May 26 Roundup: OpenAI files for IPO, Anthropic eyes first profit, Nvidia cedes China to Huawei, and AI cracks an Erdős problem
The AI industry's financial story took center stage this week: OpenAI races toward a September stock market debut worth hundreds of billions, Anthropic quietly crosses into the black for the first time, and Nvidia's CEO admits the China AI chip fight is already over. Meanwhile, an 80-year-old mathematics conjecture fell to an AI reasoning model — a milestone that redefines what "AI-generated knowledge" actually means.
1. OpenAI Races to September IPO — Potentially the Largest in Tech History
OpenAI is preparing to confidentially file a draft IPO prospectus as soon as this week, working with Goldman Sachs and Morgan Stanley on what could be one of the largest public market debuts in history. The company, valued at over $850 billion by private investors, is targeting a public debut as early as September 2026, according to reporting confirmed by CNBC, Reuters, Bloomberg, and the Wall Street Journal.
"As part of normal governance, we regularly evaluate a range of strategic options. Our focus remains on execution." — OpenAI spokesperson
OpenAI CFO Sarah Friar had previously signaled the move, telling CNBC that it is "good hygiene" for a company of OpenAI's size to "look and feel and act like a public company." The confidential filing gives OpenAI time to refine its prospectus before entering a mandatory public review window. Reuters reported that Musk v. Altman trial proceedings — which sought to relitigate OpenAI's nonprofit origins — have largely concluded, clearing the legal path forward. With annualized revenue estimates in the hundreds of billions and a ChatGPT user base rivaling Google's Gemini at roughly 900 million, the IPO math is compelling on paper. The harder question is whether public-market investors are willing to price in billions in annual compute spending and the existential uncertainty of a race to AGI.
An OpenAI IPO isn't just a liquidity event — it's a governance reset. Going public forces the company to be more explicit about what "pursuing AGI for the benefit of humanity" means in quarterly earnings language. For enterprise buyers, a public OpenAI actually reduces some risk: disclosure requirements, audited financials, and board accountability create a level of transparency that a $850B private company never had. But expect volatility. The compute cost structure is brutal, and the first quarter of disappointing growth numbers will be a blood sport. For businesses evaluating OpenAI as a long-term platform bet, the IPO is a green light to deepen the relationship — the company isn't going anywhere.
Source: CNBC · Reuters · Bloomberg
2. Anthropic Is About to Hit Its First Profitable Quarter — Revenue Doubling to $10.9 Billion
Anthropic has told investors it will more than double revenue to approximately $10.9 billion in Q2 2026, delivering an operating profit for the first time in the company's history, according to the Wall Street Journal. This milestone, which dropped the same day OpenAI announced its IPO preparations, dramatically reshapes the competitive landscape. TechCrunch and Bloomberg confirmed the figures, noting they were shared as part of a recent funding round update.
"Anthropic is on pace for its first profitable quarter after experiencing a surge in revenue driven by demand for its artificial intelligence software." — Bloomberg
The profitability, however, may not be durable. The WSJ noted that Anthropic is scheduled to incur massive compute costs later in the year — including a reported SpaceX Colossus deal adding hundreds of megawatts of capacity — which could push it back into the red in Q3 and Q4. Anthropic's Q2 revenue surge is being driven by explosive enterprise adoption of Claude. Recent expansions include Claude for Small Business (with QuickBooks, PayPal, and HubSpot connectors), new Managed Agents features, a PwC and KPMG alliance for professional services, a SpaceX deal for compute, and a Gates Foundation partnership for public health AI. Anthropic has also topped CNBC's 2026 Disruptor 50 list at No. 1, leapfrogging OpenAI.
In a separate significant move, Anthropic acquired API tooling company Stainless on May 18, bolstering its developer infrastructure and signaling that it intends to own more of the developer toolchain stack beyond the model layer itself.
Anthropic's first profitable quarter is a signal, not a destination. The deeper story is that enterprise trust in Claude has crossed a threshold: professionals — lawyers, developers, analysts — are expressing a clear preference for Claude over ChatGPT for high-stakes work, and that's translating directly into enterprise contract growth. The Stainless acquisition is smart. SDK quality is a real moat: developers who build on top of well-designed APIs don't switch. The volatility risk is real too — $10.9B in Q2 followed by potential losses in H2 will test investor confidence, but for Anthropic's strategic position, the profitability milestone matters more than its duration. It proves the unit economics can work.
Source: TechCrunch · Bloomberg · CNBC
3. OpenAI's Reasoning Model Solves an 80-Year-Old Math Problem — Autonomously
In one of the most striking AI research demonstrations to date, an internal OpenAI reasoning model has disproved a central conjecture in discrete geometry that had stood unsolved for nearly 80 years. The problem — the planar unit distance problem, first posed by legendary Hungarian mathematician Paul Erdős in 1946 — asks how many pairs of points placed in a plane can be exactly distance 1 apart. Since Erdős's work, mathematicians had believed that "square grid" constructions were essentially optimal.
"An internal OpenAI model has disproved this longstanding conjecture, providing an infinite family of examples that yield a polynomial improvement. The proof has been checked by a group of external mathematicians." — OpenAI blog
What makes this especially significant is how the proof was found. OpenAI did not train a system specifically for mathematics, scaffold it with special proof-search strategies, or target the unit distance problem specifically. A general-purpose reasoning model, evaluated across a collection of Erdős problems as part of a broader frontier research capability test, simply solved it. External mathematicians verified the proof and published a companion paper explaining the argument. The Guardian noted that Noga Alon, a leading combinatorialist at Princeton, described the Erdős unit distance problem as "one of Erdős' favorite problems" — one he offered a monetary prize to resolve. Per Gigazine: "AI has gone beyond being just an assistant."
This isn't just a parlor trick. Solving an 80-year-old open problem in pure mathematics — autonomously, without domain-specific scaffolding — is a qualitative capability threshold. It means current frontier reasoning models are now operating at the edge of human mathematical knowledge, not just behind it. For enterprise leaders, the practical implication isn't "AI will do your math homework." It's that the same depth of sustained logical reasoning that cracked an Erdős problem is available for complex contract analysis, multi-step financial modeling, and rigorous system design. The research gap between AI and human experts in structured reasoning domains is closing faster than most enterprise planning horizons assume.
Source: OpenAI · The Guardian
4. Nvidia CEO: China's AI Chip Market Is "Largely Conceded" to Huawei
Nvidia CEO Jensen Huang made a remarkable admission on May 21: the company has "largely conceded" China's advanced AI chip market to Huawei. Huang made the comments in a CNBC interview discussing Nvidia's latest earnings, deploying data centers into space, and the evolving competitive dynamics in China. The statement follows years of escalating U.S. export controls that have progressively blocked Nvidia's most capable chips from reaching Chinese customers.
"Nvidia CEO Jensen Huang said the company has 'largely conceded' China's advanced artificial intelligence chip market to Huawei." — CNBC
The implications are significant. Huawei is projected to generate $12 billion in AI chip revenue in 2026, up from approximately $7.5 billion in 2025, based on orders already in hand. Beijing has been actively encouraging domestic firms to purchase Huawei AI chips as a tactical move to reduce vulnerability to future U.S. export controls. For Nvidia, the loss of China's AI chip market removes a substantial growth vector even as the company remains dominant globally. Separately, CNBC reported that Trump administration AI policy discussions have pivoted toward a possible federal licensing regime for AI models — a significant shift from the prior stance of opposing regulation — amid mounting public concern about frontier model capabilities.
Jensen Huang's concession reflects a new geopolitical reality: the AI chip market has effectively bifurcated into two ecosystems. Nvidia dominates outside China; Huawei dominates within it. For enterprise leaders with global supply chains or China operations, this matters enormously for AI infrastructure planning. Any AI system that needs to run in China — whether for compliance, latency, or data sovereignty reasons — will increasingly run on Huawei hardware, with different driver stacks, different toolchains, and different security profiles. Building genuinely portable AI workloads will require architectural discipline that many organizations haven't invested in yet. This is a long-term strategic design constraint, not a short-term procurement footnote.
Source: CNBC · Startup Fortune
5. Anthropic's Mythos Model Is Reshaping Cybersecurity — and the Job Market
Anthropic's Mythos AI model, which the company has declined to release publicly after discovering it has advanced capabilities in identifying previously unknown software vulnerabilities, is dominating national security discussions about AI. The model — alongside OpenAI's GPT-5.5 to a lesser degree — is described by researchers in controlled settings as a "game-changer" for its ability to find and potentially exploit software flaws.
"Anthropic has declined to release Mythos publicly after announcing the AI model had advanced capabilities in highlighting previously unknown flaws in IT systems that could be used by hackers." — The Guardian
Anthropic is working closely with the U.S. government to advance "shared priorities" and has committed to sharing Mythos cyber vulnerability findings with the Financial Stability Board (FSB), the global finance watchdog. The company is operating Mythos in a controlled access model, giving vetted security teams the ability to scan codebases, triage vulnerabilities, and generate fixes while contributing to open-source defense research under Project Glasswing.
The paradox driving policy debate: the same capabilities that make Mythos valuable for defense make it dangerous in the wrong hands. Initial fears of "unfettered hacking" prompted by Mythos have been described as overstated by Reuters analysts, who note that finding vulnerabilities is only the first step — exploitation requires additional technical knowledge and access. Nevertheless, the New York Times reported that cybersecurity is now one of the few job categories growing in the AI era, with demand for security engineers surging as AI generates more code, more surface area, and more urgency around who can find and fix the flaws first.
Mythos represents a dual-use AI capability problem that won't be resolved by containment alone. The skill set needed to exploit a vulnerability discovered by Mythos isn't particularly specialized — any motivated adversary with access to the findings is closer to exploitation than they would have been before. For businesses, the practical implication is clear: AI-assisted vulnerability discovery is now a baseline capability that attackers will eventually access through open models, leaked tools, or nation-state proxies. Enterprises that wait for a security incident to invest in AI-assisted defense are accepting a timing disadvantage they may never fully recover from. The security teams building Mythos-grade capabilities into their workflow now will be ahead by years, not months.
Source: The Guardian · Reuters · New York Times
6. Google I/O 2026: Gemini Becomes the Operating Layer — 900 Million Users, New Models, Agentic Search
Last week's Google I/O 2026 (May 19–20) delivered a clear strategic signal: Google is turning Gemini from a chatbot into an operating layer for AI. The company announced that Gemini users have more than doubled in one year to 900 million — on par with OpenAI's self-reported figures. The company unveiled the Gemini 3.5 series (with Gemini 3.5 Flash already generally available at frontier-competitive speed and cost), the Gemini Spark personal agent designed to run autonomously around the clock, and the deepest change to Google Search in the product's 25-year history.
"Google unveiled a new model family, a personal agent that runs around the clock and an intelligent shopping cart, alongside what it called the deepest change to Search in the product's 25-year history." — Forbes
Gemini 3.5 Flash — generally available as of May 19 — delivers frontier-level intelligence at 4× the speed of comparable models, priced at $1.50/$9 per million input/output tokens with a 1 million token context window. Gemini 3.5 Pro is being used internally but won't reach wider distribution until June. Google also launched Managed Agents in the Gemini API in public preview, enabling developers to build and deploy autonomous, stateful agents running in secure, isolated Google-hosted Linux sandbox environments. The new search box — the first dimensional change in 25 years — uses Gemini to answer longer, more complex queries and now includes a video-generation tool and simplified AI shopping. Google's Antigravity agent-first development platform received upgrades to orchestrate and build agents at scale.
Google's 900 million Gemini users number means the AI assistant race is effectively a two-horse competition between Google and OpenAI, with Anthropic playing the enterprise specialist role. The more strategically interesting signal from I/O is Managed Agents in the Gemini API: stateful, sandboxed agents running in Google infrastructure means Google is now in direct competition with every enterprise orchestration platform, from LangGraph to AutoGPT to homegrown agent frameworks. Any business building a multi-agent architecture today should be modeling the "what if Google is the runtime" scenario seriously. Gemini Spark as a 24/7 personal agent is early, but it's the first credible Google push into ambient AI — the territory Peter Diamandis calls the most transformative horizon for the next decade.
Source: AP News · TechCrunch · Forbes
7. EU AI Act Full Compliance Clock Is Ticking — 70 Days to August 2 Deadline
With August 2, 2026 approaching as the date when the EU AI Act's high-risk obligations under Articles 8–17, 26, 27, and 73 fully activate, enterprises operating in Europe are facing a compliance crunch. The AI Act entered into force on August 1, 2024, with a two-year full-applicability window. While prohibited AI practices and AI literacy obligations have already been in effect since February 2025, and governance rules for general-purpose AI models have been running since August 2025, the remaining high-risk provisions are the most operationally demanding. According to DigitalApplied.com's compliance guide, "most enterprises building AI agents are not ready."
"EU AI Act high-risk obligations — Articles 8–17, 26, 27, and 73 — activate August 2, 2026: 71 days from publication. Most enterprises building AI agents are not ready." — DigitalApplied.com
High-risk AI use cases under the Act — including AI in employment decisions, credit scoring, healthcare triage, critical infrastructure, and law enforcement — require robust risk management frameworks, technical documentation, human oversight mechanisms, and transparency to users. Meanwhile in the U.S., the White House postponed an executive order on May 20 that would have established a voluntary review process for AI models before release, with President Trump citing aspects he "didn't like." The administration continues to favor a deregulatory approach, while states — particularly Colorado — are pushing their own AI regulation frameworks. Colorado enacted a revised AI law this month, regulating "automated decision-making technology in consequential decisions" rather than framing requirements around "high-risk AI systems."
The EU AI Act's August 2 full-compliance deadline is real, not theoretical. For any organization deploying AI systems in Europe that touch employment, credit, healthcare, or critical infrastructure decisions, the clock is running. The gap between "we have an AI strategy" and "we have EU AI Act-compliant AI systems" is a documentation, audit trail, and human-oversight problem — not a technology problem. Companies that haven't started their compliance documentation now are looking at a very uncomfortable summer. For U.S.-headquartered companies with European operations, the practical advice is blunt: treat August 2 as a hard deadline, not a grace period, and start with your highest-risk use cases first. The Colorado precedent also signals that even in the deregulatory U.S. environment, state-level AI compliance requirements are accumulating faster than most legal teams are tracking.
Source: European Commission · DigitalApplied · CNN
Why This Week Matters for Your Business
Three interconnected forces are converging in the second half of May 2026. First, the AI industry's financial infrastructure is being formalized: IPOs, profitability milestones, and public market accountability will make AI platforms more stable and more scrutinized simultaneously. This is good news for enterprise buyers. Second, the capability frontier is genuinely advancing — not incrementally, but in jumps. A general-purpose model cracking an 80-year math problem, Mythos-grade vulnerability discovery, Gemini 3.5 Flash at frontier speed — these are not marketing claims. They are new baselines. Third, the regulatory environment is bifurcating sharply between the EU's structured compliance timeline and the U.S.'s deregulatory drift, creating real compliance complexity for any organization that operates across both jurisdictions.
The organizations that will navigate this best are the ones treating AI infrastructure decisions as strategic architecture choices — not experiments. The window between "AI is a pilot" and "AI is operational infrastructure" is closing. The companies on the right side of that window in 2026 will have compounding advantages in 2027 and beyond.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy — from EU AI Act compliance to enterprise agent architecture.
Contact SEN-X →