. A startup raised sixty-six million dollars to give AI agents their own corporate HR files because McKinsey now has twenty-five thousand AI agents working alongside sixty thousand humans and apparently nobody had thought to give the bots an employee ID. And the US Department of Health and Human Services launched a program that uses ChatGPT to audit two hundred billion dollars in suspected Medicaid fraud, which is the government using an AI famous for making things up to find things that other people made up. Let us begin.
Here is the situation. Microsoft owns Azure, one of the three largest cloud computing platforms on earth. Microsoft also owns GitHub, which is where more than one hundred million developers store, version, and collaborate on code. In a world where everything was working as planned, if GitHub ever needed more computing capacity it would simply route that traffic to Azure, which Microsoft also owns, and the problem would be solved before anyone noticed.
That is not what happened.
On June 16, Microsoft confirmed it has been routing GitHub traffic through Amazon Web Services because Azure cannot absorb the load fast enough. The reason GitHub needs more load than Azure can currently handle is AI coding agents. As of June 2026, GitHub is processing two hundred and seventy-five million commits per week. In all of 2025, GitHub processed approximately one billion commits total. Two hundred and seventy-five million per week means 2026 is on pace for fourteen billion commits in a single year. That is a fourteen times increase year over year. AI agent pull requests alone jumped from four million in September 2025 to seventeen million by March 2026.
What this means in practice is that GitHub logged nine service-degrading incidents in May 2026 alone, followed by ten in April. Platform availability fell to approximately 88.4 percent in June. Enterprise service level agreements guarantee 99.9 percent uptime. Companies are noticing the gap.
A Microsoft spokesperson confirmed the arrangement and described it as a response to “the incredible spike in agentic development that began late last year,” while simultaneously stating the company is “accelerating its move to Azure” and exploring a multi-cloud strategy. The sentence “we are moving to our own cloud but in the meantime we are on a competitor’s cloud” is not one that typically appears in official corporate statements. This one did.
To be fair to Microsoft, the technical explanation is not embarrassing so much as it is genuinely difficult. The original GitHub infrastructure was built on-premises, then partially migrated to Azure over the past few years. As of the May 2026 availability report, forty percent of monolith traffic was being served from Azure, up from eight percent in February 2026. Git traffic was at thirty percent Azure. The architecture was mid-migration when AI agent traffic exploded faster than the migration timeline anticipated.
The slightly more embarrassing part is that the company selling Azure as an enterprise cloud product had to go to the company selling AWS as a competing enterprise cloud product and say please hold some of our traffic while we figure this out. AWS agreed. You would also agree. If your biggest competitor’s flagship developer platform needed your infrastructure to stay operational, you would say yes while smiling very broadly.
The broader implication here is not just that GitHub had a rough spring. It is that AI coding agents are generating infrastructure demand at a pace nobody planned for and nobody has fully solved yet. Every AI agent making commits, opening pull requests, running CI pipelines, and triggering webhooks is doing what a human developer would do, except faster and around the clock. The collective result of millions of agents doing this simultaneously is a platform load that apparently exceeds what the second-largest cloud provider on earth can absorb without calling the first-largest cloud provider on earth for backup. This is an infrastructure problem that will not get smaller.
OPENAI SPENT THIRTY-FOUR BILLION DOLLARS IN 2025 AND MADE THIRTEEN BILLION AND IS NOW GOING PUBLIC
The audited 2025 financials for OpenAI leaked on June 16 and they are the kind of numbers that make you put down your drink and read the sentence again slowly.
Revenue: thirteen billion dollars. Up from three point seven billion in 2024, which is genuinely extraordinary growth. Costs and expenses: thirty-four billion dollars. Net loss: thirty-eight and a half billion dollars, though to be precise about it, twenty-one billion of that was operating losses and the rest was a one-time accounting charge from when OpenAI converted from a nonprofit to a public benefit corporation in October 2025. The conversion generated a forty-one billion dollar non-cash charge. So the real operating picture is thirteen billion in and twenty-one billion out, which is still losing sixty cents for every dollar of revenue earned.
Of the thirty-four billion in spending, seventeen point two billion went directly to Microsoft for Azure cloud infrastructure and research and development compute. That is roughly half of all spending going to a single vendor. Microsoft, in turn, is one of OpenAI’s largest investors. There is a circular financial relationship here that accountants describe using words like “related party transactions” and that the rest of us describe by staring at the ceiling for a moment.
The IPO plan is proceeding. OpenAI filed a confidential S-1 with the Securities and Exchange Commission, with Goldman Sachs and Morgan Stanley leading the process. Current valuation: eight hundred and fifty-two billion dollars. The thesis for going public at that valuation while losing twenty-one billion a year operationally is that revenue tripled in one year and the company expects it to keep growing faster than costs. The efficiency metric actually supports this reading. In 2024, OpenAI spent two dollars and thirty-seven cents for every dollar of revenue. In 2025, that ratio improved to one dollar and sixty cents per revenue dollar. The trajectory is in the right direction. The absolute numbers are still extraordinary.
What the audited financials also confirm is that Sora, OpenAI’s AI video tool that launched to enormous fanfare and was shut down in March 2026, cost approximately fifteen million dollars a day in inference costs while generating a total of two point one million dollars in lifetime revenue. That is not a typo. Fifteen million a day out. Two point one million total in. The product ran from November 2025 to March 2026 and the numbers worked out so badly that the team was reassigned to robotics research under a project codenamed Spud, which at least has a more honest name than Sora did.
The broader picture is a company that is genuinely growing at a historically unusual pace, genuinely losing money at a historically unusual pace, and genuinely moving toward a public listing at a valuation that would make it one of the largest companies in America on day one of trading. Whether this is a story about a transformative business in its investment phase or the most expensive scaling experiment in corporate history depends heavily on what you think AI revenue looks like in 2027 and 2028. The audited financials do not answer that question. They just make it feel more urgent.
Source: Where’s Your Ed At (confirmed by Financial Times)
On June 1, 2026, Anthropic quietly filed a confidential S-1 with the Securities and Exchange Commission. This is the first step in going public. The filing was not announced. It is required to be confidential at this stage. The valuation attached to the process: nine hundred and sixty-five billion dollars.
On June 4, 2026, three days later, Anthropic published a research paper titled “When AI Builds Itself” through its internal Anthropic Institute, authored by head of research Marina Favaro and co-founder Jack Clark. The paper proposes a globally coordinated pause on frontier AI development. The argument: AI systems are approaching the ability to recursively improve themselves, humans are losing the ability to meaningfully oversee the process, and a coordinated slowdown before this threshold is crossed is necessary and urgent.
The paper disclosed that as of May 2026, more than eighty percent of code being merged into Anthropic’s own production codebase is authored by Claude, not by human engineers. AI task completion horizons, meaning how complex a task an AI can handle autonomously, have been doubling roughly every four months. Anthropic is describing, using its own internal data, a process of AI-assisted AI development that is compounding faster than any human engineering team can track. The proposed solution is for all frontier AI labs, globally, to stop simultaneously.
To be absolutely precise about what Anthropic proposed: not a unilateral halt by Anthropic itself. A coordinated multilateral halt by every leading AI lab in the world, verified by some mechanism analogous to nuclear weapons inspection regimes, with Anthropic continuing its own development work in the meantime. Anthropic explicitly acknowledged that if only one lab stopped, competitors would race ahead. So the pause they are calling for is one that requires every major AI company, in multiple countries, in the middle of a commercial and geopolitical race, to agree simultaneously to stop. Until that agreement happens, Anthropic will continue. While also going public at a trillion dollar valuation.
Noah Giansiracusa, an associate professor at Bentley University, told Scientific American he does not think it is a genuine call to slow down. Critics pointed out that a coordinated global pause, if achieved, would freeze the competitive landscape at a moment when Anthropic is already in the top two or three AI labs globally. The company that proposes a pause from a position of strength is proposing something very different from one doing so from behind.
To be fair to Anthropic, the internal data they disclosed is genuinely alarming, and the paper is intellectually serious. The argument that a pause only works if it is multilateral is correct. The proposal to build coordination mechanisms toward that goal is a real thing that would require real work. The question is whether publishing a paper calling for something that requires global consensus, simultaneously with filing to go public at a valuation that implies your investors expect you to keep building indefinitely, is a genuine warning or remarkable brand positioning. Both things can be true at once. That is what makes it so interesting to watch.
Daniela Amodei told CNBC the same week that Anthropic continues to see “reasonably exponential” year over year performance improvements. The IPO is moving forward. The pause paper was published. These two things happened in the same week and no one at Anthropic appears to have found the timing awkward.
McKinsey has sixty thousand human employees. It also has twenty-five thousand AI agents. Those agents work alongside the humans, run tasks, generate outputs, access systems, and in some cases make decisions. The ratio of human workers to AI agents at McKinsey is currently about 2.4 to one. Goldman Sachs tested an AI coding agent called Devin as a new employee last year. Most large enterprises are building or deploying autonomous AI agents into their workflows right now.
The problem is that none of the identity and security infrastructure companies use to manage employees was built for this. When a human joins a company, they get an HR file, a laptop, an email address, an access badge, a set of permissions tied to their role, and when they leave those permissions get revoked. The systems that track all of this have been built and refined over decades. They assume the worker is a person. An AI agent is not a person. It is software that runs tasks autonomously, accesses systems, handles data, and operates continuously without going home for the night.
NewCore emerged from stealth on June 15 with sixty-six million dollars in funding to solve exactly this problem. The company is building what it describes as a workforce identity platform rebuilt from the ground up for environments where the workforce includes humans, machines, and AI agents. The platform treats AI agents as first-class identities with their own permissions, life cycle controls, audit trails, and revocation mechanisms rather than as service accounts or machine credentials, which is how most enterprise IT teams are currently handling them.
The specific security architecture NewCore built involves a split-key system that divides critical identity credentials between the customer and the platform, designed so that no single point of compromise can take down all agent access at once. This is a real engineering problem. An AI agent with access to company files, payment systems, databases, and customer data is an attack surface. If that agent is authenticated through a credential that can be stolen, the consequences are significant. The fact that most companies are currently handling this by issuing the agent a service account with broad permissions and no life cycle review is, per the NewCore founders, the main reason there is a market.
What is genuinely strange about this story is not that the problem exists. It is that we got here so fast. At some point in 2025 or early 2026, at a meaningful number of large companies, the number of autonomous software agents accessing internal systems exceeded the number of humans doing the same thing. Nobody announced this. It happened gradually through individual deployment decisions until it was the default reality. And the response to discovering that your company’s digital workforce is majority bots is apparently to raise sixty-six million dollars and build an HR department for software.
The next logical step, once AI agents have HR files, is presumably that they get performance reviews. And then they get managed. And then someone writes a corporate policy about agent conduct that the AI agents themselves will be asked to follow. We are clearly heading somewhere very specific and there is no sign anyone is going to stop before we get there.
THE US GOVERNMENT IS USING CHATGPT TO AUDIT TWO HUNDRED BILLION DOLLARS IN MEDICAID FRAUD
The Department of Health and Human Services has a program. It is called AERO, which stands for Audit Enforcement and Risk Oversight. The program scans five years of Medicaid audit records from all fifty states. The HHS Assistant Secretary for Financial Resources estimates the department has between one hundred billion and two hundred billion dollars in wasteful or fraudulent spending sitting in those records. The tool built to find it uses ChatGPT.
Let that sequence settle for a moment. The government is using an AI that is publicly documented to make up facts, invent citations, produce plausible-sounding but inaccurate summaries, and confidently assert things that are simply not true, to audit financial records in search of fraud committed by people who may have done things like make up facts, invent documentation, and produce plausible-sounding but inaccurate records.
The AERO program works by scanning audit histories across all fifty states for patterns of repeated deficiencies that were flagged in federally required audits but never addressed. States and grantees that show up repeatedly in this scan face consequences including the possible loss of federal funding across programs including Medicaid. The program was expanded in May 2026 to include what HHS describes as next-generation AI analytical tools, with ChatGPT built into the system.
The healthcare industry response has been divided into two groups. One group thinks deploying AI to find systemic patterns in enormous audit datasets is actually a reasonable application of the technology, that looking for repeated failures across fifty state records is exactly the kind of pattern-recognition task AI handles well, and that the dollar figures involved justify aggressive tooling. The other group thinks that using a hallucination-prone language model to make determinations that could result in states losing healthcare funding for their citizens is exactly the scenario where you want the most careful, verifiable, human-reviewed process possible.
Both groups have a point. AI is genuinely useful at scanning large document sets for structural patterns. And AI genuinely does make things up in ways that are convincing enough that multiple federal courts have had to sanction lawyers for submitting fabricated AI citations as real case law. These two facts coexist in the same universe and AERO lives in that universe.
What is interesting about this story beyond the obvious tension is the scale of the bet. One hundred to two hundred billion dollars in estimated fraud is not a small number. If the AI identifies a state as having a pattern of Medicaid abuse and the finding turns out to be a hallucination, the consequences are not a mildly embarrassing news cycle. They are a state losing federal healthcare funding based on a document an AI invented. HHS is aware of this risk. Whether the program has sufficient human review built in to catch it is a question the agency has not answered publicly with much detail. The fact that they launched the program and the fact that the question is unresolved are both true simultaneously.
There is also a deeply American quality to the situation. The same country that has watched courts sanction lawyers for submitting AI-hallucinated citations for three years running has now deployed the same category of AI to audit the healthcare system for fraud. This is either a story about bold institutional ambition or about an organization that read all the same news everyone else read and decided the warnings did not apply to them specifically. Possibly both.