Snowflake thinks context is king

Welcome to Runtime! Today: Snowflake hopes to make it easier for customers to share data and improve agent reliability, Microsoft jumps back in the AI model game, and the latest funding rounds in enterprise tech.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime each week.

Event horizon

SAN FRANCISCO – As everybody starts to freak out about the cost of AI, tools and services that help agents get things right the first time could start to become very important. Fresh off one of the strongest quarters in its short history, on Tuesday Snowflake outlined new ways that customers can ensure their agents will understand the full picture of their corporate data wherever it is stored.

"Being able to bring together trusted information about an enterprise with all of the context that matters to you, whether that's sitting in Gmail for enterprise or in Salesforce, creates this amazingly capable environment in which you can get many, many things done," Snowflake CEO Sridhar Ramaswamy said in a press conference Monday afternoon on the eve of Snowflake Summit 2026. Snowflake unveiled Horizon Context, a new feature in its Horizon Catalog that promises to connect the dots across business intelligence data so that agents can consistently find the right answers.

New AI agents have little or no understanding of what happened before they were created, and as they run they can generate data in slightly different ways compared to traditional applications with strict guidelines.
Horizon Context is an update to the company's Semantic Views schema, which sat between agents and databases or data lakes as an interface that helps agents understand the relationships between data produced by different sources, said Christian Kleinerman, executive vice president for product, during Tuesday's keynote.
"What really makes AI work, especially in the enterprise context, is … understanding your data, understanding context," he said.
For example, Horizon Context will allow employees to establish those relationships without having to write SQL queries and update those relationships as business conditions change without losing everything they've learned so far.

Snowflake also took another step in support of Iceberg, the open-source table format that the data industry has embraced over the last several years. New features from Iceberg v3 are now baked into Snowflake's core data tools, allowing users to both read and write data both inside Snowflake and in data lakes managed by third parties or the customers themselves.

In the not-so-distant past, if companies wanted to use data with different query engines or analysis tools they had to physically move that data or make a copy of it, which can be time-consuming and expensive depending on the amount of data in question.
When that data is written as an Iceberg table, it becomes much easier for third-party data tools that support Iceberg — which is an awful lot of them at this point — to work with data stored in a cheap cloud service like AWS's S3.
"We can make the data engineering teams be dramatically more effective, and that allows them to elevate and think about business outcomes that they're trying to drive, rather than pipelines they're trying to keep running," said Chris Child, vice president of product for data engineering at Snowflake, in an interview with Runtime.

Of course, Snowflake isn't the only company working on making sure AI agents have better context about business data (more on that in a bit). What we used to call prompt engineering is now increasingly called context engineering, which, as Anthropic described it last year, "refers to the set of strategies for curating and maintaining the optimal set of tokens (information) during LLM inference, including all the other information that may land there outside of the prompts."

The brief days of "tokenmaxxing" are clearly over; token consumption will need to be carefully managed as companies deploy agents, and those agents will run more efficiently with better context.
And a shift toward context also should pour even more cold water on the idea that the frontier model companies will take over enterprise software as we know it, given that data companies like Snowflake and Databricks as well as application companies like Salesforce and ServiceNow have much better access to that enterprise context than the model makers, Ramaswamy argued during Monday's press conference.
But he also cautioned that right now the entire enterprise tech industry is changing very quickly, and "every large change like the one that we are seeing with AI changes a lot of assumptions about how things work."

Disclosure: Snowflake paid for Runtime's travel and accommodations for Snowflake Summit.

Building blocks

Microsoft is a great example of how fast this industry is changing right now: three years ago it was the king of the AI mountain thanks to its exclusive relationship with OpenAI, which turned out to be a bit of a drag after Anthropic's models took the industry by storm last November. But on Tuesday it declared itself back in the frontier model game and also introduced several services for helping Azure customers run AI agents more effectively and efficiently.

Microsoft unveiled seven new AI models Tuesday at Microsoft Build that it compared favorably (of course) to similar models from Anthropic and Google across coding, image generation, and voice translation tasks. “We got here in six months, which is itself a remarkable achievement,” Microsoft AI chief Mustafa Suleyman told Semafor.

The company also addressed the context question by unifying four of its IQ services under the Microsoft IQ banner, which promises to deliver "the right information in forms agents can actually use, so they can reach accurate insight without drowning in noise or hallucinating answers," Microsoft said in a blog post. If enterprise customers have several top-tier models to choose as the back end for their AI agents, the competition to deliver the best experience around those models will become even more heated.

Enterprise funding

Cognition raised $1 billion in Series D funding, valuing the company behind the Devin AI coding agent at $26 billion.

DriveNets scored $410 million in Series D funding for its networking technology, which is based around the Ethernet standard and was designed for large-scale AI clusters.

XCENA landed $135 million in Series B funding for its "computational memory" chip design, which adds computing power to memory chips to improve performance for a key AI bottleneck.

Daloopa raised $47 million in Series C funding for its data infrastructure software, which was designed for financial services companies.

Gray Swan scored $40 million in Series A funding for its AI security technology, which looks to reduce hallucinations and unpredictable agent behavior.

Geordie AI landed $30 million in Series A funding for its own AI security technology, which provides better visibility into what those agents are actually doing.

The Runtime roundup

This week's software package supply chain attack hit Red Hat, after an "employee's GitHub account was compromised and used to push malicious orphan commits directly to several repositories, bypassing code review entirely," according to Aikido Security.

President Trump signed an AI executive order Tuesday that asks AI model companies to submit new models for a "voluntary review" 30 days before they are to be released, according to Politico.

Anthropic announced that it would extend access to Mythos Preview to 150 new companies and organizations in 15 countries, a significant expansion beyond the initial 50 groups involved in Project Glasswing.

The first day of token-based billing for GitHub Copilot went about as you would have expected.

Thanks for reading — see you Thursday!