The Week AI Got Very Real: Google I/O, IPOs, and the Agentic Shift

A Week That Will Be Hard to Top

If you felt like AI news moved at warp speed this week, you weren't imagining it. In just the past few days, Google held its most AI-dense developer conference in history, OpenAI moved to go public, and Anthropic revealed it's about to post its first-ever quarterly operating profit. These aren't isolated stories — they're different facets of the same shift: AI is no longer a product category. It's becoming infrastructure.

Let's break it all down.

Google Reinvents Search — For the First Time in 25 Years

The biggest news out of Mountain View this week wasn't a new chatbot or a benchmark result. It was Google announcing what it called "the biggest upgrade to our Search box in over 25 years — now completely reimagined with AI."

At the center of that reimagining is Gemini 3.5 Flash, which is Google's biggest Flash-tier model launch ever — and it shipped generally available the same day it was announced, becoming the default model in the Gemini app and AI Mode in Google Search worldwide as of May 19, 2026.

The performance numbers are notable. In an independent analysis by Artificial Analysis, it was the only model in the upper-right quadrant of the intelligence-versus-speed index, running four times faster than other frontier models on output tokens per second. According to Google's own benchmarks, DeepMind's chief technologist told reporters, "3.5 Flash offers an incredible combination of quality and low latency. It outperforms our latest frontier model, 3.1 Pro, on nearly all the benchmarks," including coding, agentic tasks, and multimodal reasoning.

But raw speed isn't the real story. What Google is actually announcing is a transition away from the search-box-as-question-answerer model entirely. Google said it is placing AI agents directly inside the Search box, capable of handling tasks such as completing purchases, checking ticket availability, and managing schedules in real time.

For ongoing projects, Search goes further still. Often, you aren't asking one-off questions — you have an ongoing task you find yourself searching for over and over, like planning a wedding or managing a home move. Search can go a step further, building custom dashboards or trackers that you can continue to come back to and make progress on.

Google is also deploying persistent "information agents" that run continuously on Google Cloud virtual machines, tracking user objectives and building personalized, interactive dashboards — a significant departure from the traditional query-and-result model. These agents are rolling out first to Google AI Pro and Ultra subscribers this summer.

For AI creators specifically, Google also announced Gemini Omni, a new multimodal world model. Gemini Omni is designed for advanced video generation and editing, positioned as capable of creating and editing videos, images, and simulations from any input. And the Google Pics canvas editing tool, built on the Nano Banana model, isolates design elements as individual objects — useful for anyone doing AI-assisted visual design work.

The scale behind all this is staggering. Google CEO Sundar Pichai revealed the company is now processing more than 3.2 quadrillion tokens per month, a massive increase from 480 trillion tokens per month at I/O 2025. Google has committed to spending between $180 billion and $190 billion in infrastructure investment this year, roughly six times its 2022 capital expenditure figure.

The IPO Wave: OpenAI, Anthropic, and SpaceX All Line Up

While Google was on stage in Mountain View, Wall Street was watching a different drama unfold. Three AI-adjacent companies are racing toward public markets in rapid succession.

OpenAI is preparing to confidentially file its IPO prospectus with the SEC as early as Friday, May 22, 2026, according to CNBC, Reuters, and Axios all reporting the same timeline. The company is working with Goldman Sachs and Morgan Stanley on the planned filing. The target listing valuation sits somewhere between $852 billion and $1 trillion.

SpaceX filed its public IPO prospectus with the SEC on May 20, 2026, targeting a listing on Nasdaq under the ticker SPCX at a reported valuation of $1.75 trillion — which would make it the largest IPO in history if achieved. One detail buried in that filing is particularly telling for the AI industry: Anthropic is paying SpaceX $1.25 billion every month through May 2029 for GPU compute. That single number tells you more about the economics of frontier AI than any benchmark ever could.

And then there's the Anthropic bombshell. Rival Anthropic recently told investors it had solved the seemingly intractable arithmetic problem of running an AI company where revenue tops costs, claiming it was on track to post a profit in the second quarter. Specifically, Anthropic is projecting an expected $559 million in operating profit on $10.9 billion in revenue in the current quarter, up from $4.8 billion in the first quarter.

How? Partly efficiency. Per financial projections seen by the WSJ, Anthropic expects to spend 56 cents on compute power for every dollar generated, down from 71 cents in the first quarter. The gap between OpenAI and Anthropic here is instructive: OpenAI's compute demand is still dominated by consumer users on free tiers, whereas the Anthropic user base is largely composed of enterprise clientele. OpenAI's models are also tightly intertwined with Nvidia's architecture, which is proving more expensive than the Google TPUs and Amazon Trainium chips that Anthropic has largely leaned on.

Combined potential demand from SpaceX, OpenAI, Anthropic, Databricks, and Cerebras is estimated at $100–$200 billion, exceeding the entire 2025 US IPO market by two to four times. That's not a pipeline — that's a flood.

What the Model Releases Tell Us

Zoom out from the headlines, and the model landscape itself is sending a clear signal. Mid-May 2026 is the first month in a year where the most interesting AI release was not the highest-scoring one.

Through mid-May 2026, confirmed releases include GPT-5.5 Instant (OpenAI, May 5) as the new ChatGPT default, SubQ 1M-Preview (Subquadratic, May 5) as the first commercial subquadratic LLM with a 12M token context, Grok 4.3 (xAI, May 6), and Gemini 3.1 Flash Lite (Google, May 8).

That subquadratic model from Subquadratic is worth a second look. SubQ doesn't claim to beat GPT-5.5 on GPQA. It claims to run 12M-token contexts at one fifth the cost. The race is shifting from "who scores highest" to "who can do real work most affordably at scale" — and that distinction matters enormously for AI creators building tools and workflows on top of these models.

The practical takeaway is this: the model you picked three months ago may already be outdated. Build your product stack to swap models without rebuilding everything. API-first architecture is no longer optional.

What This Means for AI Creators

For anyone making a living — or a creative practice — with AI tools, this week is worth pausing to absorb.

Google's agentic Search isn't just about convenience. It signals Google's shift from pitching AI as a conversational tool to AI as an agentic tool. It's not just answering questions, but planning, building, and iterating on real work with minimal human input. For creators, that means AI tools are about to get considerably more capable at handling multi-step, ongoing projects — not just one-shot prompts.

The IPO wave matters too, but perhaps not in the ways you'd expect. When companies go public, their financial disclosures become public as well. We're about to learn a lot more about how the AI business actually works — who's profitable, who's burning cash, and how much of that revenue is coming from creative and consumer use versus enterprise. That transparency will shape how these companies prioritize their products going forward.

And the compute cost story — Anthropic paying $1.25B a month to SpaceX just to run its models — is a useful reality check. Frontier AI is expensive to build and run. The companies most likely to survive and thrive are the ones finding efficiency at scale, not just chasing benchmark leaderboards.

For now: if you haven't experimented with Google's new AI Mode in Search, it's worth an afternoon. And if you're building anything on top of a specific model API, start planning for the model to change beneath you. That's just the environment we're operating in.