Xenoss

Merck committed $1 billion to embed Google engineers inside its teams. Enterprise AI is no longer a software purchase. In the latest CEO Brief, Dmitry breaks down what that means for any organization signing a multi-year AI partnership, and which capabilities should never leave the building. https://lnkd.in/dJb4th96

Amazon, Microsoft, Google, Meta, and Oracle are spending over $600 billion on data center infrastructure this year. That's a 36% jump from 2025. They can't build fast enough; the market is supply-constrained. As of late 2025, there were 1,297 operational hyperscale data centers worldwide. Nearly triple the number from 2018. Another 770 are in various stages of planning or construction. This is the infrastructure behind every AI model you've used, every cloud app your company runs, every real-time analytics dashboard that processes millions of events per second. Hyperscale isn't just big. It's built to grow without redesign. You add more compute. The system keeps running. No architectural overhaul required. For companies building AI and data-intensive products, this is the foundation. Swipe through for a quick breakdown of what hyperscale means.

You can’t govern what you can’t trace. That becomes a serious problem when enterprise data moves through: - source systems - ETL pipelines - feature stores - model training - inference - AI-generated outputs If a model starts drifting, a dashboard shows the wrong number, or a regulator asks which data influenced a decision, you need more than logs scattered across tools. You need 𝗹𝗶𝗻𝗲𝗮𝗴𝗲. In our new article, we explain how 𝗱𝗮𝘁𝗮 𝗹𝗶𝗻𝗲𝗮𝗴𝗲 is changing in the AI era: - what lineage tracks - why table-level tracking is no longer enough - how model-level lineage helps debug drift - where off-the-shelf tools stop - when custom engineering is needed for proprietary ETL, IoT, SCADA, and ML pipelines Read the full article:

Xenoss CEO shared a useful take on the new Google and Blackstone AI cloud venture. The interesting part is not only that Google wants to expand access to TPUs. It is what this says about the next stage of AI infrastructure. Worth watching closely, especially for companies planning large AI workloads over the next few years.

We've been recognized as a Top AI Automation Company for 2026 by Techreviewer.co. Over the past year, we've built multi-agent systems that cut accounting costs by 55%, deployed predictive maintenance platforms for oil and gas operators, and helped enterprises run production systems at scale. We build within existing infrastructure. Fast, so that clients start seeing value in weeks. Now in high demand is the question of how fast AI systems can be operational and who's going to maintain them after launch. That's the work we're built for.

CRM, ERP, marketing platforms, payment processors, internal databases. Getting all of that into one place where you can analyze it is harder than it sounds. That's what ETL pipelines do. Extract data from sources. Transform it into a usable format. Load it into a warehouse or lake where teams can run queries and train models. The challenge is deciding how to build it. Pre-built tools like Fivetran and dbt cover a lot of ground. 500+ connectors to common data sources. SQL-based transformations. Minimal engineering overhead. But every enterprise has systems that no connector supports. Legacy databases with custom schemas. Proprietary APIs. Real-time requirements that batch processing can't handle. Compliance logic that needs to live inside the pipeline itself. The dbt-Fivetran merger is consolidating the tooling landscape. Snowflake, Databricks, and Microsoft are all building full-stack platforms. The independent, modular approach is getting squeezed. Our latest article breaks down when platforms work, when custom pipelines make sense, and what changes for teams building AI workloads. Link in comments.

Our CRO, Maria Novikova, spent a day at BTEX Toronto and came back with notes worth sharing. Her observation about the shift from "can you ship it" to "can you run it" matches what we're seeing across client conversations.  Everyone can get a proof of concept working. The hard part is keeping it running in production, with governance, with security, without the whole thing drifting off course later on. Data foundation before the model and operability after. That's where enterprise AI lives. Full takeaways in her post below.

Our CRO, Maria Novikova, was at the Salesforce Agentforce World Tour in Toronto last week. Her takeaways on enterprise AI guardrails and cross-system orchestration are worth reading, especially if you're building agentic systems right now. We've been designing multi-agent architectures with the same principles she mentions: treating drift detection and hardcoded guardrails as preconditions, not afterthoughts. Full notes in her post below.

Real-time analytics still faces the same tension it did a decade ago: the business wants answers now, but it also expects those answers to be complete, correct, and reproducible. Lambda architecture was designed to solve this by running batch and stream processing in parallel, then merging both outputs in a serving layer. Batch gives you accuracy. Streaming gives you speed. The serving layer decides which answer to show. But maintaining 2 separate pipelines means duplicating logic, testing, and operational ownership. That pain triggered the push toward Kappa architecture, where a replayable log plus a streaming engine handles everything through one codebase. Since then, medallion architecture has emerged as another option, especially in lakehouse environments built on Databricks or Snowflake. The decision comes down to one question: are your batch and streaming workloads fundamentally different, or are they the same logic applied to different time windows? If different, Lambda. If the same, Kappa. If analytics-first without real-time requirements, medallion. Full breakdown with implementation patterns in our latest article. Link in comments.