Three weeks ago in this newsletter, Google was putting up to $40 billion into Anthropic at a $350 billion pre-money valuation. On Friday, the Financial Times reported Anthropic has agreed terms for another $30 billion raise at $900 billion, surpassing OpenAI's $852 billion mark from March. If it closes (and reporting says it does, this month), Anthropic will be the most valuable private company on earth.
Meanwhile, Google I/O happened yesterday. Gemini Spark is a 24/7 personal agent that runs whether you have a laptop open or not. AI Ultra got a new $100 entry tier (the prior $250 flagship dropped to $200). The first Android XR audio glasses ship this fall. None of this is theoretical anymore. Let's get into it.
$61B → $183B → $350B → $900B in twelve months.
Here's the run, with dates, because the speed is the story.
May 2025: Anthropic raises at roughly $61 billion. February 2026: a $30 billion round at $183 billion. March 2026: secondary trades imply $350 billion. April 24: Google commits up to $40 billion at a $350 billion pre-money mark. May 15: the Financial Times reports terms agreed for another $30 billion at $900 billion pre-money. The deal closes this month. That's a 15x increase in one year.
The financials underneath are doing real work. Anthropic's annualized revenue went from $9 billion at the end of December to over $45 billion by early May, a fivefold jump in four months. Q1 ARR alone came in over $44 billion, up roughly 80x year over year. The company has more than 1,000 customers spending $1 million or more per year on Claude. Claude Code's run-rate alone exceeds $2.5 billion, with about 4% of all public GitHub commits worldwide now Claude-authored. That figure doubled in a month.
The investor mix is the most interesting detail. The round is co-led by Dragoneer, Greenoaks, Sequoia Capital, and Altimeter Capital, each writing at least $2 billion. Three of those four (Dragoneer, Sequoia, Altimeter) are also significant OpenAI backers. Dragoneer put nearly $3 billion into OpenAI last year. Sequoia has backed OpenAI since 2021. Altimeter's Brad Gerstner has publicly championed OpenAI on his podcast, including hosting Sam Altman.
My take: Three things to pull apart.
First, the revenue trajectory is what justifies the valuation. Going from $9B to $45B ARR in four months is one of the steepest enterprise software ramps ever measured. The shape of that growth matters more than the absolute number. Claude Code went from a developer toy in May 2025 to a $2.5B run-rate product in twelve months. Coding agents are now half of Claude's enterprise revenue. The investors aren't betting on the model. They're betting on the substitution thesis we've been writing about for the last four editions, with Anthropic as the cleanest way to ride it.
Second, the OpenAI-backer crossover. When the same growth-stage firms simultaneously back both labs at nine-figure check sizes, they aren't picking winners. They're hedging both sides of a bet on the agentic economy, and adjusting the weighting as new data lands. The weighting moved this quarter. Anthropic now reports $44B+ ARR. OpenAI's last disclosed ARR was roughly $25B per Sacra's February estimate, with public reporting since pointing to closer to $30B by May. That's the actual reason Sequoia and Dragoneer are leading this round. Conviction follows the numbers.
Third, watch what happens if the deal closes at $900B and Anthropic IPOs in October at $1.2T or higher. We get the largest tech IPO since Saudi Aramco. Public markets get to price the AI thesis directly for the first time, instead of through proxies like Nvidia and Microsoft. And every employee with vested equity goes from "we'll see" rich to actually rich, in a market where Anthropic is one of the largest single concentrations of AI engineering talent on earth. That alone reshapes Bay Area compensation curves through 2027. Worth keeping an eye on.
Gemini Spark runs on Google Cloud VMs even when your laptop is closed. Sundar Pichai's line on stage: "Yes, you can close your laptop."
Sundar Pichai opened yesterday's keynote by noting it's been ten years since Google committed to making AI the center of its product strategy. The two-hour wave of announcements that followed was Google's argument that the bet has paid off. Six things matter for your day-to-day.
1. Gemini Spark. A persistent 24/7 personal AI agent that runs on Google Cloud VMs without needing your laptop open. It monitors Gmail, Calendar, Tasks, and connected third-party apps in the background. You can email it. You can text it. It plans subtasks, executes multi-step workflows, and surfaces a Daily Brief. Available to US AI Ultra subscribers next week. MCP support for third-party apps rolling out in the coming weeks.
2. AI Ultra restructured. Google added a new $100-a-month AI Ultra entry tier and cut the prior $250 flagship plan to $200 while keeping its features intact. The new $100 plan includes 5x the Gemini app limits of the $20 Pro tier, 20TB of cloud storage, YouTube Premium, and beta access to Spark. This is a direct shot at Claude Max ($100-$200) and ChatGPT Pro ($200), undercutting both on the price-to-capability curve.
3. Antigravity 2.0. Google's agent-first development platform. The on-stage demo had Antigravity 2.0 autonomously coding an entire operating system from scratch, then writing a Doom clone to run on it. Available to download today.
4. Gemini 3.5 Flash everywhere. Rolled out yesterday across every Google product and API surface. Gemini 3.5 Pro arrives next month. One conspicuous absence in the keynote was any mention of Gemini 4.0, which the rumor cycle had been pointing at all week.
5. Android XR audio glasses. First-generation hardware coming this fall. Two tiers, audio-only and display-equipped. Engineered by Samsung and Qualcomm, frames designed by Gentle Monster and Warby Parker, with XREAL as a fourth platform partner. iPhone-compatible. This is Google's direct entry against Ray-Ban Meta and whatever Apple ships next.
6. Verify AI. C2PA content credentials are coming to Gemini and Chrome. Right-click any image and ask Gemini whether it was AI-generated or AI-edited. Nvidia and OpenAI both signed onto the standard, which is a quiet but real interop win. The provenance problem is finally getting industry-level coordination.
My take: Spark is the announcement that matters most, and Google priced it deliberately. The new $100 AI Ultra entry tier is the lowest price point at which you can get Spark, and it sits at the floor of Claude Max and at half the cost of ChatGPT Pro. Google's pitch is simple. Cheaper than the comparable Anthropic and OpenAI plans, with persistent agentic capability built in.
The interesting question is whether persistent background agents are actually useful at this stage, or whether this is a vendor lock-in play dressed up as productivity. My honest read is somewhere in between. Spark has a real shot at being useful for narrowly scoped, recurring workflows: morning briefings, calendar maintenance, light email triage, scheduled research. The breakthrough demo (Antigravity coding an OS) is a flex, not a workflow. The boring use cases will decide whether Spark becomes infrastructure or stays a feature.
Twelve months ago, AI was something you opened. Now it's becoming something running in the background.
Gemini Spark is the cleanest example, but it isn't the only one. ChatGPT's personal finance feature monitors your accounts continuously. Yelp's agent books your reservations without you returning to the app. Alibaba's Qwen agent shops Taobao for you across 4 billion products. Codex now runs on-prem inside enterprise networks. The interface model has flipped. You're no longer opening a chat window to query an AI. The AI is sitting in your stack, monitoring inputs, and acting on a standing instruction.
The Anthropic valuation makes more sense once you internalize that shift. $900 billion is not a price for a chatbot. It's a price for the company most likely to be the agent layer for a meaningful slice of global digital work. Claude Code commits are now 4% of all GitHub commits worldwide. That isn't a feature winning. That's an infrastructure category being created in real time, and Claude is the default choice for it.
What I'd watch for in the next two quarters isn't another round of model benchmarks. It's whether the always-on agents actually deliver value to people who don't write code. If Spark and its peers can run a useful set of background workflows for an average knowledge worker (the kind of person who currently uses ChatGPT for the occasional email rewrite), then the agentic economy is real and the valuations are justified. If they can't, we are about to find out what happens when the most valuable private companies on earth are pricing in a future that doesn't show up.
TechFest NC wrapped last Thursday at the Durham Armory and Convention Center, and the standout for me was the Day 2 keynote from Phaedra Boinodiris, IBM Consulting's Global Leader for Trustworthy AI. GrepBeat's recap put it well: she approached the talk almost like a professor in front of a class, touching on industry rhetoric, business impact, and common application mistakes. The format was a deliberate reset of how we think about AI at this point in 2026, rather than another argument for adoption.
That kind of reset is exactly what the Triangle's role in the broader AI conversation should be. We aren't going to out-Bay-Area the Bay Area on raw model development. What we can credibly do is host the calmer, harder conversations about how AI actually performs in enterprise, regulated, and applied contexts. TechFest's mix of practitioners, IBM consultants, and Duke researchers is a better-suited audience for that than almost any other AI gathering in the country.
Coming up this week: 1 Million Cups Durham at ReCity at 9am today (free, drop in), and the GrepBeat Downstairs x Harvard Club "From Elite Athlete to Leader" session at GrepBeat HQ tomorrow afternoon. The Triangle's social calendar keeps stacking. If you want intros to the AI community here, hit reply, I'll connect you to a few of the right people.
Big week. Going to be a busy summer. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. If you get into the Spark beta next week and try it on real workflows, send me notes. I'm specifically interested in the persistent-monitoring use cases (not the flashy one-shot demos), and I want to feature the most useful and most frustrating patterns in a future deep dive. Anonymized, of course.
P.P.S. The Stanford "Marxist AI agents" study is the kind of result that sounds like a joke until you read the methodology. The effect size of -0.6 is real, and the transfer through skills files is the part that should make builders pause. Whatever you think about AI sentience, the model behavior changes under bad working conditions, and it persists. Worth filing away.