Wave Theory Notes: Anthropic's $900 Billion Approach
Anthropic lines up a $30B raise at a $900B+ valuation with KPMG and PwC behind it, Google I/O drops the premium AI subscription floor by 20%, and OpenAI's reasoning model cracks an 80-year-old Erdős conjecture. (May 18-24, 2026)
This week the through-line was distribution. Anthropic lined up a $30 billion raise at a $900 billion valuation while quietly placing Claude inside KPMG’s 276,000 staff and PwC’s 364,000-person workforce. Google used I/O 2026 to push Gemini into search, Android, Workspace, and YouTube at the same time, and cut its top consumer plan by 20 percent. OpenAI started lining up a trillion-dollar IPO and, in a different press release the same week, said one of its internal models had cracked a math problem that had been open since 1946. The frontier is no longer just bigger models. It is who can reach the most desks and the most workflows fastest.
Hey Wave Riders!
Welcome back. This issue covers May 18 through 24, the busiest seven days I have written up since starting the newsletter. Grab a coffee.
Anthropic’s $900 Billion Approach
The headline number is that Anthropic is closing a roughly $30 billion primary round at a post-money valuation above $900 billion, with Sequoia, Dragoneer, Altimeter, and Greenoaks expected to co-lead at about $2 billion each. The round is reported to close as soon as the week of May 25. If it prints at that price, Anthropic enters the conversation about trillion-dollar private companies less than four years after launching its first commercial product.
The valuation is being underwritten by three things that all moved this week. The first is revenue. Anthropic told investors it expects $10.9 billion in Q2 revenue, up 130 percent from $4.8 billion in Q1, and the company guided to its first ever operating profit of $559 million in the quarter. Ramp’s May AI Index showed that, for the first time, more US businesses pay for Anthropic than for OpenAI: 34.4 percent versus 32.3 percent, with Claude Code the main driver. Claude Code now authors about 4 percent of all public commits on GitHub, double its share from a month earlier.
The second is distribution. Anthropic signed a global alliance with KPMG to roll Claude out to all 276,000 KPMG employees and embed it inside the KPMG Digital Gateway, starting with tax and legal. Two days later, Anthropic and PwC expanded their own partnership, with PwC training roughly 30,000 staff on Claude tools and a broader rollout planned across its 364,000-person global workforce. Two of the Big Four professional services firms have now committed to making Claude the default AI surface for their consulting and audit teams. That is the kind of contract structure investors price as recurring revenue, not seat licenses.
The third is talent. Andrej Karpathy, OpenAI co-founder and one of the most cited names in modern deep learning, joined Anthropic’s pre-training team. The signal matters more than any single hire. Anthropic has been pulling senior researchers from rivals for the better part of a year, and the names willing to publicly cross the floor keep getting bigger. Combine that with a separately reported $1.25 billion-per-month, $45 billion compute commitment to SpaceX’s Colossus clusters through 2029, and you have a company that looks well capitalized for the next training cycle and the next two after it.
My read: this is the week the enterprise tier of the AI market stopped being a thesis and became a fact. OpenAI is still the largest consumer brand by a wide margin, and that matters for a different scoreboard. But if you sell into the Fortune 500, your shortlist for the next twelve months is now Anthropic, Google, and Microsoft, in that order. The interesting question is what OpenAI does in response. The IPO filing news later in the week is one answer. It is unlikely to be the only one.
What to watch next: whether the round prices at exactly $900 billion or quietly clears a trillion, and whether any other Big Four firm signs before the end of June.
Google I/O 2026 Drops the Pricing Floor
Google used I/O to make Gemini the default intelligence layer across its consumer surface. Gemini 3.5 Flash is now generally available at $1.50 per million input tokens and $9 per million output tokens, with a 1 million token context window and about 4 times the speed of the prior Flash. Gemini Omni, a native multimodal model that accepts image, audio, video, and text and outputs grounded video, is rolling out to the Gemini app, Flow, and YouTube Shorts. Gemini Spark, a personal agent that runs on cloud VMs and takes actions inside Gmail, Docs, and Workspace, ships next week to AI Ultra subscribers in the United States.
The move I think people will still be talking about in a month is the pricing. Google cut the top AI Ultra plan from $250 to $200 a month, a 20 percent reduction, and introduced a $100 entry tier. OpenAI Pro and Anthropic’s Max plans sit at $200 and $200 respectively. Google has just told the market that the ceiling on a premium consumer AI subscription is $200 and the floor is $100. Everyone else now has to either match that or justify a premium with a meaningfully different product.
For anyone running an AI budget at work, this is the week to redo your seat-cost spreadsheet. Per-seat economics on the premium tier are going to compress through the back half of the year. The vendor pitches will follow within a quarter.
OpenAI’s Two-Track Week: $1T IPO Prep and an 80-Year Math Problem
OpenAI began the week preparing to confidentially file its IPO prospectus with the SEC, targeting a valuation above $1 trillion, with Goldman Sachs and Morgan Stanley leading. The public listing window is reportedly between Labor Day and Thanksgiving. The same week, OpenAI also launched a self-serve Ads Manager inside ChatGPT for advertisers and is reportedly targeting $2.5 billion in ad revenue this year. A federal jury in Oakland separately rejected every claim Elon Musk brought against OpenAI and Sam Altman, after under two hours of deliberation following eleven days of testimony.
The other OpenAI story this week is the one I keep coming back to. An internal general-purpose reasoning model produced a result that disproves the planar unit distance conjecture, a problem Paul Erdős posed in 1946 about the maximum number of unit-distance pairs among n points in a plane. For nearly eighty years, mathematicians had assumed the square grid arrangement was essentially optimal. The model found a new infinite family of constructions, reached for through deep algebraic number theory and infinite class field towers, that gives a polynomial improvement. Nine independent mathematicians verified the work and co-authored a companion remarks paper.
Two things make this matter beyond the math department. First, the model that did this was a general-purpose reasoning system, not a math-specialized one. That is a different claim than what DeepMind made with AlphaProof, and it lands closer to the future that Jack Clark described at Oxford this week when he predicted AI will co-author a Nobel-worthy discovery within a year. Second, it is the first time a model has resolved a named open problem in pure mathematics under its own steam. The pace at which this kind of result starts to appear in chemistry, biology, and physics is now a fair thing to forecast.
Trump Postpones the AI Executive Order
Hours before signing, President Trump pulled an executive order that would have required pre-launch security evaluations for frontier AI models, saying he did not want regulation to slow the US lead over China. The order would have built on the Center for AI Standards and Innovation framework I wrote about last week and formalized the pre-deployment review process for any high-risk frontier release.
The practical effect is that the voluntary CAISI agreements remain the operative regime, at least for now. The political reading is that the administration is willing to talk publicly about AI risk and federal review and then defer the binding step when the calendar gets close. If you are building anything that will eventually have to clear a frontier-model review, the window stays open longer than the press from last week suggested. Watch for whether Europe moves first instead.
Quick Hits
Meta laid off roughly 7,800 employees, about 10 percent of its workforce, with leaked all-hands audio capturing Mark Zuckerberg discussing training AI models on the work of staff who were being cut. NextEra Energy agreed to acquire Dominion Energy for $67 billion in stock, the largest US energy deal since 1998, with the merger framed around AI power demand expected to double by the late 2030s. Iren signed a $3.4 billion, five-year AI cloud contract with Nvidia to run Nvidia’s own internal workloads, underwriting Iren’s plan to deploy up to 5 gigawatts of GPU capacity. Nvidia put its Vera Rubin platform into full production with seven new chips, including the Rubin GPU, Vera CPU, NVLink 6, and an integrated Groq 3 LPU, with AWS, Google Cloud, Microsoft, and OCI as the first deployers. Brett Adcock’s new AI hardware startup Hark raised a $700 million Series A at a $6 billion valuation, with Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures, Salesforce Ventures, and Brookfield in the round. Microsoft AI chief Mustafa Suleyman said publicly that most white-collar computer work will be automated inside 12 to 18 months, naming accounting, legal, marketing, and project management as the first to feel it. Salesforce launched Agentforce Coworker in beta, an AI teammate embedded inside agent workflows that can pull CRM context and act on the user’s behalf. Anthropic and the Gates Foundation announced a $200 million, four-year partnership to develop AI tools for healthcare, education, agriculture, and economic development in underserved regions.
This Week’s Template: The 5-Part AI Brief
Three of this week’s stories involve organizations rolling AI out to hundreds of thousands of employees. The single biggest predictor of whether those rollouts work is not the model. It is whether the people pressing Enter know how to write a brief instead of typing a question. This is the template I lean on hardest, and it is the one I promise viewers in the Friday TikTok every Week 1 of the cycle.
The 5-Part AI Brief Template
Most people type a question and hit send. Professionals write a brief. Paste this structure at the top of any complex prompt and you have done 80 percent of the work before you type a word of the actual request.
ROLE
Who is the AI for this task?
Example: “You are a senior copywriter who specializes in B2B SaaS landing pages.”
AUDIENCE
Who is reading or using the output?
Example: “The audience is mid-level marketing managers at companies with 50 to 500 employees. They are analytical and skeptical of hype.”
TONE
How should it sound?
Example: “Direct and confident. No filler phrases. Short paragraphs. No bullet points unless they genuinely aid clarity.”
CONSTRAINTS
What to avoid. What to include.
Example: “Avoid generic advice, passive voice, sentences over 20 words. Include one concrete example per section and a clear next action at the end.”
SUCCESS CONDITION
How will you know the output is right?
Example: “This is ready when I could hand it to a colleague and they would immediately understand what to do and why.”
The people getting the most out of AI are not better at prompting. They are better at briefing. That is a professional skill. It is learnable, and you can install it across a team in an afternoon.
Wrapping the Waves
The thread running through this week is that the AI industry is starting to settle into shapes that look durable. A two-horse frontier race at the top. A handful of agent and infrastructure winners in the middle. A power and chip layer being repriced upward by the quarter. Hundreds of thousands of professional services employees getting an AI surface installed at their desk in the next twelve months. Next week, watch for Anthropic’s round to actually close and price, for Nvidia’s fiscal Q1 2027 earnings reaction to flow through the supply chain, and for any signal that European regulators are about to match the CAISI framework now that Washington has paused on it.
Find me on YouTube and follow @WaveTheoryAI on X for daily AI coverage. See you next Sunday.