# Ambient Advantage — May 25, 2026 *Monday · May 25, 2026 · [Episode page](https://podcast.ambient-advantage.ai/episodes/2026-05-25.html) · [Audio](https://storage.googleapis.com/ambient-advantage-podcast/2026-05-25-ambient-advantage.mp3)* [AVA] Anthropic just told investors it expects its first ever profitable quarter — half a billion dollars in operating income. But the accounting behind that number? Let's just say it deserves a very close read. [JON] Oh, we're going there. Welcome to Ambient Advantage — I'm Jon, and this is Ava. It's Monday, May 25, 2026, and here's what matters in AI today. We've got Anthropic's contested profit milestone, Cursor's jaw-dropping revenue ramp, a Trump executive order that could delay your next model upgrade, Starbucks pulling the plug on an AI tool that couldn't tell oat milk from whole milk, and a bigger picture argument about why the real AI battle has shifted from models to infrastructure. Ava, let's start with the lead. [AVA] So Anthropic disclosed to investors last week that it expects Q2 2026 revenue of ten point nine billion dollars. That's more than double Q1's four point eight billion. And they're projecting their first operating profit ever — roughly five hundred and fifty-nine million dollars. [JON] That's a massive number. First profitable quarter for a frontier AI lab. That should be huge news, full stop. [AVA] It should be. And in some ways it is. This is the first credible evidence that building frontier AI models can actually generate operating income, not just burn through investor capital. For the industry narrative, that matters enormously. [JON] But you said "should be." What's the catch? [AVA] The catch is timing. This profit window coincides with a discounted ramp-up period on Anthropic's massive compute contract with SpaceX — one point twenty-five billion dollars a month. During the ramp phase, Anthropic isn't paying full freight on compute. Once Q3 hits and full costs kick in, that profit could evaporate. [JON] So the profit is real in a technical accounting sense, but it might not be structurally sustainable? [AVA] Exactly. There's a sharp critical analysis I'll drop in the show notes that walks through the accounting contradictions methodically. My advice to anyone preparing a board deck or a vendor evaluation — do not cite these numbers without reading that piece first. [JON] Now this landed the same week OpenAI filed confidential IPO paperwork. That's not a coincidence, right? [AVA] Not remotely. OpenAI generated about five point seven billion in Q1, roughly a billion ahead of Anthropic in the same period. Both companies are now in an arms race to prove financial viability before going public. Anthropic's profit disclosure is as much a signal to investors and enterprise buyers as it is a financial milestone. The subtext is: we're not just a research lab, we're a real business. [JON] So what should enterprise buyers actually take away from this? [AVA] Two things. First, financial staying power matters when you're committing core workflows to a vendor's API. Anthropic showing any profit, even a contested one, is better than perpetual losses. Second, scrutinise your dependency. If Anthropic's economics shift dramatically in Q3 when compute costs normalize, that could affect pricing, capacity, and roadmap decisions that ripple into your stack. Ask your account team hard questions now, not in October. [JON] Clear. Alright, let's move into the rundown. We've got a handful of stories to get through quickly. Ava, start us off. [AVA] Cursor hit three billion dollars in annualized revenue in late April. Up from two billion just two months earlier. They now have over three thousand enterprise customers paying at least a hundred thousand dollars a year, and sixty-four percent of Fortune 500 companies use the tool. SpaceX holds the right to acquire Cursor for sixty billion after its June 12 IPO. [JON] A million to three billion ARR in roughly three years. Is that the fastest enterprise software ramp ever? [AVA] It is. It dwarfs Salesforce's decade-long climb to the same milestone. The signal for enterprise leaders is that agentic coding — what people used to dismissively call vibe coding — is no longer experimental. It's becoming critical developer infrastructure. And there's a smart analysis I'll link in the show notes arguing SpaceX doesn't just want Cursor's revenue, it wants to own the developer interface layer. The last mile where AI meets code. [JON] Alright, next up — the White House. Trump signed an AI executive order. [AVA] He brought in the CEOs of Alphabet, Amazon, Meta, Microsoft, OpenAI, Anthropic, Apple, and SpaceX. The order establishes a voluntary framework — key word, voluntary — requiring labs to share new frontier models with the government ninety days before public release for national security testing. [JON] Ninety days is a long time in AI. What does that mean practically? [AVA] It could delay model availability. It creates a new compliance layer. And critically, it gives labs that already have deep government relationships — OpenAI and Anthropic primarily — an asymmetric advantage over challengers and open-source alternatives who may not have the Washington infrastructure to navigate this. For enterprise buyers building on frontier APIs, model release timelines just got less predictable. [JON] Speaking of security, Anthropic's Mythos model had quite a week. [AVA] It did. Claude Mythos Preview, their agentic cybersecurity model, has autonomously discovered over ten thousand critical vulnerabilities across various AI systems and infrastructure, including a serious FreeBSD remote code execution flaw. The NSA is reportedly already using it. Anthropic is briefing the Financial Stability Board. [JON] So AI is now better at finding security holes than human red teams? [AVA] At scale, yes. Your next penetration test might be cheaper, faster, and more thorough if run by an agent. But here's the uncomfortable flip side — adversaries have access to similar capabilities through open models. The implication for CISOs is blunt: patching speed just became your single most important security metric. It's no longer about whether someone finds the vulnerability, it's about how fast you close it after a model already has. [JON] Now for a story that I think every enterprise AI team should pin to their wall. Starbucks. [AVA] Oh, this one stings. Starbucks deployed an AI-powered inventory counting tool across all eleven thousand-plus North American stores. Nine months later, they've pulled the plug. The reason? It couldn't reliably distinguish between types of milk. Oat milk, whole milk — too visually similar for the system. [JON] Nine months. Eleven thousand stores. And it couldn't tell milk apart. [AVA] An internal memo confirmed employees are going back to manual counting. The lesson isn't that AI can't do inventory. It's that pilots in controlled lab conditions are a terrible proxy for messy operational reality. Real stores have inconsistent lighting, different carton designs by region, worn labels... The system presumably passed benchmarks. It just couldn't handle the real world. [JON] Alright, one more for the rundown. DeepSeek made a big pricing move. [AVA] DeepSeek permanently cut V4-Pro API pricing by seventy-five percent, enabled by deep optimization for Huawei Ascend 950 chips. This is a direct workaround to U.S. GPU export controls. And they've warned prices could drop further once Huawei supernodes roll out at full scale in the second half of this year. [JON] So a frontier model from China that doesn't depend on Nvidia hardware, and it's permanently cheaper. [AVA] It structurally lowers the cost floor for AI inference globally. For enterprise procurement teams, this is immediate leverage. If you're in contract renewals with OpenAI, Anthropic, or Google, you now have a pricing benchmark from a capable competitor that's moving in one direction — down. Use it. [JON] Let's step back now. The bigger picture. Ava, you've been connecting these dots all morning. What's the thread? [AVA] The thread this week is one most coverage is missing. We are entering what I'd call the infrastructure ownership era of AI. The decisive competitive battles are no longer about which model scores highest on benchmarks. They're about who controls the compute stack, the developer interface, and the enterprise data layer. [JON] Unpack that with this week's stories. [AVA] Look at the board. Anthropic's one point twenty-five billion dollar monthly SpaceX compute deal — that's an infrastructure lock-in play. Cursor's sixty billion dollar SpaceX acquisition option — that's controlling the developer interface. Cerebras raising the size and price of its IPO this week, driven by conviction that agentic AI needs fundamentally different compute — that's the chip landscape diversifying beyond Nvidia. DeepSeek optimizing for Huawei chips — that's China building a parallel infrastructure stack. These are all moves on the same board. [JON] So what does that mean for the enterprise buyer sitting in a conference room trying to make smart decisions? [AVA] It means you are not just choosing an AI vendor. You are choosing an infrastructure consortium. And switching costs will compound faster than most procurement cycles can accommodate. A Stanford economist flagged this exact point last week — calling it the weak link problem. The bottleneck to AI ROI is rarely the model. It's the surrounding infrastructure: data quality, integration layers, change management, legacy systems. [JON] So the organizations that map their dependencies now... [AVA] Will have materially more negotiating leverage in twelve months than those who wait until renewal season. Map your compute dependencies. Map your API dependencies. Map your developer tooling dependencies. Do it this quarter, not next year. [JON] Alright, what should people be watching this week? [AVA] Two things. First, the Cerebras IPO could price as early as this week at a hundred fifty to a hundred sixty dollars a share — up significantly from the original range. That's a real-time market signal on investor conviction in agentic compute. Second, watch for OpenAI's confidential IPO filing. With the Musk lawsuit dismissed last week — jury found it was past the statute of limitations — the legal cloud is gone. That filing could drop any day, and it'll reshape the competitive narrative heading into summer. [JON] Two IPOs, one week. The AI industry is growing up fast. [AVA] Growing up, and consolidating. That's the story of 2026. [AVA] That's your Ambient Advantage for Monday, May 25, 2026. [JON] Share it with a colleague figuring out what AI means for their business. See you tomorrow.