Fortinet Inc. announced its intent to acquire Virtue AI Inc. Monday, and though the financial aspects are interesting, what’s more significant is how this will help businesses with their next big security challenge: How do you secure agentic artificial intelligence?

Generative AI has raised new concerns about data leakage, unsanctioned use and inaccurate outputs. Agentic AI raises the security stakes because it’s significantly harder to secure, and the impact of breaches is greater. These systems can retrieve information, call application programming interfaces, trigger workflows, interact with applications and increasingly act on behalf of people. That makes an AI agent less like a “chatbot” and more like a digital employee with access credentials, permissions and the ability to affect business processes.

The industry’s response cannot be limited to securing the model or filtering prompts. Organizations need to secure the full lifecycle of autonomous AI: what data an agent can access, which tools it can use, what decisions it is permitted to make, how those actions are monitored and when humans must intervene. That is why Fortinet’s acquisition of Virtue AI, which brings AI runtime protection, automated AI validation and agentic AI security capabilities to the company, is strategically important.

Why agentic AI changes the security equation

Traditional application security was designed for relatively predictable systems. Code may be updated periodically, but it generally operates within defined parameters. Users authenticate, applications make known requests and security teams can establish policies based on expected behavior.

Agentic AI operates differently. An agent can dynamically decide which tools to invoke, what information to retrieve, and what next step to take based on the context it receives. It may interact with multiple systems within a single workflow, drawing on customer data, internal knowledge bases, cloud services and business applications. In some cases, one agent may delegate tasks to another.

That flexibility is what makes agentic AI valuable. It is also what makes it difficult to secure. The core challenge is that agents operate in continuously changing environments. A model may be updated. A prompt may be modified. A new API connector may be added. The data source for a retrieval system may change. A business team may expand an agent’s role from answering questions to initiating transactions. Each change can alter the system’s risk profile.

This raises several security concerns. Prompt injection can cause an agent to ignore its intended instructions or manipulate it into taking an unintended action. Excessive permissions can turn a minor error into a significant incident. Connected tools can create indirect paths into critical systems. Sensitive data can be exposed through poorly governed retrieval or workflow execution. Because agents can operate at machine speed, a flawed decision can escalate much faster than a traditional human-driven security event.

The biggest mistake enterprises can make is treating AI agents as just another application category. They are dynamic actors in the environment. They have identities, permissions, behavioral patterns and connections to other systems. Security controls need to reflect that reality.

Why continuous AI protection is needed

Most security programs still rely heavily on point-in-time assessments. An application is tested before going into production. A vendor is reviewed during procurement. A configuration is audited periodically. These processes remain important, but they are insufficient for autonomous AI.

An AI agent can behave differently tomorrow even if no one changes the underlying business process. A new data set may introduce unexpected content. A model provider may alter the model’s behavior. A tool integration may change its capabilities. An adversary may find a new way to influence the agent through malicious content, instructions or inputs.

That is why continuous AI protection is so important. Continuous protection means validating an agent before deployment, monitoring it while it operates and reassessing it whenever its model, data, tools, permissions or role changes. It also means moving beyond simple guardrails at the model interface and gaining visibility into the agent’s decisions, tool calls and outcomes.

The objective should not be to stop enterprises from using agentic AI. That would be unrealistic and strategically counterproductive. The goal should be to enable organizations to deploy agents with confidence by establishing meaningful guardrails for their behavior.

A practical, continuous-protection approach should include several elements:

• Clear inventory of all AI agents, their owners and their business purpose.
• Visibility into the data sources, applications and tools each agent can access.
• Least-privilege permissions that limit agents to only the actions they need.
• Continuous testing for prompt injection, unsafe tool use, data leakage and policy violations.
• Runtime monitoring that detects unusual behavior or actions outside expected parameters.
• Human approval requirements for high-impact tasks, such as financial transactions, system changes or access modifications.
• Detailed logs that connect prompts, agent decisions, tool calls, identities and outcomes.

This is where security will need to evolve from protecting infrastructure and applications to governing behavior. The ability to see that an agent connected to a particular application is useful. The ability to understand why it acted, what data it used, which tool it called and whether the action complied with policy is far more valuable.

How Virtue AI fits Fortinet’s security platform story

Fortinet has long emphasized the convergence of networking and security through its Security Fabric. The company’s core argument is that enterprises are better served when security controls share intelligence, policy and enforcement across networks, clouds, endpoints and operations, rather than operating as disconnected point products.

The addition of Virtue AI extends that thesis to agentic AI. Fortinet already has AI-related security capabilities, including tools to protect AI interactions and broader efforts to integrate AI into security operations. Virtue AI adds a more specific layer focused on validating and protecting AI agents as they operate.

Securing traffic to and from a large language model is necessary, but it is only one part of the problem. Enterprises also need to secure what the agent does after it receives an answer from a model. Can it access sensitive data? Can it send an email? Can it create a ticket? Can it change a configuration? Can it trigger a financial workflow? Can it invoke another agent?

Virtue AI’s capabilities should help Fortinet broaden its platform story from protecting AI access to protecting AI actions. If integrated effectively, the technology could provide Fortinet customers with a way to apply consistent policy and visibility across the network, applications, endpoints and increasingly autonomous systems operating within those environments.

That is especially important for enterprises already struggling with security tool sprawl. Few organizations want to deploy a separate point product for each new AI-related risk. They want AI protection that fits into existing operational workflows, uses existing telemetry and can be managed alongside broader security controls.

The proof of this will come from execution. Fortinet will need to demonstrate how Virtue AI integrates with the broader Security Fabric, how customers can operationalize it without adding unnecessary complexity, and how the technology performs across multivendor environments. The most successful AI-security platforms will not be those with the most ambitious marketing claims. They will be those that make autonomous AI safer without slowing the business value it can create.

Recommendations for security pros

Security leaders should treat agentic AI as a new class of privileged digital worker, not merely another software-as-a-service application or chatbot. Start by inventorying every AI agent already in use, including departmental experiments, and document the data it can access, the tools it can invoke and the actions it is authorized to perform.

The foundational controls are the ones most security pros use today. These include least-privilege access, strong identity governance, segmentation, logging and human approval for high-impact actions. The difference is that these controls must be evaluated continuously because an agent’s behavior can change as models, prompts, tools and connected data sources evolve.

Organizations should also establish a formal AI-agent lifecycle: Validate an agent before production, monitor its behavior at runtime, test it against adversarial inputs, and revoke or reduce permissions when its role changes. Security teams should insist on telemetry that links an agent’s prompt, decision, tool call, identity and outcome. Without that end-to-end view, investigating an AI-driven incident becomes an exercise in reconstructing an event after damage has already occurred.

Finally, enterprises should favor security architectures that can enforce policy consistently across networks, identities, endpoints, applications and AI interactions, rather than adding another isolated dashboard for each new AI use case.

Final thoughts

Fortinet’s Virtue AI acquisition reflects an important market reality: Agentic AI is likely to create more security work, not less. The enterprises that gain the most from autonomous AI will be those that pair experimentation with disciplined guardrails, continuous validation and clear accountability for agent actions.

AI may reshape the security market and force every incumbent to evolve, but it will not eliminate the need for security platforms. If anything, it raises the value of a platform capable of seeing, governing and responding across the increasingly connected environment in which AI agents operate.

Artificial intelligence is a double-edged sword when it comes to cybersecurity. It is giving defenders new ways to sift through vast amounts of telemetry, identify anomalous behavior and automate routine work. But it is also giving attackers the ability to discover vulnerabilities faster, build more convincing social engineering campaigns, and execute multistage intrusions at a scale that would have required large, well-resourced teams only a few years ago.

That shift is at the heart of “frontier AI,” a term that is increasingly important to security leaders. Frontier AI refers to the most capable, general-purpose AI models: systems able to reason through complex problems, analyze code and data, plan multistep tasks and, increasingly, invoke tools to act. Agentic AI is a particularly consequential evolution because it can plan, decide and act on a user’s behalf rather than simply generate an answer or summarize information.

For cybersecurity teams, the concern is not that every frontier model is inherently malicious. The issue is that these models can lower the cost, skill threshold and time required for adversaries to conduct sophisticated operations. Attackers can use AI to automate reconnaissance, identify exposed assets, analyze software for flaws, adapt phishing lures, write or refine exploit code, and coordinate activity across many targets simultaneously. Tasks that once unfolded serially and often required distinct specialists can now be compressed into a faster, more scalable workflow.

That compression changes the economics of defense. An organization that leaves a vulnerability unpatched for days or weeks has always carried risk. In a frontier AI environment, however, the window between vulnerability disclosure and exploit development with widespread targeting can be much shorter. The security challenge is no longer just finding weaknesses and responding to alerts. It is anticipating how a weakness, a misconfiguration, an identity control failure and a poorly segmented application environment can be combined into a viable attack path.

Security’s growing speed problem

Many enterprise security architectures were designed for an era when attacks moved comparatively slowly. Security teams collect alerts from endpoint, identity, network, cloud and application tools. Analysts investigate those alerts, determine priority, coordinate with infrastructure teams, and eventually adjust policy or remediate the issue. That process can work well when threats are isolated, predictable and slow-moving. It is far less effective against adaptive attacks that change tactics based on the environment they encounter.

Frontier AI amplifies several security challenges:

• Vulnerability discovery at machine speed: AI-assisted attackers can assess large numbers of internet-facing assets and software components faster than human-led teams. As a result, technical debt and forgotten exposures become more dangerous.
• Adaptive multistage attack chains: A successful intrusion is rarely a single event. Attackers may start with a phishing email, a stolen identity credential or an exposed service, then move laterally, escalate privileges and seek sensitive data. AI can help tailor each stage to real-time results.
• Security tool fragmentation: Enterprises often have extensive security stacks, yet their data and enforcement points are scattered. A cloud security tool may see one signal, an endpoint product another, and the network team a third. Fragmented visibility makes it difficult to understand the full attack path.
• Alert overload and delayed response: Security operations centers are already overwhelmed by telemetry. An AI-enabled adversary can act faster than it takes to investigate a single high-priority alert, making traditional detect-investigate-remediate workflows increasingly inadequate.
• Identity and access exposure: As more applications, APIs, SaaS services, and AI agents connect to corporate resources, identity becomes an even more central attack surface. Excessive permissions, compromised credentials, and poorly governed agent access can turn a modest foothold into a broader breach.

The key point is that frontier AI does not eliminate the need for security fundamentals. It raises the cost of getting those fundamentals wrong. The U.K. National Cyber Security Centre notes that frontier AI makes it easier, faster and cheaper to discover and exploit weaknesses, while emphasizing that strong cybersecurity basics remain the most effective foundation for resilience.

Cato’s case for agentic prevention

Cato Networks Ltd. is addressing this problem with Cato Agentic Threat Prevention, a new capability built into its cloud-native secure access service edge platform. The company’s thesis is similar to mine. If attackers can use AI to move at machine speed, enterprises need defensive AI systems that can do more than detect threats after they begin to unfold. It’s a matter of fighting fire with fire.

Cato Agentic Threat Prevention uses autonomous agents to predict likely attack paths within a specific customer environment and generate protections to stop attacks before they escalate. It combines network and security telemetry with customer activity and threat intelligence to model risk across users, applications, traffic patterns, assets and exposures.

This is an important distinction from conventional exposure-management or attack-path analysis products. Those tools can identify vulnerabilities, prioritize risks, or show potential paths through an environment. Cato is extending the concept from analysis to action. It aims to determine how an attacker could chain techniques, exploit control gaps or evade existing defenses, then enforce preventive controls tailored to that environment.

The company benefits from operating a converged network and security cloud. Because Cato’s platform unifies networking, security and access, it has visibility into more of the context needed to make a useful prediction. More importantly, Cato says protections can be enforced globally through its points of presence, avoiding the service chaining and enforcement gaps that can slow response times in a collection of disconnected tools.

Cato is pairing this prevention capability with its Agentic CVE Mitigation technology, which it says can autonomously assess and apply protection for newly disclosed vulnerabilities in as little as 45 minutes. The two capabilities address different but closely related problems: Common vulnerabilities and exposures mitigation narrows the exposure window after a new vulnerability is disclosed, while Agentic Threat Prevention focuses on predicting how an adversary may exploit a broader set of weaknesses.

The strategy reflects a broader shift in security operations. Detection and response will remain necessary, but they cannot be the sole line of defense against AI-assisted attacks. Security teams must increasingly use context-aware automation to reduce exposure and disrupt the likely path of an attack before the adversary reaches critical systems.

Recommendations for security leaders</4h>

Security professionals should view agentic defense as an enhancement to, not a substitute for, cybersecurity discipline. Five actions should be priorities:

1. Unify visibility across security and network domains. Attack-path prediction is only as good as the context it is based on. Organizations should reduce blind spots across identity, endpoints, cloud workloads, applications, branch locations and remote users.
2. Focus on attack paths, not on vulnerability counts. A long list of CVEs is not a risk strategy. Prioritize exposures based on reachability, privilege, asset criticality, compensating controls and an attacker’s likely ability to chain weaknesses.
3. Automate low-regret prevention actions. Establish guardrails for automated policy updates, isolation, access restrictions and virtual patching. Begin with controls that are reversible, well-tested and clearly owned, then expand automation as confidence grows.
4. Treat identity as part of the attack path. Enforce least privilege, multifactor authentication and continuous access evaluation. Inventory the permissions granted to AI agents, automation tools and service accounts with the same rigor applied to human identities.
5. Measure and track time to protection. Mean time to detect remains useful but is insufficient. Chief information security officers should track how long it takes from a new exposure or credible threat intelligence to effective prevention across the enterprise.

Final thoughts

Frontier AI is making cyberattacks faster, more adaptive and more accessible. The response cannot be to add another dashboard or generate more alerts. The security industry needs systems that can understand context, predict how attacks will progress and enforce defenses before an attack becomes a breach.

Cato’s Agentic Threat Prevention exemplifies that emerging model. Whether enterprises use Cato or another platform, the strategic lesson is that, in the age of AI-assisted attacks, prediction must be a core component of prevention.

Just prior to last month’s British Open, Cisco Systems Inc. announced a multiyear partnership with The R&A, a worldwide golf governing body. While there are many tech vendor–sports league marketing partnerships, this is much more than that. Cisco and The R&A are entering into a strategic infrastructure deal that should attract the attention of any IT executive looking to modernize operations across complex, distributed environments. It’s not just about powering The 2026 Open Championship; it’s about creating an AI-ready network architecture that spans tournaments and a new headquarters in St Andrews, offering lessons that translate directly beyond sports. At its core, the agreement designates Cisco as The R&A’s official network supplier for The Open, the AIG Women’s Open and other key championships, as well as for The R&A’s new global headquarters in St Andrews. This move aims to standardize on a single, secure, AI-ready platform across front-of-house fan experiences and back-office operations. Karen Lyttle (pictured, right), director of technology at The R&A, described Cisco as “our network and infrastructure partner” and said the formal partnership “will ultimately allow us to elevate that relationship and spend more time getting to know each other and understanding where there are real opportunities” across championships and daily operations. For executives, the lesson is that vendor relationships should extend beyond equipment procurement to co-innovation and joint opportunity discovery. Cisco is also positioning the deal as part of a broader global strategy. Gordon Thomson, Cisco’s Europe/Africa/Middle East leader, noted that after years of working with the US Golf Association, signing with The R&A was “a fundamental step in terms of us showing the world that we’re serious about the integration of talent and technology.” That framing places the partnership squarely within Cisco’s inclusive future narrative and its intent to use top-tier sports not only to demonstrate technology but also to broaden access and engagement.

“Pop‑up town”

For decision-makers, the most compelling part of this story is the operating environment. A championship golf venue is, effectively, a temporary town that must be built, secured and operated on a living landscape for a week — then taken down. That town hosts roughly 50,000 spectators per day, plus media, players, officials and staff. Unlike a fixed stadium, fan density shifts constantly, following star players from hole to hole. The result is a highly dynamic demand pattern; connectivity must be reliable everywhere, but capacity must surge when marquee groups draw thousands of followers to a single hole. Ed Green, Cisco’s Field chief technology officer, described this as a “shoal of fish” effect: The network must absorb sudden, localized spikes in usage as crowds cluster, check leaderboards and replays, share content and upload media. This mirrors many industries where people and devices move unpredictably, such as factory workers across the floor, clinicians between wards, guests throughout a large resort or students across a campus. Network and infrastructure planning can no longer assume static density; it must be designed for mobile hot spots. There’s also a shift in the traffic profile that matters for digital strategy. At The Open, Cisco is seeing more upload than download traffic as media and fans push content and devices flush cached data when they connect to strong Wi-Fi. That’s a forward-looking pattern that echoes what many enterprises are experiencing as sensors, cameras and employees continuously send data to the cloud. IT professionals should read this as a signal that uplink capacity, not just downlink streaming, is becoming a first-class business requirement.

Invisible but resilient

Brand and customer experience are central to The R&A’s mission. That extends from the visual purity of its championship venues to the quality of the fan experience. The partnership demonstrates that infrastructure can be both invisible to end users and highly resilient. Visually, The R&A and Cisco are careful about where and how they deploy hardware. Access points and cabling must be tucked into grandstands, hospitality structures and tee complexes so they don’t detract from the course. As Green put it, “the best networks are the ones where people don’t talk about them and don’t see them.” For executives responsible for high-end retail, hospitality, cultural venues, or public spaces, this is a useful model: Infrastructure can be designed to blend into the environment without compromising its capabilities. For resilience, Cisco is applying environmental hardening, such as weather-sealed access points and flexible mounting options, without creating entirely separate product lines. The goal is to make deployment in harsh or unconventional settings straightforward, so operational teams can focus on outcomes rather than bespoke engineering. That principle applies directly to industrial sites, logistics yards, and other outdoor or semi-permanent environments where downtime has real business impact.

Exporting the model

What makes the R&A–Cisco story particularly interesting is that the same architecture is being extended to a new global headquarters designed for modern, collaborative work. The R&A is consolidating seven older office locations into a purpose-built campus in St Andrews. Lyttle described this as a complete transformation in how the organization works: open-plan “neighborhoods,” a wide range of meeting spaces, work-anywhere zones and integrated security and collaboration technology from the fiber entry point through cameras and conference devices. The philosophy is “work anywhere,” with a secure, robust network underpinning every space and use case. This demonstrates how a single infrastructure stack can connect rugged field deployments with headquarters environments. The same AI-ready network that supports fans and media on a course can also support hybrid work, digital operations and collaboration in a corporate setting. This reduces complexity, simplifies governance and positions The R&A to roll out future applications, including AI-driven experiences, on a consistent backbone.

Takeaways beyond sports

Looking more broadly, there are four lessons IT professionals across other industries can draw from this partnership:

• Think in terms of ecosystems, not individual sites. The R&A and Cisco are designing a single, AI-ready architecture that spans championship venues and headquarters. In the AI era, it’s critical to plan infrastructure and security as a unified ecosystem covering plants, offices, stores and remote sites, rather than as isolated projects.
• Design for dynamic, mobile demand. The “shoal of fish” on a golf course is a proxy for shifting patterns across workplaces, campuses and public spaces. Capacity plans and investments should assume that people and devices will move and cluster, and that infrastructure must adapt in real time.
• Balance brand experience with operational robustness. The care taken to hide hardware on the course while delivering high-density connectivity shows you can respect aesthetics and still meet performance targets. IT leaders in consumer-facing industries should demand both: infrastructure that aligns with the brand and delivers quietly.
• Anchor AI and modernization in operational outcomes. Both The R&A and Cisco emphasize tangible outcomes: improved monitoring and alerting, smoother fan journeys, more efficient operations and a headquarters that enables new ways of working. Successful modernization will follow the same path — tying AI and new infrastructure to specific, measurable outcomes that matter to the business and its stakeholders.

In that sense, what’s happening at The Open is more than a sports technology story. It’s an example of how a heritage organization and a global technology company can co-design an infrastructure platform that honors tradition, elevates the experience, and prepares the business for a data- and AI-driven future. Any executive grappling with distributed operations, complex customer journeys, or hybrid work can look to this partnership as a roadmap worth studying.

Advanced Micro Devices Inc.’s Advancing AI 2026 keynote this week was Chief Executive Lisa Su’s bid to redefine the company from a “graphics processing unit alternative” to a full-stack artificial intelligence infrastructure vendor and to make this the year the central processing unit officially rebounds as a first-class AI platform.

The message from Su was ambitious and often compelling, but it also sharpened the competitive contrast with Nvidia Corp. and Intel Corp., raising as many questions as it answered.

Though the obvious theme of the event was that AI is moving from pilots to production, that has been the theme of every event I have attended this year. Beyond that, there were several other sub-themes. Here are the five most notable:

1. Helios and MI450: Finally a credible rack-scale alternative — with a catch

Su opened by turning Instinct MI450 and the Helios rack into a single, rack-scale product story aimed squarely at Nvidia’s system-level dominance. In AMD’s benchmarks, Helios delivers “an average of 10% to 15% more performance than the competition” at fixed rack power on “the highest throughput workloads” and “leading inference modes,” and she translated that into “up to 30% more tokens per dollar than the competition.” Though Nvidia has set the standard for the systems approach, if AMD can deliver on the savings it claims at comparable performance, it can use that to position itself as a credible alternative.

The partner lineup was the strongest evidence that these claims are real. Su said demand for Helios is “extremely strong… from the largest AI labs to hyperscalers,” and she highlighted OpenAI as “one of our deepest and earliest partners deploying Helios,” noting that joint engineering teams are already running GPT-class workloads. OpenAI’s infrastructure lead described AMD and OpenAI engineers “working side by side to optimize the software stack” and said they expect to deploy Helios “at massive scale, starting towards the end of this year, and then accelerating toward 2027.”

The catch is that “average of 10% to 15% more performance” and “30% more tokens per dollar” remain AMD-run numbers against unnamed “competition” and unspecified model mixes. Until cloud instance specs, public benchmarks and customer case studies show comparable gains in the wild, Helios is only a strong narrative and a promising design, not yet a proven market-share shift. Nvidia still owns software mindshare and the incumbent installed base; AMD must convert a handful of flagship design wins into a durable ecosystem.

2. CPU rebounds: Venice turns agentic AI into a three-tier compute story

One of the more interesting and underappreciated parts of the keynote was AMD’s aggressive effort to reset the CPU narrative in AI. For the past few years, CPUs have been cast as glorified I/O controllers in GPU boxes. Su pushed back hard on that, arguing that AI infrastructure is splitting into three CPU roles:

• GPU servers where “the CPU’s job is basically to drive the GPUs.”
• Dense “agent servers or what we call agent sandboxes” where “the priority is actually density and the highest-performing cores per watt to run thousands of agents at once.”
• Traditional genera-purpose servers where it’s “all about efficiency” for databases, data services and enterprise apps.

Venice, AMD’s new Epyc family on Zen 6 and TSMC ‘s two-nanometer process, is the company’s attempt to own all three tiers. Su called it “one of the largest generational gains in the history of Epyc,” claiming “up to 1.8 times more performance than Turin” and up to 512 threads per socket. She then broke Venice into a family: Venice HF for GPU host nodes at up to 5 GHz; a 256-core Venice with “the highest compute density in the industry” for agent sandboxes; and a 128-core part tuned for enterprise performance per dollar, plus variants like Venice X and Verano for HPC and AI host interconnect.

Her punchiest line of the keynote, “Epyc is the only CPU portfolio that leads across all use cases,” was clearly aimed at both Intel’s latest Xeons and the growing crop of Arm server CPUs. She claimed that Venice delivers more than twice the agents per watt for agent sandboxes versus x86 and up to 3.3 times more performance per watt at the rack than unnamed Arm competitors in a 100-kilowatt rack. She also pointed out that x86 software compatibility still matters when you’re “adding thousands of employees to your enterprise” in the form of agents.

The critical angle: AMD is right that agentic AI gives CPUs a second life, but this space is now intensely contested. Intel is not standing still on core count, memory bandwidth, or AI offload, and Arm vendors are pushing hard on efficiency and custom silicon for cloud providers. Venice’s generational gains are impressive on paper, but AMD still has to win OEM designs, cloud footprints and independent software vendor certifications at scale, all while customers are also considering Arm and custom accelerators tailored to their agent workloads.

3. Jeetu Patel and Santosh Janardhan: CPUs and GPUs become ‘conjoined things’

If Su provided the product narrative, Jeetu Patel from Cisco Systems Inc. and Santosh Janardhan from Meta Platforms Inc. provided the architectural reality check, and they largely backed AMD’s thesis that CPUs are back in the spotlight.

Patel argued that “it’s not just a GPU game anymore. It’s CPUs and GPUs. If anything, I think CPUs are becoming at least as important, if not more.” His view is grounded in enterprise knowledge, which Cisco has more of than all the other companies on stage combined. Long-running agent workflows, tools, databases and networks all must be orchestrated around frontier models.

He framed compute as a heterogeneous fabric where “you must think about CPUs and GPUs as conjoined things. You hand off workloads depending on other workloads. You employ the right hardware.” That is exactly the kind of messaging AMD needs enterprise CIOs to hear to become a more strategic vendor.

Meta’s Janardhan pushed the same idea at the data center scale. For him, the AI problem is no longer about squeezing every percentage point out of a single chip; it’s about “the whole data center as one integrated system — servers, hardware, networking, cooling, power.” He noted that data centers and silicon “take years to build,” and argued that the industry needs to be “sitting down in a room, co-designing today for what we need to deploy in 2027 and 2028.” In other words, CPUs, GPUs, memory, networking and power are now co-equal design levers.

It was great to see these companies on stage with Su, because they aren’t second-tier logos; they’re two of the largest AI and networking players on the planet, validating AMD as a co-design partner, not a backup supplier. For the established players, that’s a warning that their traditional lock-in at the CPU and GPU levels may not survive an era in which hyperscalers want multi-vendor, co-designed systems to manage risk, cost and power.

4. ROCm.AI and Hyperloom: AMD tries to leapfrog on AI-native tooling

On software, AMD went straight at its perceived weakness, the software gap versus Nvidia’s stack, with a different approach: Let AI write and optimize more of the GPU code.

Senior Vice President Vamsi Boppana introduced ROCm.AI as “an agentic AI platform that brings the capabilities of AI-assisted GPU programming to developers.” Built on AMD’s open software stack, ROCm.AI adds an AI optimization layer called Hyperloom that can “analyze the workload, tune configurations, select and tune kernels, adjust parallelism strategies and iterate towards performance goals.”

Internally, AMD has already pushed “a suite of 14,000 models through Hyperloom,” generating optimizations that would have been “impossible even with a large team of engineers before.” In a live example, an AI agent targeting MiniMax M3 on MI355s with VLLM identified an opportunity to write a more optimized Mixture-of-Experts GEMM kernel, delivering a 38 percent improvement in tokens per second. Engineers onstage admitted that some AI-generated kernels are “shockingly good, sometimes better than the most manually tuned versions.”

AMD’s competitive bet is that an open stack plus AI-assisted optimization can close the software-ecosystem gap with Nvidia faster than traditional hand-tuning ever could. The risk is that this is still early-stage technology: automatic code generation must be safe, reproducible and debuggable at scale, and customers will want to see real-world workloads, not staged demos. Nvidia will not sit still here either; it has every incentive to build its own AI-native tooling on top of CUDA and its closed ecosystem.

5. Co-designed, open systems: AMD plays the long game against lock-in

The final through line in Su’s keynote was an embrace of co-design and openness as AMD’s strategic wedge against incumbents. OpenAI’s infrastructure chief, Sachin Katti, described a future where AI becomes “a problem at a data center scale,” not just a “rack-level problem,” and stressed the need to co-design “from CPUs to GPUs to memory, networking, storage, power distribution and the cooling systems that go with it.”

Cerebras CEO Andrew Feldman, announcing a joint solution that marries AMD’s Helios racks with Cerebras’ wafer-scale engine, argued that customers who used to choose between “high throughput” and “extraordinary speed” can now get both — “five times the throughput while continuing to deliver this extraordinary speed” for ultra-low-latency inference.

Su tied these threads together by leaning hard into open software. Because AMD’s compiler stack and drivers are largely open, partners and even AI agents can see all the way down to the ISA. OpenAI’s Philippe Tillet credited that openness with enabling “very, very significant performance gains” and faster portability of GPT-class models to AMD hardware. In a world where “recursion,” that is AI helping design the next generation of AI systems, is becoming a reality, AMD is betting that openness will let it harness that flywheel more effectively than a closed stack.

The competitive implication is that while Nvidia still has the deepest, most entrenched software ecosystem, AMD aims to be more open to attract a broader base of partners and developers. Intel has the x86 incumbency but has struggled to convert it into AI mindshare. AMD is positioning itself as the third pole: not just cheaper GPUs, but an open, co-designed CPU-plus-GPU platform tuned for the agentic, data-center-as-a-system future.

Whether that bet pays off will depend on execution in silicon delivery, software quality, and ecosystem traction, not just on keynotes. But if Advancing AI 2026 is any indication, AMD has stopped talking like a fast follower and started acting like a company that expects to set the terms of the AI infrastructure debate.

Final thoughts

AMD used Advancing AI 2026 to make a statement: It is no longer content to be “the alternative” to Nvidia, and it now has credible hardware at the rack level, a resurgent CPU portfolio tuned for agentic AI, and a software story that leans into AI-assisted optimization and open ecosystems. Helios plus MI450 gives hyperscalers and frontier labs a rack-scale option that can be argued on performance per watt and tokens per dollar, while the Venice Epyc family targets the emerging three-tier AI compute stack, comprised of GPU hosts, dense agent sandboxes, and general-purpose enterprise, at a moment when CPUs are quietly becoming the control plane and workhorse for agentic workflows.

On software, AMD has stopped pretending it can out-CUDA Nvidia head-on; instead, it is betting that an open stack plus AI-native tools like ROCm.AI and Hyperloom can compress the time it takes to get real workloads performant on its silicon.

The uncomfortable reality for AMD is that none of this makes Nvidia “trail” in the broader AI race. Nvidia still owns the dominant software ecosystem, the bulk of the deployed AI accelerator footprint, and a deeply integrated toolchain. In most large AI shops, AMD remains the second platform to be qualified rather than the assumed default.

Openness is AMD’s best strategic tool against that incumbency, lowering switching costs and inviting partners and, eventually, AI agents to co-design and tune all the way down to the instruction set, which differs from Nvidia’s model. The question is whether that openness, combined with competitive performance and TCO, is enough to move AMD from “necessary diversification” to “first choice” for a meaningful share of new deployments.

Right now, openness looks necessary but not sufficient, and AMD will need multiple cycles of flawless execution on silicon, software and the ecosystem to turn this impressive keynote positioning into durable market power.

By extending Security Hub to Microsoft Azure and adding artificial intelligence-specific protections, Amazon Web Services Inc. is positioning its security stack as the foundation for securing enterprise AI at scale.

Amazon Web Services’ latest Security Hub updates announced earlier this month acknowledge two realities its customers already live with every day: AI is now the fastest-growing attack surface in the enterprise, and almost every serious AI initiative is, or soon will be, multicloud. The result is a Security Hub release that extends to Microsoft Azure and delves into AI workloads on AWS, with the clear goal of turning the service into a unified control plane for cloud and AI security.

Security Hub, which began as a cloud security posture management or CSPM tool for AWS, has steadily evolved into the aggregation point for detections across native services and partners. The new release expands that role. I recently discussed the update with Michael Fuller, director of security services at AWS, who explained, “Customers have gone multicloud, and have been pretty loud and clear: Don’t make me buy a different tool when I already have Security Hub.”

What AWS announced

The Security Hub announcement bundles four main capabilities:

• Azure resource monitoring
• GuardDuty AI Protection for AI workloads
• GuardDuty AI-powered investigations (preview)
• Security Hub AI inventory

Regarding Azure, Security Hub can now discover and monitor Azure Virtual Machines, container images in Azure Container Registry, Function Apps and identities. It evaluates those resources for misconfigurations, internet exposure and software vulnerabilities, including checks aligned with the CIS Microsoft Azure Foundations Benchmark. The findings appear side by side with AWS findings, using the same format, prioritization logic and automation workflows. That means a security engineer can finally see AWS and Azure posture through a single lens instead of juggling separate consoles and rule sets.

Under the covers, AWS is leveraging one of its core security primitives: AWS Config. It has quietly gone multicloud and now powers near-real-time, change-triggered evaluation of Azure resources, rather than the 24-hour polling cycle typical of many third-party CSPM tools. Fuller emphasized that customers don’t have to wire any of this together themselves. “We’ve replicated the one-click experience customers love from AWS,” he said. “You make a tenant-level decision in Azure, give us read-only access once, and we do the rest under the covers.”

Pricing has been deliberately simplified. Security Hub consolidates what used to be multiple line items into a per-resource-per-month model and charges only for four resource types — virtual machiness, container images, functions and identities — across both AWS and Azure. Fuller framed the philosophy as making Security Hub a “no-brainer” alternative to building your own detection and posture pipeline from cloud primitives, rather than a race to the absolute lowest line-item price.

AI workload protection: The strategic move

In my opinion, the most significant part of this release is squarely in the AI stack. AWS is introducing GuardDuty AI Protection, AI-powered investigations and a new AI inventory within Security Hub to help enterprises secure model-driven workloads on services such as Amazon Bedrock and SageMaker.

GuardDuty AI Protection targets threats unique to AI services, including anomalous model invocations that signal abuse of an inference endpoint, prompt-injection attempts (via integration with Bedrock Guardrails),and what AWS calls “cost harvesting” attacks. Cost harvesting is essentially the AI-era version of crypto-mining abuse: an attacker gains access to credentials, spins up large volumes of inference and converts your AI budget into free compute. Fuller noted that this is not an entirely new category so much as a new expression of an old pattern. From the earliest days of the cloud, compromised accounts were abused to run unauthorized workloads; AI inference is simply the latest, and sometimes most lucrative, target.

GuardDuty AI-powered investigations, currently in preview, directly address the number one pain point security teams cite: alert fatigue. When GuardDuty raises a finding, the investigation engine automatically retrieves related activity, context on the affected resources and relevant threat intelligence. It then generates a confidence score, MITRE ATT&CK mapping, supporting evidence and recommended actions — such as whether to suppress, contain or fully investigate the issue. Analysts still make the final call, but they no longer must manually stitch together logs and events for every alert.

Critically, AWS isn’t throwing a generic large language model at the problem. Fuller noted that GuardDuty has been live since 2017, and AWS has nearly a decade of experience helping customers and its own internal teams triage findings. That history has been used to teach the AI investigator which questions to ask and what context to look for, whether that’s a Palo Alto Networks Inc. firewall behind an EC2 instance or traffic routed through a customer’s on-premises data center.

The new AI inventory in Security Hub ties everything together by providing a single view of AI assets — Bedrock models, SageMaker endpoints and agents built with Amazon AgentCore — and mapping them to the underlying compute, networking, IAM and data stores they depend on. That makes it much easier for security teams to answer basic but critical questions: What AI workloads do we actually have? Where are they running? If one is compromised, which systems and data are in the blast radius?

How this helps accelerate enterprise AI

Most organizations aren’t slowing AI pilots because they lack models or GPUs; they’re slowing down because risk and compliance teams don’t have guardrails they trust. This release is AWS’ attempt to move “secure by default” for AI workloads from slide decks to something closer to operational reality.

There are a few concrete accelerators here.

• A unified view across clouds. If your security operations can see misconfigurations and vulnerabilities across AWS and Azure in a single console, using a single set of workflows, it becomes much easier to sign off on hybrid AI architectures spanning multiple clouds. That directly shortens the approval loop for new AI services.
• AI-aware threat detection. Treating AI services as first-class citizens in GuardDuty prevents AI from becoming a blind spot. Anomalous model activity, cost harvesting and prompt injection are surfaced alongside more traditional network and endpoint threats, rather than in a separate, under-instrumented silo.
• Triage at AI speed. AI-powered investigations are less about replacing analysts and more about compressing the time it takes to distinguish real attacks from background noise. Fuller acknowledged that customers don’t want fewer alerts; they want all of them, but ranked and explained. The investigations preview is an early step in that direction.
• An integrated partner ecosystem. Security Hub Extended now brings 21 curated partner solutions across nine categories into a more unified experience, using OCSF to normalize findings. That gives enterprises a way to keep best-of-breed tools like CrowdStrike, Okta, Cyera and Splunk in the mix without owning all the integration work. As Fuller put it, customers are tired of the “false choice” between a single-vendor suite that isn’t best-of-breed and a best-of-breed stack they have to wire up themselves.

For enterprises trying to industrialize AI, that integration layer is increasingly where projects stall. When cloud, identity, endpoint and AI signals all land in a common findings plane, it becomes much easier to make AI a first-class citizen of existing security operations rather than an exception process.

What IT and security professionals should do next

For IT professionals and security architects, the question isn’t whether this is interesting — it’s what to do with it over the next 12 to 18 months.

First, if you’re already invested in GuardDuty and Security Hub, treat AI protections and the AI inventory as table stakes for any Bedrock or SageMaker deployment. Make enabling GuardDuty AI Protection and registering services in the AI inventory policy requirements before AI workloads move into production. That gives risk and audit teams a consistent framework for assessing AI risk with tools they already know.

Second, if you’re an AWS-first shop with a growing Azure footprint, you should pilot Security Hub’s Azure monitoring on a representative subscription and compare it with your existing CSPM tool, especially regarding latency and operational friction. You don’t need to rip and replace overnight, but you should have a clear target state for where cloud posture management lives. Increasingly, it will be difficult to justify a fractured approach that looks different in each cloud.

Third, use the AI-powered investigations preview to address alert fatigue in a measured way. Start by enrolling a subset of GuardDuty findings, track mean time to triage before and after, and use the confidence scores and recommendations to refine your suppression rules and playbooks — not to bypass analyst judgment. This is a low-risk way to get your SOC comfortable with AI “in the loop” while keeping humans firmly on the hook for final decisions.

Finally, zoom out to assess your longer-term SIEM and data strategy. AWS’s direction of travel — consolidated findings in Security Hub, shared data stores for security and observability, and partner tools querying common data — chips away at the traditional model in which every security tool insists on owning all of your raw data. You don’t need to blow up your SIEM today, but it’s worth aligning your next renewal with a plan for how Security Hub, GuardDuty and your partner tools will share data, rather than each demanding its own silo.

Final thoughts

AWS is doing what customers have been asking for: using its weight to collapse security seams across clouds and between traditional workloads and AI. For organizations trying to move faster on AI without losing control of risk, Security Hub’s new multicloud and AI-focused capabilities are worth more than a casual look — this is a blueprint for how cloud security will need to work in the AI era.

On the surface, this week’s Vera Rubin launch is another major platform moment for Nvidia Corp., as the company maintains a steady drumbeat of artificial intelligence infrastructure innovation.

Nvidia is positioning Vera Rubin as a full-stack system designed to improve performance per watt and reduce token costs, with production ramping across a broad global partner base, including cloud and AI infrastructure providers. Nvidia says the platform spans seven co-designed chips, integrates new networking and is already being deployed by partners such as CoreWeave, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure.

However, the bigger story isn’t that Nvidia launched another AI system, since that’s nothing new for the king of AI. Rather, it’s that the company is making an aggressive case that the AI era, especially the rise of agentic AI, requires rethinking the central processing unit, the network and the system architecture as a single, interdependent design problem rather than a pile of best-of-breed parts. That is why Vera (pictured) matters.

From cloud economics to agentic bottlenecks

For years, the CPU roadmap in the data center was largely shaped by cloud economics. Hyperscalers wanted more cores, higher throughput and lower costs, which rewarded chiplet-heavy designs optimized for scale-out efficiency. Nvidia’s analyst briefing, and another for reporters last week, framed that era as one in which core counts grew roughly nine time and single-thread performance gains doubled, leaving the market with processors well-suited for classic cloud workloads but less suited to the latency-sensitive, branch-heavy behavior of AI agents. As Nvidia’s Hannah Coutand explained in the briefing, “AI is asking for a new CPU,” because “agentic AI is putting CPU back on the critical path.”

That framing is important. In an agentic system, the graphics processing unit performs the reasoning, while the CPU is constantly in the loop, handling tool calls, code execution, queries, orchestration and data handling between reasoning steps. These workloads create continuous loops of reasoning, acting, observing and evaluating, which put simultaneous pressure on single-thread performance, memory bandwidth, latency and scaling efficiency. In other words, throwing more generic cores at the problem is not enough if the real bottleneck is the sequential work between GPU inference passes.

Inside Vera: A CPU designed for the AI loop

This is the central rationale behind Vera’s new CPU architecture. Rather than extending the conventional server CPU playbook, Nvidia built a custom Arm-based processor around its new Olympus core. Vera features 88 custom cores, 176 hardware threads, up to 1.2 terabits per second of LPDDR5X memory bandwidth, 164 megabytes of unified L3 cache, and up to 1.8 TB/s of coherent CPU-GPU bandwidth via NVLink-C2C.

Nvidia says the design target is what it calls the “max single-threaded CPU at scale,” meaning a processor that preserves strong per-core responsiveness while still scaling across highly concurrent agent workloads. The analyst briefing made the design intent unusually clear. Coutand stated that Nvidia “didn’t set out to go win CPUs,” but instead recognized that “the CPU was becoming a bottleneck in the AI factory” and designed a better processor to improve “AI factory economics.” That is classic Nvidia strategy: Start with the system bottleneck, then build the silicon needed to remove it.

Nvidia’s Ian Finder provided the technical details behind that claim. He described Olympus as a ground-up custom core featuring a 10-wide decode engine, aggressive reordering logic and a graph prefetcher tuned for pointer-chasing patterns common in compilers, graph structures and agent runtimes. His point was that CPU performance in this new era is less about chasing clock speeds and more about increasing the amount of useful work each cycle can do. That emphasis on instructions per cycle, branch handling and memory behavior is exactly what you would expect if the target is agent orchestration rather than old-school enterprise middleware.

Fabric and memory: Making data movement part of compute

Just as important, Vera is a reminder that in AI infrastructure, the network is no longer a peripheral technology but part of the compute architecture.

Inside the CPU, Nvidia’s second-generation Scalable Coherency Fabric serves as the data-movement backbone, linking cores, caches, LPDDR5X controllers, I/O and NVLink-C2C interfaces with multiterabyte-per-second bandwidth. Nvidia contrasts this monolithic fabric with chiplet-based designs that incur a “chiplet tax” in the form of higher latency and lower effective bandwidth as traffic crosses die boundaries. For agentic workloads, which are sensitive to loaded latency and cross-core data sharing, those differences translate directly into GPU utilization and end-to-end responsiveness.

The memory subsystem follows the same philosophy. By pairing LPDDR5X with an enterprise-ready module form factor, Vera aims to deliver high bandwidth per core and better bandwidth-per-watt than conventional DDR-based servers. In an AI factory with thousands of deployed servers, shaving tens of watts from the CPU-plus-memory envelope while increasing bandwidth frees more of the power budget for GPUs and high-speed networking.

Beyond the rack: The role of Spectrum-X

Outside the CPU, the platform extends to the rack and the cluster. Within AI infrastructure there is a significant distinction between scale-up and scale-out networking. NVLink connects GPUs within a rack, enabling them to act as a unified accelerator with all-to-all bandwidth and in-network compute. Spectrum-X Ethernet provides the scale-out fabric that ties those racks together across the AI factory.

Spectrum-X is more strategically important than many realize. In traditional enterprise infrastructure, Ethernet can be treated as a largely modular layer. In AI factories, the network directly affects token throughput, latency, utilization and ultimately economics. If mixture-of-experts models and agentic systems create much heavier east-west traffic and more distributed coordination, generic Ethernet becomes a tax on the entire system.

Nvidia’s answer is a purpose-built Ethernet stack: 102.4T Spectrum-6 switches, 1.6T ConnectX-9 SuperNICs, adaptive routing, congestion control, telemetry and open software, all tuned for RDMA and AI traffic patterns. The goal is to make Ethernet behave more like an AI-specific fabric while preserving operational familiarity, so that scale-out networking enhances AI factory performance rather than undermining it.

Extreme co-design as a competitive weapon

This brings us to Nvidia’s “extreme co-design.” The company says that Vera Rubin NVL72, the Vera CPU rack, BlueField-4 infrastructure processors, Spectrum-6 switching, and the rest of the platform were engineered as a single system rather than assembled from separate off-the-shelf products.

With agentic AI, infrastructure services such as networking, storage, telemetry, security and context handling are now part of the inference pipeline itself. That means CPUs, GPUs, DPUs and switches need to be tuned together to keep expensive accelerators fed and productive without burning host CPU cycles on infrastructure work.

Most semiconductor vendors can compete credibly in one layer of the stack; a few can reach two. NVIDIA now has meaningful assets across GPUs, CPUs, scale-up networking, scale-out networking, DPUs, interconnect software and system design. That breadth lets it optimize for delivered AI output — tokens per watt, cost per token and usable throughput — not just component specs.

Nvidia’s next share gain story: CPUs

The most interesting industry implication of Vera is that CPUs may become Nvidia’s next share-gain story. Nvidia is not trying to displace x86 across every general-purpose data center workload. It doesn’t need to. The company’s own sizing suggests a large incremental CPU opportunity tied specifically to AI-driven workloads and infrastructure patterns. If the CPU’s role in the AI factory is increasingly to orchestrate agents, feed GPUs, manage memory movement and support low-latency tool execution, then Nvidia can leverage its GPU dominance to pull its own CPU into the design.

This is the same playbook the company has used elsewhere: win the control point, then expand adjacencies. Because Nvidia already owns the strategic budget line in AI infrastructure through GPUs, it is uniquely positioned to define what the surrounding CPU, network and data processing unit should look like. Vera’s tight integration with NVLink-C2C, BlueField and Spectrum-X means buyers considering Rubin-class systems are not evaluating the CPU in isolation. They are evaluating a full AI factory architecture.

In the near term, that will matter most for AI clouds, hyperscalers and large model builders with agentic or reinforcement-learning-heavy workloads. Over time, though, the definition of a “good” data center CPU may shift more broadly. If Nvidia is right, the next important CPU category will not be the cheapest cloud workhorse or the highest-core-count generalist. It will be the processor that best removes friction from the AI loop.

Final thoughts

The fundamental tenet of my research has always been that rapid market share shifts occur when markets transition, and the CPU industry hasn’t seen a significant transition in a long time. But this is the AI era, and it’s seemingly redefining all industries.

Nvidia is making the case that the future CPU is no longer a standalone component decision. It is a systems decision, tightly coupled to GPUs, memory, networking and infrastructure processors — and Nvidia intends to own as much of that system as possible. This week, AMD is holding its own “Advancing AI” summit, and we should get a good look at how it plans to address the challenges Nvidia laid out above.

Neocloud provider QumulusAI said today that it will starting trading Thursday as a publicly traded company on Nasdaq under the ticker symbol QMLS via a direct listing.

For those unfamiliar with the process, the typical initial public offering takes time and requires an investment banker, whereas a direct listing does not create new shares. Instead, existing shareholders sell their shares to the public without an underwriter.

IPOs are ideally suited for companies that need to raise capital, while the speed of a direct listing is better for highly liquid companies that have sufficient cash on hand but want to provide an easy way for investors or employees to turn shares into cash.

Though the QumulusAI move is a financial transaction, there’s a broader story. The neocloud model, artificial intelligence-first infrastructure built around graphics processing units and power availability rather than generic compute, is maturing into a distinct layer of the enterprise stack. For information technology leaders, the story isn’t about a listing but about the kind of cloud you’ll need to put AI into operation over the next three to five years.

Unlike hyperscalers that offer a broad portfolio of services, neoclouds, such as QumulusAI, are explicitly focused on the infrastructure that powers AI in the enterprise. The company’s value proposition is to bring high-end GPU capacity online in months rather than years, and to do so where there is real, available power. In a world where many enterprises can get all the AI tools they want but struggle to secure predictable capacity at large scale, the timing of the direct listing gives the company access to more capital to move faster.

A neocloud built for the AI bottleneck

The current AI wave has made one reality painfully clear: The limiting factor isn’t demand but infrastructure. Hyperscalers are pouring hundreds of billions of dollars into AI-related capital spending, yet customers still complain about limited access to the latest Nvidia chips, long lead times and opaque capacity planning. At the same time, utilities and regulators warn that data center growth is outpacing available grid capacity in several key markets.

QumulusAI sits in that gap. The company has evolved from a crypto-infrastructure heritage into a GPU-centric cloud designed for high-performance AI workloads. Instead of committing to mega-campuses that take years to bring online, QumulusAI leans on a mix of existing colocation facilities and modular, roughly 50-megawatt-class data center footprints. That approach allows it to deploy GPUs on a quarterly cadence and turn capital into billable infrastructure much faster than with greenfield hyperscale projects.

On the hardware side, QumulusAI is closely aligned with the AI ecosystem enterprises already trust. It deploys the latest Nvidia GPU generations — Hopper and Blackwell — alongside familiar data center brands for servers, storage and networking. The company doesn’t try to build its own AI framework or MLOps stack; instead, it focuses on delivering reliable, high-performance infrastructure that integrates with the platforms customers already use. That’s a notable point of differentiation from some AI-first clouds that blur the line between infrastructure and platform.

Why go public now?

The obvious question is why a company at this stage of its evolution chooses to go public rather than raise another round of private capital. For QumulusAI, there are three overlapping answers: capital, credibility and timing.

First, the model is capital-intensive by design. Scaling from a few hundred to thousands — and then to tens of thousands — of GPUs requires consistent access to financing for both hardware and power. QumulusAI has been methodical in building a capital stack that doesn’t rely entirely on dilutive equity. It relies on asset-backed convertible notes, equipment leases tied to specific GPU clusters, and customer prepayments that fund part of each deployment upfront.

Going public doesn’t replace that structure; instead, it adds optionality. A publicly traded equity currency gives the company more flexibility in future financings, partnerships and potential acquisitions without having to renegotiate its entire balance sheet.

Second, public-company status matters to the customers QumulusAI wants to serve. Multiyear, take-or-pay infrastructure contracts are no longer the exclusive domain of hyperscalers and colos. As enterprises and AI platforms commit to three-year GPU deals for training and inference, they want the governance, transparency and durability signals that come with a public listing. Audited financials, an independent board, detailed risk disclosures and capital structure visibility make it easier for procurement and risk teams to justify signing with a neocloud that isn’t a household name — yet.

Third, there is a genuine “right now” window in AI infrastructure. The first phase of the current cycle was defined by scarcity: Whoever could get H100s first won. The next phase will be defined by scale, utilization and power. QumulusAI is already showing the kind of trajectory you’d expect from a company trying to win that race. It has materially expanded its deployed GPU base over the last year and locked in a meaningful book of forward-looking, multiyear revenue through signed contracts. Early revenue growth numbers, while still off a relatively small base, show that its pivot from crypto to AI compute is working.

Going public while that growth curve is steep lets QumulusAI invest ahead of demand, while the market is still repricing AI infrastructure as a strategic asset. Waiting another two or three years would risk ceding share to better-capitalized rivals or getting caught in a potential cooling of AI hype that could make financing large infrastructure bets harder.

Neocloud differentiation: GPUs, power and geography

The neocloud business is becoming crowded, with several well-funded players positioning themselves as AI-first alternatives to general-purpose clouds. They share characteristics such as next-generation GPUs, highly tuned networking and storage, and a focus on AI and machine learning workloads, but they don’t all look the same.

QumulusAI’s differentiation lines up around three themes:

• Infrastructure, not platforms. QumulusAI is focused on infrastructure. It doesn’t claim to be the place to build, fine-tune and serve models end-to-end under a single proprietary interface. Instead, it offers bare-metal and virtualized GPU clusters, exposed through the control surfaces infrastructure teams expect: Kubernetes integration, reserved clusters and on-demand pools. That makes it appealing to enterprises and AI platforms that already have their own software stack and simply need predictable, high-performance capacity.
• Time-to-capacity is a core metric. The company’s mantra of bringing GPU capacity online “in months, not years” is more than a tagline. By targeting smaller, geographically distributed sites, QumulusAI can often avoid the longest queues for power and permits that plague mega-campus projects. Faster deployment cycles also translate into faster capital turns: hardware starts generating revenue sooner, allowing the company to reinvest that cash into the next wave of sites and GPUs.
• “Pockets of power” as a strategy. The constraint in AI infrastructure is increasingly electricity, not floor space. QumulusAI treats the hunt for available power as a first-class problem, working with utilities, colocation partners and regional stakeholders to identify locations where it can secure megawatts of capacity without waiting half a decade for grid upgrades. That opens up markets where hyperscale players might not bother to build, but where regional enterprises, AI startups and platform partners still need high-end GPU capacity.

Behind those differentiators is a go-to-market approach that blends direct enterprise relationships with channel-driven demand via AI platforms and marketplaces. Multiyear, take-or-pay agreements with AI inference platforms provide QumulusAI with both revenue visibility and utilization assurance, while marketplace partnerships help it backfill demand across a broader customer base. The result is a model that aims to solve both sides of the AI infrastructure equation: securing scarce GPUs and power on one side and keeping utilization high on the other.

Advice for IT leaders

For technology leaders, the rise of QumulusAI and its peers doesn’t mean you should abandon hyperscalers. It does mean you should start thinking about AI capacity in portfolio terms and ask sharper questions about where different workloads belong.

A few practical recommendations:

• Segment your AI workloads by capacity profile. Frontier model training, bursty experimentation and steady-state inference behave differently. Hyperscalers will continue to dominate elastic, spiky workloads and tightly integrated services up the stack. Neoclouds like QumulusAI are more interesting where you have stable, high-duty-cycle GPU demand — think production inference, long-running fine-tunes, or internal platforms that serve multiple business units — and where reserved capacity with clear economics matters more than access to the widest catalog of services.
• Make power and geography part of your RFP. When evaluating infrastructure providers for AI, don’t stop at GPU SKUs and hourly rates. Ask exactly where the clusters will be located, what the power situation looks like at each site, and how that aligns with your latency, data residency and resilience requirements. Providers that can show a pipeline of sites and power arrangements, rather than a single flagship campus, may be better aligned with distributed AI use cases.
• Probe utilization and contract structure. Take-or-pay contracts with multiyear terms can be powerful tools for cost predictability — but only if you can keep the GPUs busy. When you talk to QumulusAI and other neoclouds, ask how they help your teams drive utilization: what telemetry they expose, how they integrate with your orchestration and MLOps stack, and what options you must shift workloads across clusters or sites as your portfolio evolves.
• Treat neoclouds as strategic partners, not just vendors. Companies such as QumulusAI are still early in their journey, so their product roadmaps and site strategies are more adaptable than those of hyperscalers. If you have a clear view of your AI roadmap, you can shape where and how these providers build capacity — potentially even co-designing sites or contractual structures that align more closely with your needs.

QumulusAI’s public listing highlights a broader trend: AI is prompting enterprises to rethink their infrastructure, and a new class of cloud providers is emerging to meet those needs. Whether QumulusAI ultimately becomes a category leader or a specialized complement, its Nasdaq debut underscores a shift CIOs can’t ignore: The cloud for AI will be as much about GPUs and gigawatts as about APIs and services.

There remains a common misconception that Zoom is a meeting app. Though Zoom Communications Inc. arguably does meetings better than anyone, it has evolved into a broad work platform. Mount Sinai Health System in New York treats Zoom as core infrastructure.

Zoom recently held a webinar with Mount Sinai to discuss the deployment. What I found interesting is that, despite the rapid growth of cloud communications across most industries, healthcare has remained a holdout thanks to the critical nature of communication in that industry. I recently had a conversation with a U.K. healthcare organization, where the chief information officer told me he needs to ensure clinicians can make calls regardless of any macro issue. The Mount Sinai deployment can serve as a proof point for Zoom and its cloud peers in mission-critical environments.

Over the past two years, Mount Sinai has rearchitected its use of Zoom across telehealth, clinical voice, nursing mobility, education and artificial intelligence-assisted productivity — effectively turning the platform into a collaboration fabric for care delivery and operations.

“We focus on providing exceptional medical care to our communities,” said Rudra Hirlau, head of Workplace Technology at Mount Sinai. “For us, that starts with communication.” On that premise, Zoom evolved from a video client into a core enabler of how the health system delivers care.

Adding to the story is that the design decisions Mount Sinai made under healthcare constraints — Epic integration, HIPAA alignment and clinician workflows — map directly onto the challenges CIOs and chief technology officers are grappling with across industries.

Healthcare: Objectives before features

To its credit, the Workplace Technology team at Mount Sinai didn’t approach the project with a feature checklist. Instead, it started with outcomes. “Any project or initiative we embark on, we start by looking at our business objectives and the desired outcomes,” Hirlau explained.

Before any new Zoom initiative moved forward, they anchored it in four objectives: simplify and transform telecommunications, increase “impact velocity” for new capabilities, optimize costs (including staff time), and improve the patient and clinician experience by keeping technology simple and friction-free.

Because the health system was already a long-time Zoom Meetings and Webinar customer, the information technology organization could rely on existing familiarity. “We were already a user of Zoom, and the ease of use with the platform was already there among our colleagues,” Hirlau said. “So, we wanted to leverage that as we tried to simplify and move toward a single platform.”

That discipline translates directly to other sectors. Whether you’re running a global contact center or a distributed factory network, the pattern remains the same: define the business objectives first, then work backward to platform and architecture choices. Collaboration tooling is no longer just about audio and video; it’s about impact velocity and the frontline experience.

From telehealth to Sinai Phone

On the clinical side, Mount Sinai integrated Zoom Meetings with Epic to enable telehealth across primary care, ambulatory clinics and urgent care. Epic remains the system of record, but the “front door” experience runs through Zoom, which proved especially important for adoption and completion rates during the health system’s switch to a telehealth platform. Heidi West of Zoom recalled that telehealth completion rates improved to the 80% to 90% range almost immediately.

The bigger story is Zoom Phone. Mount Sinai made a deliberate choice not to treat Zoom Phone as a one-for-one PBX swap. In the Patient Transfer Center or PTC, which coordinates patient movement and access to specialty care across the system, Zoom Phone now supports up to 10 simultaneous calls per user, recording for continuous improvement, and highly customized queues tailored to clinical workflows.

“This was our first major Zoom Phone rollout, and it was by no means simple or easy,” Hirlau said. “But at the end, this was a very successful rollout. We were able to decrease our wait times, reduce abandonment rates, make callers happier and overall improve the time to care.” The PTC’s 50 users now handle around 40,000 calls per month on Zoom Phone, with trouble tickets from the legacy system dramatically reduced.

The “Sinai Phone” initiative goes further. Every nurse receives a mobile device loaded with key clinical applications and the Zoom client, anchored by click-to-dial integration with Epic. “The solution we came up with was a mobile device for each nurse that’s loaded with all the key clinical applications they need on a day-to-day basis, and on top of that include the Zoom application,” Hirlau said. Nurses can tap a number in the EMR to place a call over Zoom Phone and receive calls directly on their mobile devices. Adoption has been strong enough that Mount Sinai plans to double the deployment by year’s end, with nursing leaders publicly calling the change “life-changing” for their workflows.

Strip away the clinical language, and this looks a lot like what banks want in branch operations, what utilities want in field service, and what manufacturers want on the plant floor: programmable voice integrated directly into systems of record and frontline workflows, not a standalone system sitting off to the side.

Ecosystem, not ‘petting zoo’

Mount Sinai is also a Microsoft shop, but it didn’t turn the Zoom-versus-Teams debate into a religious war. Instead, the Workplace Technology team talks about avoiding a “petting zoo” of overlapping applications and about designing ecosystems.

“When I talk about consolidating functionality and applications into a platform, you have to think of it as an ecosystem,” Hirlau said. “We shouldn’t be, when it comes to tools and applications, a petting zoo with one of everything or two of everything. Look at the ecosystems you really want to build.”

The approach:

• Use Epic as the clinical source of truth, with Zoom augmenting workflows via telehealth and Zoom Phone integration.
• Coexist with Teams and Microsoft 365 by integrating and maximizing complementary capabilities rather than choosing a winner.
• Extend Zoom into medical education and public events, all the way to Zoom-enabled cameras embedded in clinical equipment like surgical lights.

Most large enterprises already have multiple collaboration platforms in use. The lesson is to stop obsessing over which tool “wins” and instead define roles: which platform owns which ecosystem, how they integrate with systems of record, and where you can safely retire redundancy.

AI Companion: Security first, then scale

Mount Sinai’s rollout of Zoom’s AI Companion is another example of how healthcare constraints are driving the development of a reusable template. The health system has enabled AI features for every licensed employee — but only after a rigorous pilot with cybersecurity and compliance teams.

“First off, going back to Zoom and ensuring that the Zoom AI component was HIPAA-compliant was obviously very important to us as a health system,” Hirlau noted. “We also conducted a very in-depth analysis with our cybersecurity and compliance teams before we even considered rolling this out to everyone else.”

A key sticking point was ensuring that meeting content didn’t become training data. “We wanted to ensure that it never leaves our tenant or is ever used to train anything from a third-party perspective,” he said. Mount Sinai also chose not to enable features such as automatically starting AI Companion or auto-generated meeting summaries; instead, end users decide when a meeting is appropriate for AI assistance.

Once those conditions were met, the organization shifted its focus to value: reclaiming minutes in the day through summaries, follow-ups and action item extraction. In regulated industries, that arc — pilot, verify data handling, then scale — is becoming the norm for AI-driven collaboration.

Rounding: Getting out of the dashboard

One of the most powerful practices Mount Sinai has institutionalized is “rounding” — sending IT, engineering and vendor teams to the clinical floors to observe workflows, speak with clinicians, and see how technology performs in context. Zoom joined these rounds, meeting with nurses, physicians and educators across departments, including labor and delivery and mother-baby units.

“These sessions are so valuable because we collect information, hear feedback from our clinicians, doctors, and nurses, and witness the workflows firsthand,” Hirlau said. “Since that session, there’s been a lot of good work that’s come out of the feedback we collected.”

For enterprises outside healthcare, rounding is a low-tech practice that yields high dividends. Whether you call it branch walks or shop-floor visits, the idea is to stop designing collaboration and communications solely from dashboards and instead co-design with the people who actually use the tools.

Continuous improvement, not a big-bang cutover

Despite multiple successful deployments, Mount Sinai is clear that it’s still on the journey to modernize and migrate off its legacy PBXs. “Even though we finished a lot of these migrations, in our mind it’s a learning process,” Hirlau said. “In the spirit of continuous process improvement, do we need to go back and look at anything?” Each rollout is followed by a cycle of reviewing statistics, tuning queues, adjusting workflows and identifying new use cases — across clinical, educational, and public engagement.

That continuous improvement mindset should be baked into CIOs’ and CTOs’ collaboration and voice strategies. The “go-live” date isn’t the finish line; it’s the start of the optimization loop. Success metrics aren’t just seats migrated — they also include ticket reductions, improvements in key performance indicators such as abandonment rates, and frontline satisfaction.

Three takeaways for IT leaders

1. Treat collaboration as infrastructure, not a utility. Mount Sinai’s story works because Zoom is embedded in telehealth, clinical voice, nursing workflows and education — not just meetings. Design your stack so voice, video and messaging are part of how work happens, not bolted on around the edges.
2. Design around systems of record and frontline workflows. Epic remains the source of truth; Zoom augments it with telehealth and click-to-dial Sinai Phone experiences. In your world, that may be CRM, ERP, or core banking, but the principle is the same: integrate collaboration where people already live.
3. Build an ecosystem strategy and a playbook for AI. Avoid the “petting zoo” of overlapping apps, define roles for each platform, and plan coexistence where it makes sense. For AI, adopt the healthcare pattern: Pilot, verify data handling and compliance, configure conservatively, then scale where you see clear productivity gains.

Couchbase Inc. is trying to solve one of the hardest problems in enterprise artificial intelligence today: turning brittle, chat-style pilots into production-grade agents capable of remembering, reasoning and acting on live operational data. With the launch of its AI Data Plane, the company is betting that the real bottleneck for “agentic” AI isn’t the model — it’s the underlying data architecture.

Industry discussions about what’s holding back AI often focus on security, graphics processing unit availability and other infrastructure-level issues, but in most cases, data is what keeps chief information officers up at night. At the recent HPE Discover event, I asked Dallas Cowboys CIO Matt Messick about the challenge of bringing disparate data sets together, and he said it’s a major challenge and the thing he thinks about most right now – all day, every day, as he put it.

Couchbase’s announcement aims to address this headache with a scalable data platform.

Couchbase takes the covers off an AI data layer for agents

Couchbase’s AI Data Plane is a unified data infrastructure layer for enterprise AI agents across cloud, edge and lakehouse environments. It combines persistent Agent Memory, an Agent Catalog of discoverable tools, and an enterprise-supported MCP server to standardize how models access context and tools. The offering consolidates prior Couchbase deployment models into a single architecture spanning Couchbase Capella and self-managed environments. It is paired with Enterprise Analytics 2.2 for Apache Iceberg lakehouse federation, along with a Trino adapter expected in the third quarter of 2026.

At a high level, Couchbase is positioning the AI Data Plane as an operational data foundation for the agentic enterprise. These are organizations where AI agents are woven into front- and back-office workflows rather than operating as isolated pilots. The goal is to consolidate today’s separate vector databases, caches, document stores, and operational databases into a single governed layer that can feed AI agents at sub-millisecond latency and at scale.

Why this matters: AI is hitting a data wall

The subtext of this launch is that, like the Dallas Cowboys, most enterprises are finding that their first wave of generative AI projects doesn’t fail because of model quality; it fails because the data plane can’t keep up. Though Messick’s comment can be viewed as anecdotal, in the press release IDC’s Devin Pratt noted that roughly 80% of agentic AI use cases will require real-time, contextual and widely accessible data — exactly the opposite of how most enterprises’ fragmented data stacks evolved.

In today’s architecture, a typical agent pipeline includes:

• A vector store for embeddings.
• Multiple caches for short-lived context.
• One or more operational databases for transactional state.
• A data warehouse or lakehouse for analytical context.

Every new AI project tends to bolt on yet another specialized store, increasing integration tax and governance risk. Couchbase’s argument is that you can’t scale agents across the business if every workload requires stitching together bespoke data stacks with inconsistent latency, security and observability.

By making agent memory and context retrieval core capabilities of the database, Couchbase is drawing a line between AI infrastructure designed for agents and the rest of the market, which still treats memory as an afterthought. For CIOs and heads of platform engineering, that’s a meaningful differentiation: memory, context and retrieval become shared services rather than per-project plumbing.

Agent Memory: Closing the reasoning-memory gap

The most interesting part of the announcement for practitioners is Couchbase Agent Memory. In many enterprises, early agents work well within a single interaction but fail when they need to carry state across sessions, understand historical context, or coordinate with other agents and systems over time. Couchbase frames this as the gap between what agents can “reason” about and what they can “remember,” and it has become a critical bottleneck as teams move beyond prototypes.

Agent Memory aims to close this gap by providing a unified persistence layer that:

• Treats conversational context, structured operational data and state as a single service, rather than forcing teams to integrate separate caching, vector and document stores.
• It is framework-agnostic and validated with LangGraph, CrewAI and LlamaIndex, so teams can switch or combine orchestration frameworks without rewriting the memory layer.
• Delivers sub-millisecond latency at the decision point while scaling to billions of vectors and tens of millions of transactions per second.

That combination matters because agentic workloads are far more demanding than traditional request/response applications. Each agent action typically triggers context retrieval, memory writes and state synchronization across thousands of concurrent sessions. Without an integrated data plane, these operations introduce unpredictable latency and failure modes that directly degrade the user experience.

For organizations building complex workflows, it’s important to consider multi-agent systems that orchestrate customer journeys, field operations or financial processes. Having a single place to manage memory and state can dramatically shorten time-to-production and simplify compliance.

From cloud to edge: operational AI where the work happens

Couchbase is also targeting the edge, where much of the inferencing will take place. Agents don’t just live in the browser or the data center; increasingly, they operate on mobile devices, in stores, factories, stadiums and other distributed environments where connectivity may be intermittent. In fact, I recently ran a survey that found 60% of generative AI transactions occur on mobile devices, a trend many information technology organizations have ignored.

The AI Data Plane is designed to meet this full set of requirements by:

• Extending the operational data platform so agents in mobile and edge environments can access replicated data and perform local vector search, even when disconnected.
• Building on Couchbase’s multimodel architecture, which supports JSON documents, key-value, SQL for JSON, full-text search, eventing and vector search in a single distributed system.
• Delivering specific edge capabilities, including Couchbase Lite 4.1 with peer-to-peer Bluetooth sync and automatic Wi-Fi failover; Edge Server 1.1 with client-level access control and expanded Windows/ARM support; React Native 1.1 with Turbo Module integration; and Sync Gateway 4.1 for cloud-to-edge synchronization and non-disruptive rolling upgrades.

For scenarios such as retail associates using AI copilots on mobile devices, field technicians working in low-connectivity environments, or stadium operations relying on local AI agents for crowd management, this edge-aware data plane is a differentiator. It ensures agents can retain memory and context near where the work occurs and then sync back efficiently when connectivity resumes.

From an industry perspective, this aligns with the broader shift toward distributed, event-driven architectures for AI: data doesn’t just flow into a central lake; it circulates through a mesh of devices, microservices and agents. Platforms that can push trusted, governed data and memory to the edge while keeping analytics and governance consolidated will be better positioned as AI becomes part of the operational workforce.

Lakehouse federation: Bridging operational and analytical AI

Couchbase is also refreshing its analytics stack to align with how enterprises are standardizing on open lakehouse technologies. Enterprise Analytics 2.2 introduces Apache Iceberg lakehouse federation, enabling teams to query real-time operational analytics from Couchbase alongside existing Iceberg tables without complex ETL or data duplication. This gives organizations adopting Iceberg for its governance and ecosystem benefits a way to treat operational and analytical data as a single logical layer for AI workloads.

The roadmap goes further with a Trino adapter expected in Q3, providing in-place SQL access to Couchbase operational data from Trino-based platforms, including AWS Athena, Amazon EMR, Google Dataproc and Starburst. This eliminates the need to replicate live data into separate analytical stores just to make it accessible to AI and analytics workflows, a persistent source of cost and complexity.

Additional analytics enhancements, such as Google Cloud Storage support, JWT authentication, Oracle and SQL Server change data capture, asynchronous queries, index advisor, index-only plans, and SQL++ UPDATE support across multiple SDKs, round out the platform by giving teams more governed analytics within their existing tools and languages. The implicit message is that AI agents shouldn’t require a parallel analytics stack; they should be able to tap into the same operational-analytical fabric the business already uses.

For organizations trying to measure and optimize AI value, this matters. If agents can read and write to both the operational system of record and the analytical lakehouse without duplication, it becomes much easier to:

• Instrument AI-driven processes end to end.
• Analyze impact on efficiency, revenue and customer experience.
• Iterate quickly on prompts, tools and workflows based on real usage data.

In other words, it tightens the feedback loop between AI experimentation and business outcomes.

Governance, cost control and accelerating AI value

Finally, Couchbase is bringing governance and cost control into the conversation with Capella iQ enhancements. The natural-language query assistant now supports multi-model provider selection across AWS Bedrock and OpenAI, governed by organization-level policies that determine which models are available to which teams. This allows administrators to keep inference costs, compliance, and data residency within guardrails while still giving developers the flexibility to choose the right model for each workload.

Together, the AI Data Plane, Agent Memory, edge extensions, lakehouse federation, and policy-controlled model access form a broader thesis: enterprises will unlock AI value at scale only if they treat the data plane as a shared, governed platform rather than a sprawl of point solutions.

From an industry perspective, we should expect the following over the next few years:

• Database and data platform vendors should compete not on raw performance alone, but on how natively they support agent memory, tool integration and cross-environment consistency.
• AI infrastructure stacks to converge on unified data planes that bridge the operational, analytical, edge and lakehouse worlds instead of reinforcing their silos.
• Governance, observability, and cost control will become table-stakes features of AI data platforms, not bolt-ons.

Couchbase’s AI Data Plane is an early example of this trajectory. If it delivers on the promise of a single governed data layer with integrated memory, context, and analytics from cloud to edge, it will give organizations a way to move from pilot to production faster — and, more importantly, to measure and scale AI value with far less integration friction.

Final thoughts

Couchbase’s AI Data Plane is well timed as enterprises move from isolated generative AI experiments to agentic systems that sit directly in the path of revenue and operations. The company is betting that the winning architectures will treat data as a first-class capability for agents, not a bolt-on, and that CIOs will favor platforms that turn memory, contex, and retrieval into shared services rather than bespoke integrations.

For IT leaders, the takeaway is that the AI conversation has to move beyond models and GPUs to focus on the data plane design that will either unlock or limit value from age.

Artificial intelligence played a prominent role at this week’s Bio International Convention in San Diego, the largest biotech event with vendors spanning the full ecosystem of companies in this industry.

Today in a special address, Kimberly Powell (pictured), vice president and general manager of healthcare and life sciences at Nvidia Corp., made the case that agentic AI is about to do for biotech what it just did for software — and the company’s BioNeMo is the stack that turns generic large language models into working “AI scientists” that are both faster and cheaper to run.

Nvidia wants to make ‘AI scientists’ mainstream in biotech

Powell opened her presentation by outlining where the industry is now. “We are witnessing the fastest platform shift the life sciences industry has ever seen,” she said. She compared AI to the microscope, X-ray crystallography, and gene sequencing, calling them a new class of scientific instruments. This time, the instrument doesn’t just see or measure; it reasons, plans and acts.

At the event, Nvidia announced its BioNeMo Agent Toolkit, a software stack that turns large language models into domain-specific AI agents capable of executing end-to-end biology and chemistry workflows — from literature review to protein design to lab automation — while optimizing for performance and cost.

From generative to agentic AI for science

Powell’s core thesis is that the life sciences, a $300 billion annual pharmaceutical budget (global R&D is reaching $3.8 trillion), have quietly been preparing for this inflection for a decade. On one side, there has been an explosion of AI research in biology, chemistry, imaging and genomics. On the other hand, Nvidia has been building the infrastructure to operationalize that research: GPUs, networking, CUDA-X libraries and domain platforms such as MONAI, Parabricks, cuEquivariance and BioNeMo.

What has changed in the last 12 to 18 months is the emergence of agentic AI, systems in which a large language model “brain” is wrapped in a harness that manages tools, memory, security policies and multistep workflows. Nvidia’s NeMo Curator and NemoClaw framework and open-source harness are generic versions of that pattern; the BioNeMo Agent Toolkit is the life-sciences-optimized edition.

“Agents are becoming the modern application layer in life sciences,” Powell said. “Every single one of the thousands of companies in life sciences is about to become an agent builder.” That’s a very different framing than “just another model.” It says the next application tier in biotech won’t be GUIs and pipelines, but rather networks of specialized agents coordinating work across digital and physical labs.

BioNeMo as the scientific toolbox — tuned for speed and cost

Nvidia’s announcement positions BioNeMo as the science that sits behind those agents. In practice, the BioNeMo Agent Toolkit does three important things for biotech teams:

• Packages proven life-science models, such as protein folding, molecular docking, generative chemistry, genomics and imaging, into agent-callable tools with clear schemas: what each tool does, what inputs it requires, what outputs to expect and how to troubleshoot.
• Exposes those capabilities via NIM microservices that can run on-premises, in the public cloud or across hybrid environments, so pharma and biotech can place compute where data and regulatory constraints demand.
• Optimizes for token efficiency and computational cost, not just raw accuracy, by giving agents access to highly accelerated libraries and models, so they spend fewer tokens and less wall clock time hunting for the right tool or rerunning failed steps.

Powell specifically addressed the historical cost-performance trade-off. She described BioNeMo’s skills and tools as “the knowhow” that lets agents complete complex workflows with “strong task completion, workflow accuracy, and reduced token expense — that means less compute, more reliable results.” In other words, a BioNeMo-enabled agent doesn’t just produce better science; it does so with fewer LLM calls and more efficient graphics processing unit usage, making cost and performance optimization possible at the same time.

Powell emphasized that BioNeMo is agent-agnostic. The same toolkit can serve agents built on OpenAI, Anthropic, in-house LLMs or Nvidia’s own Nemotron models. That matters for buyers who don’t want their next decade of drug discovery workflows locked to a single model vendor.

What an AI ‘co-scientist’ looks like in practice

To ground this in something beyond architectural diagrams, Powell walked through a protein-binder design workflow targeting MCL1, a protein that helps tumor cells survive. Traditionally, that path — understanding the target to generating binders, predicting structures, scoring candidates and deciding what to synthesize — takes months of specialized human effort.

A generic agent can attempt that workflow but will burn time and tokens “searching for the right tools, figuring out how to call them and oftentimes completely failing to complete the task.” With BioNeMo, Powell said, a scientist gives a single goal such as “Design a binder for MCL1,” and the agent:

• Retrieves or predicts the target structure and its binding region.
• Generates candidate binders using BioNeMo generative models.
• Folds the target and binder together, then evaluates docking poses using accelerated structural engines.
• Ranks and returns the top candidates for human review — “all done without human intervention.”

This is the “AI scientist” pattern many startups are pursuing. The key nuance is verification. Panelist Andrew White, co-founder and chief technology officer at Edison Scientific, noted that as agents improve, “the era of humans writing questions and agents taking the test is over. We really do need this kind of lab-in-the-loop.” His takeaway: The true bottleneck is shifting from reasoning about existing literature to running new experiments, which is exactly where closed-loop digital and robotic labs come in.

Why this matters for biotech and pharma

For biotech leaders, the strategic implications are less about any single toolkit and more about the operating model shift Powell and the panelists described:

• Compression of timelines. Powell argued that agents will “take scientific discovery and shrink the timeframe” — work that took years moves to months, and months to days. Josh Meier, CEO of Chai Discovery, gave a concrete example. Antibody design success rates have risen from one in 1,000 to 10% to 15% in just a few years, driven by improved models and faster iteration.
• Rising expectations on wet-lab speed. As in-silico design compresses from months to hours of GPU time, lab workflows become the new bottleneck. Meier pointed out that many assays were never optimized for speed because there was no incentive; now, tightening that loop is a competitive necessity.
• New collaboration patterns: Powell sees pharma shifting from primarily “deep scientific relationships” to partnerships that integrate frontier AI labs, tool providers, and platform companies within closed-loop systems — where every experiment feeds back into proprietary foundation models and agents. Benchling CEO Sajith Wickramasekara echoed this, arguing that electronic lab notebooks are evolving from retrospective records into “systems of action” co-authored by AI.
• Lowering barriers and de-siloing science. Powell believes tools like BioNeMo will let biologists tap into advanced modeling “in a natural language way, instead of having to get into any type of coding at all,” breaking down silos between disciplines and making modern AI tools accessible to more of the bench.

That last point is worth watching. If AI agents can reliably orchestrate highend modeling and workflow automation behind a conversational front end, the practical distinction between “computational biologist” and “wetlab biologist” starts to blur.

Reading the signal for the road ahead

From an industry watcher’s perspective, BIO 2026 is less about Nvidia “entering” life sciences, since it has been here for a decade, and more about standardizing the agentic stack for biotech before others do. The BioNeMo Agent Toolkit turns Nvidia’s existing beachheads, such as MONAI, Parabricks, cuEquivariance and BioNeMo models, into a coherent runtime that any agent harness can plug into, with clear value props for speed, accuracy, and cost.

The open-source angle is also notable. Powell made it explicit that the toolkit is available on GitHub and is designed to work with both open- and closed-frontier models, giving pharma and biotech the option to build their own domain-specific “brains” on top of Nvidia’s toolbox. In a world where IP, data residency and regulator trust are existential concerns, that flexibility will matter.

Powell closed with an ambition that neatly captures Nvidia’s posture: “Agentic AI has revolutionized coding — that’s a done deal. Now this ecosystem is assembling to revolutionize science as we know it.” For biotech leaders, the question is no longer whether AI can help science, she argued, but “does AI have the right instruments to run science?” With the BioNeMo Agent Toolkit, Nvidia is betting that the answer for a growing slice of the industry will be yes.

When Amazon Web Services Inc. held its New York Summit last week, Vice President of Agentic AI Swami Sivasubramanian as usual was the headline act, delivering the opening keynote.

Sivasubramanian made the case to enterprise leaders that the artificial intelligence conversation has moved beyond pilots and productivity hacks into a world where the real advantage lies in compounding momentum across work, security, software delivery and data. For IT pros, that means your architectural decisions over the next 12 to 18 months will determine whether AI agents become a force multiplier or a new source of chaos.

Here are five big ideas from Sivasubramanian’s keynote and what they mean for those responsible for building and operating enterprise technology:

1. From ‘faster search bars’ to compounding agents

Sivasubramanian’s main critique of the first generation of AI assistants is that they never broke out of chat-window gravity. They sit on top of tools, answer a question and then forget. “We gave them chat windows and connected them to our tools,” he said. “They answer one question, and then they forget. The promise was intelligence, but what we got was a slightly faster search bar. Faster search doesn’t compound; it flatlines.”

The alternative he laid out is an agentic model in which every completed task feeds the next. “What you really need is agents that actually change the way you work, not just speed up the steps, but completely eliminate them,” Swami argued. “If humans are still forced to be the orchestration layer, your momentum actually has a ceiling.” In his framing, “every task that their agents complete makes the next one smarter,” creating “compounding momentum” and widening the gap between early adopters and those who wait.

That’s the design center for Amazon Quick, an AI assistant that “states the outcome you want and figures out how to get there across all your systems, all your data and all your context,” powered by a knowledge graph that reasons across people, documents, communications and data lakes. In the live demo, Quick assembled a marketing report by pulling data from Slack, Google Drive and OneDrive in about 20 seconds — work, Sivasubramanian said, “would have taken probably hours of actual research” before.

Implications for IT pros: This model assumes your collaboration and data platforms are open to agent access and governed by strong identity and policy controls. The job shifts from choosing yet another assistant to curating an ecosystem where agents can safely traverse silos. Connectors, metadata and policy enforcement become as important as model choice. This is a vastly different role for IT pros, but one that’s critical for companies that succeed with their agentic initiatives.

2. Security: Ending the ‘walled garden vs. wild garden’ tradeoff

On security, Sivasubramanian highlighted a dilemma many chief information security officers will face. On one side, “agents that work inside their own walled garden only see what’s inside their own productivity suite. The moment you need something outside the wall, you are back to being the orchestrator.” On the other hand, open tools “do not offer the level of security, compliance and governance that enterprises demand. You traded the walled garden for the wild one.”

“This is a false choice,” he said. “Quick doesn’t ask you to choose. No walls, no copy-and-paste bridges, and every action it takes carries its own governance. Who acted on it, what data they touched, where it went, and whether the policy allowed it.” That theme continues with AWS Continuum, a suite of agent-driven security capabilities spanning penetration testing, threat modeling and code vulnerability assessment. Chet Kapoor, who leads security, observability, search and governance products, described the shift from “telemetry, storage, query and dashboards for humans” to “telemetry to context to reasoning to actions for agents.” Telemetry without context is “noise,” he said; with context, it becomes a “signal” agents can act on.

Customer stories were included to make the stakes concrete. Swami cited GoDaddy using Amazon Quick to eliminate “15,000 hours of manual work annually.” He also highlighted the NBA’s use of Quick to structure 25 years of prospect data into interactive leaderboards and comparisons.

Implications for IT pros: Security operations are headed toward agents taking actions under policy, not analysts staring at dashboards. That raises the importance of policy as code, identity boundaries, least-privilege design, and clear “rails” for where agents can operate. The conversation with the CISO is no longer “Should we use AI?” but “What will we allow AI to do, and under what guardrails?”

3. Software delivery as a closed loop

If the first wave of generative AI was about coding copilots, this keynote reframed the narrative around end-to-end software delivery loops. “Write it right, ship it fast, keep it modern – not three tools, one continuous loop, always running, always compounding,” he said. That loop is already in production at Amazon Stores, where teams behind the retail experience saw a “median 4.5x improvement in how fast correct code reaches production, with some teams hitting up to 17x,” and “AI-generated code changes landing with 95% accuracy, higher than the human baseline.”

Kiro is the engineering agent that anchors the “write it right” part of the loop. You give it a prompt, and it generates “clear requirements, structured design docs, implementation tasks, and validated tests before a single line of code is generated.” It then uses agents and property-based testing to implement and verify. Swami pointed to fintech startup Dhan, which needed to support more than 170 complex trading indicators. Without agents, it estimated “over a dozen engineers in a period of 12 to 24 months;” with Kiro, “all this was built by a single engineer in just eight weeks.”

The loop extends into operations. AWS DevOps Agent started as an incident-response companion used by customers like T-Mobile and United Airlines; now AWS is adding release management. It can project production risk from a code change, explore an application such as an end user, score releases, and feed its report “directly to your coding agent to start implementing those fixes automatically.”

On the other side of the loop, AWS Transform moves from one-time modernization projects to “continuous modernization,” performing “continuous state analysis and remediation at machine speed, always watching, always fixing across every code base you own.” AWS says customers have already used Transform to eliminate 1.6 million hours of manual modernization work.

Implications for IT pros: This is an opinionated pipeline: spec, code, test, release, modernize, repeat, with agents in each phase. To benefit, enterprises will need to standardize how they organize their Git repositories, pipelines and quality gates so agents can act safely across services and to make a cultural shift that treats modernization and reliability work as continuous flows, not project-of-the-year initiatives.

4. Southwest Airlines: A playbook for a ‘modern fleet’ of systems

The most compelling customer story came from Lauren Woods, executive vice president and chief information officer at Southwest Airlines. She linked technology choices directly to lessons from Winter Storm Elliott. “It wasn’t our systems that were failing, but they were not designed to keep up with the pace and the level of complexity happening across the operation all at once,” she said. To run like a modern airline, “we need technology that operates like a modern fleet.”

Southwest chose AWS as its preferred cloud partner for a “secure, scalable foundation” and access to innovation. Regarding AI, Woods said she uses Amazon Quick every day, describing a shift from “looking at data after the fact to interacting with it in real time” across fare and revenue analysis and call center behavioral trends. The impact has been faster decisions, closer to the point of action.

For engineering, Southwest scaled Kiro to “more than 2,700 developers, about two-thirds of our engineering organization,” using it for unit test generation, infrastructure as code, and faster onboarding. The Southwest.com platform, which is mission-critical and built on legacy architecture, had a long modernization roadmap. Using Kiro, “our teams have accelerated that modernization significantly, pulling the original timeline in by three years,” Lauren said. “We’re making it easier to build on, evolve and scale as our business changes.”

Implications for IT pros: Southwest is an excellent case study. AI-augmented decision-making across the business, agents embedded in the SDLC at scale, and modernization and transformation running in parallel. It’s also a reminder that the key performance indicator for AI initiatives will increasingly be operational resilience and customer satisfaction, not just developer productivity.

5. Agent platforms: Harness, guardrails and context as first-class primitives

The final act of the keynote shifted from AWS-built agents to the agents that customers will build themselves. Sivasubramanian noted that “the agents that will matter the most are the ones for your business that only you can create,” but many are “stuck between prototype and production” because teams are re-implementing basics: authentication, memory, tool access, security and governance.

Amazon’s answer is AgentCore, which provides “core components to build agents” and includes a managed runtime, built-in identity, session memory, observability, evaluations and access controls. It is designed to work with any agent framework and model. Over the past six months, Swami said, “the number of tasks performed by agents in AgentCore has grown by 15x,” and customers such as PGA TOUR, Nasdaq and Visa are building production agents in weeks instead of months.

Two concepts are important here. First, the harness. Sivasubramanian described the model as the “brain” and the harness as the “body” that provides “state persistence, error recovery, context management, [and] session isolation.” AgentCore Harness can turn a model into an agent in minutes with three application programming interface calls. Second, Agent Core Policies define what agents can and cannot do and are enforced “outside the agent’s code, where the agent can’t bypass it,” including detection of prompt attacks, harmful content, and sensitive data. AWS plans to ingest signals from third-party security providers into that policy layer.

Underpinning this is context. AWS Context automatically builds a knowledge graph across structured and unstructured data and exposes it to agents at runtime. Swami pointed out that within Amazon, the semantic knowledge store behind Q processes “over 1.8 million requests” per day, mapping business semantics (“escalations” vs. “tickets”) and relationships across systems. In the enterprise, that graph spans public web data via managed search tools, organizational content in S3, SharePoint, Confluence, and Google Drive, and structured data in lakes and warehouses.

Implications for IT pros: This is the AI platform north star: an agent runtime/harness, a policy and guardrail layer outside prompts, and a governed context service — often graph-based — that encodes how your business works. Whether you adopt AWS’ stack or assemble your own, success will come down less to prompt engineering and more to how well you design skills, policies and knowledge graphs that reflect your domain.

Final thoughts

Sivasubramanian’s core point is that agents aren’t a feature toggle but an architectural choice. The advantage goes to organizations that design for compounding momentum across work, security, software delivery and data, rather than to those that simply switch on Amazon Quick, Kiro or DevOps Agent.

For information technology leaders, that means treating agent access, guardrails and context as platform services, embedding AI more deeply in delivery and operations, and copying the Southwest playbook: Start with a high-impact domain, align business and engineering on outcomes, and let agents handle the undifferentiated heavy lifting while your teams focus on domain-specific decisions.

Artificial intelligence dominated headlines and keynotes at every event I’ve attended this year, including the recent Cisco Live 2026. Though the thirst for AI has been insatiable for a couple of years, customer feedback at the event showed that the era of AI curiosity has given way to AI urgency.

Information technology and business leaders are no longer satisfied with conversational chatbots or basic AI scribes that merely summarize meetings or draft text. They want systems that proactively identify and resolve problems across their massive, complex IT estates.

The industry is rapidly moving toward autonomous, agentic artificial intelligence — that is, systems that can observe, reason, plan and execute tasks across distributed environments without human intervention. Yet doing this at an enterprise scale is proving remarkably difficult.

Production-ready agentic AI is something startup Fabrix.ai has been developing for a couple of years. At Cisco Live, I stopped by the AI Village to get an update on the vendor’s progress. At the booth, Fabrix.ai demonstrated a multi-vendor, multi-agent platform designed specifically for enterprise operations that customers can run today.

The underlying crisis forcing the shift to AgenticOps

To understand why what Fabrix.ai is building is important, it’s vital to understand the state of traditional operations. For decades, IT teams have operated in a strictly reactive mode. The typical enterprise uses seven to 10 monitoring tools per department. When an outage or performance degradation occurs, these fragmented point tools unleash a storm of alerts. What follows is the notorious “swivel-chair” choreography: subject matter experts jumping from console to console, interpreting logs, reading dashboards and manually correlating issues across network, security and cloud silos.

This model has not and will never scale. Traditional AIOps suffered from a critical “last-mile” problem. It was highly effective at generating and clustering alerts, but it still left the actual analysis and manual remediation to humans. This friction strains organizations and stalls digital transformation.

According to data cited by Fabrix.ai, failed IT modernization initiatives drain an astonishing $2.3 trillion annually, and 70% of digital transformation programs fail to deliver their promised outcomes. The industry requires a fundamental evolution from AIOps to agentic operations. IT pros need AI agents that not only alert that a fire has started but also autonomously trace the root cause, assess the blast radius and execute remediation before a human analyst even opens a ticket.

The four debts blocking the agentic control plane

If the value of AgenticOps is so obvious, why hasn’t every enterprise deployed it? The reality is that building a unified control plane capable of steering autonomous agents is an architectural nightmare. In fact, fewer than 5% of enterprises have achieved measurable return on investment from their AI initiatives, and only 13% feel truly ready for AI, according to Cisco’s own Readiness Index.

Enterprise architectures are blocked by four compounding debts:

1. The hallucination and governance gap: Large language models are inherently nondeterministic. In a marketing or copywriting use case, a minor hallucination is harmless. In network engineering or cybersecurity operations, an autonomous agent making a nondeterministic choice can inadvertently take down a core data center fabric. Without strict operational governance, trust frameworks and guardrails, agents cannot be unleashed in production.
2. The siloed telemetry problem: Agents are only as good as the data they can parse. Dumping raw, unorganized telemetry data into an LLM context window doesn’t make it smarter; it only accelerates hallucination. Agents do not need more volume; they require structure — a unified semantic data layer that maps relationships, identities and causality across disparate tools.
3. Context degradation in multi-agent orchestration: Complex enterprise troubleshooting requires multiple specialized agents working in parallel. However, maintaining context purity across these boundaries is incredibly difficult. If a network agent and a security agent act on shared infrastructure using fragmented or contradictory data, the operational context breaks down, leading to erroneous or destructive actions.
4. The lack of universal connectivity: Autonomous agents are trapped by what they cannot reach. Static API catalogs become obsolete the moment an enterprise updates its stack. True operational intelligence demands dynamic, schema-aware connectivity that can interact directly with devices and software at runtime.

How Fabrix.ai bridges the agentic value gap

Fabrix.ai is tackling these hurdles head-on with a vendor-neutral, full-stack AgentOps platform. Rather than forcing companies to undergo expensive rip-and-replace migrations, Fabrix sits atop an organization’s existing software estate via a unique Robotic Data Automation Fabric or RDAF layer.

“At the core of the Fabrix platform is its multi-agent, Mythos-ready orchestration and Reasoning Layer, which coordinates specialized digital workers across disciplines. Instead of relying on a single, massive generic model, Fabrix uses domain-aware, AI-engineered hierarchical agents, specifically for ITOps, SecOps and NOC use cases,” explained Shailesh Manjrekar, chief marketing officer for AI strategy.

The platform’s architectural pillars map precisely to the challenges mentioned above:

• Agentic data federation: Fabrix connects to more than 1,900 enterprise data sources and creates run-time MCP wrappers for any data source. The federation agents perform in-place data discovery and continuously link metrics, logs, traces, topology and CMDB metadata into a single semantic data layer, providing agents with a clear, hallucination-resistant view of operational states.
• The multi-domain context engine presents only curated data to agents across domains, preserving tokens with a shared state.
• FinOps and agent governance: To address trust and cost issues, Fabrix features a granular FinOps and spend management engine. Organizations can enforce individual AI quotas per user, departmental limits, and LLM-specific cost caps. More importantly, it embeds an evaluation and guardrail layer that enforces strict, predictable execution limits.
• Pre-built digital worker catalog: Rather than forcing enterprises to build agents from scratch, Fabrix offers an out-of-the-box Orchestrator AI Agents Catalog. This catalog includes specialized digital workers such as Root Cause Analysts, SecOps Compliance Monitors, and Auto-Remediation Techs that can be deployed in weeks.

CollabOps: Bringing autonomous agents into the team meeting

One of the most interesting components demonstrated at the event was CollabOps. Most enterprise collaboration tools use AI defensively, primarily as a passive scribe on the sidelines, generating transcripts. Fabrix.ai flips this script by making Voice AI agents active participants in the conversation.

With an ambient listening pattern, a Fabrix digital worker can be invited directly into meeting rooms and channels across Webex, Microsoft Teams, Zoom, and Slack. During a live incident bridge, engineers don’t need to leave the call to query data. They can simply speak to the ambient agent: “Hey Fabrix, check the health of the wireless network in Building C” or “Run an RCA on incident CFX-2026.”

The agent processes the request through the semantic data layer, performs automated root cause analysis, runs safe diagnostic checks and drops the live interactive link directly into the channel chat in real time. Furthermore, Fabrix highlighted that this ambient architecture is extending directly into front-line Cisco Webex Contact Center environments to assist agents with live case reconciliation and sentiment triage.

Certified sovereign AI for Cisco Secure AI Factory

For highly regulated verticals such as healthcare, financial services and the public sector, moving operational data to public cloud LLMs is out of the question due to compliance and data sovereignty constraints.

To address this, Fabrix.ai announced at the show that it has become a certified independent software vendor for the Cisco Secure AI Factory and Unified Edge. For customers seeking a fully sovereign, air-gapped AI infrastructure, Fabrix can deploy its entire AgentOps platform natively on-premises on Cisco AI PODs, using local LLMs/GNNs.

By running locally on GPU-optimized, Cisco-validated compute (including UCS Series and Nexus Dashboard infrastructures), enterprise buyers gain the full power of cross-domain agentic reasoning and real-time cluster observability, with their proprietary telemetry data never leaving their physical control. Fabrix estimates that this on-premises architecture can reduce total cost of ownership by 30% to 40% compared with equivalent public cloud deployments.

Real-world results: Proof in production

The proof, as always, lies in the production metrics. Fabrix showcased several customer case studies across diverse verticals, demonstrating that this architecture has moved beyond the experimentation phase:

• Telco/service providers: An enterprise customer reduced NOC alert noise by 85% by deploying autonomous 5G RAN agents to isolate faults across more than 500,000 network elements. BizOps agents span OSS/CRM systems, connecting records, contracts, and case data to enable instant, governed decisions.
• Energy: A Fortune 500 energy company used Fabrix DEXOps (Digital Employee Experience) agents to proactively detect and analyze VPN peer losses and wireless authentication failures. Campus hotspot failures were isolated in under two minutes, reducing combined OT/IT downtime by 35% without a single human ticket being opened.
• FinTech and SOC: Automated triage of billions of daily financial transactions reduced SOC alert noise by 90% using explainable AI reasoning.

Advice for IT pros: How to turn the ‘autonomy dial’

The transition to agentic operations will fundamentally change the day-to-day realities for IT professionals. For engineers and operational leaders seeking to navigate this shift successfully, I offer the following advice:

• Stop fighting telemetry volume; demand an ontology: Stop spending budget on adding more disconnected point-monitoring tools that dump raw data into isolated buckets. When evaluating platforms, prioritize data liquidity and semantic layers. Your AI strategy will stall if your agents cannot natively resolve identities across Cisco and non-Cisco tools.
• Look for an extensible harness, not a closed box: Avoid vendors pushing closed, single-ecosystem agent frameworks. True enterprise environments are complex composites of multiple clouds, legacy software and multi-vendor networks. Look for open control planes that embrace standards such as the Model Context Protocol to orchestrate smoothly across your entire ecosystem.
• Ease into autonomy with human-in-the-loop controls: You don’t have to hand over the keys to the kingdom on day one. Use platforms with a flexible “autonomy dial.” Start by configuring your agents to operate in an advisory capacity — generating root-cause narratives and drafting runbooks. Once an agent has consistently earned your trust in specific error categories, promote those actions to fully automated remediation.

The shift from reactive dashboards to proactive, autonomous operations is no longer a futuristic concept. Platforms such as Fabrix.ai demonstrate that with the right data federation and governance models, agentic operations can deliver substantial, measurable efficiency today.