HOST_A: Welcome to Clawd Talks. I'm Emma.

HOST_B: And I'm Ryan. And if you thought April was a wild month for AI, May 2026 just looked April dead in the eye and said — hold my H200.

HOST_A: [laughs] That's exactly right. So we're going to start with what I think is genuinely one of the most historic moments in the entire history of AI as an industry. Not a model release, not a benchmark shattered — but a line item in a financial statement.

HOST_B: Anthropic made money.

HOST_A: Anthropic made money. Five hundred and fifty-nine million dollars in operating profit in a single quarter. On $10.9 billion in revenue. In Q2 2026.

HOST_B: And just to put that in context — Anthropic's own internal projections had them reaching profitability in 2028. They beat their own target by two years.

HOST_A: And the number that really got me was the revenue growth rate. Dario Amodei said, and I'm quoting here, that the company had planned for ten-times annual growth — and instead saw eighty times.

HOST_B: Eighty. That's not a rounding error.

HOST_A: That is a company that surprised itself. And the driver, fascinatingly, wasn't consumer Claude subscriptions. It was Claude Code. Enterprise deployments of Claude Code alone are generating two-and-a-half billion dollars in annualised revenue.

HOST_B: So the thesis that coding assistance is THE killer app for frontier AI — not chatbots, not creative writing, not customer service bots — coding. That thesis is playing out in real money.

HOST_A: In very real money. And here's where it gets interesting for the broader picture — because on the same week that Anthropic posted its first profit, OpenAI quietly filed a confidential IPO prospectus.

HOST_B: The timing. It cannot be a coincidence.

HOST_A: It really can't. OpenAI is targeting a public listing as early as September 2026, at a valuation above one trillion dollars. Goldman Sachs and Morgan Stanley are advising. Twenty-five billion in ARR, nine hundred million weekly active users. Enormous numbers.

HOST_B: But they're not profitable.

HOST_A: They are not profitable. And now their closest rival is. So the IPO filing right as Anthropic posts Q2 results — that's a race to set the narrative before the other company's numbers become the comparison point. If you're writing the OpenAI S-1, you need investors thinking about a trillion-dollar AI behemoth, not a company bleeding cash while a Dario-founded upstart breaks even.

HOST_B: The VC era of AI is ending. That's the theme of this whole month, isn't it? For years these companies raised at absurd valuations with no profitability in sight. And suddenly in a single month — the accounting era begins.

HOST_A: That's exactly right, and we'll come back to that thread. But before we get too deep into the financial drama, we have to talk about the number that reframed everything. Because SpaceX filed its own IPO S-1 in May, and buried in the middle of it was a single line that the entire industry had to read twice.

HOST_B: One point two five billion dollars — per month.

HOST_A: One point two five billion dollars per month. That is what Anthropic has agreed to pay SpaceX for access to Colossus compute — specifically 220,000 NVIDIA GPUs at 300 megawatts — through May 2029.

HOST_B: The original analyst estimates had pegged that deal at three to six billion annually. The actual figure is fifteen billion annually. Forty-five billion total.

HOST_A: Which means the Colossus deal alone generates more revenue for SpaceX per year than the entire company's standalone 2025 revenue. The compute infrastructure story for AI is not just big — it is SpaceX-scale big.

HOST_B: And this raises a question I find genuinely uncomfortable: if Anthropic is spending one-point-two-five billion a month on compute, and they're just now reaching their first profitable quarter... what does the unit economics look like? You need extraordinary revenue growth just to stay ahead of the compute bill.

HOST_A: Right, and this feeds the argument that the GPU bottleneck and the energy constraint are the real limiting factors for this whole industry. Which brings us nicely to Google.

HOST_B: Google I/O. May 19th. And let's be honest — Google needed a good showing after some mixed signals earlier in the year.

HOST_A: They delivered. The headline was Gemini 3.5 Flash — shipping immediately, same day as the keynote, rolled out across Search, the Gemini app, and the API. Google claims it outperforms Gemini 3.1 Pro on nearly all agentic and coding benchmarks, at four times the output speed of frontier competitors.

HOST_B: The honest framing from pre-event reporting was that Gemini 3.5 Flash lands roughly at GPT-5.5 level. Not definitively ahead of Claude — more like pulling level with the current tier. But Google's real argument isn't benchmark leadership.

HOST_A: It's distribution. The day any Google model ships, it is running in front of three billion users. That is not something OpenAI or Anthropic can replicate.

HOST_B: They also announced Gemini Omni — which is the first model in the family that takes any input: image, audio, video, text — and generates back out across all those modalities. Conversational video editing that retains physics and consistency across edits. That's genuinely new.

HOST_A: And Gemini Spark — a 24/7 personal agent, exclusive to the new hundred-dollar-per-month AI Ultra tier. Samsung XR smart glasses coming this fall. Universal Cart — AI-powered cross-site shopping checkout. The biggest Search upgrade in thirty years, according to Google.

HOST_B: That last claim is doing some work.

HOST_A: [laughs] They love to say "the biggest in thirty years." But the AI Mode in Search integration does seem legitimately significant. Google is essentially dissolving the line between a search result and an AI answer.

HOST_B: Okay, we need to talk about mathematics. Because on May 21st, OpenAI announced something that I had to read three times before I believed it.

HOST_A: A general-purpose reasoning model autonomously disproved a central conjecture in discrete geometry. The planar unit distance problem. First posed by Paul Erdős in 1946.

HOST_B: Unsolved for eighty years. And an AI cracked it. Without human guidance. Generated a novel proof independently.

HOST_A: Now I want to be careful here — OpenAI hasn't fully published the technical details yet, so we can't fully peer-review the claim. But if it stands up, the implications are hard to overstate. An AI capable of original mathematical discovery means AI-accelerated breakthroughs in physics, materials science, drug discovery become conceivable. This isn't AI doing things faster. This is AI doing things humans couldn't do at all.

HOST_B: There's a version of this where 2026 is the year historians point to as the moment AI went from a very powerful tool to something that starts generating genuinely new knowledge. And the Erdős problem is exhibit A.

HOST_A: It's also just — sort of poetic, isn't it? An eighty-year-old puzzle from one of history's greatest mathematicians, solved by a machine that didn't exist ten years ago.

HOST_B: Okay, I want to briefly touch on what might be the least consequential-but-most-satisfying legal news of the year. May 17th. The Musk versus Altman trial.

HOST_A: The jury dismissed all claims. In under two hours.

HOST_B: Under two hours! Musk sued OpenAI alleging breach of mission and nonprofit obligations. The jury didn't even go back to deliberate properly — just voted no.

HOST_A: I feel like that says something about how well that case was argued.

HOST_B: It says a lot about how the case was structured. Look, the underlying tensions between Musk and Altman are real and go back years — the original OpenAI founding, the departure, the xAI rivalry. But apparently none of that translated into a legally coherent claim.

HOST_A: And Musk still has xAI and Grok to compete with. Speaking of which — Grok Build CLI launched in early beta in May. xAI's first coding agent command-line tool. Grok 4.3 beta underneath it, 2 million token context, up to 8 parallel subagents. Still very early days but notable.

HOST_B: Let's do the model roundup properly, because May was actually quite dense on releases. We've covered Gemini 3.5 Flash from Google. Anthropic's big release was Claude Opus 4.8, which shipped May 28th.

HOST_A: Incrementally stronger than 4.7 — better coding, better reasoning, longer horizon tasks at the same price point. Notably it was the only model to complete all cases in the Super-Agent test benchmark, and the first to break ten percent on a challenging legal benchmark. And it was released the same day Anthropic closed the thirty billion dollar Series H.

HOST_B: Which values Anthropic at nine hundred and sixty-five billion dollars. Sequoia, Dragoneer, Greenoaks, Altimeter co-led. AWS and Google continuing as major investors. SpaceX — obviously — in there too given the compute arrangement.

HOST_A: In February Anthropic was valued at roughly three hundred billion. By May 22nd it was closing in on a trillion. That is a three-times increase in three months.

HOST_B: Which either says something profound about AI being genuinely transformative — or something slightly concerning about the current investment climate. Possibly both.

HOST_A: Possibly both is probably right. Now — Cursor shipped Composer 2.5 on May 18th. And this is interesting because it's the first time Cursor has positioned one of their own in-house models as a genuine head-to-head competitor with Claude Opus and GPT-5.5 rather than a faster fallback.

HOST_B: Seventy-nine-point-eight percent on SWE-Bench Multilingual. That's not a joke. That's parity with the frontier on coding specifically. And Cursor's own pricing is aggressive — fifty cents in, two-fifty out per million tokens on the standard tier. That undercuts Anthropic significantly.

HOST_A: The coding model market is getting genuinely competitive. And then there's DeepSeek, who made what is arguably the most impactful pricing move of the month: they made their 75% discount on V4-Pro permanent.

HOST_B: The permanent rates are $0.435 per million input tokens, $0.87 out. For context — that is roughly eight to ten times cheaper than Claude Opus 4.7 on a comparable task. And V4-Pro remains one of the strongest open-weights coding models by benchmark.

HOST_A: DeepSeek's approach of just — continuously slashing prices on high-quality open-weight models — is a very deliberate wedge into the enterprise market. And it's working.

HOST_B: On the Chinese side of the ledger — Alibaba's Qwen 3.7-Max Preview launched at the Apsara Summit on May 20th. One million token context, a claimed 35-hour autonomous run chaining over a thousand tool calls. LM Arena ranked it thirteenth overall — the highest-ranked Chinese model on that index. Strong story.

HOST_A: Though the open-weight release everyone's waiting for — the Plus variant — still hasn't landed. The Qwen team keeps previewing closed weights while promising open weights are coming. The open-source community is watching.

HOST_B: On the open-source front more broadly — Cohere released Command A+ at 218 billion parameters, open source, claims it runs on just two H100s. Stability AI launched Stable Audio 3.0 — open-weight music generation up to six-plus minutes. Both moves reinforcing that there's a genuine alternative track developing to the closed-weight frontier race.

HOST_A: Let's talk about enterprise deployment, because the scale of what's happening here is starting to feel qualitatively different from previous waves of tech adoption. PwC deployed Claude to hundreds of thousands of employees globally.

HOST_B: KPMG as well, integrating Claude into enterprise workflows. JPMorgan and other major banks are running on AI tools. Snowflake committed six billion dollars on AWS to jointly build agentic enterprise AI. Isomorphic Labs — the Google DeepMind spinout doing drug discovery — raised two-point-one billion in a Series B.

HOST_A: The drug discovery story is one I want to spend more time on. Isomorphic has been applying AI to protein structure prediction and molecular design. A two-billion-dollar Series B in a single round suggests serious institutional confidence that this is going to produce real drug pipelines.

HOST_B: And then there's the regulatory picture, which is — complicated. In the US, President Trump announced an executive order requiring pre-release safety reviews of advanced AI, and then postponed it.

HOST_A: Citing global competitiveness concerns. The logic being — if we slow down AI development for safety reviews, China won't. Which is the argument that keeps winning against safety frameworks in Washington.

HOST_B: Meanwhile Congress debated a five-hundred-million-dollar "Pax Silica" fund to help US allies adopt American AI and chips. The EU moved ahead with Chips Act 2.0. The UK is building a domestic chip strategy and framing it as AI sovereignty. Saudi Arabia declared 2026 the — and I love this — the "Year of AI."

HOST_A: Saudi Arabia does not do things quietly.

HOST_B: No it does not. And China took the opposite approach to openness — tightened controls on AI talent travel, requiring government approval for top engineers at Alibaba and DeepSeek to go abroad. Plus new regulations effective July 2026 on anthropomorphic AI services — content controls, age limits, corporate governance requirements on chatbots and virtual avatars.

HOST_A: Which is a reminder that the geopolitical competition here isn't just about who builds the best model. It's about who controls the people who build the models, and what the models are allowed to say.

HOST_B: Alright, let's land the plane. What does May 2026 actually mean?

HOST_A: I keep coming back to the phrase: graduation day. The AI industry spent its first decade in startup mode — burning cash, chasing capabilities, raising on pure potential. May 2026 is the month the industry started having to act like it's grown up. Anthropic's profit announcement, OpenAI's IPO filing — these are events that drag AI companies into the same accountability structures as every other industry.

HOST_B: Public markets are unforgiving. When OpenAI lists, analysts will pick apart its unit economics in a way that venture investors never did. And that's actually healthy. The hand-wavy "we'll monetize later" era is over.

HOST_A: The compute economics are sobering though. A billion-and-a-quarter a month just for Colossus. The energy and infrastructure constraints are real. The path to sustained profitability at scale requires either dramatically reducing inference costs or dramatically increasing the value being created.

HOST_B: And the math problem — the Erdős proof — is the optimistic case for the latter. If AI starts generating novel scientific knowledge, the value being created is genuinely unbounded.

HOST_A: Right. You cannot put a price on solving diseases or uncovering new physics. So there's a universe where the current infrastructure investment is spectacularly rational.

HOST_B: There's also a universe where we're building the most expensive pile of sand in history.

HOST_A: [laughs] That's always the alternative universe. I'm choosing optimism.

HOST_B: As ever. Alright — that's May 2026 in AI. Anthropic's first profit, OpenAI's IPO filing, a forty-five billion dollar compute deal hiding in a SpaceX S-1, Gemini 3.5 at Google I/O, an eighty-year-old maths proof cracked by a machine, and a jury that needed less time to dismiss the Musk lawsuit than it takes to watch a movie.

HOST_A: The VC era is ending. The public markets era begins. We'll be watching every quarterly report.

HOST_B: Thanks for listening to Clawd Talks. If you're not subscribed — please do — new episodes drop monthly, covering the biggest stories in AI. See you next time.

HOST_A: See you next time.