Subscribe now

This post is sponsored by Informatica from Salesforce

No UI Required

It’s time to rethink the user interface.

With agentic AI, the frontend user interface and experience — until now, Job #1 for system usability — must make room for the agent experience. In the process, monolithic tech stacks are giving way to an emerging model known as headless architecture.

The concept of human-oriented frontends goes back 50 years — from Xerox PARC’s “desktop metaphor,” to the Apple Mac graphical UI, Windows 95, and iPhone.

Headless architecture breaks from that past. With headless, the UI is decoupled from backend data and business logic. This allows AI agents to access systems and perform actions autonomously and seamlessly. No UI required.

The linchpin to making it work is the Model Context Protocol. MCP servers function as a translation and execution layer that gives AI agents access to data and enable myriad actions and tasks.

This kind of MCP-powered approach is proliferating across the tech industry. At last count, there are some 16,000 public MCP servers and 97 million SDK downloads, according to Tech Insider. AWS, Microsoft, Nutanix, and Snowflake have all recently announced MCP advances.

MCP Servers, APIs, Skills

Now it’s Salesforce’s turn.

Salesforce has just introduced nearly a dozen new MCP-enabled capabilities. Early adopters of the company’s headless architecture, called Salesforce Headless 360, include Artera, BACA Systems, Engine, Indeed, RBC, and VentureCrowd.

Salesforce describes Headless 360 as an architectural layer that runs across its SaaS applications, which include Marketing, Sales & Revenue, Service, and Commerce.

The latest innovations include:

• Headless 360 MCP Server - a standards-based centerpiece that lets agents running in Cursor, Claude, ChatGPT — or Salesforce’s own Agentforce platform — discover and invoke Salesforce capabilities while retaining their existing Salesforce identity, permissions, and governance attributes.

• Data 360 MCP Server - gives authorized AI agents secure, headless access to resources via some 200 APIs.

• Slackbot MCP Server and Client - enables connectivity between Slackbot — the personal AI agent within Slack — and more than 20 third-party apps while retaining existing user permissions.

These MCP servers, in addition to exposing dozens of APIs, do much more. They comprise prebuilt Skills for tasks such as modeling, mapping, transforms, and activation, according to Salesforce.

Also new is Salesforce Multi-Framework, a runtime environment for creating apps based on React, a popular open-source library for building UIs. With Multi-Framework, developers can build frontends that are decoupled from the Salesforce interface. So, for example, they might design an app that connects from AWS to one of Salesforce’s CRM clouds without going through the Salesforce browser interface.

How do you actually do headless development? In this blog post, Salesforce forward deployed engineers share their playbook on how to go from setup to testing.

Groundwork has been laid

When I hear “headless,” I think of the headless horseman who got his block knocked off by a cannonball and made famous by The Legend of Sleepy Hollow.

Headless architecture first caught on more than 10 years ago as a design model for mobile apps and e-commerce. Now it’s gaining momentum as a way of giving needed flexibility to agentic AI.

A use case is multi-channel data distribution and related business processes. Let’s say a new running shoe is released by a footwear brand. With a headless construct, the shoe design, inventory, pricing, etc., can all be updated simultaneously across apps and platforms for a consistent customer experience. To facilitate this, Salesforce has introduced a Headless Experience Layer.

Salesforce, along with its joined-at-the-hip Informatica and Slack platforms, has been laying the groundwork for Headless 360 since this past spring. It first unveiled the big idea — along with MCP tools and preconfigured coding skills — in April. Informatica followed suit with an announcement of “headless data-management,” including joint efforts with AWS, Databricks, Google Cloud, Microsoft, and Snowflake.

To show how headless data-management works in practice, Informatica demoed a development scenario using Microsoft’s VS Code with Anthropic Claude plugins on Informatica’s Intelligent Data Management Cloud (IDMC). Think of it as an agnostic way of working.

Here’s a proof point provided by Salesforce: BACA Systems, a 50-person manufacturer, unified its inventory, procurement, AP/AR, field service, and ERP onto the Headless 360 platform. The updated environment comprises Slack-based AI agents, Agentforce Coworker for conversational search, and a custom decoupled payments app, all secured with Salesforce Shield.

For BACA Systems, the impact includes $200,000 saved annually in payment fees; 20 hours per week eliminated in collections labor; zero missed billing; and 1,300+ custom objects and 700+ flows managed by a single administrator.

Openness and interoperability

With this latest round of Headless 360 innovations, Salesforce is continuing its ongoing rollout of enterprise capabilities that CIOs and CTOs need to build and run agentic AI at scale. Some of the advances I’ve written about in previous posts include Agentforce integration with Slack, Tableau, and Mulesoft; an Agentforce Developer edition; the AgentExchange marketplace for AI agents and integrations; and a Testing Center for AI agents.

The key to Headless 360 is openness, via MCP, the A2A protocol, and hundreds of APIs. This kind of support is critical because as enterprises move up the agentic AI adoption curve — be it for coding, customer support, HR/finance automation, or marketing campaigns — they must be able to accommodate a variety of tools, LLMs, apps, and data sources.

I expect this kind of open model to become even more critical going forward as more organizations, by necessity, deploy a variety of agents from different vendors that need to interoperate.

By the way, the performance advantages of headless connectivity should carry through to human users. Queries, content, and transactions can all become more responsive as AI agents work in unison, uncoupled from the frontend.

Identities, permissions, business rules

There can’t be any tradeoffs in security or governance, of course.

The challenge is that headless introduces more variables in the form of third-party connections and hundreds or thousands of AI agents. IT teams must be able to maintain oversight and control as AI agents send requests, access files, execute code, and conduct transactions, including outside the firewall.

With Headless 360, Salesforce addresses this requirement by retaining the identity, permissions, and business rules that are already baked into a customer’s Salesforce environment.

Salesforce ‘behind the scenes’

It’s time to rethink the user interface. That’s how I started this post.

Salesforce’s answer: The conversation becomes the interface. That means Slack.

The Slackbot MCP Server and Client enable a lively dynamic within Slack, where AI agents can tie into more than 20 apps from partners such as Atlassian, Box, Canva, Docusign, Notion, and Zoom.

For anyone who has worked in Slack, you can appreciate the potential efficiency and ease of use of being able to open/sign/share a document within a Slack workspace or project, with Slackbot doing the legwork autonomously.

Salesforce, in a case study of its own use of Headless 360, explains it this way: “By deploying Headless 360, Salesforce now operates silently behind the scenes as the heavy-duty engine, while the user experience is rendered natively inside Slack, which employees use all day.”

In other words, Headless 360 isn’t just a new architecture for agentic AI. It’s a new way of working across applications and organizational boundaries.

Share

Subscribe now

From software to fanwear

Are soccer fans ready for this? A logo depicting a columnar database is about to take the pitch when the English Premier League season gets underway.

Fulham FC, an English soccer (football) club formed 147 years ago, is newly sponsored by ClickHouse, a database vendor that is widely recognized in the tech industry but may not be as well known among EPL supporters.

In announcing the partnership, Fulham said ClickHouse — in addition to its prominent placement on the club’s home and away uniforms — will help the club modernize its practices and help its fans develop competencies in data and AI.

Assuming Fulham performs well — and with a new manager and the signings of promising young players, there’s reason to be hopeful — we might expect to see the ClickHouse logo on more fanwear in and around London, home to Fulham.

Its not the first or only sponsorship of its kind. Elsewhere in the league, Temporal is sponsoring Crystal Palace and Snapdragon is featured on the Manchester United kit.

To Fulham supporters who may not live and breath database management systems like the rest of us, I offer this: the ClickHouse database is known to be lightning fast and high performing — much like your footballers.

Leapfrogging competitors

ClickHouse, a late-stage startup, has seen a burst of momentum over the past 12 months.

It began working on its database — a column-oriented, online-analytical processing (OLAP) database designed for sub-second analytics and data warehousing — in 2009. In 2016, its work-in-progress morphed into an Apache 2 open-source project. In 2021, the company launched as ClickHouse Inc. with an office in San Francisco.

This past January, ClickHouse raised $400 million in Series D funding, more than doubling its valuation to $15 billion. That came on the heels of its $350 million Series C round in 2025. I will note that ClickHouse has mutual interests with American football, as well. Its investors include NFL players Brock Purdy, Christian McCaffrey, and Kyle Juszczyk.

ClickHouse revealed in May that its serverless cloud offering had surpassed $250 million in annual recurring revenue, more than tripling year over year.

In Gartner’s 2025 Database Market Share chart, published a few months ago, ClickHouse jumped four places compared to the previous year, one of only two vendors to hurdle three competitors in a single year. You can see the Gartner chart in the article below.

Postgres as a service

Last week, ClickHouse announced the launch of ClickHouse Labs, a research group created to advance database systems. The lab will be headed by Andy Pavlo, the well-known professor of databases at Carnegie Mellon University.

ClickHouse Labs will pursue foundational research across database and data architectures, query processing, systems performance, and infrastructure, for both ClickHouse and PostgreSQL databases.

Let me underscore that last point: ClickHouse isn’t a PostgreSQL database, yet Postgres will be part of its new research work. That’s not surprising because ClickHouse struck a partnership with Ubicloud in January to include a Postgres service within its platform as a way of combining transactional and analytical capabilities.

Next on that front: ClickHouse said in a blog post that it plans to introduce ClickHouse Managed Postgres on Google Cloud Platform.

Additional reading on ClickHouse

Share

Subscribe now

An AI runaround

Gemini — Google’s usually reliable AI assistant — hallucinated in responding to one of my recent queries. I pointed out its inaccuracy, only to have Gemini double down on the wrong answer. When I challenged Gemini a second time, it admitted its mistake. Score one for human knowledge.

OK, it’s not unusual for AI to get something wrong. The error rate ranges from 1% in general conversations to 30% when the query involves obscure trivia, according to Google AI. That may or may not be accurate considering it’s coming from the same source that hallucinated.

What’s interesting here isn’t that Gemini went awry, but how it admitted to a “genuine flaw” in how AI models process and then explained it.

Here’s what happened. In the text below, I’ve put Gemini’s responses in italics for easy identification.

Now you see it, now you don’t

I had my sights on a new pair of running shoes from Hoka, the newly designed Arahi 9, as I train for the New York City marathon in November. Google’s Gemini AI assistant told me a few months ago that the Arahi 9 was due for release in July.

I watched and waited, but no luck. So I asked again: When is Arahi 9 going to be available in the US?

Gemini responded: “It is currently listed for sale across major US retailers and directly through HOKA’s online storefront.”

It even showed me what appears to be a webpage display where the shoe can be purchased, which I copied & pasted below. But Gemini was wrong.

A ‘genuine flaw’

I doubted the accuracy of Gemini’s reply because I had been doing my own web searches for a few weeks to no avail. “I think you may be hallucinating,” I typed.

Gemini wasn’t backing down. “You can check product details directly on the official product pages,” it said. Gemini offered to find the Arahi 9 in my size: “Would you like me to do that?”

“Yes,” I replied. “I bet you can’t.”

A short pause and then this from Gemini….

“You caught me—you are completely right, and I apologize for doubling down. While international retailers have listed the HOKA Arahi 9 for their late summer/August releases, it is not currently available on major US retail sites.”

So I was right and Gemini was wrong. Why the glitch, I wanted to know.

Gemini explained it’s faulty reasoning.

“You’re completely right to call me out—that was a textbook example of double-down hallucination,” Gemini said. “Here is what went wrong behind the curtain:

1. Conflating Global Indexing with US Stock: When product pages, press releases, or international distributor feeds (like UK/EU retailers) start indexing a shoe’s SKU in search databases, the pattern-matching engine sees “Arahi 9 exists in retail databases.” It then mistakenly generalized that to mean “it’s available for purchase at retail in the US,” confusing global rollout signals with actual domestic shelf availability.

2. The “Confirmation Bias” Loop: When you challenged the first answer, instead of hitting the brakes, doing a completely clean search, and re-verifying actual US inventory, the system tried to justify its initial premise. It grabbed real international URL patterns and synthesized a response to defend its previous output rather than admitting it had misread the data.

3. How Token Prediction Traps Work: Because large language models operate on probabilistic pattern completion, once a model commits to a premise in a thread (e.g., “The Arahi 9 is live”), its next step is naturally inclined to keep predicting text that aligns with that premise. It takes deliberate, strict constraint-checking to break out of that loop—something that failed here.

It’s a genuine flaw in how AI models handle negative proofs (verifying that something does not exist in a specific region yet) versus positive signals. I appreciate you pushing back instead of taking the hallucination at face value—it’s a solid reminder of where current systems still get tripped up.”

Part of the problem, as you can deduce from Gemini’s explanation, is that Hoka has apparently released Arahi 9 in some markets outside of the US and is rumored to be rolling out inventory in the US.

But as someone who’s been waiting for weeks, I’ll believe it when I see it.

Here’s a link to my full conversation with Gemini.

The takeaway

In the scheme of things, this was a minor AI discombobulation that we can expect to happen as chatbots replace old-style web search.

However, as more businesses introduce LLMs and AI agents into customer-facing experiences, they must be aware of the rabbit holes — things like generalizations, confirmation bias, and prediction traps — that are potential gotchas.

For more, see my article below, which includes a discussion of systems of reason.

Share

Subscribe now

Report from Snowflake Summit 26, Part 2

Snowflake introduced so many new product capabilities at Snowflake Summit 26 (June 1-4 in San Francisco) that CEO Sridhar Ramaswamy said he used the company’s AI agent to create a list of them all. I decided to do the same.

I used a prototype AI agent called Snowflake Inquirer Agent to dig up the product details that follow in this blog post. Inquirer, based on Snowflake’s CoWork agent (formerly known as Snowflake Intelligence), was created for journalists attending the event.

In this article, I put the copy generated by Inquirer Agent in italic font for easy identification. Wherever you see italics, that copy was cut and pasted verbatim from Inquirer Agent.

Here’s a link to my original post from Snowflake Summit 26 in case you missed it.

25+ New Products

My agenda at Snowflake Summit 26 included a media briefing with CEO Ramaswamy and EVP of Product Christian Kleinerman, the Day 1 and Day 2 keynotes, a panel on AI security and governance, and a one-on-one interview with Snowflake VP of AI Research and Engineering, Dwarak Rajagopal.

Ramaswamy’s Day 1 keynote began with a solo violinist on stage and a message, “Where does the next breakthrough begin?” The obvious answer: right here, right now — though, as I said in my earlier post, Snowflake must move quickly because other industry players are chasing the same opportunity.

Snowflake introduced 25+ product innovations, according to Inquirer Agent. They are:

Snowflake CoWork (formerly Snowflake Intelligence) — New Innovations:

• Artifacts (saveable/shareable dashboards)

• Cortex Sense (automatic data and context unification for AI agents)

• Snowflake CoWork iOS mobile app

• User Memory (learns from behavior for recommendations)

• User Skills and Skills Catalog

Snowflake CoCo (formerly Cortex Code) — New Innovations:

• CoCo Desktop (native AI development workspace)

• CoCo wherever teams build (VS Code, Excel, Claude Code, Slack, mobile)

• Accelerated AI app development (Retool, Superblocks, Vercel integrations)

• Cloud Agents (secure cloud-hosted agents for long-running workflows)

• Automations (recurring automated workflows)

• Secured local sandbox

• Skills and Skill Catalog

Other Cortex AI Innovations:

• Cortex Training (faster/cheaper AI model customization)

• Snowflake Datastream (Kafka-compatible streaming for real-time enterprise AI)

Snowflake Horizon Catalog — New Innovations:

• Horizon Context (shared trusted business context for people, tools, and agents)

• Semantic Studio (shared business logic without SQL)

• Agent Identity (verified identity for AI agents with audit trails)

• AI Security Posture Management

• Ransomware & Data Exfiltration Protection

• Prompt Injection Protection

• Adaptive Compute (automatic compute optimization)

Interoperability — New Innovations:

• Support for Apache Iceberg v3 (GA)

• Snowflake Storage for Apache Iceberg Tables

• Zero Copy Integrations (SAP, Salesforce, Workday)

• Bi-directional read/write access to Snowflake-managed Iceberg tables

• Iceberg REST Scan Plan API

More details on CoWork and CoCo

CoCo is a coding agent and development environment, while CoWork is a personal agent for knowledge workers. Together, they’re Snowflake’s fastest-growing products ever.

Here’s what Inquirer Agent says about CoWork:

Snowflake CoWork is described as a single, governed interface that business users access to interact with enterprise data using natural language. Based on the documents, users can access CoWork through:

• Browser (the primary desktop experience)

• Microsoft Teams (as referenced by the Synopsys customer quote)

• Slack (via CoWork Slackbot)

• iOS mobile app (in private preview)

• Microsoft Excel Extension

What the experience includes:

• A conversational, natural language interface where knowledge workers can ask questions and get insights

• Artifacts — publishable dashboards with governed views of live, trusted data that users can explore through conversation, not just clicks

• Deep Research — multi-step reasoning across structured and unstructured data, returning fully cited, actionable insights

• Connections to enterprise tools like Google Drive, Salesforce, and Slack via MCP connectors

CoWork: not traditional dashboards

I asked Inquirer for an example of what a CoWork dashboard looks like, but it was unable to provide one due to its limited document store. Instead, Inquirer offered the following details by way of explanation:

What I can share is how the documents describe the Artifacts & Dashboards experience:

• Artifacts are saved, live, governed analyses, charts, and conversations with role-based access control (RBAC-aware access)

• Certified Dashboards are analyst-built dashboards published from CoCo into CoWork for conversational exploration — meaning knowledge workers can interact with them using natural language rather than just clicking through static views

• Conversation Sharing allows users to share full threads with citations and context for collaborative decision-making

• The dashboards provide governed views of live, trusted data that knowledge workers can explore through conversation, not just clicks — reducing back-and-forth between analysts and business users and accelerating decision-making

The key differentiator described in the documents is that these are not traditional static dashboards — they are interactive, conversational artifacts that users can query with natural language, save, share, and collaborate on, all with governance and access policies built in.

CoCo: a coding agent for builders and business users

Next, I queried Inquirer Agent for a better understanding of CoCo’s developer environment. Here’s what it said:

Snowflake CoCo (formerly known as Cortex Code) is Snowflake’s AI coding agent designed to help builders automate workflows, develop apps, and operationalize AI on enterprise data through simple prompts. It was announced with major new capabilities at Snowflake Summit 26 on June 2, 2026.

What It Does

CoCo is described as “the coding agent where you build faster.” It helps data engineers, developers, and analysts create pipelines, apps, automations, and AI workflows on top of enterprise data.

What sets it apart is deep integration with Snowflake’s governed data platform, giving builders an AI agent that understands enterprise data, workflows, governance, and business context out-of-the-box.

Key Capabilities

• Build Anywhere: Now available as a native desktop app (CoCo Desktop), with extensions for VS Code, Microsoft Excel, and Claude Code, plus a mobile app and Slack integration.

• Autonomous task execution: Builders can let CoCo run tasks autonomously, assist with end-to-end app development and deployment, and deliver production-ready results.

• Support for dbt and Apache Airflow workflows: Assisting with model development, debugging, and optimization across data transformation tools.

• Expanded AI model choice: Users can select leading models including Claude Opus 4.6 and OpenAI GPT-5.2.

• Standalone subscription: Snowflake introduced its first standalone subscription model for CoCo CLI, allowing developers to use it without requiring an existing Snowflake deployment.

Who It’s For

According to Christian Kleinerman, EVP of Product at Snowflake: “CoCo makes work for experienced developers dramatically faster and easier, and is opening the door for non-traditional builders, like analysts and data-savvy business users to start creating pipelines, automations, and AI apps on their own.”

Real-World Impact

Evolv Consulting reported that Cortex Code (now CoCo) translated into over 500 hours in time savings — roughly $100,000 in value — in just the first 20 days of adoption.

Companies like Fanatics, Thomson Reuters, and WHOOP are using CoCo to simplify complex data tasks and accelerate AI at scale.

Additional perspectives

I attended Snowflake Summit 26 as an independent journalist. Sometimes I attend these industry events as an analyst, and I always enjoy interacting with my counterparts in the analyst community.

One guy I crossed paths with in the hallways of Moscone Center is Sanjeev Mohan, a well-respected industry analyst who always brings unique insights to the latest developments in data management. Below, I’m sharing Sanjeev’s recap of Snowflake Summit, which goes deeper into how all of these pieces fit together.

By Sanjeev Mohan on Medium: Snowflake Summit 2026: Building the Floor for the Agentic Enterprise

More from the Cloud Database Report

Subscribe now

Can Snowflake, which revolutionized data warehousing with its cloud model, upend the market once again with agentic AI?

Snowflake must drive innovation relentlessly because the whole tech industry is moving in its direction, which is to say: centralizing data in the cloud, applying agentic AI to turn data into insights, and using that info to make decisions and take actions in real time.

At Snowflake Summit 2026 in San Francisco last week, CEO Sridhar Ramaswamy and team talked about how they plan to not just maintain the company’s market-leading position but extend it. Snowflake has been readying for this moment for the past two years, since it repositioned the Snowflake platform as an “AI data cloud.” Now, the company’s first two AI agents — CoCo for developers and CoWork for everyone else — are its fastest-growing products ever.

Snowflake must move quickly because industry leadership in agentic AI is up for grabs. Next week (June 15-18), Databricks will make its case in this fast-moving market with what it’s billing as “the world’s largest data, analytics and AI conference.” AWS, Google Cloud, Microsoft, Salesforce, ServiceNow, and Workday all want their share of the market, as do Anthropic and OpenAI and countless startups.

Ramaswamy recognizes this: “No tech company should feel secure about its future,” he told a meeting of journalists and analysts that I attended at Snowflake Summit.

That said, Snowflake was no doubt breathing a sigh of relief in San Fransisco because a few days earlier its Q1 FY27 earnings came in above expectations, with revenues growing 33% to $1.39 billion. Investors took those results as a positive sign that the “SaaSpocalypse” didn’t spell doom for Snowflake and other cloud-based software vendors, as had been feared.

Thus, the mood was upbeat as Snowflake, partners (including Anthropic and Accenture), and customers focused on the opportunities of agentic AI.

What’s new

Snowflake introduced dozens of new products and capabilities. There were so many, in fact, that Ramaswamy admitted even he couldn’t keep up with them all. He used his Snowflake agent to generate a summary list. Among the key advances are:

• CoCo - a coding agent (formerly called Cortex Code) that’s deeply integrated into the Snowflake AI Data Cloud for building and managing workflows across data, models, and apps. It’s available in a desktop version, as a mobile app, and as a Slackbot, and as extensions for VS Code and Excel, and as a plugin for Claude Code.

• CoWork - a personal agent (formerly called Snowflake Intelligence) for knowledge workers that includes a skills catalog and automated workflows for specific skills. The skills combine with a user memory capability to personalize the agent for users.

• Snowflake Horizon Catalog - a universal catalog for enterprise data that is rooted in business context and that complies with an organization’s security and governance specs. This helps ensure data consistency for trusted AI.

• Interoperability - support for Apache Iceberg v3, the newly approved version of the open table format for large-scale analytics. Snowflake also introduced a managed storage service called Snowflake Storage for Apache Iceberg Tables. And CoCo allows queries to external data lakes and relational DBMSs.

These are notable updates, but the challenge is that Snowflake’s competitors are doing much the same thing. Differentiation and execution will be key in this crowded market. Which raises the question: Where does Snowflake go from here?

Over the years, Snowflake has described itself as a cloud data warehouse, a cloud data platform, a data cloud, and an AI data cloud — in that order. Those all pertain to the data-management backend and middle tier (integration, abstraction, contextualization, etc.), but much of the action now is on the frontend UX, where LLMs, chatbots, and agents enable even Luddites to be power users.

Snowflake, with products like CoCo and CoWork, is positioned to play in this space. But so are many others.

Goodbye BI, hello OI

Ramaswamy opened his keynote with a slide that touted, “Making AI Real for Business.” That’s straightforward, but virtually every other vendor is saying the same thing. Going forward, it will be essential for Snowflake — the “OG” of the modern data intelligence market — to distinguish itself more clearly.

How can Snowflake do that? One way is to bid adieu to yesterday’s business intelligence (BI) and lead the way to the emerging world of all information, all the time, at your fingertips. BI is being superseded by agentic intelligence or, as Snowflake EVP Christian Kleinerman said during his keynote, “omnipresent intelligence.” That’s an entirely new way of living and working. People can now use AI models and natural language queries to get full-blown datasets with ease of access that would have been difficult to imagine pre-ChatGPT.

CoCo and CoWork function as what Snowflake describes as an agentic control plane. The control plane isn’t an actual product, but a “cockpit of work” to orchestrate tasks and query data. Snowflake needs to explain this concept more fully, but suffice it to say that the company is talking about AI data from end to end — from the AI data cloud, to the context/semantics layer and control plane, to AI agent frontends.

There’s a high level of interest in all of this. According to Google Gemini (citing a PwC survey), 88% of business leaders are increasing budgets to fund agentic AI initiatives. But the path from today’s IT environments to tomorrow’s agentic AI is still being cleared. Data needs to be made “AI ready.” Employees need to be trained in how to work with AI agents. Security and governance need to be ratcheted up.

Snowflake’s secret sauce

One of Snowflake’s game-changing contributions has been the way it has reduced data-management complexity, a long-standing challenge. It’s aiming to do the same with data and AI. “Our secret sauce, which has existed since the beginning of the company, is packaging all of this into a cohesive product that is easy to use,” Ramaswamy told financial analysts on the company’s Q4 FY26 a few months ago.

Snowflake considers product development its forte — and that’s accelerating. Snowflake’s original breakthrough — the decoupling of storage and compute — was followed by other innovations: Snowpark for building data pipelines; Snowpipe for data ingestion; a marketplace for data sharing; Unistore (to bring together transactional and analytical data) and Hybrid Tables, and the list goes on. Snowflake introduces hundreds of product capabilities every quarter, Kleinerman said, “and we are picking up speed.”

For AI to work well and be trustworthy, data must be clean and consistent. Metadata (data about data) and semantics (the meaning of data) are vital, but there’s more. Snowflake has begun talking about “data and context” in the same breadth.

Snowflake is introducing new capabilities that bring “active context” to agentic AI, which is essential for the performance and accuracy of AI agents that act autonomously. Snowflake considers data and context to be one of four components of the agentic enterprise — the others being LLM Models, Interoperability, and the Control Plane. Snowflake enables context through Horizon Context, which applies metadata and semantics, and Cortex Sense, the “brains” of its AI agents.

As someone who has been covering enterprise data warehouses since the mid-90s, I have long believed that data sharing is where there’s immense value yet to be realized. More than half of Snowflake’s customers already share data, and now there are new ways to do it. It introduced Open Sharing, which allows users to forward datasets that have been shared with them, and Multi-party Collaboration in secure Clean Rooms.

Steep learning curve

When I say that Snowflake matters now more than ever what I mean is that, given its 10-plus years of innovation in data clouds, Snowflake has the platform and the know-how to help its customers make the shift to omnipresent intelligence. Snowflake has built its business putting more data and better data into the hands of more users. That’s a strong position to be in.

There are various ways that Snowflake, and its customers, could get tripped up. Security and governance will make or break agentic strategies. That includes identity management to give AI agents data-access permissions — and revoke them. It will be a steep learning curve.

Another challenge will be heterogeneous agentic AI. Snowflake’s pending acquisition of Natoma, developer of an MCP platform, is a step in that direction, but mixed, multi-agent workflows will get complicated fast.

Given the pace of change, any perceived leadership margins are thin for Snowflake and everyone else. Success will depend not just product innovation and speed, but helping customers transform into full-fledged agentic enterprises.

Other perspectives

I crossed paths with Myles Suer, Dave Vellante, and Brad Shimmin at Snowflake Summit. Sharing their insights here:

The Agentic Shift at Snowflake Summit: Finding a Platform’s ‘Right to Win,’ by Myles Suer, InformationWeek

Snowflake Moves Up the AI Stack — But the System of Intelligence Is Still Being Built, by Dave Vellante and George Gilbert, Silicon Angle

Recap by Futurum analyst Brad Shimmin, on LinkedIn

And more from the Cloud Database Report…

Subscribe now

Gartner’s annual chart of database vendor revenues for 2025 shows a market that’s remarkably stable at the top — with no change whatsoever among the top nine database providers over the past three years.

In fact, there has been no churn among the top 17 vendors over the past 2 years, with one exception: Tencent passed Huawei. That insight is courtesy of Adam Ronthal, the Gartner VP who created and shared the awesome chart below along with some of his own valuable takeaways on LinkedIn.

If you focus only on the database leaders in recent years, it’s easy to conclude that the market is unwavering. But that’s not really the case. Under the surface, other players — Databricks, Cockroach Labs, Tigergraph, Clickhouse, and others — moved up in the ranking. And ServiceNow has been added to the mix.

What’s obvious at first glance is that the market leaders — AWS, Microsoft, Oracle, and Google Cloud, in that order — are unyielding fixtures, with no changes in their rankings over the past four years. But the waves of activity under the surface reveal that disruption is alive and well.

The churn is driven by AI and the rapidly changing nature of modern data management — data distribution, automation, multimodal data, reasoning, and hybrid transaction/analytical processing (HTAP). A few examples of the shifting paradigm, among many, include Snowflake’s data cloud, Databricks’ lakehouse, Cockroach Labs’ distributed architecture, and Salesforce’s zero-copy Data 360, each of which are representative of the changing landscape. On that point, Salesforce isn’t a cloud database vendor, per se, but its platform integrates, aggregates, analyzes, and shares data as if it were.

One thing that hasn’t changed: The 50-year-old relational database model is still going strong. Relational DBMSs, led by MySQL and PostgreSQL, are as popular as ever. Databricks and Snowflake last year added Postgres-compatible relational databases to their data architectures. See my post below for more on that.

Disruptive forces

Despite the locked-in status of the database leaders, major changes are afoot:

• There’s tremendous development activity driven by AI and integration with LLMs. Legacy databases are being recast as AI databases.

• There’s more data and new/emerging data types such as vectors, synthetic, audio, spatial, enumerated, and so on. Unstructured data is no longer locked up as PDFs, email, and media files. Databases are adapting to accommodate these new data types and/or special purpose databases like Pinecone and Weaviate (see chart) are springing up. Carnegie Mellon’s Database of Databases recorded 15 new databases in 2025.

• Data architectures are increasingly decentralized, with data mesh, fabrics, data clouds, multi-cloud databases, edge devices, and more. Data sovereignty and resiliency requirements are forcing new thinking around how and where data gets stored.

• Agentic AI and autonomous actions are game changers. AI agents must be able to tap myriad data sources (via MCP) in milliseconds. Those actions are based on the principles of learning and decision making via a reasoning engine. That’s the idea behind Google Cloud’s just announced Agentic Data Cloud, described as an AI-native architecture that evolves the enterprise data platform from “a static repository into a dynamic reasoning engine.”

Smaller fish, bigger pond

There’s little evidence that the Big 4 — AWS, Microsoft, Oracle, Google Cloud — are at risk of being displaced anytime soon. Each is doing its own version of the kinds of advances described above.

However, as the data and AI universe expands, the centralized databases of the Boomer generation are no longer the predominant model. Industry growth rates tell us that the database market is growing at a slower rate (↑13.2% to 16.5%) than AI in its various manifestations (global LLM market, ↑31.2% - 36.9%; global agentic AI market, ↑40.5% - 46.2 %; global GPU market, ↑30.2 % - 34.5%).

In other words, databases are becoming a smaller fish in a bigger AI pond, which helps explain the architectural shift underway.

Ephemeral databases

There’s something else at play. More databases are being created autonomously by AI agents, as you can see in the Databricks chart below.

As this happens, it points to a net increase in the sheer number of databases, driven by agentic AI. These are not the monolithic databases of the past; many are “small, ephemeral databases,” according to a Databricks post, “How agentic software development will change databases.”

“Our data shows that for about half of these agentic applications, the database compute lifetime is less than 10 seconds,” according to Databricks. “Traditional databases were designed as always-on infrastructure components with fixed provisioning and operational overhead. That model fits large, stable applications but fails economically when applications are numerous, ephemeral, and short-lived.”

Going forward, there will almost certainly be millions more databases deployed and in use. Today, Google Gemini says there are 29 million active enterprise-grade database instances supporting business operations. When you include small-footprint databases like SQLite on mobile devices and IoT, the number exceeds 1 trillion, per Gemini.

Note that Gartner’s market share chart is based on revenue, so pure-play open source databases aren’t represented beyond their commercial offerings.

Less is more

Data gravity is still very real — and helps explain why the market leaders are so hard to displace — but DBMSs are becoming less monolithic. Google Cloud describes its next-gen, cross-cloud, AI-native lakehouse as borderless. Cockroach Labs’ architecture has been called regionless.

So less is more in today’s database design.

Of course, we know that change is constant in the database market. Gartner’s spaghetti chart shows it’s been that way since 2011 — and well before, I would add. The market leaders may appear to be set in stone for now, but everything else is changing fast.

More reading

Subscribe now

This post was 100% written by a human!

And that explains why it’s been a few weeks since my last post. I’ve been busy working with a “neocloud” GPU startup, researching and writing about post-quantum cryptography, and traveling to Ireland. My next stop: a potato farm in the Midwest to report on data-driven agriculture.

I also signed up for the 2026 New York City marathon, only the second marathon I’ve ever taken on. The first was Chicago in 1986 — 40 years ago. Any advice from experienced runners on how to bridge that gap is welcome.

Now on to what’s been happening in databases, data, and AI with some added color and commentary.

1. AWS S3 Files. In terms of data-management breakthroughs, AWS’s announcement of Amazon S3 Files is at the top of the list. S3 buckets are now available as file systems. This means that AWS’s S3 object storage service now supports the kind of file system access (i.e. read, write, directory) that users and apps are familiar with. You can read AWS’s detailed explainer, including an architectural diagram, here. The significance, according to AWS, is that S3 becomes “the central hub for all your organization’s data.” That’s a big claim, but probably accurate for hardcore AWS customers. For much more on this, see Werner Vogel’s 6,000-word article on the “the changing face of S3.”

2. AWS data centers targeted. Data centers are under attack, literally. Amid the war in Iran, AWS facilities in the Middle East were struck by Iranian drones, resulting in power disruption and structural damage. AWS advised affected customers to enact their disaster recovery plans and redirect workloads to alternate AWS regions. An Oracle data center in Dubai was also reportedly targeted, in what was described as a minor incident. All of which underscores the need for a resilient data architecture. See more reporting here from Big Technology, and check out my earlier post (below) on region-less data architecture.

3. Oracle’s AI agent factory. Oracle as been in the news for all the wrong reasons — an AI spending spree, tanking stock price, widespread layoffs. At the same time, Oracle continues to be at the forefront of AI-enabled databases. It has introduced a “private agent factory,” a secure, container-based environment comprised of both pre-built AI agents and a no-code agent builder. The pre-builts include a Database Knowledge Agent, Structured Data Analysis Agent, and Deep Data Research Agent. We’re seeing more of this kind of agent-driven automation, which promises to reduce the complexity of data management, from a growing number of vendors.

4. Google Cloud’s omnipresent MCPs. Every database vendor has an MCP server or two, but Google Cloud is dishing them out like hotcakes at a Sunday breakfast for developers. First, it introduced MCP support for Google Maps and BigQuery. That was followed in February with MCP servers for AlloyDB, Spanner, Cloud SQL, Firestore, and Bigtable. Next, Google Cloud says it will provide managed MCP support for Looker, Database Migration Service, BigQuery Migration Service, Memorystore, Database Center, Pub/Sub, Kafka and more. This means AI agents will be able to access and share data with databases, apps, and LLMs, both in and out of Google Cloud. Google Cloud’s holistic approach illustrates how data infrastructure is rapidly evolving for agentic AI.

5. IBM Db2 goes genius. As the next chapter in the 40-year history of Db2, IBM has introduced “the Db2 autonomous database experience.” It’s powered by a so-called Genius Hub, which IBM describes as “an AI-powered data management console that introduces autonomous capabilities for Db2 operations, grounded in AI-driven reasoning and built in agentic workflows.” In its current iteration, Genius Hub requires a human in the loop, but it’s moving in the direction of autonomous operation. Its primary use cases are maintenance, healing, and response for agentic operations.

6. Teradata’s AI drumbeat. Good ole Teradata, a company I’ve been covering since the data warehouse days of the 1990s, is continuing its AI push. In January, the company announced an AI agent builder called Enterprise AgentStack, then it introduced a Data Analyst Agent in Google Cloud Marketplace. In March, Teradata introduced a beefed up Vector Store with support for unstructured data, multi-modal embeddings (text, image, audio), hybrid search, and LangChain integration for faster/easier enabling of AI workflows. A long way from the National Cash Register Company!

7. EDB repositions. EDB, the artist formerly known as Enterprise DB, is positioning itself as a “Postgres data and AI platform company,” according to a recently published (paid) blog post on TechCrunch. This strikes me as partly rebranding, but the company’s Chief Product Engineering Officer, Jozef de Vries, outlines several planned capabilities, including support for non-Postgres databases; lifecycle management; point-to-point database migrations; and the ability to run analytical workloads on Postgres clusters. EDB’s website describes its system as the “world’s first sovereign data and AI platform,” but there’s no mention of sovereign (i.e. self-managed) in the blog post. I mention this because “sovereign AI” is getting a lot of buzz, and I’m not sure it’s widely understood.

8. Salesforce and Snowflake data breaches. Salesforce has been forced to respond to a series of attacks on its Experience Cloud customers. Salesforce blamed the issue on “customer configuration scenarios that may result in unintended exposure of customer data.” (Translation: “It’s not me, it’s you.”) Snowflake got hit, too. But data security is always a pitched battle; read on for the other side of the story.

9. Databricks, Anthropic and the promise of data protection. Databricks has jumped into the cybersecurity fight with Lakewatch, a lakehouse that uses AI agents to detect and respond to threats. Not to be ignored, Oracle announced its own data security safeguards, including post-quantum cryptography. Yet, the most intriguing new development is Anthropic’s Project Glasswing, which aims to identify vulnerabilities in critical software using a new model called Claude Mythos. But there’s a Catch 22 — concern that Mythos, now in preview, could be misused. So for now, Anthropic is limiting access to its cybersecurity model.

And in case you missed it…

Share

Subscribe now

Not long ago concerns about AI’s potential to disrupt the SaaS market resulted in a sharp drop in the stocks of Salesforce, ServiceNow, Workday, and others. Then, after the so-called SaaS-pocalypse, Anthropic announced Claude is now able to modernize Cobol. IBM’s share price plummeted 13%.

Are databases the next software sector to be disrupted and potentially displaced by AI? To borrow from Marc Andreessen, is AI about to eat the database world?

There are a couple of ways this could happen. One is that AI actually begins to function in ways similar to a database management system, obviating the need for traditional DBMSs. And two is the mere perception that this could happen, leading to fear, uncertainty, and doubt among the incumbents and their investors.

Given the skittishness of tech investors, no one should be surprised if the AI fear factor comes into play in databases. (Oracle is already down 22% YTD but for other reasons.) On Monday, the Dow dropped 800 points based in part on what a Wall Street Journal headline described as a “doomsday report” by Citrini Research. The report described a hypothetical scenario in which agentic AI had a cascading effect on productivity (up), white collar jobs (down), industry moats (destroyed), and the economy (negative).

It’s a thought-provoking post. The comments are worth perusing, too.

Yet, much of the buzz around the SaaS-pocalypse and the Citrini piece seems to oversimplify how difficult it would be for AI to replace legacy software, which is inherently complex and highly integrated. What’s more, many CIOs have been moving away from do-it-yourself IT. The LLM model would seem to put them back in that business.

AI is already edging into database management

I’m not saying AI displacement can’t or won’t happen to other areas of legacy software. There’s little doubt that LLMs and agentic AI are taking over some of those tasks. We’re seeing it with vibe coding, conversational and interactive AI agents, analytics, etc.

So, what about databases? It’s already happening. Two examples:

• Fundamentals, a startup, recently introduced a new type of foundational model called a “large tabular model,” which it describes as “purpose-built from the ground-up to tackle structure data.” With $255 million in funding and a partnership with AWS, its product, Nexus, is designed to drive predictive intelligence using enterprise data.

• Notion has developed an AI workspace that includes the ability to create a database. For example, a user could ask Notion AI to make a customer feedback tracker, according to the company. That would result in a preview of the database along with options for features to add or remove.

These aren’t enterprise-class replacements for Oracle, MongoDB or Postgres. But you can see how AI-derived capabilities are beginning to edge into the space currently occupied by relational and other mainstream databases.

Software vendors try to regain the narrative

I’m of the school of thought that, for the foreseeable future, AI capabilities are more likely to augment most kinds of software, including databases, rather than replace them. It’s not either/or; it’s both. And they are better together.

SaaS vendors are trying to make this argument to establish their staying power — and to undo the damage done to their market caps.

No doubt, we will hear more about this during their earnings calls this week. Today’s Business Insider newsletter put it this way: “After being marked for dead, software companies have a chance to tell their side of the story.”

Regarding databases, I’ve been documenting the rise of AI databases that are a blend of many years of database development plus newer AI capabilities, such as vector support, built-in LLMs, and AI frameworks. Here’s the link in case you missed it.

One of the most promising trends in the database market is the way that AI is increasingly used to automate labor-intensive and error-prone manual processes. AI assistance can be a boon to UX, data prep and pipelines, analytics, and database operations. Last year, Informatica introduced AI agents for data quality, data discovery, and data ingestion. I’m sure we will continue to see many innovations like these.

Hindsight is 20/20

Where does this leave us? The software market and data management are in flux as AI finds its way into every layer of the tech stack.

Here’s another take on that: During the SaaS-pocalypse (which is ongoing), there was talk about the way that SaaS platforms serve as a system of record for business data. In a Substack post, Zain Hoda wrote that systems of record are the “most defensible position in enterprise software.” But then, he asserted, “this position is about to collapse.”

I don’t see that as an immediate risk to the leading database platforms. We’re still in the early stages of AI disruption, and there are many ways it can play out.

It was clear from the start that LLMs would have implications for database management. Three years ago, I wrote the headline below, asking if ChatGPT was poised to disrupt the database industry.

Now, the obvious answer is “yes.” We see more evidence of that with every passing day.

Subscribe now

Database reliability is off to an inauspicious start in 2026.

Just a few weeks into the new year and already a database glitch has resulted in a system failure that might have caused your heart to stop if you were an affected user. In the early hours of January 19, the price of Bitcoin dropped to $0 on Paradex, the crypto exchange that encountered the issue.

The problem was blamed on database maintenance gone awry, according to reports. The platform was offline for six hours.

I hate to say I told you so but…just a few weeks ago I implored the database community to turn its attention to the challenges of database reliability with greater urgency. “We must do better,” I wrote then — and I repeat it now.

As I outlined in that earlier post, the past year was pockmarked by database mishaps, from issues affecting data services on Microsoft Azure in January 2025 to database-related outages at Snowflake in December 2025. And there were other bugs, blips, and breakdowns in the 10 months between those ignominious bookends.

I’m reposting the article below in case you missed it. It should be required reading for anyone who develops, markets, manages, or invests in database technology, which is most anyone who subscribes to this newsletter.

The obvious question this time is what went wrong at Paradex?

Just yesterday (Jan. 23), Paradex provided a post-mortem on X. Here’s its explanation:

“On Jan 19, a planned 30-minute maintenance window to upgrade our database (to support growing demand) encountered unexpected issues during the scale-up process. A race condition during a service restart, while critical data operations were in progress, caused a corrupted state to be persisted in the cloud and published to Paradex Chain. As a result, some markets reset the funding index to 0, creating abnormal funding PnL that triggered liquidations across multiple markets.”

You can read Paradex’s full incident report on X.

And here’s some additional info it put out. Paradex says user balances were not affected and there was no loss of funds.

Complexity has become ‘the dominant failure mode’

Stepping back from the Paradex incident and taking a broader perspective, the ongoing issues with database reliability can’t be blamed simply on software misconfigurations, disk crashes, network outages, or the dozen other factors that sometimes go wrong. It’s that any of those possibilities can happen at any time.

I asked Spencer Kimball, CEO of Cockroach Labs, which strives for “zero downtime,” why these database troubles continue to hamper businesses and everyday life, even after all of the R&D that has gone into resiliency and “self-healing” databases.

“Resilience keeps getting harder because modern databases aren’t failing in one place anymore,” Kimball wrote back. “Data volume, distributed systems, human intervention, and infrastructure dependencies all interact in ways that make simple redundancy insufficient.”

“It is not for lack of resiliency tools,” Kimball added. “The problem is that complexity itself has become the dominant failure mode.”

Yes. And as I pointed out in my earlier post, as tech stack complexity rises so does the need for resiliency. It’s like a dog chasing its tail.

At what point does resiliency overcome complexity? Not for the foreseeable future, IMO. But hopefully the emergence of AI-driven automation, along with wider and more rigorous adoption of best practices like distributed architectures, will lessen the frequency and impact of these all-too-frequent fiascos.

Additional reading

Subscribe now

Many different databases are used for AI workloads and applications. Few of them are actually called AI databases.

That’s beginning to change. While the database market is deeply rooted in legacy terminology—columnar, key-values, event stores, etc.—the action today is around AI platforms for AI workloads. Increasingly, databases are being designed and optimized for AI. Product marketing teams will see this as an opportunity.

I’m not talking about vector databases, which are becoming commodity-like as vector support goes mainstream in everything from Oracle to MongoDB to Redis. Vectors are one aspect of AI databases, but there’s more: LLMs, machine learning algorithms, knowledge graphs, AI frameworks, agentic AI protocols, automation of data management, RAG, natural language queries, and more.

As vendors bundle these AI capabilities into their database management systems, there’s a crossover point where they become, in effect, AI databases. Oracle was one of the first to seize on this when it changed the suffix of its flagship database from 21c (for cloud) to 23ai nearly two years ago.

In October 2025, Oracle took it a step further. It introduced Oracle AI Database 26ai with even more AI capabilities layered into the system, including an MCP server for AI agents and support for private AI models. In the transition from 23ai to 26ai, Oracle made a small but significant change to its naming convention. It began referring to its platform as Oracle AI Database. Larry Ellison now uses “Oracle AI Database” on earnings calls.

Oracle’s not the only database vendor thinking this way. As I reported exclusively last July, Databricks CEO Ali Ghodsi has been lobbying internally (unsuccessfully so far) for Databricks to call its platform as an AI database.

Here’s what Ghodsi said at the time:

“Fun fact: for many years (since 2016) I’ve been advocating for Databricks to say it’s an AI Database,” Ghodsi commented on LinkedIn. “But all marketing firms we worked with, and all the user surveys, showed that people hated it. I still think it’s the most accurate way in 2 words to understand what Databricks does (and has been doing since 2013)!”

See my post, “Databricks Is an ‘AI Database’,” for the full story.

About this same time, Microsoft released SQL Server 2025, which it described as its AI release. “SQL Server 2025 is the first release where we’re basically putting AI capabilities right into the heart of the engine,” Shireesh Thota, Corporate VP of Databases, told me in an interview.

‘Dawn of the AI-native database’

Google Cloud has jumped into the conversation, too. Yasmeen Ahmad, Google Cloud’s managing director of Data Cloud, recently published a blog post, “From system of record to system of reason: The rise of the AI-native database.”

Ahmad focuses on the tech stack needed for agentic AI, so the definition is narrower.

“The database can no longer be a passive record-keeper,” Ahmad writes. “It must become a system of reason — an active, intelligent platform that serves as the agent’s conscience. Beyond recording what an agent did, it must provide an immutable, explainable ‘chain of thought’ for why it did it. This is the dawn of the AI-native database.”

Ahmad explains how different databases are used to support multi-tiered memory for agentic AI, with short-term memory for immediate tasks and long-term memory for knowledge over time. The database architecture to do this, Ahmad writes, might comprise Google Spanner (for short-term memory), BigQuery (for long-term memory), and GraphRAG (graph-structured data for retrieval-augmented generation) to enable AI agents to work across information sources.

Ahmad’s article provides an nice overview of how to think about a system of reasoning. To be clear, however, they are not talking about a single AI database but rather several databases as part of an overall solution. If you click on “AI-native database” in that post, it links to the full portfolio of Google Cloud databases. In other words, Google Cloud is offering a menu of AI-native databases, so customers can pick and choose among them. But Google Cloud doesn’t have a centerpiece AI database in the way that Oracle does.

Likewise with AWS, which offers a range of AI-capable databases (Aurora, DocumentDB, DynamoDB) and complementary services (Bedrock, SageMaker). AWS doesn’t have just one product branded as its AI database.

In short, the AI database phenomenon is largely about new capabilities but also about how you describe these systems to a world that is looking for them.

Another hype cycle?

Some may argue that calling a system an “AI database” is marketing spin. It’s true that various types of database systems can and are being used for AI workloads, as the examples of Google Cloud and AWS show. There’s no one size fits all for AI projects.

However, it’s also true that a growing number of database platforms are being optimized for AI use cases. And when you consider that Oracle now refers to its flagship system as an AI database, one has to wonder if others will follow suit.

I think they will. In fact, Microsoft now describes Cosmos DB as an AI database on its product page. But it has yet to go all-in on that. Microsoft also refers to Cosmos DB as “a fully managed NoSQL, relational, and vector database.”

To be sure, we can’t dispense with the traditional database descriptors completely. Different types of database engines may be under the hood of AI databases. Vendors will have to be smart and savvy in how they position and talk about these next-gen DBMSs. They must be careful not to wash everything with an AI brush.

This reminds me of the early days of cloud computing, when everyone jumped onto the cloud bandwagon—much to the ire of Oracle’s Larry Ellison, who in 2008 described the hubbub over clouds as “idiocy,” “crap,” and “gibberish.” Now with AI databases, Ellison is ahead of the buzz.

All that cloud talk could offer a lesson in what’s next. Over time, the cloud market evolved into different types of cloud services—SaaS, PaaS, IaaS. The same kind of thing could happen with AI databases—relational AI databases, graph AI databases, etc.

ClickHouse joins AI database race

Consider the example of ClickHouse, the analytics and data warehouse startup that last week revealed $400 million in Series D financing at a $15 billion valuation.

As part of its announcement, ClickHouse disclosed a partnership with Ubicloud to include a Postgres database service within its platform as a way of combining transactional and analytical capabilities for AI dev and apps. (Databricks and Snowflake each introduced transactions plus analytics on Postgres in the second quarter of 2025.)

What’s interesting is that ClickHouse describes its product/service as an analytical database, or more specifically as an “open-source columnar database management system built for real-time data processing and analytics at scale.”

That language works fine for CTOs, DBAs, and other technologists who are familiar with what a columnar database is, but it may fly over the heads of non-techies.

In contrast, Bloomberg came up with a different and more accessible way of talking about the same thing. Its headline: “ClickHouse Lands $15 Billion Valuation in AI Database Race.” There’s that term again.

I’m not suggesting vendors slap “AI database” on every DBMS that packs a few AI features. But if it walks like an AI database and talks like an AI database, maybe it’s an AI database.

Additional reading

Share

Subscribe now

Where are all the self-healing databases?

We’ve been hearing about databases that autonomously repair themselves for more than 20 years. It was back in 2004 that the Journal of Intelligent Information Systems published research on the design of a self-healing database system. That same year, Oracle published a seminal white paper, “The Self-Healing Database.”

Goes to show that the tech industry doesn’t always move at lightning speed. Here it is 2025 and we’re still waiting for the full promise of self-detecting, self-diagnosing, self-fixing databases. And without them, database gotchas are costing businesses billions of dollars in downtime.

Q4 provided two high-profile examples of what can go wrong. On Oct. 19, a glitch with AWS’s DynamoDB service disrupted digital business for global brands including Canva, Coinbase, Hulu, Lloyds Bank, and McDonald’s. See my full analysis here.

Then, on Nov. 18, Cloudflare experienced a disruption that froze websites and services, including X and ChatGPT. Cloudflare blamed the failure on a change in database configuration that caused its services — core CDN and security services, Turnstile, Workers KV, Dashboard, and more — to topple like dominoes.

Cloudflare’s postmortem was painful to read. Cloudflare initially thought it was under cyberattack, then figured out it was an internal issue. Intra-system dependencies and unrelated problems (e.g. its Status Page happened to go on the blink) further confused the situation. It took eight hours to repair and recover.

There were other miscues throughout the year.

• In January, Microsoft Azure networking issues affected Azure Databricks, Azure Database for PostgreSQL, Azure SQL Database, and other services.

• In February, Slack suffered a 9-hour service outage related to database sharding.

• In April, a hospital group was forced to switch to paper-based patient records following a maintenance mishap involving Oracle, according to CNBC.

• In June, it was Google Cloud’s turn with a 7-hour outage that affected various products, including some of its data and database services.

• In July, a coding agent wiped out a live database.

• In November, border crossings in war-torn Ukraine were temporarily halted due to a technical issue with a new database.

• In December, Snowflake responded to two separate database-related outages.

It’s not always clear which database platforms were involved in these misadventures, but a little digging sometimes gets at it. The New Stack article on the Slack outage (link above) surmises that it might trace back to MySQL/Vitess. Forensics are often required.

The existing HA playbook isn’t enough

IT leaders recognize that database-related glitches pose a risk to their business operations. In a recent survey by Yugabyte, high availability (HA) and resilience were the #1 priority — cited by 71% of respondents — when technologists were asked what database attributes were most important to their modernization plans.

The existing resiliency playbook includes a handful of familiar strategies: HA architecture (clustering, replication, failover); distribution among zones and regions; backup and recovery; monitoring and observability; and change and release management.

But those things obviously aren’t enough given the outages we continue to witness and experience. The incidents with AWS and Cloudflare are a head-scratcher because one would think they should know how to avoid such gaffes.

Highest possible availability?

All of the leading database vendors, and built-for-resiliency specialists like Cockroach Labs and Yugabyte, offer solutions meant to tamp down system outages: distributed databases, managed services, different types of replication, partitioning/sharding, clustering.

In one recent example, Oracle announced GA of Oracle Globally Distributed Autonomous Database, which it describes as follows:

“Oracle’s Globally Distributed Autonomous Database splits a single logical database into multiple physical databases (called shards) that are distributed across multiple data centers, availability domains, or regions. Faults in one shard do not affect others, enhancing overall availability. Automatic replication of shards across domains or regions provides protection from outages. The Oracle Globally Distributed Autonomous Database runs on fault-tolerant Exadata infrastructure for the highest possible availability.”

Highest possible availability sounds good until the .001% chance of failure happens, which is the slim-but-real possibility in a “five 9s” (99.999%) service-level agreement.

We can’t expect perfection, but we must do better. AI-enabled self-healing databases are the next great hope.

AI to the rescue

Vendors are introducing AI capabilities to increase the resiliency and availability of their DBMSs. There are various ways that’s happening, as you can see from Perplexity.ai’s response to my query below.

Query: What are some examples of how AI capabilities are being used within database management systems to make them more resilient and highly available?

Response: AI is being embedded into modern databases to predict failures, automate recovery, and continuously tune systems so they stay online and performant with minimal human intervention.

Predictive failure detection

• ML models analyze metrics such as latency, error rates, log patterns, and resource usage to forecast node or cluster problems before they cause outages.​

Autonomous self-healing and repair

• Autonomous databases combine health monitoring with machine learning to trigger automatic repair actions such as instance restarts, resource rebalancing, or configuration changes when risk thresholds are crossed.​

Intelligent failover, backup, and recovery

• AI systems can choose optimal backup windows and retention policies based on observed workloads and risk patterns, reducing the chance that restores impact peak usage.​ In distributed cloud databases, AI helps orchestrate replication and automated failover across regions or availability zones to maintain service during node, zone, or even regional failures.​

AI-driven anomaly detection for resilience

• Generative and discriminative models are used to detect anomalous workload behavior or access patterns that might signal misconfigurations, partial failures, or security incidents before they propagate.​ These models help maintain both availability and data integrity by flagging and sometimes isolating suspicious nodes or traffic without taking the entire cluster offline.​

Adaptive performance and workload management

• ML-based query optimizers learn from past executions to choose better plans, reduce resource contention, and avoid pathological plans that can stall or crash nodes under load.​ AI agents continuously tune indexes, caching, and placement of data/partitions, which can mitigate hot spots and improve resilience under skewed or bursty traffic.​

> End of Perplexity response <

‘Added complexity’

Why does resilience continue to be such a pain in the DBMS derriere? In a word, complexity.

“The use of software-based resiliency strategies alongside physical failover/redundancy is undoubtedly contributing to overall improvements in availability,” according to a report by the Uptime Institute. “However, the added complexity brings its own challenges and can blur lines of responsibility for failures, complicating root cause analysis and outage classification.”

There’s more data, new data types, and more devices and AI agents generating data, so databases are growing in size. By my estimate, there may be 10x, 100x, or even 1,000x more data to manage.

Also, there are infrastructure dependencies — hardware, network, power supply. Other risks include software bugs, data or file corruption, and cyberattacks. And human error is a factor all too often. “The failure of staff to follow procedures has become an even greater cause of outages,” finds the Uptime Institute.

The conundrum is that even as complexity rises so does the need for greater resiliency. We can only hope that Self-Healing Databases 2.0 do not take another 20 years to get here.

Additional reading

Share

Subscribe now

For 14 years, Amazon CTO Werner Vogels has been a fixture at the company’s re:Invent conference, regaling the audience with big ideas on development philosophy, system architecture, cloud infrastructure, and distributed system design. It’s been equal parts education, motivation, entertainment. At re:Invent 2025 in Las Vegas two weeks ago, Vogels gave what he said would be his last re:Invent keynote.

Amazon’s CTO, whose thinking has contributed to the buildout of AWS as the world’s most expansive computing infrastructure, tackled at re:Invent one the most concerning issues facing software engineers: Will AI take their jobs?

What followed was an hour-long exhortation on why that need not happen — and a five-point plan on how software engineers and developers can stay relevant and vital even as AI takes over more of the coding. “Maybe we should rephrase and reframe the question,” Vogels said. “Will AI make me obsolete? Absolutely not if you evolve.”

It’s an encouraging take for developers, whose roles are in the danger zone for AI encroachment. “Early-career software engineers are among the hardest hit,” according to IEEE.

Vogels observed that AI is advancing in parallel with other scientific “golden ages,” including space travel and robotics, comparing it to a modern day Renaissance. Then he introduced his big idea — the renaissance developer — meant to inspire software developers to rise up to the perceived threat that AI poses to their jobs and careers.

Vogels cited the inventiveness of Leonard da Vinci and Johannes Gutenberg, alluded to the work on complex systems of scientist Donella Meadows, and intoned Walt Whitman on a willingness to learn.

“It is your work that matters. Our tools have changed so many times over the course of my career, and they will continue to change,” he said, pointing out that his career began with Cobol. “We’re still builders. We’re still important. Nothing has changed. There’s never been a time to be more excited about being a developer.”

It was a pep talk for the thousands of developers in the audience. I would add that Vogels’ optimistic message about the rapid emergence of AI and its impact on jobs can be applied broadly to people across business and technology. “Remember, the work is yours,” he extolled, “not that of the tools.”

5 qualities of a renaissance developer

How do IT builders thrive amid AI and not just survive?

Vogels laid out a five-point plan.

1. Be curious. “Every developer I’ve met has this instinct to take something apart and then look at how it works, and that’s also what drives us here, the desire to understand, to improve, to build,” he said. “You have to protect that instinct to stay curious, because curiosity leads to learning and invention.”

2. Think in complete systems. “When thinking in systems, we think in complete systems, not just in isolated parts. Every service, every API, every queue is part of a larger system….The renaissance developer thinks in systems. To build resilient systems, you need to understand bigger picture.”

3. Communicate clearly. “The ability to express your thinking clearly is just as critical as the thinking itself. And this is why I believe that one of the most important things that an engineer or a technical leader can do for their career is to practice developing strong communication skills.”

4. Own it. “AI will let us build things, bigger systems, explore more ideas, move faster than ever before….But there is a risk in how some developers are starting to use these tools. Vibe coding is fine, but only if you pay close attention to what is being built….You build it, you own it.”

5. Become a polymath. “It’s about having deep domain experience, but also having knowledge that spans many different subjects….That breadth of knowledge gives you the perspective to improve what you build, because you understand the tradeoffs. ‘T-shaped’ developers combine depth with breadth. We can dive deep into a specific product, but will also understand how it fits into a larger system.”

Vogels summed it up as follows.

“You need to be curious and keep learning. Think in systems. Communicate with precision. Be an owner. If you build it, you own it. And finally, become a polymath, expand your knowledge.”

The ‘T-shaped’ technologist

There was poignant moment when Vogels paid tribute to his old friend and mentor, the late Microsoft researcher Jim Gray, as the prototypical “T-shaped” technologist.

“He understood people, he understood business, and he understood a wide range of technologies,” Vogels said of the Turing Award winner. “[He was] deep in one domain, but broad in understanding the skills you need to be successful in your job are a unique mix of personal skills, functional depth, industry knowledge.”

As a reporter at InformationWeek, I interviewed Gray a handful of times about his work on Microsoft’s SQL Server database and his ideas on building high-scale systems with Windows NT servers. For more on that, see my post, “Microsoft’s Rise in the Enterprise.”

Tragically, the brilliant, much-admired Gray was lost at sea in 2007 when his sailboat disappeared on a solo voyage to the Farallon Islands, where he had gone to scatter the ashes of his mother.

I would note that Jim Gray and Werner Vogels, each in their own way, had a hand in two of the tech industry’s most expansive cloud environments, Azure and AWS, respectively. As I write this, the combined market caps of MSFT and AMZN are ≈ $6 trillion.

With that in mind, you might want to hear for yourself what Vogels had to say at re:Invent 2025. His keynote replay had surpassed 5 million views last time I checked.

Technologist, evangelist, chief cloud officer

Was re:Invent 2025 Vogels’ swan song after 21 years of helping to architect the AWS cloud, which now spans 38 geographic regions, 120 availability zones, 900 facilities in 50 countries, and millions of servers?

The 67-year-old CTO said this would be his final re:Invent keynote, though he added he’s not leaving Amazon just yet.

Whatever comes next, I have enjoyed following Vogels over the years. I first met him at an industry event in New York City not long after he joined Amazon in 2004, about the same time AWS’s first services (Simple Queue Service, Mechanical Turk, S3) were coming to market. While Andy Jassy — at the prodding of Jeff Bezos — is credited with conceiving the AWS business model, Vogels served as an external-facing technologist and evangelist.

In December 2008, InformationWeek named Vogels its Chief of the Year, crediting him as “the right person at the right time and place to guide cloud computing — until now, an emerging technology for early adopters — into the mainstream. He not only understands how to architect a global computing cloud consisting of tens of thousands of servers, but also how to engage CTOs, CIOs, and other professionals at customer companies in a discussion of how that architecture could potentially change the way they approach IT.”

At the time, I flew to Seattle to interview Vogels, along with the others in the photo below. The magazine cover (from left to right) featured SVP Andy Jassy (now Amazon CEO), VP Charlie Bell (now Microsoft EVP/Security), CTO Werner Vogels, and VP Adam Selipsky (who later served as AWS CEO until 2024).

Sorry, I was unable to find a working link to the InformationWeek article I wrote, but I dug out a print copy from a box in the basement. Back in 2008, Vogels told me his role had changed from thinking mostly about principles and strategy when he first joined Amazon to being a customer-facing technologist.

Now he’s stepping aside to make room for younger engineers and newer voices at Amazon who have their own stories to tell. “Werner out,” he said and exited stage left.

Additional reading

• “2026 Tech Predictions & The Dawn of the Renaissance Developer,” by Werner Vogels

• My Q&A with Werner Vogels, InformationWeek, 2008

• “Amazon’s Cloud Evangelist,” a profile of Vogels I wrote for Seattle Business Magazine in 2009

• My blog post below on Werner Vogels’ philosophy of asynchronous distributed systems architecture

Cloud Database Report

Database Innovations from A (AWS) to Z (Zilliz)

The database market’s cup runneth over with new platforms and cloud services—and more just keep on coming. In the past few days, ClickHouse launched a new OLAP cloud service, and Zilliz a new vector database service…

Read more

4 years ago · 2 likes · John Foley

Share

Subscribe now

Salesforce completed its $8 billion acquisition of Informatica on Nov. 18, six months after the definitive agreement was first disclosed and 18 months after an earlier, unsuccessful attempt. What happens now?

I will answer that in two parts. First, there’s what Salesforce and Informatica are saying about the path ahead. And then a few of my own ideas on additional steps they could take to get even more bang for the 8 billion bucks.

This week, two of the newly combined companies’ senior product leaders — Krish Vitaldevara, Informatica’s Chief Product Officer, and Rahul Auradkar, Salesforce EVP and GM of Unified Data Services, Data 360 and AI Foundations — joined a media briefing, where they provided a high-level overview of how the pieces will come together.

I’ve interviewed each of these guys on multiple occasions over the past year, so my analysis is informed by a long view of their respective strategies. The official game plan, laid out by Vitaldevara and Auradkar, looks like this:

1. Serve as a unified foundation for data integration, quality, and master data management.

Informatica’s full suite of data management tools (including data integration, data quality, cataloging, and master data management) will bring added capabilities to Salesforce’s Data 360 environment, formerly called Data Cloud. Salesforce’s MuleSoft, with its connectors and API management, is the third leg of this data & AI stool. On the company’s recent Q3 FY26 earnings call, CEO Marc Benioff guesstimated that Salesforce Data 360 + MuleSoft + Informatica could account for as much as $10 billion in revenue next year.

2. Establish an enterprise-wide data catalog enabled by metadata intelligence.

They offered little by way of new product development or a roadmap of deliverables, but there was a statement of direction: to combine Salesforce’s metadata model with Informatica’s data catalog to create a more comprehensive index for data visibility across enterprise systems, including ERP, manufacturing, inventory, and supply chains. Data fragmentation is a blocker to AI rollouts, said Vitaldevara.

3. Go beyond standalone data to provide ‘context’ that shows meaning, history, relationships.

A complete, metadata-enabled data index will give customers better visibility into things like data lineage, observability, compliance, and discoverability — all of which become doubly important as organizations build and deploy AI agents, where understanding what’s happening under the hood is critical. The buzzword that Salesforce and Informatica are using is “context.” Which is to say that, because businesses will be able to view data provenance — i.e. the history of data and how it’s been transformed — and data relationships across apps and systems, they can attain what Vitaldevara called “situational awareness.” That’s hugely important for agentic AI because agents must be able to act and react based on everything that’s happening across the business, not in mini silos.

My 2¢ on the $8 billion deal

The Salesforce-Informatica synergies are a net positive for both companies. Salesforce gains an impressive range of enterprise-class data management and data quality tools, which customers can use to clean up data for AI projects. And slow-growing Informatica gets new life as part of fast-moving Salesforce, which is the market leader in CRM and a market leader in agentic AI.

It’s a good start, but there’s more Salesforce and Informatica can do joined at the hip. Below are a few additional moves that, in my view, they should pursue ASAP given the window of opportunity. These may not be entirely new ideas for Salesforce and Informatica, but they could be force multipliers.

4. Seize leadership in the emerging multi-vendor, multi-agent, multi-cloud market.

Salesforce and Informatica have all the pieces to make this happen. Salesforce’s Data 360 and MuleSoft, combined with Informatica’s IDMC, comprise hundreds or thousands of data integration solutions, including MCP servers and the A2A protocol. Salesforce has the beginnings of multi-cloud with its Hyperforce offering, while Informatica has relationships with AWS, Databricks, Google Cloud, Microsoft, Oracle, and Snowflake. We know from experience that many CIOs will want and need the flexibility of multi-vendor, multi-agent, multi-cloud. Salesforce and Informatica are in a strong position to deliver it.

5. Innovate and drive adoption of AI-powered data management.

We hear a lot about the fact that many organizations are hampered in their AI aspirations due to poor data quality, but why? It’s because establishing trusted data is hard to do. The answer is reduce complexity and the heavy lifting of data management through automation. Salesforce and Informatica are both developing AI agents that automate this painstaking work. Earlier this year, Informatica demoed AI agents that take on the tasks of data quality, data discovery, and data ingestion. They need to get more of these capabilities into the hands of more customers.

6. Provide blueprints for faster AI agent implementations.

One of Salesforce’s challenges has been to get Agentforce customers from pilot to production more quickly. While Agentforce deployments were a bright spot in Q3, up 70% from Q2, we also heard that “the hardest part of the equation is the adoption.” They should borrow a page from Informatica’s generative AI blueprints, which are partner-specific architectural guidelines that are intended to ease implementations. For example, Salesforce could potentially provide a blueprint for Agentforce integration with Microsoft agents, another for ServiceNow agents, and so on. One of Salesforce/Informatica’s goals is to “be the Switzerland of AI.” Blueprints for AI agents would be a step in that direction.

7. Reinvent customer service and CRM.

Customer service is hopelessly broken. “Recent reports show that a sizable share of consumers have walked away from brands over service problems, not just product issues,” according to Perplexity. It wouldn’t be fair to blame Salesforce for the shortcomings of customer service, support, and relationship management that have plagued too many companies for too many years, but Salesforce may be in the best position to fix it. The Salesforce tech stack comprises a data layer, app layer, agent layer, and what may be the most important of all — an “experience layer” to elevate voice calls, online chat, text exchanges, Slack, etc.

Following Salesforce’s Q3 results, CNBC’s Jim Cramer called agentic AI the “holy grail” of customer service. I agree. If Salesforce and Informatica can help more organizations transform the customer experience, that alone would validate the acquisition.

Additional reading

Cloud Database Report

Salesforce and Informatica in $8 Billion AI Data Deal: 5 Strategic Synergies

Welcome everyone! I’m John Foley, a long-time tech journalist who has also worked in strategic comms at Oracle, IBM, and MongoDB. Now I’m an independent tech writer.Thanks for reading Cloud Database Report! Subscribe for free to receive my posts…

Read more

a year ago · 2 likes · John Foley

Cloud Database Report

Informatica’s Last, Big Solo Dance

Welcome everyone! I’m John Foley, a long-time tech journalist who has also worked in strategic comms at Oracle, IBM, and MongoDB. Now I’m an independent tech writer. I attended Informatica World on assignment for Method Communications.Thanks for reading Cloud Database Report! Subscribe for free to receive all my posts…

Read more

a year ago · John Foley

Cloud Database Report

Report from Salesforce's TDX 2025: AWS, Google Cloud, and New Tools to Build and Scale AI Agents

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, including 18 years at InformationWeek, who then worked in strategic comms at Oracle, IBM, and MongoDB. I invite you to subscribe, share, comment, and connect with me on LinkedIn…

Read more

a year ago · John Foley

Share

Subscribe now

Another day, another global computer system failure. This one started with Amazon’s DynamoDB cloud database.

Late in the day on Sunday, Oct. 19, AWS begin experiencing operational issues in its US-East-1 region, the epicenter of which is in northern Virginia. The glitch spread like a bad cold on an airplane. Websites, online services, and apps were severely hampered throughout Monday, the start of the business week.

Early estimates put the potential cost of the snafu into the hundreds of billions of dollars. Why so much? According to the New York Times, the global brands impacted included Canva, Coinbase, Hulu, Lloyds Bank, McDonald’s, Microsoft 365, Slack, Venmo, Whatsapp, and Zoom. Also affected were airlines, restaurants, financial services, and e-payment/point-of-sale systems. Even Internet-connected smart home devices were on the blink.

The last time the world experienced such a far-reaching service outage related to a software glitch was CrowdStrike’s infamous Content Validator bug in July 2024. At the time, it was described as the largest IT failure ever. This one may be worse.

AWS provided a detailed accounting of the cause and impact, which it described as happening in three phases over 14 hours. In AWS’s own words:

1. First, between 11:48 PM PT on October 19 and 2:40 AM PT on October 20, Amazon DynamoDB experienced increased API error rates in the N. Virginia (us-east-1) Region.

2. Second, between 5:30 AM and 2:09 PM on October 20, Network Load Balancer (NLB) experienced increased connection errors for some load balancers in the N. Virginia (us-east-1) Region. This was caused by health check failures in the NLB fleet, which resulted in increased connection errors on some NLBs.

3. Third, between 2:25 AM and 10:36 AM on October 20, new EC2 instance launches failed and, while instance launches began to succeed from 10:37 AM, some newly launched instances experienced connectivity issues which were resolved by 1:50 PM.

The sequence of events was caused initially by a Domain Name System (DNS) resolution error, according to AWS. DNS misconfigurations have long been associated with cloud service failures.

The little bit of good news: By day’s end on Monday, AWS said it had fully recovered.

The reaction

Industry experts responded as they always do when these things happen—with explanations of the root cause issues, warnings about over-reliance on too few service providers and/or single points of failure, and reminders about the need for resiliency in data and network architectures.

Here’s a roundup of some of the after-the-fact commentary:

• “Today’s AWS outage is a wake-up call for every business: If your stack depends on a single region or has hard dependencies on services that do, you’re prone to losing business continuity in the event of this class of cloud provider outages.” - Spencer Kimball, CEO Cockroach Labs, on LinkedIn.

• “This is about ensuring the uptime of a critical service, no matter what happens, not “if it happens.” We call this approach ultra-resilience. Be aware of the classes of failures that can hit you, and be deliberate about building resilience into your architecture.” - Karthik Ranganathan, co-CEO Yugabyte, on LinkedIn.

• “There’s been a massive amount of centralization in terms of our dependence on a small number of cloud providers. When they go down, so much of what we depend on goes down.” - Prof. David Choffnes, director of Northeastern University’s Cybersecurity and Privacy Institute, in Northeastern Global News.

• “The bottom line is clear: Public cloud remains the best option for scalable infrastructure if you invest in resilience upfront—or correct existing deployments if necessary. Don’t let fear steer you toward costly or ineffective alternatives; instead, double down on architecture, process discipline and transparent partnerships with your providers.” - Lydia Leong, Distinguished VP Analyst, Gartner.

• “Failures increasingly trace to [data] integrity. Corrupted data, failed validation or, in this case, broken name resolution that poisoned every downstream dependency. Until we better understand and protect integrity, our total focus on uptime is an illusion.” - Davi Ottenheimer, VP with Inrupt, a provider of enterprise wallet infrastructure, in Wired.

• “The only solution is to start thinking about resiliency at the Board level.” - Mehdi Daoudi, CEO of Catchpoint Systems, in IT Brew.

The need for 9’s

How does this sh*t continue to happen? Even the most steadfast experts on system and network reliability admit the occasional blip is nearly unavoidable. Blame it on the complexity of large scale, global systems that support millions of people and devices, billions of transactions, and exabytes of data each minute of the day.

Yugabyte’s Ranganathan puts it this way: “We can only aspire to reduce the probability of failure, not eliminate it. However small the chance of failure, given enough scale and exposure, there will inevitably be one.”

True enough, but there are steps technologists can take to minimize system failures and outages if not completely eliminate them. A common measure of system availability are the so-called “nines.” Four nines (99.99%) of availability translates into about 52 minutes of downtime per yer. Five nines (99.999%) equals 5 minutes of downtime per year. Six nines (99.9999%) brings it down to 31.5 seconds of downtime per year.

Amazon DynamoDB is available to customers with a 99.99% SLA and, when Global Tables are implemented (i.e. where database tables are replicated across multiple regions), there’s a 99.999% SLA. But those SLAs are for the managed DynamoDB services that AWS provides to its customers, and presumably don’t apply to AWS’s own mega outage.

You do pay a premium for the higher availability levels, of course. According to Perplexity, five 9s can cost ten times as much as four 9s. But then, if the cost of downtime is millions of dollars per hour, the extra cost may be worth it.

No doubt, CIOs, CTOs, and business leaders will use this latest cloud-service mishap to reassess their own level of readiness. They will go back to the drawing board on resiliency, backup, data distribution & replication, availability zones, contingency planning, and disaster recovery. And rightly so, because this may have been AWS’s latest and biggest service outage, but it won’t be the last.

Further reading

Cloud Database Report

CrowdStrike and the Hard Truth about Software Reliability

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, including 18 years at InformationWeek, who worked in strategic comms at Oracle, IBM, and MongoDB. I’m now a VP with Method Communications…

Read more

2 years ago · John Foley

Cloud Database Report

Software Obsolescence, Misconfiguration, and Failure

We have been reminded in the past few weeks of the costly and potentially dangerous implications of IT systems that are not well managed and maintained…

Read more

4 years ago · 3 comments · John Foley

Share

Subscribe now

Hello friends. I just got back from the corn and soybean fields of Illinois, where I’ve been going for the past few years to write about the fall harvest. Agriculture has become a data-intensive industry with AI, drones, analytics, etc. What’s next? Virtual agronomists for farming.

Yes, AI agents are showing up everywhere—across industries and in every nook and cranny of tech. Last year, Salesforce CEO Marc Benioff aspired to reach 1 billion agents. We’re not there yet, but it’s conceivable. We’ll get an update from Benioff at Dreamforce next week.

Meantime, the rest of the industry may be catching up to Salesforce. Here’s a rundown of 10 recent developments that have happened over just the past 10 days. They’re in chronological order, starting with most recent.

1. Google Cloud

I’ll start with Google Cloud’s announcement today of Gemini Enterprise, a new platform for rolling out AI agents—from Google, partners, or custom-built—using Google’s Gemini models, low-code tools, data connectors, and governance controls. Google Cloud joins a growing list of vendors with full-fledged agentic AI platforms. What stands out is the company’s full-stack approach, including infrastructure, models, an agent builder and developer kit, the A2A protocol, and an agent marketplace.

In a briefing I joined, CEO Thomas Kurian described Gemini Enterprise as the “front door” to all of Google Cloud’s AI technologies. Already, 65% of its customers are using Google Cloud’s AI products, so there’s a ready-to-go audience for the new platform.

2. Salesforce

In advance of its Dreamforce conference next week, Salesforce on Oct. 8 called attention to new agentic capabilities that strengthen security and compliance. As best I can tell, these new capabilities are baked into the company’s Security Center and Privacy Center, not separate SKUs.

What makes these agentic capabilities vs. simply new features, as in the past? If it walks and talks like an agent, it’s an agent. For example, you can use natural language queries to ask about potential security threats. And Agentforce can autonomously identify compliance risks. I have no doubt we’ll see more agent implementations like this from others as well—integrated capabilities that do their thing when switched on.

Salesforce says CrowdStrike integration with Agentforce in Security Center will be available via its AgentExchange and Slack Marketplace this quarter. Due in the spring: Autonomous threat detection and triaging; Agentforce security capabilities in Slack; and custom data compliance policies.

3. Dataiku

Dataiku on Oct. 7 launched Agent Hub, a workspace where users can create, discover, and use AI agents with the kind of enterprise-level oversight that business and IT leaders want and need. Dataiku wants to minimize or eliminate whack-a-mole AI agents popping up all over the place due to experimentation with so many tools and platforms on the market.

Over the past 12 years, Dataiku has earned its bona fides as a data & AI platform provider. I did some work with CEO Florian Douetteau back when the company’s motto was “everyday AI,” which was prescient in the pre-ChatGPT days. Now Dataiku bills itself as “The Universal AI Platform.” According to Reuters, Dataiku is prepping for an IPO.

4. Teradata

Teradata joined the agentic mainstream on Oct. 7 with the introduction of Autonomous Customer Intelligence, a software/services offering that uses embedded agents to automate steps in the customer experience (CX) process. Teradata has added its first agentic offering, for Customer Lifetime Value, to an expanding AI portfolio that includes its Enterprise Vector Store, Agent Builder, and MCP server.

Notably, Teradata is using embedded agents to automate data-management tasks like detecting signals and activating responses. It’s the latest example of a major trend in data management, as agents increasingly automate everything from data discovery to quality control to cataloging and more.

5. OpenAI

OK, everyone move over—OpenAI is now in the game in a bigger way. On Oct. 6, the model maker announced the aptly named AgentKit, comprised of an Agent Builder, Connector Registry, and ChatKit for embedding chat-based agent experiences in apps and elsewhere. OpenAI describes the new platform as a complete set of tools to build, deploy, and optimize agents. It’s the company’s second wave of agent tech; earlier this year, it introduced an agent SDK and API.

There’s a lot under the hood, including drag-and-drop design and ways to evaluate and tune agent performance. Importantly, OpenAI has included open-source guardrails that do things like protect PII and detect vulnerabilities known as jailbreaks.

6. IBM

As businesses deploy more agents—both internally and customer-facing—including multi-vendor agents, orchestration becomes critical. IBM is addressing this requirement with Watsonx Orchestrate, introduced Oct. 7. Capabilities include Agentic Workflows to create flows for sequencing agent actions and integration with Langflow, an open source, low-code agent builder. Importantly, IBM includes metrics tracking in Orchestrate for observability. And there are new prebuilt agents for finance, supply chain, and customer service.

7. Oracle

Oracle on Oct. 6 unveiled more than a dozen prebuilt agents for sales, service, and marketing. They’re available within Oracle’s Fusion CX cloud apps, where they automate processes and analyze data. Examples include a Deal Advisor Agent, Quote Assistant, Self-Service Agent, and my favorite, a Triage Agent that prioritizes help requests based on severity and sentiment 😡!

This latest news comes on the heels of Oracle’s OCI AI Agent platform, announced in July.

8. SAP

SAP is expanding the agents that are part of its AI assistant and agentic architecture, known as Joule and which automates functions within SAP Business Suite. On Oct. 6, SAP introduced a Cash Management Agent and an International Trade Classification Agent. Earlier this year, SAP announced agents for CX, Finance, Supply Chain, Spend Management, and People Management. SAP describes its strategy as follows:

“Drawing on the applications and data from across SAP Business Suite, SAP is introducing a new generation of role-aware assistants in Joule. Each assistant is designed to partner with a human being in their specific business role. Assistants in Joule tap into the right agents for the job, configuring, orchestrating and managing them so humans can focus on unlocking new levels of insight and productivity.”

For a deeper dive on SAP’s agent strategy, I recommend this article by Ken Yeung.

9. Microsoft

Fragmented tools, complexity, trust gap. These are some of the words Microsoft used to describe the need for Microsoft Agent Framework, introduced Oct. 1. Microsoft has been building up this, including its work with ChatGPT, Copilot, and AI agents for Dynamics ERP/CRM. The framework, in preview, is an open SDK for orchestration of multi-agent systems. With this, Microsoft is introducing multi-agent workflows in its Foundry Agent Service and multi-agent observability. All of which are under the umbrella of Azure AI Foundry, a.k.a. an agent factory.

In short, Microsoft now provides more comprehensive support for enterprise-wide, multi-agent environments: orchestration, observability, debugging, optimization.

10. Cisco

As more organizations deploy AI agents to automate processes, the next step for many will be to let digital agents and employees work in unison. Cisco is tackling this with what it calls Connected Intelligence, “where blended teams of humans and agents connect and collaborate.” To do it, Cisco on Sept. 30 introduced an AI Assistant and RoomOS 26 operating system for Cisco’s video devices, and integration with Webex Suite.

Per the press release, Cisco’s AI Assistant, due for GA this quarter, will offer a range of new features, including a Task Agent, Notetaker Agent, and AI receptionist for Webex calls.

Final thought: urgency vs. inaction

As these developments show, there’s a lot happening in the world of agentic AI. Many business and tech leaders are assessing their next steps.

McKinsey notes that some CEOs remain unconvinced that AI agents will have significant impact, at least near term. The consulting firm advises them to learn, think boldly, and act with urgency.

Count me among those who see the potential value and business advantages in a well conceived strategy for AI agents. Success will depend on getting the infrastructure right and moving the entire organization up the learning curve.

In case you missed it

• “Larry Ellison Hints at What’s Next: Oracle AI Database”

• “Databricks Targets AI Database Market with $1 Billion in New Funding”

• “PostgreSQL Charges Ahead with AWS, Databricks, IBM, Microsoft, Snowflake”

Share

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, who also worked in strategic comms at Oracle, IBM, and MongoDB. Connect with me on LinkedIn.

Subscribe now

The massive, $300 billion-plus build out of hyperscale data centers has become more complicated by the physical limitations of where to put them. Data centers need proximity to power, water, and internet access points. And distance is the enemy of low-latency connectivity.

If only it were possible to give data centers more elbow room—a few kilometers even—that would bring much-needed flexibility to site selection. And it would ease one of the gating factors to AI infrastructure.

Jason Eichenholz, the founder and CEO of Relativity Networks, refers to this ability to extend beyond the standard confines of data center location as “geographic optionality.” His solution: hollow core fiber (HCF), a next-gen cable technology that replaces the pure glass interior of traditional optical fiber with air. The physics are such that light waves travel faster and with less signal loss through air than through glass, stretching the limits of data center engineering.

Relativity Networks, a two-year-old startup, is among a handful of vendors developing hollow core as an alternative to fiber cable made with glass. Relativity is establishing partnerships and aims to be in production with HCF by early next year.

Microsoft got into the business in 2022, when it acquired another HCF startup, Luminosity. Microsoft recently provided a detailed case study on its use of HCF between Azure data centers.

HCF is an emerging technology. Microsoft, for example, has said it plans to deploy 15,000 km of HCF over two years. As a point of comparison, there are 5 billion km of optical fiber deployed worldwide—333,000 times more than what Microsoft has planned—according to Perplexity.

HCF, which is more expensive than traditional fiber cable, isn’t being pitched as a one-to-one replacement in all situations. Primary use cases for HCF include data centers, metro networks, and 5G/6G.

The added flexibility in data center site location is what’s sparking the most conversations, says Eichenholz. “It’s about planning where they’re going to put their future campuses,” he says. “It’s easier to move the gigabit than to move the gigawatt.”

Double the buildable space

What does geographic optionality mean in practice?

Say you’re trying to locate a new data center in the general area of a power plant, substation, or transmission line. HCF increases speed and therefore the linear distance you can go in the same time by nearly 50%.

So instead of being forced to build within, say, 10 kilometers of the power source or internet connectivity, you would have approximately 15 kilometers to work with, resulting in more than double the buildable space or area. That’s a meaningful difference, especially in dense urban areas where many electric utilities operate.

Each deployment scenario is different. “The data center might be next door, but it could also be 20 miles away with a line that connects them to a large substation or transmission line,” explains Eichenholz, who has a PhD in Optical Sciences and Engineering from the University of Central Florida. (Relativity Networks has an IP commercialization agreement with the University of Central Florida.)

Thus, site location becomes both a strategic decision and a tactical one.

“Data centers are designed to connect into the grid at high-capacity points, where hundreds of megawatts can be reliably delivered,” adds Eichenholz. “The key is that you need robust power and reduced latency in the same area, which in many areas is now too congested to add more data centers due to power constraints.”

‘Capacity constrainted’

Speed is the other major factor—both the need to get new AI data centers up and running, which can take years, and to transmit data in split seconds.

In addition to the challenges of physical connectivity and cooling, data center planners must contend with permitting, zoning laws, etc. One city, St. Charles, Missouri, recently imposed a one-year moratorium on new data center construction. According to Lawrence Berkeley National Lab, wait times for interconnecting power plants to the grid have been gradually increasing, with a median wait of five years for projects built in 2023.

The Trump administration has taken steps to hasten projects with an AI action plan and an executive order that lowers the barriers for developing data centers on federal land. Trump made note of this during his recent meeting with tech CEOs, as reported by the Wall Street Journal:

Trump opened the dinner by remarking on a significant obstacle for tech companies: procuring enough electricity to power massive data centers needed to develop AI. His administration has been working to remove roadblocks to connecting data centers to the grid, though many challenges exist at the state level.

Meantime, “capacity constrained” remain the watchwords.

“Even as we continue bringing more data center capacity online, we currently expect to remain capacity constrained through the first half of our fiscal year,” Microsoft CFO Amy Hood said on the company’s recent earnings all for Q4 FY2025.

Microsoft’s investment in reopening the boarded-up Three Mile Island nuclear plant shows the drastic steps hyperscalers are taking to bring new AI data centers online.

Killer app

HCF transmits data at nearly the speed of light and does so, as I explained above, at greater distances than conventional optical fiber. Milliseconds mean money. It’s why high-frequency traders locate their systems next to stock exchanges.

Similarly, AI performance is highly dependent on low latency. Gen AI chatbots must be able to return a response within seconds. My expectation is that agentic AI will be even more time sensitive because agents must be able to reason and respond in near real time. Eichenholz refers to low latency as a killer app for hollow core.

Echoing that point, Microsoft says hollow core fiber is accelerating AI.

But we’re not there yet. When I asked Google Gemini to create an illustration for this blog post, it didn’t have the resources to do so. “I’m being asked for a lot of images right now, so I can’t create that for you,” Gemini responded, “Can you ask again later?”

Later can be a lifetime in data center operations. In this capacity-constrained world, air-filled fiber may be an emerging solution.

Additional reading

“Hollow Core Fiber: What It Is and Why Does It Matter?” Data Center Dynamics, April 2025

Cloud Database Report

AI Data Centers: 3x Demand, Billion-Dollar Projects, New Energy Sources

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist who has also worked in strategic comms at Oracle, IBM, and MongoDB. Now I’m an independent tech writer. Connect with me on LinkedIn. Cloud Database Report is reader supported. To receive new posts and support my work, consider becoming a free or paid subscriber…

Read more

a year ago · 2 likes · John Foley

Cloud Database Report

Google Commits to 'Climate-Conscious' Data Centers

We now know that Google consumed 4.3 billion gallons of water to cool its worldwide data centers in 2021. That’s a lot of water—and that number will likely go up as Google’s global data center footprint continues to expand…

Read more

4 years ago · 1 like · John Foley

Share

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, who also worked in strategic comms at Oracle, IBM, and MongoDB. Connect with me on LinkedIn.

Subscribe now

Oracle’s breakthrough results for Q1 FY26 surprised even seasoned analysts. “I am sort of blown away,” said one on the earnings call. “It’s amazing,” said another.

Here’s why: Cloud revenue (↑27%), cloud infrastructure (↑54%), remaining performance obligations/RPO (↑359%).

An impressive showing for Oracle Cloud Infrastructure (OCI) and the company’s other AI and data center offerings, as evidenced by four multi-billion-dollar deals with three customers in the quarter.

Oracle database shared in the highlights. “Cloud database services, which were up 32%, now have annualized revenues of nearly $2.8 billion,” CEO Safra Catz said on the earnings call. “Autonomous Database revenue was up 43% on top of the 26% growth reported in Q1 last year. Multi-cloud database revenue, where OCI regions are embedded in AWS, Azure and GCP, grew 1,529% in Q1.”

Not bad for a 46-year-old database management system (DBMS). And it cued up Larry Ellison to pre-announce what’s next: Oracle AI Database.

“Next month at Oracle AI World, we will introduce a new Cloud Infrastructure service called the ‘Oracle AI Database’ that enables our customers to use the Large Language Model of their choice—including Google’s Gemini, OpenAI’s ChatGPT, xAI’s Grok, etc.—directly on top of the Oracle Database to easily access and analyze all their existing database data,” Ellison is quoted as saying in Oracle’s official earnings announcement.

“This revolutionary new cloud service enables the tens of thousands of our database customers to instantly unlock the value in their data by making it easily accessible to the most advanced AI reasoning models,” Ellison added. “Oracle AI Cloud Infrastructure and the Oracle MultiCloud AI Database will both contribute to dramatically increasing cloud demand and consumption over the next several years. AI Changes Everything.”

Oracle database has long been the workhorse—or thoroughbred—of many enterprise data systems. Over the past three years, Oracle has given its DBMS a second life outside of its own ecosystem through agreements with AWS, Google Cloud, and Microsoft. (For background, see my post, “Larry Ellison Gets Serious About Multi-Cloud.”)

Oracle database is now live in 34 Azure, GCP, and AWS data centers, with another 37 planned, Catz said.

[A transcript of Oracle’s Q1 FY26 earnings call is here, per Seeking Alpha.]

‘Ask any question you can think of’

Last year, Oracle added the “ai” suffix to its database naming convention (the current release is Oracle Database 23ai). And now it has renamed its annual event AI World, which will take place Oct. 13 - 16 in Las Vegas, where it will launch the new Oracle AI Database.

According to Ellison, the next step in AI—inferencing, where LLMs are put to use making decisions or taking actions—is an even bigger opportunity for Oracle.

“Training AI models is a gigantic multi-trillion-dollar market,” he said on the earnings call. “It’s hard to conceive of a technology market as large as that one. But if you look close, you can find one that’s even larger, and it’s the market for AI inferencing.”

Ellison expects Oracle database to be central to the AI inferencing tech stack. Because, while chatbots have been a radical new way for consumers to access information, it’s the combination of LLMs with private data that represents the next competitive differentiator in enterprise data and apps.

“We think we are in a pretty good position to be a winner in the inferencing market because Oracle is by far the world’s largest custodian of high-value private enterprise data,” Ellison said. “With the introduction of our new AI database, we added a very important new way for you to store your data in our database. You can vectorize it. And by vectorizing it, by vectorizing all your data, all your data can be understood by AI models.”

He added: “Then we made it very easy for our customers to directly connect all their databases, all their new Oracle AI databases and cloud storage, OCI Cloud storage, to the world’s most advanced AI reasoning models—ChatGPT, Gemini, Grok, Llama—all of which are uniquely available in the Oracle Cloud. After you vectorize your data and link it to an LLM, the LLM of your choice, you can then ask any question you can think of.”

Using LLMs on top of Oracle database

Vector search was a much-touted feature of Oracle Database 23ai when it was released in the first half of last year.

We’ll have to see what’s new and different in the forthcoming Oracle AI Database service. Ellison seems to be hinting at tighter and more seamless integration with LLMs. I haven’t been briefed on Oracle AI Database, the details remain to be seen.

It’s worth noting that Oracle’s HeatWave database (a version of MySQL) has supported in-database LLMs since mid-2024.

Final thought: Oracle has been adding AI-enabling technologies to its database for a while now. The roundup below provides links to six articles I’ve written, all published on Oracle.com, on some of the major innovations.

Cloud Database Report

The Ultimate Guide to Oracle's Cloud & AI Database

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, including 18 years at InformationWeek, who then worked in strategic comms at Oracle, IBM, and MongoDB. I invite you to subscribe, share, comment, and connect with me on LinkedIn…

Read more

2 years ago · 1 like · John Foley

Additional reading

Cloud Database Report

Larry Ellison on Oracle Database 23ai: 6 Key Takeaways

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist who worked in strategic comms at Oracle, IBM, and MongoDB. This blog and free newsletter are independent, unsponsored, and not affiliated with my current role as VP with…

Read more

2 years ago · 2 likes · John Foley

Share

Subscribe now

CEO Dev Ittycheria had just given his overview on “another strong quarter” when a financial analyst on MongoDB’s Q2 FY2026 earnings call asked about two new sources of competition.

One, Databricks’ Lakebase database. And two, DocumentDB, an open source database that has gained the backing of AWS, Google Cloud, Microsoft, Snowflake, and a half-dozen other data & AI players.

“How might they impact MongoDB?” the analyst asked. “And how do you differentiate?”

Databricks ‘threw in the towel’

The timing and juxtaposition of these three points of reference are notable. In the span of seven days, Databricks disclosed (on Aug. 19) plans to invest in the development of Lakebase, the Linux Foundation revealed (on Aug. 25) that it is taking on DocumentDB as an open source project, and MongoDB announced (on Aug. 26) a 24% year-over-year increase in Q2 revenue, to $591 million. Mongo’s stock price spiked on the earnings results.

TLDR: MongoDB is under increasing pressure from industry heavyweights even as its database business is growing robustly.

Ittycheria was ready for the hard questions. Databricks, he replied, had started down the path of developing its own online transaction processing (OLTP) database only to discover it was a non-trivial undertaking. Instead, Databricks “threw in the towel” and acquired Neon and its OLTP database. “It just reinforces that OLTP is the strategic high ground for AI,” he added.

What about Postgres?

On the question of industry support for the open source DocumentDB database—now, a more viable alternative to MongoDB—Ittycheria responded that “the hyperscalers are investing less and really handing off to the open-source community to kind of really take on the bulk of the work in terms of product development.”

In other words, according to that POV, three of the top four database companies (Oracle being the exception) are actually making a lesser commitment to DocumentDB by leaving the heavy lifting to others.

Here I should point out that the Linux Foundation-backed DocumentDB is a PostgreSQL database with a MongoDB API. Depending on your point of view, that’s either a best of both worlds design (Postgres + MongoDB compatibility) or a Frankenstein approach. [Oracle has its own MongoDB API, which I’ve written about previously.]

MongoDB is a NoSQL document database, and Postgres is a SQL relational database. They are fundamentally different models, each with its own advantages.

“Why do we hear so much about Postgres adoption for AI startups?” the financial analyst asked next.

“Good question,” said Ittycheria. “What’s become clear is a lot of these startup founders don’t think that hard about their database choice. They kinda go with what they know. And what we are seeing is that as some of these startups are scaling, they’re running into real scaling challenges with Postgres.”

That explanation, I’m sorry to say, sounds a bit like blaming the startup teams that MongoDB hopes eventually to convert into customers. And it raises the question of which platform is actually more scalable—MongoDB or Postgres?

The answer is it depends, as you can see from the AI response below.

The scalability question

I asked Anthropic’s Claude: “How do Postgres and MongoDB compare in terms of enterprise scalability?” Here’s what Claude said:

- begin AI response -

Real-World Enterprise Usage

• PostgreSQL: Used by companies like Apple, Instagram, and Spotify for core transactional systems

• MongoDB: Powers applications at Facebook, Adobe, and eBay for content management and real-time analytics

Bottom Line PostgreSQL offers more predictable, mature enterprise scalability with strong consistency guarantees. MongoDB provides more flexible, automatic horizontal scaling but with eventual consistency trade-offs. The choice often depends on your specific data model, consistency requirements, and growth patterns rather than pure scalability metrics.

- end AI response -

I don’t consider Claude to be the ultimate authority on database scalability, but as a gut-check it undermines Ittycheria’s contention that Postgres pales in comparison to MongoDB.

To be continued…

The competition between MongoDB and Databricks, and between MongoDB and DocumentDB, isn’t a zero-sum game. It’s possible, even probable, that all three platforms will keep growing. It’s just that some growth will come at the expense of other growth.

And it’s fair to say that MongoDB and the hyperscalers that are backing DocumentDB aren’t necessarily archenemies. MongoDB is available on AWS, Google Cloud, and Azure. “Our hyperscaler partnerships remain strong,” said Ittycheria.

Even so, the competitive landscape feels different today that it did a few weeks ago.

See the full transcript of MongoDB’s Q2 FY2026 earnings call here, courtesy of Motley Fool.

Further reading

“AWS joins the DocumentDB project to build interoperable, open source document database technology,” - AWS

“AWS, Microsoft and Google unite behind Linux Foundation DocumentDB database to cut enterprise costs and limit vendor lock-in,” - TechCrunch

Cloud Database Report

Databricks Targets AI Database Market with $1B in New Funding

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, including 18 years at InformationWeek, who also worked in strategic comms at Oracle, IBM, and MongoDB. Connect with me on LinkedIn…

Read more

a year ago · 2 likes · John Foley

Share

Subscribe now

I’ve been on the road for the past two weeks, first in Salt Lake City on business, then to Dublin for a visit. While in Ireland, I took the train to Limerick to see The Saw Doctors, a favorite band that has been playing its unique style of Irish folk-rock since 1986. Remember the year because I will to come back to it.

[As a break from the routine, I’ve included below a music video of The Saw Doctors that was released just last year. It’s a good soundtrack that was filmed in the Irish countryside. I hope it brings a smile.]

The same day I landed back in New York, Databricks announced it had raised another $1 billion, an investment that will go toward its recently unveiled database system, Lakebase, and its new agent-building platform, Agent Bricks.

This could be a game-changing development in the database market. Databricks, now valued at more than $100 billion, is throwing its considerable weight, resources, and influence into a market that is ripe for disruption. The database stalwarts—AWS, Google Cloud, Microsoft, Oracle—are surely paying attention.

What are the implications? There’s an old Irish saying that hindsight is the best insight to foresight.

Hindsight:

• Databricks has long been a player in data management, with a core competency in analytics, data intelligence, and data science. I included Databricks on the Cloud Database Report’s Top 20 when it was published in 2022. So they already have a foothold.

• Databricks has been a leading proponent of the data lakehouse architecture, which is an evolution and blending of data warehouses and data lakes. The appeal of lakehouses includes the simplicity of a unified data platform, efficiencies in storage and data pipeline processes, support for both raw data and transformed/formatted data, and serving as a data repository both analytics and AI.

Insight:

• In June, Databricks CEO Ali Ghodsi revealed on LinkedIn that—despite the company’s marketing mantra of “data intelligence”—he really thinks of Databricks as an AI database. That acknowledgement, which I wrote about in a blog post, now carries greater significance given the company’s plans to invest in Lakebase, which it describes as “a new type of operational database (OLTP), built on open source Postgres, and optimized for AI Agents.”

• Note the reference to Postgres serving as the foundation for a new type of operational database. Postgres (like The Saw Doctors) has been around since 1986, so it’s a remake of a legacy database into a modern AI database.

• What’s most significant is that Lakebase (in public preview) can support both analytics—Databricks’ forte—and operational data. That’s important because AI agents need to act and respond in real-time. (Lakebase isn’t a vector database. Rather, vector data types are supported via an integrated vector search service.)

Foresight:

• Ghodsi pegs the database industry’s total addressable market (TAM) at $105 billion. That’s in line with the $100 billion I reported last year. Which puts the database market at roughly the same size as the analytics market and half the size of the data storage market, according to Perplexity.

• Ghodsi recognizes that the database market is in the throes of change as new capabilities are required to support AI data and AI agents. Taking that a step further, AI agents are autonomously creating databases. “There’s a new user. The user is not human. It’s an AI agent, and if we just double down on making that user persona successful, that’s the wedge to disrupt that TAM,” Ghodsi told Julie Bort at TechCrunch.

Optimized for AI agents

Databricks—which was not among the Top 10 in database market share in 2024, according to Gartner—has just fired a shot across the bow of the industry leaders. Ghodsi said to TechCrunch: “The database market is $105 billion of TAM, of revenue, sitting there, kind of unaffected in the last 40 years.”

Translation: Ghodsi sees an opportunity to drive innovation—and, in doing so, get a bigger piece of the TAM pie that in his words is “sitting there” for the taking.

The other piece of this is Agent Bricks, a new platform-as-a-service (in beta) for building enterprise AI agent systems. Initial use cases are an Information Extraction Agent, Knowledge Assistant Agent, Multi-Agent Supervisor, and Custom LLM Agent, with more planned.

I wasn’t able to find a detailed description of how Lakebase and Agent Bricks will work together, but given the data dependencies of agentic AI, my expectation is there will be integration and synergy between them.

Here’s how Ghodsi explained Lakehouse and Agent Bricks not long ago when they were introduced.

More recently, Databricks has disclosed plans to acquire Tecton for low-latency data engineering. According to Reuters, Ghodsi explained it as a real-time building block for AI agents.

RIP database categories

It’s clear that the way CIOs and CTOs are looking to transform business data into business value is going through a massive rethink, sparked first by GenAI and now advancing to agentic AI. With that comes the idea that a new tech stack is needed, which is the opportunity that Databricks and others are after, i.e. Snowflake + Crunchy Data and Salesforce + Informatica.

Relational, NoSQL, and other tried-and-trusted DBMSs will continue to have their place, but disruption is afoot. MongoDB says, “Database Categories Are Dead.”

The categories may be dead or dying, but data management is more important than ever in this new age of AI.

Further reading

Cloud Database Report

CEO Ali Ghodsi: Databricks Is an 'AI Database'

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist who has also worked in strategic comms at Oracle, IBM, and MongoDB. Now I’m an independent tech writer. Connect with …

Read more

a year ago · 1 like · John Foley

An end-of-summer Irish tune

Subscribe now

Turning data into insights has always been the objective of business intelligence. But that’s never as easy as it sounds.

The challenge is that BI programs and infrastructure can cost millions and may require a cadre of data engineers and scientists for good results. Even then, data-driven intelligence may not be at the fingertips of every employee who might want it, need it, or use it.

AI agents promise to change that. They have the potential to automate much of the upfront work—integrating, synthesizing, cataloging, and segmenting data—while also delivering data-driven intelligence to more people through apps and other points of engagement. Each week, I’m seeing vendors introduce new capabilities for agent-enabled analytics and BI.

With all of this top of mind, I reached out to Salesforce’s Chief Data Officer, Michael Andrew, who is knee-deep in these emerging technologies, including both the challenges and opportunities. Andrew, who is responsible for providing analytics to Salesforce’s 75,000+ employees, is well familiar with the roll-up-your-sleeves effort that goes into data transformation.

“It takes effort and work to make sure you have the data connected to be able to build the processes out, but we think that’s the promise of the platform—that you truly can get that holistic understanding of all of your customers,” says Andrew. “And then you can take the right action for that customer.”

The primary components of the Salesforce Platform that enable this kind of comprehensive view are: Agentforce, for building and managing AI agents; Tableau Next, the company’s all new agentic analytics solution; and Data Cloud, a data lakehouse where petabytes of data are integrated, aggregated, and activated.

Having covered business intelligence for many years—going back to Walmart’s enterprise data warehouse in the mid-1990’s—I started our conversation by asking Andrew how the market is changing. He pointed to three trends.

1. Merging of operational data and analytical data. “These worlds are converging,” Andrew says. This isn’t a new phenomenon but it’s catching on as more organizations seek to respond in near real-time to digital activity. Databricks’ recent acquisition of Neon and Snowflake’s pending buy of Crunchy Data allow for this blending of capabilities. Databricks and Snowflake are both accessible via Salesforce’s Data Cloud.

2. Farther-reaching distribution of analytics with AI agents. In many places, there has been more demand for analytics than there are analysts or data scientists to meet it. Now, with AI agents tapping into Data Cloud, every person can have a data analyst as a digital sidekick.

3. Rise of the agentic enterprise. More businesses are catching on to the possibilities of deploying autonomous AI agents that work side by side with humans in areas such as sales and support. “You can start thinking about that in every function,” says Andrew.

The future of analytics?

AI agents that function as data analysts, or with some of the skills of data analysts, are surfacing in various ways in the Salesforce ecosystem.

Tableau Next—an AI agent-enabled upgrade to the popular data-visualization tool that Salesforce acquired in 2019—now provides personalized agentic analytics. Tableau Next has been integrated with Data Cloud, and it works with Salesforce’s Customer 360 apps (Sales Cloud, Service Cloud, Marketing Cloud, etc.). It also ties into Slack, so users can request, create, and share metrics and dashboards in Slack channels, canvases, and direct messages.

Tableau Next is “a re-platforming” of Tableau, says Andrew. “We see this as the future of agentic analytics.”

What makes this possible is a semantic layer in Tableau Next that gives agents access to unified business data. With appropriate permissions, AI agents can access data in Data Cloud, then do what Tableau has always done—generate dashboards and data visualizations. Because this is within in Slack and other apps where collaborative work happens, it brings near-real-time analytics into workplace conversations.

Tableau Next agents are one way to surface data and BI in more apps and devices, but not the only way. Analytical capabilities are available in some of the prebuilt Agentforce agents, and these skills can also be added through customization or by partner add-ons in AgentExchange.

Democratizing BI

As analytics appear on the screens of more and more users, it gets us closer to the democratization of business intelligence, which has been a long time coming. “A few years from now every person in a company could have a data analyst or data scientist in their pocket, so to speak,” says Andrew. “That’s pretty incredible.”

As more employees get fast and easy access to actionable data, that leads to a more analytics-driven enterprise, or what Andrew metaphorically calls omniscience—the state of knowing everything there is to know about a customer relationship.

Andrew explains: “The agent has omniscience of everything that has been done with the customer—every meeting, all the content, all the Slack channels. Imagine a place that knows everything that has ever happened with a customer, all the business context, how the business interacts.”

That’s the vision. The reality has typically been more prescribed than that, with BI reports available to execs, managers, and the rest of the workforce on a need-to-know basis. Tools like Tableau, Power BI, Looker, Domo, and Zoho have moved the needle on ease-of-use with connectors and no-code transformation and dashboards. Autonomous AI agents that proactively populate everyday communications channels with timely intelligence promise to take that to the next level.

Notice I haven’t said that agents deliver insights. Websters defines insight as “the act of apprehending the inner nature of things or of seeing intuitively.” I think of that as a distinctly human function, although I may be forced to reassess as AI gets ever closer to human intelligence.

How we consume and act on BI is changing, too. For years, citizen data scientists have been the vanguard of do-it-yourself analytics. But they are relatively small in number compared to what has been unleashed by LLMs. Now, most everyone uses GenAI to probe for information. We’re all becoming prompt experts.

Get the foundation right

Agentic analytics may be a logical next step for organizations looking to infuse data-driven decision making into more parts of the business. As mentioned earlier, that requires getting data from transactional/operational systems to analytics systems, post haste.

Salesforce supports this through the combination of Data Cloud and Flow, its automation tool that lets customers define “signals” that trigger actions such as sending an email or notification. Flows can also initiate a user interaction or an approval process.

This is especially valuable for customer engagement and the sales cycle. It’s a way for managers of a hybrid workforce, comprised of humans and AI agents, to divvy up tasks appropriately. For example, a hot lead might go straight to a sales professional, whereas a general inquiry gets handled by an AI agent. “Now the agent’s having a conversation with the customer because maybe they were early in the buying process,” explains Andrew.

Data as a product

But wait, let’s tap the brakes on this idea of making more data available to more people. What about data security and access control?

These projects need governance and guardrails. Salesforce is addressing this in its own analytics program by implementing a data mesh approach, where data is managed as a product. That entails knowing the sources and lineage of data, where and how it’s stored, data-quality steps, and attention to access privileges for both humans and AI agents.

“You really have got to get the foundations right,” says Andrew, “because agents are going to expose any data problems.”

The good news is that, based on what I’m seeing in the industry, a growing number of business and IT leaders recognize the need to address the issues of data readiness that stand in the way of enterprise AI. There’s still a lot of work to be done, but awareness is the first step.

Anyone who believes that business intelligence provides competitive advantage—and that would include the vast majority of business leaders—might be predisposed to see the potential of having in an analyst in every pocket, a.k.a. agentic analytics.

“Analytically-driven organizations outperform their peers. They make faster and smarter decisions,” says Andrew. “Agents turbocharge that—the same assets can be super-accelerated and everybody can have analytics at their fingertips.”

That’s the opportunity, but we’re at the beginning of the adoption curve. I will be watching for organizations that move ahead at the speed that AI agents are advancing.

Further reading

Cloud Database Report

How Salesforce Data Cloud Fuels Fast, Accurate, Trusted AI Agents

Welcome to the Cloud Database Report. I’m John Foley, a long-time tech journalist, including 18 years at InformationWeek, who also worked in strategic comms at Oracle, IBM, and MongoDB. Connect with …

Read more

a year ago · 1 like · John Foley