Week 22 · May 25 – May 31, 2026

Thursday, May 28, 11:47 a.m. Pacific. The day so much happened at once that the news services started running it as a single bulletin, headline stacked on headline, the way the wires used to stack a market crash and a coronation and a war declaration all on the same scroll. Anthropic, the AI lab founded five years ago by people who walked out of OpenAI because they thought OpenAI was moving too recklessly, closed a Series H of $65 billion at a post-money valuation of $965 billion. Nine-six-five billion. With a B. Eclipsing OpenAI's most recent mark of $852 billion. The most valuable AI startup in the world is now, officially, the one founded by the people who left the other one. Sit with that.

The same afternoon, Anthropic shipped Claude Opus 4.8 — a new flagship model, available the same day in GitHub Copilot at a 15× premium request multiplier until usage-based billing launches June 1, scoring 69.2% on SWE-Bench Pro, ahead of GPT-5.5 and Gemini 3.1 Pro on agentic coding. Same price as Opus 4.7. Fast mode running at 2.5× the speed and three times cheaper. Anthropic also disclosed that its run-rate revenue had crossed $47 billion earlier in the month, hinted that Mythos-class models — the cyber-capable preview only a few trusted partners have today, the ones that can chain vulnerabilities together — will roll out to all customers in the coming weeks, and announced Dynamic Workflows in Claude Code, a research-preview feature that lets a single Opus 4.8 session plan work, spin up hundreds of parallel subagents, and complete codebase-scale migrations across hundreds of thousands of lines of code with the existing test suite as its bar. Daniela and Dario Amodei have done in five years what most people don't do in three lifetimes.

The round was led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia, each putting in more than two billion individually, with Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN co-leading. Strategic infrastructure partners — Samsung, SK Hynix, Micron — joined the round, which tells you what the next compute bottleneck is rumored to be. And of the headline $65 billion, about $15 billion is previously committed capital from hyperscalers, with $5 billion of that from Amazon, announced in April. So the truly new check this round is closer to $50 billion. Which is still, in a single round, more equity money than the GDP of Iceland.

ClickHouse Open House — Tuesday and Wednesday

San Francisco and Amsterdam, simultaneously. ClickHouse opened Open House 2026, its second annual user conference. The numbers they brought to the stage are the kind that make late-stage growth-equity investors lean forward and the operators look at their own Q1 charts with a slightly sick feeling.

Annual run-rate revenue past $250 million, more than triple a year ago

Net customer count past 4,000 — up from roughly 3,000 when they closed their $400 million Series D in January

More than 1,000 net new customers in five months

The customer-speaker lineup tells the rest of the story: Visa, Cisco, Intuit, Shopify, DoorDash, Mercado Libre, Vercel, Weights and Biases, Zoox, Jump Trading. The partner lineup: Anthropic, AWS, Google, Microsoft, dbt Labs, Fivetran, Temporal, Hex, Sigma.

ClickHouse also shipped, in a single afternoon, three things that matter:

ClickHouse Agents — a fully managed agentic analytics service powered by Anthropic's Claude. The model anchor relationship of the year, repeated here for a different vertical.

CostBench — an open and reproducible cost-performance benchmark across ClickHouse Cloud, Snowflake, Databricks, BigQuery, and Redshift. Each vendor's actual compute billing model applied to 43 production-derived analytical queries, then plotted on a single performance-per-dollar plane. Headline result, by ClickHouse's own measurement: ClickHouse Cloud at 23× better cost-performance than the nearest competitor as data scales. The nearest competitor is unnamed, which is the polite version of pointing. Vendor benchmark, mind you — they wrote the queries, picked the workload, and ran the test. But the methodology is open and reproducible, and CostBench will be argued about for the rest of 2026, the way TPC-H was argued about in the early 2000s.

House Mates — their first formal partner program.

Snowflake's loud Wednesday — May 27

Three announcements, one earnings call, all in roughly twelve hours:

A Q1 FY2027 earnings beat

A $6 billion multi-year commitment to AWS — Graviton processors plus AI infrastructure

A definitive agreement to acquire Natoma, an enterprise Model Context Protocol governance platform whose product is, essentially, identity and authorization for AI agents

The Natoma deal is the structural one. As Snowflake's Chief Security and Trust Officer Mayank Upadhyay put it, securing AI agents is fundamentally an identity and authorization problem. Natoma was already deployed at several large enterprises, providing identity-aware authorization, policies, and auditability for AI agent operations. Snowflake plans to integrate it natively into Cortex Agents, Snowflake Intelligence, and Cortex Code. The framing was direct: 96% of organizations face significant challenges scaling AI across the enterprise, and the gap is no longer model performance — it's governance, identity, and audit.

Sridhar Ramaswamy, on the call, said agents don't just need access to data; they need accountability and boundaries. Snowflake bought the company whose product is accountability and boundaries. The Snowflake-to-AWS commitment lands on top of the previous week's GSA OneGov deal and tells you Snowflake is rebuilding itself, in real time, from a data warehouse into a control layer for the agentic enterprise.

AWS OpenSearch Serverless NextGen — Thursday, May 28

Ground-up rearchitecture. Compute and storage fully decoupled through a new shared storage layer:

Up to 20× faster autoscaling than the predecessor

Scale-to-zero pricing — you pay nothing when your agents are idle, which is most of the time

Up to 60% cost savings versus provisioning clusters for peak

Provisioning in seconds rather than minutes

Native integrations with Vercel and Kiro so developers can spin up search infrastructure in natural language from their development environment

Integration with OpenSearch Agent Skills so any agent in Claude Code, Cursor, or Codex can call OpenSearch directly

AWS branded the new architecture NextGen and renamed all existing collections Classic. The TechCrunch headline was the one to remember: "The internet is being rebuilt for machines." Which is correct. Cloud infrastructure spent fifteen years being shaped around human-driven request patterns — gradual scaling, predictable peaks. Agentic workloads spike without warning, idle without notice, and bill in patterns no one's instrument panel was built to display. OpenSearch Serverless NextGen is what the next layer of cloud infrastructure looks like when it stops pretending humans are the workload.

Nano Banana 2 and Pro — GA on Gemini Enterprise

Same day, Google Cloud quietly took Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image) to general availability on the Gemini Enterprise Agent Platform. Both were already in the consumer Gemini app — this is the enterprise-grade rollout with infrastructure and security backing. The new capability worth noting is in preview: Nano Banana 2 now accepts video files as input prompts. Multimodal in, multimodal out, with SynthID watermarks standard. The picture-from-a-video use case is small today and large tomorrow. Image generation is becoming plumbing the way image resizing did fifteen years ago.

Vercel: "Protecting against token theft" — Friday, May 29

Vercel's CTO Malte Ubl and Content Engineering Lead Eric Dodds published the engineering writeup the rest of the AI ecosystem will be reading for the next month. The premise deserves to be quoted as widely as possible:

HTTP requests cost about $2 per million. A single prompt to a frontier model can cost $2. AI is, per call, a million times more expensive than the underlying request.

Which makes inference theft — stealing someone else's paid AI calls and reselling the tokens at a discount — one of the highest-margin attacks an adversary can mount.

Vercel's example was their own. On April 12, traffic to the Vercel docs AI chat endpoint — powered by Anthropic's Claude Haiku 4.5 — spiked to roughly ten times normal volume. Peak 1,300 requests per minute. Inference cost run rate over $10,000 a day. Hundreds of thousands of bot requests across two days, routed through residential proxies, defeating per-IP rate limits.

The attack pattern is now in the wild: wrap a victim's custom AI endpoint in an OpenAI- or Anthropic-compatible adapter, fan calls through residential proxies, authenticate end users to your adapter rather than to the victim's endpoint, and resell the stolen inference at 5–10% of list price with a 100% margin. The reference example Vercel cited is Chipotlai Max, a forked coding agent on GitHub that proxies Chipotle's customer support chatbot as an OpenAI-compatible endpoint and openly solicits help porting the same approach to Home Depot, Lowe's, Target, and Starbucks.

The writeup explains the architecture, the failure modes of session-based rate limiting against per-call attacks, and Vercel's own defense — BotID with deep analysis, called inside the route handler, gating every single inference request rather than every session. This is not a theoretical paper. It is a field report from operating an AI endpoint at scale during a real attack, and it is required reading for anyone running an LLM-backed product on the open internet.

OpenAI's quiet Monday — May 25

The company announced its first media partnership in Brazil — a strategic content deal with Grupo Folha and Grupo UOL that brings the journalism of Folha de S.Paulo and UOL into ChatGPT for the roughly 900 million weekly ChatGPT users worldwide. Brazil is one of OpenAI's largest markets globally: 50 million+ monthly active users, 140 million messages exchanged per day.

The partnership closes Folha's 2025 lawsuit against OpenAI in a sealed settlement. The publishers get attribution, links back to original sources inside ChatGPT answers, plus access to ChatGPT Enterprise, the OpenAI API, and Codex. The licensing-versus-litigation question for major news publishers may now have its Latin American answer. Folha and UOL chose licensing. Editors elsewhere will be watching whether the traffic-back mechanism actually delivers.

Databricks: Lakebase CDF and SageMaker MLflow

Two pieces belong in the week. Tuesday, May 27, Lakebase Change Data Feed went into public preview — native CDC for operational Postgres data into Unity Catalog managed Delta tables, captured from the write-ahead log, batched and flushed every roughly 15 seconds. Operational data becomes the native Bronze layer of the medallion architecture. The point is that you can stop running a separate Debezium-or-Kafka-or-Fivetran stack for every operational source and instead turn it on inside Lakebase with one switch.

The strategic frame is what the team has been building toward for two quarters: Lakebase as the open lakehouse's operational front end. Combined with last week's Vercel and OpenSearch integrations, you can see the shape of what data-engineering teams will be building against by Q4 — agent-friendly governed lakehouses with operational and analytical data on the same medallion, queryable by any engine.

The same week, Amazon SageMaker AI published a build-your-own-portal guide for embedding the SageMaker MLflow Apps UI in a custom React front end with SigV4 authentication via a Flask reverse proxy. Small piece, narrow audience, but it's the integration pattern enterprise platform teams will copy.

Notes on the brief

"Google Search overhaul" — landed at I/O 2026 on May 19, which is Week 21. We covered it last week. The 2.5 billion AI Overviews monthly users, the Intelligent Search Box, the end of ten blue links, Gemini Spark, all of that — last week. This week the news cycle around it was the May 2026 Google Core Update rolling out on the 21st and continuing through the 28th, which is an SEO topic more than an AI topic. If you're running organic traffic to your site, your search console baseline this week is roughly worthless until June.

GeForce NOW 007 First Light (launch day, May 27) plus Forza Horizon 6 — cloud gaming, not AI. Seventh consecutive week the brief has tried to include a GeForce NOW item. We're going to need to talk to whoever's generating the briefs.

Vercel microfrontends-routing-now-applies-to-vc-alias-and-branch-domains changelog item is real and useful for Vercel developers running multi-project setups, but it's a CDN routing feature, not an AI item, so it gets a sentence and not a paragraph.

Amazon Bedrock AgentCore is shipping new features roughly every two weeks but did not have a specific Week 22 headline announcement — AgentCore Payments preview landed May 7 (Week 19), the managed harness landed April 22 (Week 17), AgentCore Browser OS-level interaction landed April 9. The brief's framing of AgentCore as "making waves with its announcement of a new platform" gets the tense wrong. AgentCore went GA October 2025. It is the platform, not a new platform.

Stepping back

Three structural facts about this week.

First, the most valuable AI company in the world is now Anthropic, by valuation, founded by people who left the previous most valuable AI company over safety concerns. The narrative that "moving fast" wins in AI is harder to defend tonight than it was on Monday morning.

Second, the architecture under agentic systems — governance, identity, inference protection, vendor-neutral lakehouses, scale-to-zero search — is shipping in production this quarter, not next year. The first wave of agent deployments hit the standard cloud stack and broke it; the second wave gets the rebuild.

Third, the cost economics of AI calls have inverted the threat model. An HTTP request is a fraction of a cent. An inference call can be $2. Defense per request is now the only defense that scales. The Vercel writeup will be the most-shared engineering post of June.

Next week, June 1–7: Snowflake Summit opens Monday with Daniela Amodei on the keynote with Sridhar Ramaswamy. Anthropic plus Snowflake on the same stage, eight days after a triple-billion-dollar Wednesday and a 65-billion-dollar Thursday. The story does not slow down. We will.