DEV Community: Valentin Podkamennyi

Enterprise AI Infrastructure and Organizational Learning

Valentin Podkamennyi — Sat, 31 Oct 2026 20:34:42 +0000

This article explores the evolution of corporate artificial intelligence and how organizations are moving beyond basic model capabilities to focus on system performance. It highlights the transition from generic tools to specialized architectures that capture organizational experience as a permanent asset for long-term competitive advantage.

Transitioning from General Models to System Performance

The history of technology provides a clear parallel for the current state of corporate intelligence. Before the late nineties, search engines like AltaVista and Lycos relied on simple word counts and sold results to high bidders. Portals like Yahoo and AOL focused on capturing user attention to serve massive amounts of advertising. These early attempts failed to grasp the deeper connections within the digital landscape.

Larry Page and Sergey Brin changed everything when they introduced PageRank. They were doctoral students who understood the value of citation indices in academic literature. They applied this logic to the web, treating links as social constructs of importance. This move turned an underused signal into a foundational metric that organized the modern internet.

Corporate intelligence today stands at a similar crossroads. Many businesses currently treat these tools as individual agents providing answers to specific questions. However, the needs of a modern corporation are far more complex than simple information retrieval. Companies require persistent memory, state management, and integrated feedback loops that operate across the entire organization.

The current challenge lies in the infrastructure used to support these technologies. Most firms are building their AI environments using legacy tools and languages like Python. This often leads to security vulnerabilities such as prompt injection or weak isolation. These issues arise because the goals and guardrails are not natively part of the platform.

Instead of building individual applications in isolation, forward-thinking companies are looking for coherent platforms. These systems must provide inherent security and logging that cannot be manipulated. The goal is to move away from twentieth-century scaffolding toward a more unified architectural approach.

Performance Optimization Through Compound Systems

Research suggests that the focus is shifting away from simply making models larger. At institutions like Berkeley, experts now advocate for compound AI systems. These architectures coordinate multiple capabilities to achieve specific business goals more effectively than a single massive model ever could.

Small, specialized models are often outperforming their larger counterparts when they are part of an optimized pipeline. Projects like DSPy show that refined system design allows smaller models to compete with the most expensive proprietary options. This shift democratizes high-performance computing for businesses that cannot afford massive hardware investments.

Cost efficiency is another major driver for this architectural change. FrugalGPT has demonstrated that using model cascades can achieve results comparable to top-tier models at a fraction of the price. By intelligently routing tasks to the appropriate model, companies reduce their operational expenses by nearly 100 percent in some cases.

The true breakthrough comes from the ability to learn across every execution. When every interaction informs the next decision, a company builds a substantial lead over its competitors. World models like Dreamer allow systems to anticipate the consequences of actions before they are even taken. This capability transforms a static tool into a dynamic, predictive asset.

Reinforcement learning further enhances this process by using experience to improve decision-making over time. This optimization mechanism ensures that the system is constantly refining its approach based on real-world outcomes. It moves the focus from what a model can do in theory to how the system performs in practice.

Retaining Organizational Experience as a Corporate Asset

A significant advantage for any business is the specific experience it gains through daily operations. While generic models are available to everyone through APIs, the unique data generated by a company belongs to that company alone. This includes insights into customer retention, supplier relationships, and inventory management.

Accumulated learning must be processed correctly to provide value. Simply storing data is not enough; it must be integrated into the core intelligence layer of the business. This ensures that the training happens at a level that the organization fully controls and manages.

The ultimate test for a corporate AI strategy is the portability of its intelligence. If a company decides to change its model provider, it should not lose its operational history. Learned policies and demonstrated gains must remain with the business rather than being locked into a specific vendor’s platform.

This distinction defines the difference between renting intelligence and owning it. Companies that rely on standard, off-the-shelf solutions are essentially renting their capabilities. Those that build their own internal logic and retain their experience are creating a proprietary asset that grows in value over time.

Optimization is the key to becoming a truly modern enterprise. Businesses that stop waiting for the next big model release and start focusing on their internal learning processes will lead their industries. They understand that their most valuable asset is the intelligence they build and keep within their own walls.

The future of the industry depends on this shift in perspective. When a business controls the learning asset, it no longer requires ownership of every underlying model. This flexibility allows for rapid adaptation and ensures that the company remains at the forefront of technological innovation while maintaining its unique competitive edge.

GitHub Copilot Redesign Focuses on Multi-Agent Collaboration

Valentin Podkamennyi — Fri, 30 Oct 2026 19:16:14 +0000

GitHub is transforming its Copilot platform from a simple coding assistant into a comprehensive environment for multi-agent software development. The company launched a redesigned Copilot app and a new Projects feature at its annual GitHub Universe conference. These updates aim to coordinate AI agents across complex development workflows.

Centralizing Development with Agentic Workspaces

The newly redesigned Copilot app represents a shift in how developers interact with artificial intelligence. While previous versions focused on localized code changes and pull request creation, the update positions the app as a central command center. Developers now use this workspace to manage broad AI-driven tasks and connect agents with various development tools. This change allows teams to maintain a consistent flow without leaving the platform for peripheral tasks.

Beyond mere code generation, the app enables users to work on technical specifications and participate in design discussions. It integrates application telemetry, giving agents a deeper understanding of the software lifecycle. This integration means developers do not have to provide minute instructions for every line of code. Instead, they provide high-level goals based on a comprehensive set of project data and organizational context.

The transition to a unified workspace may significantly change daily routines for software engineers. Amit Kumar Jena, an AI development leader at Kanerika, notes that this overhaul minimizes the need to jump between different management tools and editors. By delegating broader objectives to cloud-based agents, developers can focus on higher-level logic. These agents execute work independently and then submit their results as pull requests for human verification.

Standardization is a primary benefit for organizations looking to scale their AI efforts. Pareekh Jain of Pareekh Consulting explains that a single platform makes it easier for leadership to establish governance. It allows for better oversight of agent activity across diverse teams. This central approach also addresses the issue of tool sprawl, where developers use too many disconnected AI utilities that complicate the security landscape.

Governance and Control Mechanisms

Granting agents the ability to modify repositories independently requires strict administrative controls. Technology leaders must enforce branch protection and ensure that no agent-generated code enters production without human approval. Without these safeguards, the speed of AI development could introduce unexpected vulnerabilities.

Monitoring these automated sessions is essential for maintaining accountability. Integration with security information and event management systems allows teams to track which agents accessed specific files. Audit logs provide a trail for every modification made by an AI entity. This level of visibility is necessary as enterprises move from experimental AI use to full-scale automated production.

Coordinating Parallel Development via Copilot Projects

The introduction of Copilot Projects moves the scope of AI assistance from individual files to entire software ecosystems. This feature allows developers to set goals that span multiple repositories at once. By pulling context from Slack conversations, meeting notes, and design mockups, the system creates a holistic view of the required work. Copilot then distributes these tasks among several agents running in parallel.

This capability is particularly effective for large-scale modernization efforts. For example, teams can use parallel agents to handle repetitive tasks like updating dependencies or migrating old software development kits. It also simplifies version upgrades for frameworks like Java or .NET. Automated agents can sweep through hundreds of files to apply these changes, significantly reducing the time required for maintenance.

However, running multiple agents simultaneously introduces new layers of complexity in terms of integration. Abhishek Satapathy, an analyst at Avasant, warns that agents might work with outdated information or make conflicting assumptions. If two agents modify the same file or rely on different API contracts, the resulting merge conflicts could be severe. Developers must be prepared to resolve these architectural discrepancies manually.

A study titled CooperBench suggests that AI agents often perform better individually than in collaborative settings. The research found that success rates dropped by roughly 30 percent when agents worked together on shared tasks. While this study was not specific to the new GitHub features, it highlights a fundamental hurdle in automated teamwork. Managing the output of ten parallel sessions requires a high degree of human orchestration.

Resource Management and Cost Efficiency

Every automated session consumes resources, making cost management a priority for engineering managers. Poorly defined tasks often lead to repeated execution and wasted AI credits. Organizations need to assign human owners to monitor the progress of these agents in real-time. This oversight ensures that runaway processes do not inflate operational budgets without delivering value.

Continuous integration pipelines also face increased pressure from the surge in pull requests. Every agent-generated change must undergo testing and validation, which uses significant computing power. Leaders should set strict credit limits for projects to prevent overspending. This balanced approach allows for innovation while keeping infrastructure costs within predictable boundaries.

Advanced Code Review and New Licensing Models

Security and quality control receive a boost through the new Copilot Max tier. This version utilizes multiple AI agents to perform deep reviews of pull requests. Unlike traditional models that use a single perspective, Max employs several independent agents powered by different AI providers. An aggregation agent then compares these results to find defects and resolve any disagreements between the reviewers.

This multi-model approach is designed to catch subtle errors in authentication logic and shared libraries. By looking at the code through different algorithmic lenses, the system is more likely to spot security vulnerabilities. However, Satapathy emphasizes that this is not a total replacement for human expertise. Developers must still validate the findings of the aggregation agent, especially for high-risk components.

The financial aspect of these reviews varies based on the level of scrutiny required. Basic “Lite” reviews are inexpensive, while the “Max” reviews cost significantly more due to the compute power involved. Managers must decide which parts of their codebase require the highest level of AI analysis. Sensitive security modules might justify the cost, while routine UI updates might not.

GitHub is also simplifying how companies pay for these features. By including Copilot in standard Enterprise subscriptions, the company is removing barriers to entry. The shift to a credit-based model allows organizations to pay only for the AI capacity they actually use. This flexibility is vital for teams that have fluctuating development cycles and need to scale their AI usage up or down.

The Evolving Role of the Developer

As AI agents take over the bulk of manual coding, the role of the software engineer is shifting toward high-level architecture. Instead of writing every line, developers are becoming reviewers and task-setters. They are responsible for providing the correct context and ensuring that the final output aligns with business requirements. This change requires a different set of skills focused on systems thinking and prompt engineering.

The focus is now on the validation of results rather than the mechanics of syntax. Engineers spend more time analyzing telemetry and discussing design than debugging basic errors. This evolution could lead to faster shipping cycles, provided that the human-in-the-loop remains effective. The success of this new model depends on how well teams can integrate these agents into their existing culture.

Ultimately, the overhaul of Copilot signifies a move toward a more autonomous future for software engineering. By treating AI as a teammate rather than just a tool, GitHub is betting that agents will soon handle the majority of routine maintenance. This allows human talent to focus on the creative and complex problems that still require a deep understanding of user needs and business logic.

IQM Achieves Full Chip Entanglement on 54 Qubit Processor

Valentin Podkamennyi — Thu, 29 Oct 2026 19:44:38 +0000

IQM Quantum Computers successfully demonstrated simultaneous entanglement across its entire 54-qubit Emerald processor. This milestone proves that the hardware maintains high computational quality even when every qubit on the chip functions at once. The achievement provides cloud users with a more powerful environment for running complex, large-scale quantum algorithms on a single processor.

Advancing Quantum State Benchmarks

The recent demonstration utilized the Greenberger-Horne-Zeilinger (GHZ) state, which serves as a rigorous industry benchmark for assessing large-scale entanglement. By preparing this state across the full architecture, the company confirmed that its hardware can handle the complex interdependencies required for advanced computing. The 54-qubit result represents a significant leap in the ability to manage a high volume of qubits without losing coherence or falling victim to systemic noise.

Measurement data from the experiment showed a fidelity of 0.516 across the entire processor after applying readout error mitigation. This specific number is vital because it sits above the 0.500 threshold required to prove that the qubits are truly entangled rather than just reacting to random noise. Achieving this level of precision across 54 units simultaneously validates the stability of the underlying superconducting architecture.

Before this breakthrough, users working on similar hardware typically limited their algorithms to approximately 30 qubits to maintain reliability. Many researchers in the field frequently use only small sections of a chip, often ranging from 10 to 20 qubits, even if the total capacity is much higher. This demonstration changes that dynamic by showing that an entire processor can function as a unified computational resource.

Optimization Through Automation

The success of the 54-qubit entanglement relies heavily on new automated routines for state preparation. These routines are designed to choose the most efficient two-qubit gates by balancing the fidelity of the gate against the potential for qubit idle time. By carefully selecting the order in which gates are applied, the system minimizes the period during which qubits remain inactive and vulnerable to decoherence.

These automated systems represent a shift toward more intelligent hardware management. Rather than relying on manual tuning, the software determines the optimal path for building the GHZ state across the chip. This minimizes the build-up of errors that usually occurs as the complexity of a quantum circuit increases. It ensures that the final state remains viable for actual computational work.

Enhancing Gate Fidelity

A major factor in reaching this milestone was a year-long effort focused on hardware refinement and calibration. Technical teams introduced an updated methodology for calibrating the system and improved the primary gate sets used for operations. These changes, combined with error-reduction tools like dynamical decoupling, allowed the team to refine how the qubits interact with one another.

The cumulative effect of these refinements was a reduction in error rates for quantum operations by as much as 50 percent. This reduction is essential for scaling up to larger qubit counts because errors tend to compound quickly in quantum environments. By halving the base error rate, the system can support the longer and more intricate circuits necessary for full-chip entanglement.

Implications for Cloud Computing Users

The transition from lab demonstration to cloud availability means that external developers now have access to this enhanced performance. Users no longer have to restrict their designs to small clusters of qubits to ensure success. The ability to use all 54 qubits in a single, coherent circuit opens the door for more ambitious experiments in materials science, chemistry, and cryptography.

By offering this level of performance through a cloud interface, the company democratizes access to state-of-the-art quantum hardware. Researchers around the world can utilize the same tools and calibration routines that made the 54-qubit entanglement possible. This availability encourages the development of new algorithms that take full advantage of large-scale processor architectures rather than being limited by hardware instability.

The integration of these tools into the cloud platform also simplifies the user experience. Developers do not need to be experts in hardware calibration to achieve high-quality results. The automated systems handle the underlying complexities of gate selection and idle time management, allowing researchers to focus on the logic of their algorithms and the outcomes of their data.

Future of Automated Calibration

The work on the Emerald system is part of a broader industry trend toward agentic and AI-driven calibration. Automating the tuning of quantum systems is necessary for the next generation of processors, which will likely feature hundreds or even thousands of qubits. As systems grow in size, manual intervention becomes impossible, making these automated routines the foundation for future scaling.

The use of AI-driven agents to manage hardware ensures that the processor remains in an optimal state for longer periods. It also allows for faster recovery from environmental disturbances that might otherwise knock the qubits out of alignment. This continuous, intelligent monitoring is a key component in transforming quantum computers from experimental devices into reliable tools for industry.

Scaling Quantum Architectures

The achievement on the 54-qubit chip serves as a blueprint for scaling to even larger arrays. Because the methods used–such as smart gate ordering and error mitigation–are not limited to this specific processor size, they can be applied to future chips with higher densities. Proving that full-chip entanglement is possible at this scale provides a clear path forward for the development of even more powerful superconducting processors.

The move toward full-chip utilization also highlights the importance of connectivity within the processor. For 54 qubits to entangle simultaneously, every physical connection on the chip must function with high precision. This success confirms that the manufacturing and design processes for these processors are maturing, resulting in hardware that is consistently functional across its entire surface area.

Technical Milestones and Performance Metrics

The distinction between noise and genuine entanglement is a critical hurdle in quantum physics. By surpassing the 0.500 fidelity mark, the Emerald system proves it has the clarity required for meaningful quantum work. Even at a 40-qubit count without the benefit of error mitigation, the system maintained a fidelity of 0.515, showing that the core hardware is inherently stable.

Readout error mitigation (REM) was used to reach the final 54-qubit result, a common practice that helps account for the small inaccuracies that occur when measuring the state of a qubit. The fact that the system maintained its performance as it scaled from 40 to 54 qubits indicates that the noise does not overwhelm the signal as the system grows. This stability is a promising sign for the future of large-scale quantum integration.

The reduction of idle time played a central role in these metrics. In quantum computing, any moment a qubit is not being actively manipulated is a moment it can lose its quantum state. By shrinking these gaps, the IQM team essentially protected the qubits from their own environment. This strategy is now a standard part of the operational routine for the Emerald system.

Comparing Industry Standards

The results reported for the Emerald system are among the strongest currently available for commercial quantum systems. While many companies claim high qubit counts, the ability to use those qubits simultaneously in a single entangled state is the true measure of a processor’s power. IQM is positioning itself as a leader by focusing on the quality and usability of its qubits rather than just the total number.

This focus on full-chip performance addresses a common frustration among quantum developers who find that only a small portion of a purchased or leased processor is actually usable for complex work. By ensuring that all 54 qubits can work together, the company provides a more honest and effective tool for the scientific community. It sets a high bar for what customers should expect from cloud-based quantum services.

Evolution of Quantum Operations

The journey to this 54-qubit success involved a complete overhaul of how quantum operations are handled. The deployment of an improved gate set allowed for more efficient communication between qubits, which is the foundation of entanglement. These technical upgrades were not just incremental changes but represented a fundamental shift in how the processor manages its internal logic.

As these technologies continue to evolve, the gap between experimental laboratory results and commercial cloud performance is shrinking. The Emerald system’s recent performance is a clear indicator that the industry is moving into a phase where full-processor utilization is the standard rather than the exception. This progress brings the field closer to solving real-world problems that require the massive parallel processing power of a fully entangled quantum chip.

The ongoing commitment to automating these complex processes ensures that the hardware remains accessible to a wide range of users. As quantum computing moves out of the purely academic realm and into the commercial sector, the reliability and ease of use demonstrated by the Emerald system will be the primary drivers of adoption. The ability to entangle an entire chip is a major step toward that goal.

Oracle Cloud Infrastructure integrates Google TPU accelerators

Valentin Podkamennyi — Wed, 28 Oct 2026 18:17:08 +0000

Oracle announced a major collaboration with Google to integrate Tensor Processing Units into its cloud environment. This partnership gives enterprise clients a new way to handle large-scale artificial intelligence tasks. The integration focuses on providing specialized hardware for training and inference within Oracle Cloud Infrastructure.

Expansion of Enterprise AI Compute Options

The agreement between these two technology giants marks a significant shift in the cloud industry. For the first time, Google is allowing its proprietary AI silicon to operate within the data centers of a direct competitor. Oracle will manage the physical infrastructure required to support these units. This includes providing the necessary power, cooling systems, and high-speed networking to keep the chips running efficiently.

This move targets businesses that are struggling with the high costs of traditional graphics processing units. As AI models grow in size, the demand for specialized hardware increases. Oracle aims to position itself as a flexible provider by offering diverse hardware choices. By adding these accelerators, the company helps users manage the massive data requirements of modern machine learning.

Performance and Efficiency Benefits

The hardware in question is designed specifically for neural network processing. Unlike general purpose chips, these units excel at the mathematical operations required for deep learning. This specialization often results in better performance per watt of electricity consumed. Enterprises look for these efficiencies to lower their total cost of ownership for AI projects.

Inference tasks represent the bulk of day-to-day AI expenses for most organizations. Every time a user interacts with a digital assistant or a model generates text, it uses compute power. Specialized chips can handle these repetitive tasks more effectively than standard processors. This efficiency allows companies to deploy more complex agents without a linear increase in costs.

Strategic Market Positioning

The decision to host Google hardware on Oracle infrastructure reflects a broader trend toward multi-cloud environments. Companies often want to use specific technologies from different providers without moving their entire data stack. This partnership simplifies that process for Oracle users. They can now utilize Google’s advanced silicon while maintaining their existing data and security configurations on Oracle’s platform.

This strategy helps Oracle compete more aggressively with other major cloud platforms. By offering a unique hardware mix, the company attracts developers who prefer specific software stacks. It also provides a buffer against supply chain issues that often affect the availability of high-end GPUs. Having multiple types of accelerators ensures that customers have options during periods of high demand.

Technical Capabilities of Modern Accelerators

Google has spent more than a decade refining its custom silicon. The latest generation of these chips includes two distinct architectures tailored for specific phases of the AI lifecycle. One version is optimized for the training phase, where models learn from vast datasets. The other version focuses on inference, which is the process of using a trained model to make predictions or generate content.

The training-focused architecture, known as the TPU 8t, offers substantial performance gains over earlier iterations. It can drastically shorten the time required to develop frontier models. What used to take months of compute time can now be completed in weeks. This speed is critical for companies racing to bring new AI features to market.

Optimized Inference Workloads

On the other side of the spectrum, the TPU 8i is built for the iterative nature of inference. Modern AI applications often involve multiple agents calling models repeatedly to finish a single task. This creates a high volume of small, fast computations. The inference-optimized hardware handles these bursts of activity with low latency.

Low latency is essential for maintaining a good user experience. If an AI response takes too long, the tool becomes less useful for real-time applications. By using hardware designed for these specific patterns, businesses can ensure their applications remain responsive. This is particularly important for customer-facing agents and automated support systems.

Integration Challenges for Developers

While the hardware is powerful, switching to a new accelerator requires careful planning. Developers must consider their existing software tools and skill sets. Not every model architecture runs perfectly on every type of chip without modification. Teams may need to adjust their code or use different libraries to take full advantage of the new hardware.

The cost of this development work must be factored into the overall project budget. A cheaper hardware option is only a bargain if the migration effort is manageable. Organizations must evaluate whether the performance gains justify the time spent reconfiguring their workloads. Oracle and Google intend to provide documentation and support to help ease this transition for their joint customers.

Future Availability and Implementation

Despite the excitement surrounding the announcement, the technology will not be available immediately. Oracle plans to begin the rollout of these accelerators in the 2027 calendar year. This timeline gives the company enough time to build out the specialized data center environments required for the chips. The capacity is expected to increase gradually over several years.

Specific details regarding the pricing of these services remain undisclosed. Many industry analysts are waiting to see how the cost compares to traditional GPU instances. The final price point will likely determine how quickly enterprises adopt the new technology. Initial capacity will probably be targeted at large-scale early adopters before becoming more widely available.

Data Center Readiness

Preparing a data center for these high-performance units involves significant engineering. The power density required for AI accelerators is much higher than that of standard web servers. Oracle must upgrade its facilities to handle the increased heat and electricity demands. This infrastructure work is part of why the launch date is set several years into the future.

The networking component is also vital. AI training requires massive amounts of data to move between chips with almost no delay. Oracle must ensure its internal fabric can support the high bandwidth requirements of the Google hardware. This coordination between two different engineering teams is a complex task that requires rigorous testing and validation.

Long Term Outlook for Cloud Hardware

The partnership signals a shift toward more specialized cloud offerings. In the past, cloud providers mostly offered standardized hardware that worked for any task. Now, the industry is moving toward a model where hardware is customized for specific applications. This trend benefits the consumer by providing more tools for specialized problems.

As the AI market matures, the competition between hardware manufacturers will likely intensify. Enterprises will continue to seek the best balance between speed, cost, and ease of use. The availability of Google’s custom silicon on Oracle’s platform is a major milestone in this evolution. It breaks down the barriers between ecosystems and focuses on delivering the best performance for the user’s specific workload.

Oracle Business Semantic Layer for AI Data Management

Valentin Podkamennyi — Wed, 28 Oct 2026 18:16:49 +0000

Oracle is introducing a Business Semantic Layer to its Autonomous AI Lakehouse to provide AI agents with a unified understanding of corporate information. This initiative focuses on increasing the precision of AI-generated insights and supporting the expansion of autonomous applications across various organizational departments.

Standardizing Enterprise Data for AI Accuracy

The new Business Semantic Layer acts as a foundational bridge between raw data and AI applications. It utilizes catalog annotations, SQL metadata from the Oracle AI Database, and analytic views suggested by AI to build reusable data models. These models define critical components such as dimensions, hierarchies, and key performance indicators. By establishing these universal definitions, the system ensures that AI agents interpret information the same way across different business units.

Consistency is a primary goal for this technology. When different departments use varied definitions for the same metric, AI agents often produce conflicting results. This layer aims to eliminate those discrepancies by grounding natural language queries and data retrieval in a single source of truth. The company confirmed that these new features will become available to users over the next year.

The benefits of this approach extend to the development process for IT teams. Developers can use these pre-approved metrics to build more reliable analytics tools without manually mapping data for every new application. This structure allows organizations to deploy agentic workflows that can navigate complex internal databases with higher confidence.

Industry experts note that grounding AI in specific business contexts is essential for professional-grade reliability. By standardizing dozens of entities and nearly a hundred different KPIs, companies can see a massive reduction in the time required to investigate and fix data errors. This efficiency allows tech leadership to focus on scaling operations rather than troubleshooting inconsistent AI outputs.

Integrated Planning and Simulation Capabilities

Beyond data modeling, the company introduced a Business Simulation Engine designed to assist with strategic planning. This engine integrates the functionality of Oracle Essbase directly into the Autonomous AI Lakehouse. It enables users to evaluate different business scenarios and perform what-if analyses without the need to move large datasets between separate software environments.

The simulation engine supports complex operations such as dynamic hierarchy rollups and advanced allocations. Users can interact with the system using natural language to generate detailed reports or perform analysis through traditional spreadsheet interfaces. This integration simplifies the overall data architecture for a company by keeping planning and analytics in one location.

By consolidating these functions, finance teams and business planners can work more efficiently. They no longer need to wait for data transfers or worry about version control issues between disparate systems. The engine provides a platform where historical data and future projections live side-by-side, allowing for faster and more accurate decision-making processes.

For the IT department, this consolidation reduces the burden of maintaining custom integrations. Developers can spend their time creating high-value workflows that turn insights into actions rather than managing the plumbing between various planning tools. This shift helps the organization respond more quickly to market changes or internal shifts in strategy.

However, the success of the simulation engine depends on its ability to fit into existing corporate workflows. It enters a competitive market filled with established planning platforms. For many organizations, the decision to adopt this tool will hinge on how well it integrates with the legacy systems they already use for daily operations.

Expanding Open Data Formats and Portability

The company is also enhancing its support for Apache Iceberg within the lakehouse environment. This update allows data professionals to create, modify, and delete Iceberg tables using standard SQL commands. It represents a significant step toward improving data portability and flexibility for organizations that use multiple cloud platforms or data processing engines.

Apache Iceberg is an open table format that enables high-performance analytics on large datasets. By supporting this format natively, the lakehouse allows data engineers to maintain a single copy of their data. This single version can then be accessed by various tools, including Apache Spark, Snowflake, or Databricks, without the need for expensive and time-consuming data duplication.

Reducing data duplication is a major priority for modern CIOs. It lowers storage costs and minimizes the processing power required to sync information across different environments. The ability to manage these tables using familiar SQL tools also simplifies the training requirements for data teams, as they do not need to learn specialized workflows for open-source formats.

While these open formats increase flexibility, they also introduce new governance challenges. Organizations must still manage the proprietary catalogs and semantic layers that interpret the data. Even if the data itself is portable, the logic used to define business rules may still be tied to a specific vendor’s ecosystem.

The timeline for the full rollout of the simulation engine and the expanded Iceberg features remains unannounced. Organizations looking to adopt these tools must prepare their data governance strategies now. Maintaining accurate and updated business models is a continuous responsibility that requires dedicated effort from data teams to ensure the AI agents remain effective over time.

Modern Strategies for Making Enterprise APIs Agent Ready

Valentin Podkamennyi — Mon, 26 Oct 2026 17:52:00 +0000

Enterprises spent the last decade building comprehensive API programs to manage internal services. A large organization now runs thousands of internal APIs, each featuring a specific owner and authentication scheme. These systems provide the essential context agents need for customer records, inventory, and support tickets as they enter production.

Resolving Scalability Issues in Agent Connectivity

The Model Context Protocol (MCP) established a standard way to connect agents to external tools. However, the common method of exposing APIs through MCP involves creating one tool for every API operation. This specific strategy fails quickly when applied to a large enterprise estate. Typical enterprise CRM or ERP systems often contain hundreds of operations. When an agent requires several of these APIs, it becomes overwhelmed by hundreds of different tool definitions.

Three major problems occur when using this one-to-one mapping approach. First, the financial cost increases significantly. Every tool definition must load into the model context before any work begins. Every intermediate result then flows back through the model, which means the organization pays for every hop in tokens. A simple workflow like checking a shipment status becomes expensive very quickly.

Accuracy also suffers as the list of tools grows. Once the number of tools exceeds what the model can reasonably handle, selection logic degrades. The agent often picks the wrong tool or makes unnecessary calls. Finally, latency becomes a major issue for the end user. Processing more context and managing constant back-and-forth communication slows the agent down significantly.

Limitations of Traditional Integration Curation

Teams frequently attempt to fix these issues by manually trimming tool lists. They split large APIs into smaller, specialized servers to keep things manageable. While this helps in the short term, it creates a new problem. Every specific agent use case then requires its own hand-maintained integration. This leads to the same type of integration sprawl that original API programs were designed to prevent.

The Shift Toward Code Mode Interfaces

The industry is currently moving toward a different pattern known as code mode. In this setup, a server exposes an API as a single interface that resembles a software development kit. The agent writes short code snippets against this interface, which then run in a secure sandbox. Only the final result of that execution returns to the model context.

This process involves three distinct steps: search, understand, and execute. The agent searches the API surface for a specific operation and pulls the schema on demand. This keeps the full API available while ensuring only the necessary parts reach the context window. It aligns perfectly with how large language models function, as they are trained extensively on code.

Integrating Agent Traffic with Existing Gateways

Efficiency gains from using code mode are quite dramatic. Because the agent works with code, a single snippet can chain multiple calls together using loops or conditionals. This reduces the number of round trips required for complex workflows. In internal testing, a workflow that previously used 120,000 tokens dropped to just 1,200 tokens. This represents a massive reduction in resource consumption.

However, running agent-written code requires a trustworthy sandbox. Platform teams must have full visibility into every call the code makes to ensure security. Adopting this pattern means treating observability as a primary requirement. Without these safeguards, the benefits of code mode could be offset by new security vulnerabilities within the internal network.

Utilizing Centralized Policy Enforcement

Governance remains a critical concern even after solving the context window problem. Any agent calling a production API needs the same protections as any other client. This includes robust authentication, authorization, and rate limiting. Many teams mistakenly recreate these controls for every new integration. This leads to credentials being hidden in server configurations and traffic bypassing the standard gateway.

A superior strategy involves treating the context layer as a direct extension of the existing API platform. The specifications and access policies already live where the API is managed. Organizations should generate the agent-facing interface from this existing source of truth. By routing agent calls through the same gateway used by other consumers, existing policies apply automatically.

Improving Discovery Through Centralized Catalogs

Discovery is another vital component of a successful agent strategy. Each agent-ready interface should be registered in a central catalog. This allows other teams to find and reuse existing integrations instead of building new ones for the same system. Centralization ensures that the agent becomes a governed client rather than a shadow IT project.

Composing Specialized Context Surfaces

APIs are only one part of the broader enterprise data picture. Most organizations already run various servers from SaaS vendors and internal teams. A truly useful agent needs to draw on all these sources simultaneously. A well-designed context layer allows platform teams to compose these sources into a single interface.

This composition should be scoped to the specific job of the agent. For example, a support agent might need access to ticketing and shipping data but should never see payroll information. This scoping serves as a vital security control by limiting the agent’s reach. It also serves as a performance control by keeping the model focused on relevant data.

Implementation at the Infrastructure Level

Curation is the ultimate goal of a context layer. Simply providing more context does not result in a better agent. The objective is to make enterprise systems available while keeping the data sent to the model small. Many companies are now treating this context as a form of infrastructure. They generate it from managed systems and govern it with existing policies.

The evolution of APIs into the backbone of digital business happened because they became secure and discoverable. AI agents require that same level of discipline to be effective in an enterprise setting. Most organizations have already built the necessary foundation. Now, the focus must shift to making those systems accessible to autonomous agents in a controlled and efficient manner.

Managing AI and API Traffic Together

New tools like the Kong AI Gateway are designed to bridge the gap between traditional APIs and AI needs. By generating code mode interfaces, these tools allow agent access to reside on the same platform used for standard API traffic. This unified approach ensures that as agents become more prevalent, they do not disrupt the existing security architecture of the organization.

As enterprises move forward, the focus will remain on selectivity. The fastest-moving companies will be those that can deliver context on demand without sacrificing performance. By using the principles of code mode and centralized governance, IT managers can ensure their API programs remain relevant in the age of artificial intelligence.

Examine AI Exposure in Value Funds

Valentin Podkamennyi — Mon, 26 Oct 2026 17:51:51 +0000

Investors hold significant exposure to artificial intelligence through various funds, including S&P 500 index funds, growth funds, and target-date options. Recent market shifts, with chip stocks experiencing a decline, prompt a reevaluation of investment strategies. Vanguard cautions that a move towards traditional “safe” value funds might not reduce AI exposure as much as investors anticipate.

This reevaluation comes as Vanguard warns against the risks of passively chasing current AI market leaders. Vanguard’s research director, Andrew Shuman, stated that blindly pursuing recent AI buildout winners could lead to buying at inflated prices. This perspective, detailed in Vanguard’s October 21 commentary, “Navigating the AI hype cycle: A case for active,” suggests a potential disconnect between investor perception and actual fund composition.

Evaluating AI Market Trends and Investor Exposure

The current AI market exhibits characteristics reminiscent of previous technology booms, according to Vanguard. Large sums of money have flowed into technology stocks, fueling price increases and attracting more buyers. Data from Morningstar, cited by Shuman, indicates that investors poured nearly $82 billion into technology stocks over a 12-month period through August 2026, significantly surpassing inflows into other sectors. This influx creates a self-reinforcing cycle that drives valuations higher.

The volatility inherent in chip stocks further complicates the landscape. Shuman noted that semiconductor and semiconductor equipment stocks are prone to higher betas and volatility, often experiencing substantial drawdowns. For instance, chip stocks collectively fell over 12% in July, and semiconductor and data center funds dropped 10.2% between October 1 and October 16. This historical pattern suggests that the initial beneficiaries of a technology buildout, such as infrastructure providers, often do not retain the largest profits in the long term. Instead, subsequent winners tend to be businesses that innovate upon the established infrastructure, like e-commerce and social media companies following the internet’s widespread adoption.

This dynamic challenges the assumption that broad technology exposure will consistently yield high returns. As the market matures, the ability to differentiate between sustainable growth and speculative bubbles becomes crucial. Investors must consider whether current valuations accurately reflect future profit potential or if they are inflated by enthusiasm. The risk of overcapacity in AI data centers, despite trillions of dollars in investment, could lead to broadly disappointing profits and stock returns, according to Vanguard. This scenario would replicate past cycles where initial infrastructure investments outpaced subsequent revenue generation, impacting investor returns. Understanding these historical parallels is essential for navigating the evolving AI investment landscape effectively.

Hidden AI Exposure in Value-Oriented Portfolios

Many investors turn to value funds, packed with established, “old-economy” companies, as a natural hedge against volatile technology sectors. The expectation is that these funds offer stability and reduced exposure to high-growth, high-risk areas like artificial intelligence. However, Vanguard reveals a significant caveat: these value funds may inadvertently carry substantial AI exposure through their top holdings. Andrew Shuman pointed out that Amazon and Microsoft, both investing billions in AI data centers, have become two of the three largest holdings in the Russell 1000 Value Index. This situation means that a fund designed to offer diversification away from technology may still hold significant positions in some of the world’s largest technology companies.

For example, the iShares Russell 1000 Value ETF (IWD) held Amazon at 6.12%, Apple at 6.11%, and Microsoft at 5.26% as of October 8. This equates to approximately 17.5% of the fund concentrated in three major technology giants. Such an allocation can undermine the intended diversification benefits for investors who own value funds to balance out an S&P 500 fund, which already has substantial tech exposure. Therefore, investors should carefully review the top holdings of their value funds to understand their true technological footprint. Without this due diligence, they may unknowingly double down on tech exposure while believing they are diversifying.

Moreover, Vanguard expresses skepticism about whether cost savings from AI adoption by old-economy companies will translate directly into higher equity holder returns. Shuman explained that in highly competitive industries, all participants could achieve similar AI-driven savings. This widespread efficiency might result in competition driving prices down for consumers, rather than leading to increased profit margins for shareholders. This situation creates a potential misalignment between operational improvements and investment returns. The benefits of AI could dissipate across the market through lower prices, eroding the potential for stock price appreciation. This further emphasizes the need for investors to look beyond superficial classifications of funds and scrutinize underlying assets and market dynamics.

The Case for Active Management and Future AI Opportunities

While Vanguard is renowned for its low-cost index funds, the firm now highlights a compelling argument for active management in the current AI environment. Andrew Shuman suggests that skilled active managers, if accessible at reasonable costs, possess the potential to excel in navigating the complexities of both equity and fixed-income markets. This perspective arises from the increasing dispersion within individual stock performances, even as overall market volatility remains historically normal. Such conditions, where the gap between winning and losing stocks widens, create an opportune setting for expert stock selection.

Active managers can seek out companies that demonstrate the strongest potential return on investment from AI products, both for consumers (like search and personal agents) and businesses (such as cloud computing). They also prioritize companies capable of delivering these products efficiently and cost-effectively, while maintaining robust balance sheets amidst increasing capital demands from public and private credit markets. Although hyperscalers currently trade at approximately 20 to 25 times expected earnings, Vanguard warns that substantial investments in AI data centers could lead to overcapacity and subsequently disappointing profits and stock returns.

Vanguard also anticipates a leveling of the AI playing field as the cost of AI tokens decreases. This reduction in cost could diminish the advantage of large budgets, allowing smaller companies with more agile cultures and management teams to thrive by disrupting existing business models. These emerging companies, many of which are still private or have minimal weight in cap-weighted indexes, could become the next generation of market leaders. Vanguard draws parallels to the e-commerce and mobile internet companies that have since grown into multi-trillion-dollar market capitalizations. However, investors must consider the historical performance of active funds. Data from S&P Dow Jones Indices’ SPIVA report indicates that roughly 67% of active large-cap U.S. stock funds underperformed the S&P 500 in the first half of 2026, and over a 20-year period, 93% fell short. The practical takeaway for investors involves not only knowing the exact holdings of their funds but also recognizing the potential for skilled active managers to identify the future “Amazons and Googles of the 2030s,” assuming fees remain reasonable.

Anthropic sets aside five percent of shares for public trust

Valentin Podkamennyi — Mon, 26 Oct 2026 17:51:43 +0000

Anthropic recently announced a plan to allocate five percent of its equity into a public benefit trust to share future wealth. This decision addresses concerns regarding how rapid advancements in artificial intelligence might disrupt the global economy and impact the general workforce.

Financial commitments to public welfare

The move by the artificial intelligence startup represents a significant shift in how tech companies approach their social responsibilities. By setting aside five percent of its shares, Anthropic is preparing for a future where its valuation could reach astronomical heights. Current rumors suggest the company is moving toward a record-setting initial public offering. If these projections hold true, the shares dedicated to the trust will be worth billions of dollars immediately upon going public.

Over time, the value of these shares could skyrocket into the hundreds of billions. To put this into perspective, a similar move by a major tech giant two decades ago illustrates the potential scale. If Google had committed five percent of its equity to a similar trust before its 2004 debut, that stake would be worth roughly $184 billion today. This comparison highlights the massive financial impact a successful AI firm can have on global philanthropy.

Observers view this plan through various lenses depending on their perspective of the tech industry. Some see it as a sincere attempt to ensure that the benefits of automation and intelligence reach everyone, not just investors. Others interpret the gesture as a strategic move to ease the concerns of government regulators. By creating a public benefit trust, the company may be attempting to build goodwill before its technology causes significant labor market shifts.

Historical precedents for corporate giving

While the financial scale of this plan is unusual, the concept of a business serving the public good is not entirely new. Long before the rise of Silicon Valley, actor Paul Newman established a unique model for corporate altruism. At the age of 57, the Hollywood legend launched a food company that would change the landscape of charitable giving. He turned his personal brand into a tool for social improvement by donating all profits to various causes.

Newman initially hesitated to use his famous face to sell products like salad dressing. He viewed the idea as a form of self-exploitation that felt uncomfortable for a serious artist. However, he eventually decided that if he was going to be exploited by the industry anyway, he should do it for a worthy cause. He launched his brand with the intention of giving away 100 percent of the proceeds to the community.

Lessons from the Newman model

The food company started as a small project but grew into a massive success. Today, the brand is a staple in grocery stores across the country, selling everything from salsa to pasta sauce. Since its inception in 1982, the organization has generated more than $600 million in total charitable donations. This success proved that a business could prioritize social impact while remaining competitive in a crowded marketplace.

The rise of effective altruism in tech

The philosophy driving Anthropic is a modern, data-driven version of the ideas Paul Newman practiced. This mindset is known as effective altruism, and it has become common among the leaders of major artificial intelligence laboratories. The core tenet of this belief system is that individuals and companies should gain as much influence as possible to maximize their positive impact. In this framework, the pursuit of wealth is a means to an end rather than the final goal.

However, the application of this philosophy in the technology sector has faced intense scrutiny. Critics point out that the drive to amass wealth for the common good can sometimes lead to unethical behavior. High-profile scandals in the cryptocurrency world showed how a focus on future giving can be used to justify present-day deception. In the context of AI, there are worries that the pressure to increase share value for the trust might lead to compromises.

Balancing safety with financial growth

The primary concern is that a focus on valuation could distract from the necessary work of ensuring software safety. If a company feels obligated to make its shares as valuable as possible for the public trust, it might rush products to market. This acceleration could lead to the release of tools that are not fully tested for bias or systemic risks. The desire to help humanity with money could inadvertently lead to technology that harms society in other ways.

Anthropic acknowledges that its technology has the potential to cause significant disruption. The company often speaks about the risks of powerful AI systems, including the possibility of catastrophic outcomes. Because of these risks, the public benefit trust serves as a safeguard, but it is not a complete solution. Financial compensation cannot easily fix the social or creative issues that might arise from widespread automation.

Ethical dilemmas in the frontier AI space

The startup must navigate a complex landscape of creative rights and societal impacts. As AI models train on vast amounts of human-generated data, questions about compensation for creators remain unanswered. While a multi-billion dollar trust is a helpful tool, it does not address the underlying tensions regarding how data is sourced. The tech industry continues to grapple with these fundamental questions as development speeds up.

Long-term implications for the industry

The creation of the public benefit trust sets a new standard for other startups in the intelligence space. If Anthropic successfully goes public with this structure, other firms may feel pressured to follow suit. This could lead to a new era of corporate governance where a portion of every major tech company is owned by the public. Such a shift would fundamentally change the relationship between Silicon Valley and the rest of the world.

However, the success of this model depends on the survival and prosperity of the human workforce. There is little benefit in accumulating wealth for humans if the technology itself creates unmanageable risks. The industry must find a way to balance the creation of value with the protection of existing social structures. This balance is the most difficult challenge facing the leaders of the AI revolution today.

Focusing on the means and the ends

Altruism is not solely defined by the amount of money a company gives away. The methods used to acquire that wealth are just as important as the eventual donations. Paul Newman’s legacy suggests that the spirit of the work matters as much as the financial outcome. For AI companies, this means that safety research and ethical development must take priority over profit margins and share prices.

The public should remain engaged and critical of these gestures. While a five percent equity stake is a massive commitment, it does not give the company a free pass on other responsibilities. Monitoring how these trusts are managed and how the money is distributed will be a key task for future regulators. Transparency in how the funds are used will determine if the trust is a true benefit or a marketing tool.

The future of the public trust

As the company nears its IPO, the details of the trust will become clearer to the public. Investors will have to decide if they are willing to accept a slightly lower stake in exchange for a more socially conscious corporate structure. The result of this experiment will likely influence the next decade of tech development and corporate responsibility. For now, the world watches to see if this bold financial move can truly help manage the transition to an AI-driven economy.

AI infrastructure trends and technology shifts

Valentin Podkamennyi — Sat, 24 Oct 2026 19:27:48 +0000

Technological shifts often force a complete reimagining of the foundations that support global industry. Current advancements in artificial intelligence are driving a massive overhaul of power and networking systems. Major technology firms are no longer waiting for traditional providers to update the aging infrastructure that currently limits their growth.

Historical parallels in autonomous transportation

The current trajectory of artificial intelligence infrastructure mirrors the evolution of the self-driving car industry. A few years ago, the automotive world focused heavily on achieving full autonomy. During this period, established vehicle manufacturers poured billions of dollars into internal projects and external partnerships. Ford launched initiatives like Argo AI, while General Motors backed Cruise. These companies attempted to merge their long history of manufacturing with new software requirements.

Despite these massive financial commitments, many traditional programs eventually stalled or faced significant reorganization. The complexity of the task proved difficult for institutions built on legacy manufacturing cycles. In contrast, companies born in the technology sector approached the problem from a software-first perspective. These organizations prioritized computer vision, radar, and advanced light detection systems. By viewing the vehicle as a mobile computer rather than a traditional machine, they achieved milestones that legacy brands struggled to reach.

This distinction in approach led to a visible shift in who leads the market today. Vehicles from tech-centric firms are now operating in urban environments while many traditional programs have been shuttered or scaled back. The lesson is that the entities creating the new technology often must be the ones to build the systems that support it. Waiting for older industries to adapt frequently results in missed opportunities and technical bottlenecks.

The role of specialized hardware

Technological urgency pushed innovators to develop custom sensors and networking tools that did not exist in the traditional automotive supply chain. This move toward vertical integration allowed for faster iteration. It also meant that the software and hardware could be optimized in tandem. This same pattern is appearing in the current race to build the hardware backbones for large language models and generative systems.

Speed as a competitive advantage

In the autonomy sector, the primary hurdle was often the pace of development. Tech companies operated on monthly update cycles while traditional manufacturers worked on multi-year vehicle programs. This mismatch in speed meant that by the time a traditional project launched, the underlying technology was already obsolete. This dynamic is now playing out in the energy and networking sectors as they struggle to keep up with AI data center construction.

The shift toward independent power and networking

Artificial intelligence requires a level of electrical power and data throughput that the current grid and internet cannot provide. This has turned massive software companies into primary builders of physical infrastructure. They are no longer just customers of utility companies; they are becoming energy providers and network operators. This change is born out of necessity because the existing regulatory and physical frameworks are too slow to meet modern demands.

Power has become a primary constraint for the expansion of data centers. These facilities consume vast amounts of electricity to run specialized chips and cooling systems. Because local grids often lack the capacity to support these loads, tech giants are investing directly in energy generation. Some are signing long-term agreements to restart nuclear facilities, while others are exploring natural gas generation. This direct involvement ensures a dedicated supply that is not subject to the same fluctuations as the public grid.

Telecommunications is undergoing a similar transformation. The standard internet was built for consumers and small businesses, but AI training requires massive data transfers between specialized servers. To solve this, companies are constructing private optical networks that connect their data center campuses. By purchasing networking equipment directly and laying their own fiber, they avoid the constraints of traditional carriers. This allows for a customized architecture designed specifically for the unique traffic patterns of machine learning.

Resolving the energy bottleneck

The mismatch between data center construction and utility expansion is a significant hurdle. Utilities operate under strict regulations that require years of planning for new transmission lines or power plants. AI projects move much faster, often going from concept to operation in a fraction of that time. By building their own generation capacity, technology companies bypass these bureaucratic delays. This approach is becoming so common that some regions now require large data centers to bring their own power to the table.

Redesigning data distribution

Traditional network carriers are optimized for distributing content to millions of individual users. AI infrastructure needs to move massive datasets between a few high-powered locations. Because of this, the standard carrier model does not fit the needs of hyperscalers. They are moving toward a model where the network is a private asset. This gives them total control over latency and bandwidth, which are critical for synchronizing the hardware used in large-scale model training.

Future implications for infrastructure providers

The trend of technology companies building their own systems does not mean that traditional utilities or carriers will disappear. Their roles remain vital for the broader public, but they are no longer the sole gatekeepers of progress. The relationship between technology and infrastructure is being redefined. Moving forward, the institutions that manage our power and data must learn to coexist with private systems that operate at much higher speeds.

We are entering an era where the underlying architecture of the digital economy must change. History shows that technology almost always moves faster than the laws and institutions that govern it. Eventually, those institutions adapt to the new reality, but only after innovators have demonstrated what is possible. The companies that are currently building the tools for the AI economy are the same ones defining how we will generate power and transmit data for decades to come.

This transition is not about a lack of competence within traditional industries. It is a matter of differing operational structures. Regulated utilities are designed for stability and affordability over long periods. Technology firms are designed for rapid scaling and market capture. As these two worlds collide, the result is a hybrid landscape where the most critical infrastructure is increasingly managed by the companies that use it most.

Adaptation and coexistence

Traditional providers must eventually find ways to integrate these private advancements into the public sphere. There is a potential for these private energy and data projects to provide overflow capacity or backup systems for the general public in the future. However, for the immediate future, the focus remains on isolation to ensure that the rapid growth of AI does not compromise existing public services.

The emergence of the AI factory

Modern data centers are evolving into something more complex than simple server warehouses. They are becoming integrated factories where energy, cooling, and data are managed as a single ecosystem. This holistic approach allows for efficiencies that were previously impossible. By controlling every layer of the stack, from the power source to the optical fiber, these companies are creating a new standard for industrial productivity. The race to build this infrastructure will determine who leads the next phase of the global economy.

Google Agent Compute Engine Lowers AI Idle Costs

Valentin Podkamennyi — Fri, 23 Oct 2026 16:53:23 +0000

Google Cloud is addressing the rising financial burden of idle infrastructure as businesses expand their use of autonomous AI agents. The new Agent Compute Engine service allows companies to pause inactive agents and restart them instantly while maintaining their full execution state and environment context.

Efficient Resource Management for AI Agents

Modern AI agents often operate in cycles where they spend significant time waiting. They might wait for a human to approve a draft, an external API to return data, or a complex tool to finish a specific task. Traditional cloud infrastructure typically requires these resources to stay active and billable during these pauses. If an organization shuts them down to save money, they face high latency when restarting. The restart process usually involves rebuilding the entire environment and loading the context from scratch.

Google aims to eliminate this inefficiency by using hardware-isolated microVMs. These micro virtual machines allow the system to freeze the agent in its current state. When the agent is needed again, it resumes immediately from where it stopped. This approach reduces the boot time and resource requirements compared to traditional virtual machines. It provides a more streamlined way to handle the intermittent nature of agentic workflows.

The use of microVMs also brings a distinct security advantage over standard containers. Containers share the kernel of the host operating system, which can create vulnerabilities if one agent is compromised. In contrast, each microVM runs its own kernel. This creates a much stronger isolation boundary. Such a setup is vital for agents that execute code generated by AI or interact with untrusted external inputs. It ensures that a security breach in one agent does not easily spread to the rest of the infrastructure.

For IT leadership, the ability to suspend and resume environments quickly is a potential game-changer for budgeting. Many long-running research or coding agents stay active for days or even weeks. Traditional billing models were never designed for these types of long-duration, low-activity tasks. By stopping the clock on CPU charges during idle periods, companies can scale their agent fleets without a linear increase in their cloud bills. This makes it feasible to deploy thousands of agents where dozens were previously the limit.

Strategic Trade-offs and Implementation Risks

While the technical benefits are clear, adopting this new service introduces several strategic risks. One of the primary concerns for enterprise architects is the risk of vendor lock-in. The Agent Compute Engine is deeply integrated into the Google Cloud ecosystem. Moving these workloads to a different provider would likely require a massive overhaul of how the organization manages identity and state. This dependency makes it harder for companies to maintain a multi-cloud strategy or migrate if pricing structures change.

The actual cost savings also depend heavily on the specific behavior of the agents. Agents that are constantly active or perform compute-heavy tasks will see very little benefit from a suspension service. Organizations must carefully audit their workloads to see how much idle time actually occurs. If an agent only waits for a few seconds between tasks, the overhead of managing the suspension might outweigh any savings. IT leaders need to identify which specific agents hold state for long periods before committing to this architecture.

Storage costs represent another hidden financial factor. While a company might save money on CPU and memory by pausing an agent, they still have to pay to store the state of that agent on disk. When scaling to thousands of persistent environments, these storage fees can accumulate quickly. This makes it difficult for finance departments to predict exactly how much money the transition will save. In some cases, the cost of high-performance storage for suspended states could rival the cost of leaving the compute resources running.

There is also the matter of service maturity. The engine is currently in an early stage and has not yet reached a wide preview. This means that early adopters are essentially betting on a future roadmap rather than a battle-tested product. Companies must weigh the potential for early-mover advantages against the stability risks of a new platform. Without proven production performance at scale, some organizations may choose to wait until the service demonstrates more reliability in diverse enterprise environments.

Technical Hurdles for Development Teams

Developers face a unique set of challenges when moving to a checkpoint-based execution model. The core idea is to save the memory and state so the agent resumes exactly where it left off. However, this becomes complicated when an agent is interrupted during a transaction with an external system. For example, if a pause occurs after an agent sends a payment request but before it receives a confirmation, the agent may wake up in an uncertain state. It might attempt to process the payment a second time, leading to errors.

To avoid these issues, developers must write code that is idempotent. This means every action the agent takes must be safe to repeat without causing side effects. Most current agent code is not written with this level of rigor. Shifting to this model requires a change in how developers think about external API calls and database commits. It adds a layer of complexity to the development process that might slow down initial deployment times.

Environmental drift is another long-term concern for maintainers. If an agent environment stays active for several weeks through multiple suspend-and-resume cycles, it may begin to diverge from its original configuration. Files can change, temporary dependencies can accumulate, and versions can become inconsistent. This makes it incredibly difficult for engineers to reproduce bugs or conduct security audits. The environment becomes a “snowflake” that is unique and hard to manage at scale.

To combat this drift, development teams will need to implement strict expiry policies. They must periodically force agents to rebuild their environments from fresh, clean images to ensure consistency. This adds operational overhead that can negate some of the simplicity promised by the service. Furthermore, the tools for debugging a paused virtual machine are not yet fully developed. Engineers may find it difficult to diagnose problems that only appear after an agent has been resumed from a suspended state.

Reduce AI Data Center Deployment Time

Valentin Podkamennyi — Fri, 23 Oct 2026 13:55:49 +0000

Vertiv, in collaboration with Nvidia, has developed a new strategy to significantly reduce the time needed to design and deploy power and cooling infrastructure for AI data centers. This innovative approach aims to cut the deployment timeline from approximately 120 days to just 48 days, accelerating the operational readiness of advanced AI computing facilities. The initiative underscores a critical need for faster infrastructure delivery in the rapidly expanding AI sector.

This accelerated timeline is crucial for operators. The compute power of Nvidia’s newest AI systems demands substantial electrical and cooling resources, which form a bottleneck in bringing expensive GPUs online for customer workloads. Vertiv estimates that by bringing infrastructure online approximately ten weeks sooner, AI data centers could see up to a 30% increase in cumulative AI token output during the first 12 months, assuming consistent installed computing capacity. This projection stems from a modeled scenario, where the new approach provides 317 productive days compared to 245 days under traditional methods, though actual project results will vary. Following this announcement, Vertiv shares saw a rise of approximately 4%, trading around $212.51, up from a previous close of $203.56. The company is set to report its third-quarter earnings shortly, offering further insights into its financial performance and strategic advancements.

Streamlining Infrastructure Deployment for AI

Vertiv’s innovative model systematically reduces the traditional infrastructure deployment timeline from 120 days to 48 days, focusing on efficiency across two key stages. This significant reduction addresses common delays faced by AI data center operators.

The first stage, encompassing design, configuration, and virtual testing, historically took about 30 days using conventional methods. Under the new approach, this phase collapses to just three days through the utilization of pre-engineered designs. These designs are based on Nvidia’s Vera Rubin DSX AI Factory architecture, providing a standardized yet customizable starting point for infrastructure planning. Engineers can virtually test proposed setups using technology from Dassault Systmes, a French industrial software company, ensuring that computing, power, and cooling systems integrate seamlessly before any physical construction begins. This digital prototyping minimizes errors and rework, saving substantial time and resources.

The second stage, involving the physical deployment of power and cooling systems, traditionally required around 90 days. Vertiv’s strategy cuts this down to 45 days by shifting a greater portion of assembly and testing into a factory setting. This means larger, prefabricated sections of the power and cooling infrastructure arrive at the data center site ready for installation, rather than being assembled piece by piece on-site. The factory-built modules streamline the construction process, reducing labor requirements and potential delays associated with on-site assembly. This shift is particularly vital for AI operators, as GPUs generate revenue only once adequate power, cooling, and networking are in place to support customer workloads. Shortening the infrastructure schedule directly translates into more operational days for the high-value computing equipment that operators have already invested in, aligning with Vertiv CEO Giordano Albertazzi’s objective of accelerating the path from project requirements to productive AI capacity. Vertiv has been strategically strengthening its relationship with Nvidia, specifically developing power and cooling solutions tailored for Nvidia’s advanced systems and increasingly power-dense server racks.

Real-World Application and External Challenges

The effectiveness of Vertiv’s streamlined deployment model is now being tested in a significant project in Norway. Polar Data Centers, a prominent data center operator, has adopted Vertiv’s factory-built deployment approach for the initial 41-megawatt phase of an AI data center situated in Herya. This facility is engineered with scalability in mind, designed to eventually expand to a capacity of 100 megawatts. Vertiv is supplying essential components, including prefabricated power and thermal modules, along with overhead infrastructure for power distribution, liquid cooling, and network connectivity. This project provides a critical real-world proving ground for the deployment methodology.

While the Norway project applies the approach, neither Vertiv nor Polar has confirmed that the facility has achieved the projected ten-week time saving or the 30% increase in token output, as detailed in Vertiv’s internal models. The actual time saved and operational efficiencies gained in real-world scenarios depend on various factors beyond the equipment supplier’s direct control. One significant challenge involves securing a sufficient and timely electricity supply. Even with preassembled cooling and electrical equipment ready for installation, a data center can face substantial delays if grid connections are inadequate or deferred.

Recent incidents highlight this pervasive issue. Australian data center operator NextDC experienced a dramatic 41.7% drop in shares after electricity-network upgrade delays at its M3 Melbourne facility forced the postponement of approximately A$140 million in contracted customer revenue. The company now anticipates the next increase in grid supply at the site will not occur until 2028. Similarly, Microsoft faced prolonged delays in Britain, withdrawing from an agreement to anchor an Nscale data center in Loughton, England, because the site’s anticipated power connection was pushed into the 2030s. This planned facility was intended to house 23,040 Nvidia GB300 GPUs and deliver 50 megawatts of AI computing capacity. These cases underscore that grid connection delays represent a significant hurdle that Vertiv’s factory-built deployment model cannot independently resolve. Such examples also demonstrate why data center companies are increasingly investing in alternative strategies to secure electricity access more rapidly, often exploring solutions closer to the power source itself.

Expanding Vertiv’s Power Management Capabilities

Recognizing the crucial role of reliable and timely power access, Vertiv is expanding its capabilities beyond internal data center infrastructure to address broader power sourcing challenges. In September, Vertiv announced the acquisition of UtilityInnovation Group for an upfront cash payment of approximately $1.45 billion. This strategic acquisition introduces technology specifically designed to help data centers source and manage power before it reaches individual server racks. The integration of UtilityInnovation Group’s expertise marks a significant move further upstream in the power supply chain for Vertiv.

UtilityInnovation Group specializes in building microgrid controls, coordinating on-site generation, and designing “behind-the-meter” power systems. These systems are essential for enabling data centers to utilize local power generation or energy storage solutions when the utility grid is unable to provide adequate electricity quickly. This capability is becoming increasingly important as grid constraints and delays in power network upgrades impact data center deployment schedules, as evidenced by the challenges faced by companies like NextDC and Microsoft. By offering tools for deciding the origin of electricity and how on-site generation can bridge supply gaps, Vertiv can provide a more comprehensive power management solution.

Vertiv’s existing equipment primarily focuses on moving electricity from the point of entry into a facility down to individual server racks. The acquisition of UtilityInnovation Group complements these offerings by adding capabilities that influence where that electricity originates. This integrated approach allows Vertiv to connect infrastructure solutions from the initial power source all the way down to the individual chip. Furthermore, Vertiv has been consistently expanding its production capacity for equipment tailored to high-density AI facilities and developing systems specifically designed around newer Nvidia rack architectures. This broader strategy, aiming to unify infrastructure from the power source to the chip, positions Vertiv to capitalize on the escalating demand for AI infrastructure. The company’s financial performance already reflects this growth, with second-quarter revenue rising 24% year-over-year to $3.27 billion, prompting management to raise its full-year outlook. The upcoming third-quarter earnings report on October 28 will provide further updates on these strategic initiatives and the company’s financial trajectory. The Polar Data Centers project in Norway will offer early insights into how Vertiv’s factory-built approach performs in practice, with a key question remaining on how much of the projected 10-week savings will materialize once all external factors, including grid access and permitting, are factored into the overall schedule.

Network Compute for Local AI Applications

Valentin Podkamennyi — Thu, 22 Oct 2026 17:05:12 +0000

Extreme Networks is introducing a new technology that allows enterprise customers to execute AI applications locally by leveraging existing network equipment’s idle processing power. This approach addresses the growing demand for on-premises AI capabilities.

The company’s new offering, Extreme Exchange zones, pools available compute resources. These resources include unused capacity on Extreme access points and switches already deployed in various locations like branch offices and other edge sites. This pooled capacity will support AI applications and plug-ins directly at the network’s edge.

Enhancing On-Premises AI Capabilities

Extreme Exchange zones are specifically designed to tackle critical challenges faced by organizations regarding data sovereignty and control. By enabling local AI application execution, businesses gain full command over data residency and processing, regardless of where data originates or workloads run. Nabil Bukhari, CTO and president of AI platforms at Extreme, emphasized that this technology empowers customers to define their data boundaries.

The core concept behind Exchange zones involves utilizing data from Extreme’s Universal Compute Platform or a qualifying customer-owned server. It combines this with the spare processing capacity found in deployed network infrastructure. This strategy allows AI applications and their associated plug-ins to run closer to the point of data generation. The zones also integrate data from other Extreme network components, such as the Platform ONE management and security package and the Extreme AI intelligence core, creating a comprehensive local AI environment.

Extreme Networks highlights that its systems generate valuable operational data beyond routine network management tasks. This data holds significant potential for diverse applications, ranging from security and business operations to logistics. Exchange zones facilitate bringing this data on-premises, thereby unlocking use cases that cloud-only architectures currently cannot support. The default setting ensures models run locally, with cloud inference only occurring when explicitly permitted by customer policy. All instances of cloud inference are meticulously logged, ensuring transparency and control.

Strategic Advantages of Local AI Processing

While cloud infrastructure remains suitable for numerous AI workloads, Extreme Networks acknowledges its limitations for others. Certain applications cannot tolerate the inherent delay of a round trip to the cloud. Moreover, some applications generate an immense volume of data, rendering it uneconomical to transfer everything to remote cloud servers. Regulated industries and organizations with distributed operations frequently require tighter control over the precise location where AI processes execute. Exchange zones directly address these specific gaps, executing Exchange applications and their models within boundaries entirely controlled by the customer.

This initiative builds upon the foundation of Extreme Exchange, the company’s existing AI marketplace. Extreme Exchange offers a curated catalog of AI tools, agents, and applications. This platform provides enterprises with a streamlined mechanism to discover, deploy, and even create their own AI agents, microapps, and workflows, eliminating the need for extensive in-house development. The integration of Exchange zones extends this capability, enabling these advanced AI components to run efficiently on local infrastructure.

The benefits of this local processing approach extend to improved economics. By leveraging existing on-premises equipment, enterprises can reduce the reliance on costly cloud-based AI services, particularly for inferencing tasks. This model offers a more cost-effective solution for processing data that might otherwise incur significant transfer and compute charges in a public cloud environment. The ability to perform local inferencing also leads to faster response times, which is critical for real-time applications and operational decision-making at the edge of the network.

The Future of Distributed AI and Data Control

The introduction of Extreme Exchange zones signals a strategic shift towards more distributed AI architectures. As organizations generate increasing volumes of data at the edge, the capacity to process and analyze this information locally becomes paramount. This localized processing minimizes latency, enhances data security, and ensures compliance with data sovereignty regulations that vary across different jurisdictions and industries. Enterprises gain a robust framework for managing their data and AI workloads within a defined, secure perimeter.

The company has not yet provided specific details regarding the exact Extreme hardware models involved in Exchange zones. Similarly, the precise methodology for determining available processing capacity across deployed network devices remains undisclosed, given that the platform is still in its preview phase. However, the overarching vision is clear: to transform underutilized network assets into powerful compute engines for localized AI. This development empowers businesses to unlock new applications and insights directly from their operational data, without the constant need to transmit it to centralized cloud platforms.

This new technology represents a significant step towards democratizing AI by making advanced processing capabilities more accessible and controllable for enterprise customers. It emphasizes the importance of data governance and cost optimization in the rapidly evolving landscape of artificial intelligence deployment. As the technology moves beyond its preview stage, more technical specifications and deployment guidelines are expected to emerge, further clarifying its operational parameters and potential impact on enterprise network architectures.