Two stories landed last week that, taken together, settle a question the industry has been arguing about since 2023.
Can frontier AI labs pay for themselves? Anthropic just projected its first quarterly operating profit ever. $559 million on $10.9 billion in Q2 revenue. Can AI actually push human knowledge forward? An OpenAI reasoning model just disproved an 80-year-old math problem in a 125-page proof, verified by a Fields Medalist.
One proves the business model. The other proves the capability. Both proofs landed in the same week. There's also a SpaceX compute detail that complicates the profit story, and a Pope. Let's get into it.
Anthropic projects $10.9B in Q2 revenue and $559M in operating profit, its first ever. Less than a year ago, the company told investors it didn't expect profitability until 2028.
On May 20, the Wall Street Journal reported that Anthropic told its investors to expect $10.9 billion in Q2 revenue, more than double Q1's $4.8 billion, and a $559 million operating profit. Bloomberg and CNBC independently confirmed the numbers within 24 hours.
If those numbers hold when Q2 actually closes, three things follow. Anthropic becomes the first frontier AI lab to post a quarterly operating profit while still actively training frontier models. The quarterly revenue trajectory beats what Google and Facebook reported in the periods leading up to their own IPOs. And less than a year ago, Anthropic told investors not to expect full-year profitability until 2028.
The engine is enterprise. Claude Code crossed $1 billion in annualized revenue within six months of public launch. Customers spending over $1 million annually on Claude doubled from roughly 500 to more than 1,000 between February and April. Bristol Myers Squibb announced last week that it is deploying Claude across 30,000 employees for drug discovery, manufacturing optimization, and regulatory documentation. None of that looks like a chatbot pilot. It looks like infrastructure procurement.
CEO Dario Amodei told a developer audience earlier this month that the company had planned for 10x annual growth in 2026 and actually saw 80x in Q1. That is not a marketing line. It is an operational warning. The company is buying compute like it expects the curve to hold, and the financial projection assumes the curve does in fact hold through June.
There's a complicating detail. Ed Zitron noticed something pointed on Friday in his newsletter Where's Your Ed At. SpaceX's IPO prospectus, which dropped earlier in the week, revealed that Anthropic is paying SpaceX $1.25 billion every month through May 2029 for GPU compute access on Colossus-1 and Colossus-2. That's a $45 billion total contract. The schedule has a wrinkle. The May and June payments are at a "reduced ramp-up fee" before the full $1.25B/month kicks in. Those happen to be the exact months Anthropic is projecting its first operating profit.
Zitron's read is that this is "shiatsu-grade massaging of the numbers." His framing is more aggressive than I'd put it, but the underlying point is real. A back-loaded compute contract that shows reduced costs precisely during the quarter being marketed to investors is not an accident. It's how the operating-profit number was engineered to exist.
My take: Two things to pull apart.
First, the profitability claim is real even with the SpaceX caveat. Anthropic is genuinely doubling revenue quarter over quarter. Enterprise adoption is genuinely converting to seven and eight-figure annual contracts. The $559 million operating profit may be timing-dependent, but the trajectory underneath it is the kind of step-function growth that reshapes industries. Investors aren't going to discount the projection because of one quarter's compute schedule. They are going to ask whether the next three quarters can keep up.
Second, the timing detail tells you everything about why this number landed the week of the funding round. Anthropic is closing a $30 billion raise at $900 billion+ this week. Showing a profitable quarter in the same pitch deck is the difference between a $900 billion valuation and a $1.2 trillion one. The SpaceX contract terms were almost certainly negotiated with this exact narrative in mind. That is not deception, but it is choreography. And it tells you the IPO playbook, expected to land in October, is already in motion.
For 80 years, mathematicians believed the optimal arrangement was a square grid. OpenAI's reasoning model found a better one using algebraic number fields.
On May 20, OpenAI announced that an internal reasoning model had disproved the planar unit distance conjecture, posed by Paul Erdős in 1946. The model received the problem statement as a single open-ended prompt and produced a continuous 125-page chain of reasoning. The proof was independently verified by nine mathematicians, including Fields Medalist Tim Gowers and combinatorics heavyweight Noga Alon. Princeton's Will Sawin published a refinement the same day showing the improvement is polynomial, not just numerical.
If you don't track the history here, four things matter. The unit distance problem is one of the most-attempted open problems in combinatorial geometry, with eighty years of failed attacks. The model was not specialized for math, it was a general-purpose reasoning system. The proof is mathematically novel, not retrieved from existing literature (the Princeton refinement confirms this). And OpenAI's mathematician Sébastien Bubeck called it the first time AI has "autonomously produced an important result" in any research field.
There is a credibility subplot worth knowing. In October 2025, then-OpenAI VP Kevin Weil posted that GPT-5 had solved ten previously unsolved Erdős problems. Thomas Bloom, who maintains the official Erdős Problems site, demonstrated within hours that GPT-5 had simply located existing solutions in the literature. The post was deleted. Weil left OpenAI in April. So when OpenAI made this Erdős claim again on May 20, they did it differently. They published the model output alongside a companion paper signed by nine independent mathematicians, including Gowers, Alon, Shankar, and Tao. The verification was front-loaded into the announcement, not bolted on after the backlash.
My take: This is the first time in three years of "AI does math" announcements where the result is unambiguously real, novel, and verified before the press cycle started. The unit distance problem is the kind of problem where the proof method matters as much as the answer. The fact that the model bridged discrete geometry with algebraic number theory (using infinite class field towers, of all things) is the part working mathematicians are still digesting.
What does this actually mean for the rest of us? Two things. First, the "AI is just sophisticated pattern matching" line just got a lot harder to defend. The model did something that experts had not done in 80 years, using a technique pairing nobody had tried. That is not pattern matching against training data. That is generative mathematical reasoning. Second, we are now in the era where AI can plausibly contribute to scientific research as a peer, not just a tool. The peer-review questions that follow from that will define the next decade of academic policy.
The Erdős result and the Anthropic profitability claim are connected, even though they sit in different domains. Both are the first concrete proofs of things the industry has been promising for three years. Real money. Real intelligence. Both arrived in the same week. The pattern is unlikely to be coincidence.
The Vatican picked the same week to release its formal position on AI as OpenAI used to solve an 80-year-old math problem. The encyclical is dated May 15. The math announcement landed May 20. Pope Leo XIV signed his document on the 135th anniversary of Rerum Novarum, the 1891 encyclical that defined Catholic teaching on labor during the Industrial Revolution.
Read that timing again. The Church explicitly framed AI as the labor-rights moment of our century. And in the same week, a model did work no human mathematician had been able to do in 80 years.
For three years, the AI debate has run on two parallel tracks. The optimists said the models would eventually do real intellectual work. The skeptics said they were stochastic parrots that would never produce anything novel. Each side accumulated evidence that confirmed its priors and waved away the rest. The Erdős proof is the kind of result that does not survive that framing. It is novel by construction. It is verifiable. It was done by a generalist model with no math specialization. The skeptic case just lost its strongest piece of ground.
What I am sitting with this week is that we are watching the moment where two institutions much older than the tech industry (the Catholic Church and the Hungarian mathematical tradition) had to formally acknowledge that something has changed. The Pope put it in writing. Paul Erdős's most-attacked conjecture fell to a language model. Both happened on a regular Tuesday in May. The shape of the next decade just got a lot more visible.
The timing connects this week. Duke is one of 15 universities in OpenAI's NextGenAI consortium, and Duke's Deep Tech program runs what's described as the first AI-for-metascience research initiative. Ronnie Chatterji, OpenAI's Chief Economist, is also a tenured Fuqua professor. Sanford School professor David Hoffman co-leads the program with him. The Triangle's research universities have direct stakes in the question of how AI changes science, not just as observers.
The Erdős breakthrough is exactly the kind of result the Duke program was set up to study. Deep Tech at Duke is funding four projects this cycle, including Brinnae Bent's "Consilience" work on AI-augmented interdisciplinary research. That's the same shape of problem the Erdős proof represents (a model bridging discrete geometry with algebraic number theory), just done in a controlled academic setting. The Duke program is asking the meta-question: which fields get the next breakthrough, and how do we structure research workflows to capture them.
If you're at Duke, Carolina, NC State, or NCCU and working on AI applied to your specific research domain, this is the moment to engage the Deep Tech program. The next RFP cycle is reportedly opening before the end of the spring semester. Bigger picture: the Triangle's universities are positioned to be the academic infrastructure for the research-AI thesis, which is a more durable bet than chasing model performance directly. AbbVie picked Durham for an AI-native pharma campus last month. Bristol Myers Squibb's Claude rollout has clear Triangle implications too. The thesis is firming up.
Big week, even by 2026 standards. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. If you're a working mathematician or scientific researcher, I want to hear how the Erdős result is being discussed in your department this week. Specifically: is the conversation about the math, the model, or the implications for peer review? I'm working on a longer piece on what changes for academic publishing in the next 24 months, and I want it grounded in real conversations, not hot takes.
P.P.S. If your dev team uses VS Code and you haven't checked your Nx Console version yet, do that today. The compromised extension sat live for 18 minutes on May 18 and harvested anything in scope. The fact that OpenAI and Mistral were both hit tells you the attackers were targeting AI-aligned developer environments specifically. Rotate tokens. Audit Claude Code config files for unexpected MCP server connections. It's the boring security hygiene that will save you a bad month.