Every organization I talk to wants to be “AI-ready,” but few want to discuss what that entails.

It’s rarely a new model or agent. More often, it’s the unglamorous work of ripping out a decade’s worth of infrastructure debt so the network can deliver the bandwidth, consistency and data that artificial intelligence depends on.

That’s the real story behind the recently announced multiyear partnership between Cisco Systems Inc. and Notre Dame Athletics. The more interesting part is what the University of Notre Dame’s information technology team already accomplished: It replaced 955 switches across roughly 170 buildings in about five months, all while the campus never went offline.

The news

Notre Dame has standardized its wired campus with Cisco Catalyst 9300X multigigabit switches, managed via Cisco Catalyst Center with zero-touch provisioning. The deal also includes Cisco Professional Services and Cisco ThousandEyes for visibility into owned and third-party networks. The first building was cut over during spring break in March, and by Aug. 1, all but a handful of single-switch sites were completed.

The starting point is something all too familiar to anyone running a large enterprise network. John Buysse, senior director of network and telephony services at Notre Dame, told me the switches ranged from “a couple years old” to “probably 15-plus years old,” spanning roughly 28 models and more than 100 operating system versions. All of it was Cisco, a reminder that a single vendor doesn’t mean standardization.

“The level of coordination needed to make sure business in each building continues is sort of intense,” said Tracy Weber, Notre Dame’s chief information officer. Even over the Christmas break, the hotel, restaurants, and athletics keep running. “This place never shuts down.”

The demand for bandwidth continues to grow. Buysse said campus bandwidth usage “more than doubled” between fiscal 2021 and fiscal 2026, largely due to video. By 2030, the goal is “to be able to support a 100-gig connection to an end user, not across the board, but where it’s needed.” The network is planned around a 10-year lifecycle.

Weber was clear about who sets the bar. “Our students have very high expectations for network performance, and it’s not just their personal devices,” she said. “The washers and dryers in the residence halls need to connect to the app that tells them how many minutes are left.”

She also recalled a student tour guide pointing out the IT building: “That’s the IT center. They pretty much just run the network.” Her measure of success is telling: No complaints in The Observer, the student newspaper, “is a good sign because it does hit the news when they are not happy.”

Here are my five lessons that should matter to information technology leaders in industries well beyond higher education.

Get off the rolling-refresh treadmill

Notre Dame’s sprawl wasn’t the result of neglect but of a sensible-sounding policy. “Our previous approach, which led to a wide variety of hardware, was just yearly upgrades,” Buysse said. “That took a tremendous amount of our time. So we made the conscious decision to change our approach and upgrade all of campus at once,” freeing the team “in the off years to focus on other strategic goals.”

The user impact was inconsistency. “We would routinely receive service tickets about things working for users in one building, and then go to another building, where it was not working,” he said. The new goal is that “regardless of what building a user may end up in, they have a consistent experience.”

For retailers, banks with branch networks, or manufacturers with plants of different vintages, the lesson still applies. Incremental refreshes may feel prudent, but the hidden cost is drift. Run the math on a consolidated refresh, including labor, before defaulting to another year of piecemeal upgrades.

Wireless now dictates the wired design

The switch refresh was driven in part by Wi-Fi. “As we upgraded the campus to 6-gigahertz-capable, we needed improved back-end bandwidth,” Buysse said. “The multigig bandwidth has improved the wireless experience and benefited users who need wired connections.”

Power was the other constraint. “We were heavily underutilizing switches because we couldn’t fully populate them without blowing the PoE budget,” he said, pointing to cameras and phones alongside access points. IT leaders planning Wi-Fi 6E or Wi-Fi 7 should budget for the switch refresh and power over Ethernet headroom alongside it, not as a follow-on.

Consistency is a security control

A network with 100 operating system versions can’t be patched at attack speed. Notre Dame now runs three golden images in Catalyst Center, with what Buysse described as “constant auditing to show compliance that all of our switches are configured consistently with that golden image, so we know we’re at a secure level.”

The payoff comes on disclosure day: “When the vulnerabilities come in, we know exactly what our exposure and vulnerability are.” Buysse noted that users will mostly notice the new capacity services. The bigger win “is on the back end that they don’t see,” namely “improving our security posture across campus.”

AI operations need visibility first

Notre Dame isn’t running agentic network operations yet, but Buysse expects to start soon. “It’s not realistic to throw more people at it, and people can’t respond as fast as AI,” he said.

The foundation for that is ThousandEyes, deployed on main building switches and on endpoints when users report problems. “The ultimate goal is to have that visibility, know a problem is starting, and get it resolved before users see the problem,” Buysse said.

An AI agent is only as good as the data it can see, and device health alone won’t tell it what users experience. Get experience-level telemetry in place, including the cloud and software-as-a-service paths users depend on, before buying AIOps.

Sequence the work and forecast broadly

Notre Dame’s approach was to get the network in place, stabilize it, then add advanced capabilities, “taking them one step at a time,” as Buysse put it. He noted that trouble tickets rose right after cutover, as expected when “you replace 955 switches and all of those network connections.” Volume has since dropped now that the semester is underway.

Weber made an important point about planning. When the team built the forecast, “AI wasn’t as clear that it was going to be as important as it is now. But more generally, the forecasting was based on supporting research, and AI is absolutely part of that research.”

In other words, the network was sized for the institution’s mission, not a specific technology trend, which is why it held up when AI arrived. IT leaders in other industries should forecast based on business growth, not hype cycles.

Final thoughts

The takeaway from Notre Dame is that AI readiness is an infrastructure outcome, and technical debt is the biggest barrier to achieving it. My advice for IT leaders in any industry: Inventory the models and software versions you’re actually running, align the switching refresh with your wireless roadmap, treat configuration consistency as a core security control, and invest in experience-level visibility before adopting AIOps.

Do those things, and AI becomes something you can operationalize rather than just talk about.

Every vendor event has a center of gravity, and at theCUBE Research’s AI ROI in Contact Center Summit, it wasn’t a product. It was a metric this industry is finally ready to abandon.

I co-hosted the summit recently with Bob Laliberte (pictured, left), principal analyst at theCUBE Research, and spoke with executives from Cisco Systems Inc., Five9 Inc., Zoom Communications Inc., Talkdesk Inc., and Converged Technology Professionals, an implementation partner. One of the event’s notable aspects was its focus not only on artificial intelligence, now table stakes, but also on how consistently vendors pushed back on the vocabulary, metrics and deployment habits this industry built over the past two years.

Here are my five takeaways for customer experience and information technology leaders from the summit:

Agentic is here and can provide value today

The word has become a checkbox, and those building these systems know it. During his session, Vinod Muthukrishnan, vice president and general manager of Webex Customer Experience at Cisco, addressed this. “I feel all of us have bandied the word agentic around a little too liberally,” he said. “An AI agent that can fill a slot or fulfill a transaction is not agentic. Agentic is much more of a framework of tools, products and processes.”

His positioning offers an excellent way to think about the AI maturity model in CX. The industry is moving from orchestrating a conversation to orchestrating a journey to building a relationship. Each step up that ladder widens the blast radius, which is why his summary of the market lands: “It’s never in our lifetimes been easier to build an AI agent,” he said. “It’s never been harder to make it enterprise-grade.”

Pedro Andrade, vice president of AI at Talkdesk, made the same argument from a different perspective. Talkdesk now positions itself around customer experience automation, or CXA, and Andrade carefully defined it as something other than a product.

“It’s not a category of technology. It’s the system that coordinates a hybrid workforce of AI and human employees,” he said. “It’s an operating model shift. It’s not a tool purchase.” According to Talkdesk’s research, roughly 74% of enterprises already have generative AI in place, and a generative layer alone “stops at the moment when you need judgment or an action that crosses multiple systems,” at which point a human is back to carrying data from system A to system B.

Amit Mathradas, chief executive officer of Five9, drew a line between generations. The prior wave was deterministic, “effectively a decision tree masked with voice capabilities.” The new wave is probabilistic, making complex use cases viable and explaining why the governance model built for interactive voice response doesn’t transfer.

Containment is a vanity metric, and cost savings understate the value

For a decade, containment and deflection were the primary ways to sell the return on investment of contact centers. Several speakers at the summit argued that this approach is no longer valid. “We are huge disbelievers in terms like containment and deflection, which I think no customer wants to be either contained or deflected,” Muthukrishnan said, then explained why: “Containment could well be me telling you I’m never calling you again. That is not good containment.”

Ram Rajagopalan, head of product for AI at Zoom CX, described the same migration as part of Zoom’s “conversation to completion” roadmap. The company now pairs a post-call survey with what it calls implied resolution, using a large language model to grade the outcome. “We don’t want to count customers who are frustrated and drop off as a call resolution,” he said. “It is certainly contained, but we didn’t resolve the customer’s issue.”

Andrade presented data showing that the cost lens is too narrow. Talkdesk segmented its survey base by maturity, and the organizations it calls CXA leaders reported roughly 22% net promoter score gains, versus 5% for the tier below, a fourfold spread. Cost per contact improved as well, but far less dramatically, at 57% versus 48%.

“Savings alone doesn’t show the whole picture,” he said. “They understate the value that you can get with AI.” He also named the metric that must change: “You are going to measure how much time it takes from opening of a problem until it gets closed,” back office included, not just average handle time.

There’s a cost-control version of this, too. Joe Rittenhouse, co-CEO of Converged Technology Professionals, has seen agentic projects hit their targets yet still blow up the profit-and-loss statement. “Customers are often surprised by costs,” he said. “They will tell us, yes, we’re driving revenue, but my bottom line has gone up so much,” he said. Metered, conversational and outcome-based pricing are not interchangeable, and few buyers have modeled their volumes well enough to know which hurts them.

The customer-perception gap is the industry’s real problem

An interesting data point from the event came from Mathradas, citing Five9’s research: “Ninety-nine percent of business practitioners in companies who are deploying think that their contact centers and their serviceability have gotten better,” he said. “Only 66% of actual users think the contact center has gotten better.”

Sit with that. A third of customers think service has gotten worse, and most of those say the reason is simple: “I want to access a human and I don’t get that.”

Five9 calls its answer “Humantic,” and Mathradas named three conditions that should route to a person: complexity, value and vulnerability. These include a complicated financial question and a customer filing a claim after a death in the family. Muthukrishnan arrived at the same conclusion using different data. In a healthcare survey Cisco ran, nearly every outstanding customer experience described one thing in common: “They mentioned a human agent by name.”

The people don’t disappear, but the job changes. Talkdesk built a role around that with its CXA Operations Center, and Andrade described the supervisor’s new role when he explained, “the supervisor role shifts from building scripts into behavior monitoring.” You evaluate an AI agent before it goes live — “It’s like recruiting, exactly the same thing. You do interviews. Well, in the AI world, you are doing evaluations” — and then run performance reviews and course corrections on it the way you would with a person.

Orchestration is the bottleneck, not the technology

Another key finding from Talkdesk’s survey underscored the importance of orchestration. Ninety-eight percent of companies have deployed AI somewhere along the customer journey, yet only 15% combine agentic AI with cross-departmental orchestration. When I asked Andrade why the gap is so wide, he said, “Adoption is easy. The orchestration is the hardest part.”

The barriers the data surfaced aren’t related to model quality or latency. They’re enterprise readiness: compliance at 50%, security at 48%, disconnected systems at 45%, and legacy infrastructure at 44%. That aligns with what I hear from information technology buyers.

Andrade noted that 64% of organizations already run AI agents for discrete functions such as billing or identity, and that it looks like coordination until you notice “this handoff between those agents still requires someone to go there and move the data from one place to the other.” That someone is a human with a spreadsheet.

Rittenhouse shared similar thoughts. His first question to a customer is whether they have an AI committee, and he said the answer is roughly 50/50. The progress he sees is organizational: “What we have seen is a pretty significant creation of jobs in project management.” Project managers report to the C-suite and act as “the choreographer of the business.” Projects fail because no one scoped the internal time commitment. “This is not just an IT project. This is not just a CX project,” he said.

Scope discipline is now a competence. Andrade time-bound it: Intelligent decision routing can go live in two to four weeks, and “if it goes more than two months, maybe you are trying to boil the ocean.” Rittenhouse offered the best “start small” argument I heard at the event. Ask a client what happens to calls after hours, and the answer is usually “I don’t know.”

At one physical rehabilitation facility where therapists answered their own phones, patients leaving the emergency room at 9 p.m. with a broken ankle were booking with a competitor by morning. Fixing after-hours coverage, he said, “resulted in millions of dollars in revenue” — from a roughly 30-day project.

Governance and openness are now purchase criteria

Muthukrishnan’s line on risk deserves attention from all IT and CX leaders. “Autonomy without accountability is a liability.” He added, “If you’re having a ‘build-my-AI-agent’ conversation in one room and a ‘how-to-secure-manage-observe-your-agents’ conversation in another room in 2026, you’re either in the wrong year or the wrong room.”

His comments are correct. Open-loop transactions, in which an agent touches customer data, acts autonomously, and communicates with an end customer, create a threat surface that scales with the deployment’s ambition. On openness, the vendors said the right things. Muthukrishnan called interoperability “both a cultural and a technological trait,” including for systems that compete with Cisco.

Mathradas made a parallel commitment: “You will never hear us at Five9 say you have to end the technology or toolkit you’re using for our capabilities to work.” Talkdesk decoupled CXA from its own contact center platform so it can run on a rival or on-premises system because “the majority of the seats are still on-prem” and those customers “can’t afford to miss the opportunity of AI.”

I told Muthukrishnan on air that I’m not convinced the era of the walled garden is over, though I hope he’s right. His response was that the walls are getting shorter, which is probably an accurate summary of where the industry is today. Treat openness as a contractual question, not a philosophical one, because, as he put it, betting the other way means “you’re assuming for the next decade this one company is going to own the innovation roadmap for the universe.”

Final thoughts

The next phase of contact center transformation won’t be won by whoever automates the most interactions. It will be won by organizations that connect their data, measure resolution rather than deflection, keep people in the seats where they matter, and build the internal orchestration to scale a pilot that works. The 98-to-15 gap Talkdesk highlighted is the story in two numbers. The technology is ready. Most operating models are not.

Here’s a video from the event with Laliberte, Muthukrishnan and Kerravala:

In Denver this week, Cisco Systems Inc. and its Splunk unit made a bet that the scarcest resource in the agentic enterprise isn’t intelligence or graphics processing units — it’s trust.

The Splunk .conf 2026 show themes and keynotes focused less on dashboards, widgets and features and more on who or what is authorized to act in the agentic era.

Every vendor conference typically has a center of gravity that serves as a unifying theme. At .conf 26, it wasn’t a product but a word: trust. From the welcome keynote by Cisco President and Chief Product Officer Jeetu Patel (pictured) and Splunk Senior Vice President and General Manager Kamal Hathi to the product keynotes that followed, the themes were consistent: Enterprises are moving from software that answers questions to software that acts, and nobody can scale the latter without a system of record.

Here are my five takeaways from the keynotes that should matter to information technology and security leaders:

Agents are workforce, not tooling

In his opening keynote, Patel laid the groundwork for the event. “We have squarely moved now from the era of chatbots to the world of agents, conducting tasks for us almost fully autonomously,” he said, urging the audience to think of agents “not as tools but as teammates.” He framed the change around four vectors: inference is the new workload, agents are the new workforce, tokens are the new currency, and trust is the new currency of adoption.

Patel provided several supporting data points that should stick with buyers. According to Patel, in February the tokens consumed by agents exceeded those consumed by humans for the first time — “which is wild,” he said — and seven months later agents were consuming five times as many. He also cited projections that roughly 60% of global AI compute capacity this year goes to inference rather than training. This means the consumption curve is no longer gated by human attention span: agents run 24/7, spawn other agents and use materially more network bandwidth than the humans beside them.

His best line was a tongue-in-cheek remark about the state of agentic. Patel said agents “are like teenagers” — intelligent, fearless and notorious for poor judgment. Anyone who has worked with agentic knows this is true, and it’s a useful mental model for those writing AI policy in the near future.

Observability and security are merging

One of the more interesting implications from agentic AI came during a live demo in which the AI engineer and a security operations center analyst worked the same incident from two consoles. Patel explained why the two disciplines can no longer be separated: “It is actually very hard to distinguish between whether there’s a breach, or some agent was poisoned because of an external prompt, or the agent just exercised poor judgment because it was very literally following your instructions.” Observability and security, he said, are starting to merge.

Anyone who has run a SOC understands the implications. The classic triage question of “bug or attack?” becomes unanswerable with security telemetry alone because nondeterministic systems fail in ways that look identical to compromise. Splunk’s answer is to put agent traces, evaluator scores, application telemetry, network data, and security signals on a single correlated fabric, then let the same evidence trail serve both teams.

John Morgan, Splunk’s senior vice president and general manager of security, previewed an integration shipping at year-end that lets any customer with both Splunk Observability and Enterprise Security join those datasets, while acknowledging the org chart reality: “We want to respect that the observability and the security teams are different. We know they often have different budgets, but at the same time they have the same business goal.”

The reality is you can’t troubleshoot and fix what you can’t see, and that pushes the historically separated worlds of observability and security to come together.

Tokenomics turns AI from a science project into a P&L line

The product keynote highlighted a trend I’ve seen explode over the past few months: cost.

Splunk announced the general availability of Splunk Agent Observability, available in the cloud, on-premises, and as a native Cisco Cloud Control application, with a new capability called Tokenomics for tracking and forecasting token spend across agents and coding tools.

The demo showed how it works. A retailer’s shopping agent, told to increase customer satisfaction, cheerfully honored an expired $300 promotion. The engineers who granted it backend access didn’t anticipate this behavior, and the presenter estimated it could cost roughly $500,000 an hour at scale.

The fix wasn’t pulling the agent from production. It was converting a custom evaluator into a small language model guardrail that runs in line with sub-350-millisecond latency, cutting evaluation costs from about $1,000 a day on a frontier model to a few dollars. As Splunk Director of Agent Resilience Vikram Chatterjee put it, “not everything needs a model in the loop,” and runtime controls must work “without breaking the bank.”

Diganta Nayak of Verizon took the stage to discuss the economics of agentic. He said Verizon resisted the reflex to optimize spend downward: “It is not simply about reducing the number of tokens. It is about optimizing the economics of the business outcome. The cheapest agent may not be the most efficient agent.” A pricier model with a higher completion rate and fewer retries can be the better buy. His governance example highlighted the growing importance of observability, as you cannot govern what you can’t see.

In practice, security and observability teams are separate in most organizations. I asked many customers here at .conf for their thoughts on bringing them together, and the prevailing opinion is that integrating security and observability will be difficult but inevitable.

The data economics must change first

None of this works if telemetry growth prices customers out of their own platform, and Splunk knows it. Hathi acknowledged the longstanding objection directly, where customers told Cisco there was a practical ceiling on what they could afford to send to Splunk and pitched the reset in plain math: “capacity up 10x, bill flat,” with a goal of helping customers to “stop agonizing over what data you can afford.”

The enabler of change is the Cisco Data Fabric, powered by the Splunk Platform, featuring Machine Data Lake, Catalog, expanded Federated Search and AI data management. Federated Search now reaches Amazon Web Services CloudWatch and Databricks, with FedRAMP coverage slated for January 2027 and Splunk Enterprise on-premises later that year.

The demo highlight was an SPL query that joined Splunk operational data with business data in Snowflake. No movement, no duplication, with just-in-time promotion of a thin slice of DNS data rather than indexing it forever on the chance someone needs it. Existing platform customers get the evolution without buying something new.

Cisco also brought the AI stack on-premises with Cisco AI POD for Splunk, available now as part of Cisco Secure AI Factory with Nvidia, alongside self-hosted model options including Google Gemma 4 and OpenAI’s GPT-OSS 20B, with Nvidia Nemotron models coming. This should appeal strongly to regulated and sovereignty-constrained organizations.

Agents earn trust the way employees do

During the keynotes, the most practical advice came from customers. Jordan Kramer from Asurion explained that the company supports roughly 300 million customers and uses a “trust but verify” model that moves each agent “up a trust ladder as it proves itself, steadily expanding what it’s allowed to own.” Kramer also said that “like interns, agents earn trust,” continuing the young-person analogy Patel started. Their advice to practitioners: Treat agents like production applications from day one, keep humans in the loop, manage scope and don’t try to solve everything overnight because “none of it works without the right data and platform foundation in place first.”

Hossein Korsha from Constellation Energy offered data to support the shift to agentic. The mean time to detect and respond to a cloud-native identity incident fell from about 20 minutes to 39 seconds once automation was trusted enough to act. He also shared a cautionary tale: His team adopted AI before its asset framework was mature; it flooded analysts with false positives, and they had to restart from the data layer up.

That aligns with what Splunk shipped for the SOC: an expanded agentic workforce spanning detection engineering, triage, investigation and response, with Exposure Analytics enhancements that add historical change tracking and business context to risk scoring, and Cisco Live Protect deploying virtual patches for vulnerabilities customers can’t yet fix. On the observability side, Observability Studio and AI SRE detection and troubleshooting agents are now available, along with a Splunk remediation agent built with Anthropic that reads telemetry, opens a merge request and asks a human to approve it.

Final thoughts

Splunk’s .conf 2026 message was on point and refreshingly candid about the gap between demo and deployment. Patel called the acquisition one of Cisco’s most strategic ever, adding that if he did it again, “I probably could be tricked into paying more for it.” When Cisco acquired Splunk, many industry watchers questioned the value of the combined company, but that was before the AI era.

Today, AI is everywhere and part of everything, and the acquisition positions Cisco well for the long haul. In an agentic enterprise, the system of record for machine behavior becomes the control point for everything else, and Cisco owns a unique share of the stack beneath it — spanning silicon and photonics, network, compute, models and, with Splunk, data.

That’s not to say there isn’t execution risk. These were dense keynotes with many announcements, some available today and others arriving in 2027. The observability-security merger Patel described is real, but organizational boundaries move more slowly than architectures. Buyers should focus on two things: how the new Essentials and Premier editions map to what they already own, and whether the trust ladder their teams need is enforced by platform policy or runbook discipline.

For years, retail banking strategy followed a predictable script: Shift routine transactions to digital channels, downsize physical footprints and route customer service inquiries to centralized contact centers. The underlying assumption was that physical branches were dying embers of a bygone era.

Yet industry data tells a radically different story. According to the 2026 Future of Branch Banking Report by the Digital Banking Report, 71% of financial institutions consider branches mission-critical or very important to long-term growth. Only 3% plan to shrink their networks, while 97% are actively investing in, modernizing or expanding their physical footprints. Physical centers haven’t vanished; their core purpose has simply evolved.

Customers no longer visit a branch to deposit a check or check a balance — they handle those tasks on their smartphones. Today, customers walk through a branch door during life’s most complex, sensitive and high-stakes financial moments: applying for a mortgage, opening a commercial business account, navigating estate planning or resolving a critical fraud incident.

Despite millions invested in digital transformation, retail banking still faces a major operational blind spot: the disconnected branch experience. While contact centers and mobile apps have gradually consolidated into unified customer experience platforms, physical financial centers remain digitally isolated. When an information technology strategy leaves the physical branch out of the enterprise telemetry loop, both customer experience and operational efficiency suffer.

The core problem: Frontline blind spots and institutional amnesia

To understand why banking customer experience breaks down, look at the typical customer journey preceding a branch visit:

1. A customer browses home loan options on a mobile banking app.
2. Unsure about specific rates, they dial the contact center and speak to an agent for 10 minutes before getting disconnected or deciding to visit in person.
3. They walk into their local branch 30 minutes later to finish the process.

What happens when they sit down at the banker’s desk? The banker starts from scratch. “How can I help you today?” asks the representative, unaware of the app’s browsing history or the prior call. The customer is forced to repeat their story, restate their information and re-verify their identity. This phenomenon, institutional amnesia, erodes customer trust at the very moment it matters most.

This friction stems from four persistent execution gaps in banking technology architecture:

• The data gap: Customer interaction history is siloed across core engines, digital banking platforms and contact centers. Branch-level workstations rarely receive this context in real time, tracking account balances without capturing real-time intent.
• The workflow gap: Processes do not connect across channels. A mortgage or loan journey started online halts as soon as the customer walks into a physical center.
• The employee enablement gap: Legacy branch systems were designed for tellers processing transactions, not advisers delivering consultation. Frontline staff act as highly paid human ATMs rather than trusted advisors because they lack real-time guidance tools and context overlays.
• The measurement gap: While every click, scroll and call-center drop-off is meticulously analyzed by enterprise data teams, in-person interactions vanish into an analytics black hole, judged by transaction counts that no longer reflect the real value of branches.

According to the Future of Branch Banking Report, 38% of operations leaders cite inconsistent execution across markets as their top operational risk. When service quality varies by branch, the institution bears hidden operational liability.

Bridging the gap: Real-world ROI and connected architecture

Solving this challenge requires expanding the definition of “omnichannel” so that context flows seamlessly across all touchpoints, regardless of medium or physical location. Modern platform launches, such as Talkdesk Inc. for Financial Centers built on the Customer Experience Automation framework, demonstrate how cloud-native software layers can integrate physical bank branches directly into the enterprise contact center ecosystem.

By consolidating fragmented systems and surfacing AI guidance across all touchpoints, financial institutions are achieving dramatic, quantifiable returns:

• Merchants Bank: A 150-year-old community institution replaced static paper reference binders with AI-powered knowledge management and unified phone, chat, SMS, and email. The shift enabled the team to answer 90% of calls within 20 seconds, capture 50% of calls on the unified platform, and extend customer support around the clock.
• TowneBank: As TowneBank scaled across Virginia and the Carolinas, it unified customer relationship management across all departments with real-time AI assistance. By reducing administrative friction for staff, it successfully navigated major system conversions while achieving a 97% satisfaction rate and a 4.8 overall CSAT.
• Arbella Insurance: In property and casualty insurance, Arbella deployed a unified CX platform to 1,000+ employees and agency partners in ten months. They improved service levels by 20% within weeks, raised CSAT by 3%, and reduced average handle times by 15 to 45 seconds — generating an estimated $1.7 million in savings over three years.

These real-world outcomes demonstrate three immediate operational advantages of unifying intelligence across the customer service ecosystem, including physical channels:

1. Instant journey continuity: Frontline workers receive a unified workspace that surfaces recent app searches, contact center transcripts, open tickets and intent signals. The banker begins the conversation at step four rather than step one.
2. In-flight AI copilot guidance: As bankers transition to universal advisers, AI copilots run in the background, surfacing compliant scripts, next-best-action prompts and automated summary notes, enabling advisers to focus on listening rather than typing.
3. Centralized telemetry and the “neighborhood P&L”: Capturing branch visit intent, outcomes and customer feedback builds a complete 360-degree dataset. This enables institutions to move beyond counting teller transactions and adopt a “Neighborhood P&L” — giving branches credit for total ecosystem revenue generated, even when a customer signs the final loan application at home on a tablet.

Recommendations for banking IT leaders

For CIOs, enterprise architects, and technology directors in financial services, modernizing the branch network requires a clear shift in infrastructure strategy.

1. Treat the branch as an active telemetry edge

Stop treating physical branches as isolated IT footprints limited to local networks, security cameras, and core banking terminals. Modernize branch architecture by establishing event-driven application programming interfaces that feed in-person interaction data (visit reasons, appointment notes, outcomes) directly into your centralized data cloud or customer relationship management system alongside digital telemetry.

2. Unify contact center and branch workspaces

Avoid buying point solutions that create new software silos. Look for lightweight cloud layers that allow branch staff and contact center agents to operate within the same unified desktop workspace. Operating on shared context engines and AI copilots lowers training costs, reduces desktop software bloat and eliminates channel data discrepancies.

3. Enforce real-time identity resolution at check-in

Journey continuity relies on immediate identification. Implement branch check-in capabilities — via mobile app geofencing, QR codes or self-service kiosks — that instantly match arriving customers with their digital profiles and trigger real-time context alerts on banker desktops before the customer takes a seat.

4. Deploy AI with financial-grade compliance guardrails

Branch advisers handle sensitive, highly regulated financial discussions. When deploying AI-driven next-best-action engines or real-time assistant tools to frontline staff, ensure the platform includes built-in compliance guardrails that adhere strictly to local banking regulations, fair lending rules and customer privacy preferences.

Final thoughts

The physical branch is far from obsolete, but its role as an isolated, standalone brick-and-mortar building is over. Physical locations represent a structural advantage that digital-native neobanks cannot quickly replicate.

For IT leaders in financial services, the strategic mandate is clear: stop separating digital channels from physical spaces. By linking branch interactions into the broader CX automation strategy, financial institutions remove longstanding customer friction, empower frontline advisers and transform physical footprints into connected relationship hubs.

The next phase of artificial intelligence in legal technology will not be defined by a better chatbot or a faster document-summarization tool. It will be defined by whether AI systems can take responsibility for real work that unfolds over days or weeks, spans multiple systems and communication channels, and returns control to an attorney at the moments when human judgment matters most.

Supio is building toward this next operating model using long-horizon agents. The vision is much larger than automating individual steps in a personal-injury case; it aims to create an intelligent operating layer for the firm. This system understands the case, the firm’s institutional knowledge, the status of work in progress, and the next actions needed to move matters forward. Supio describes this broader platform as a “Firm OS,” a system of action — not merely another system of record.

This distinction is important to understand. Most legal AI products have focused on point solutions: summarizing medical records, preparing a demand letter, searching discovery materials or answering questions about a single case. Those capabilities are valuable, but they leave lawyers and staff to coordinate the workflow around them. They must recognize that an action is needed, identify the right data, navigate communication channels, follow up and document the outcome. To date, this is why most legal AI has improved task-level efficiency but had minimal impact on firms’ bottom line.

Long-horizon agents seek to take on that orchestration work. If Supio can execute this model, it can change not only law firms’ cost structure but also their capacity to handle more matters, bring senior-level expertise to more decisions, and improve the client experience.

Defining long-horizon agents

A long-horizon agent is an AI system designed to pursue an objective over an extended period, rather than to generate a onetime answer or complete an isolated task. It can maintain context, recognize follow-up work, use multiple tools or channels, make bounded decisions and escalate exceptions or judgment calls to a human.

Consider the difference between asking a chatbot, “What do I need to know about this client’s upcoming treatment?” and asking an agent to manage the treatment process, including finding the provider, scheduling the appointment, communicating the appointment to the client, and returning the records after the treatment.

The latter is not a single interaction. It is a multistep workflow with dependencies, shifting conditions and a real-world outcome. During a briefing with Supio, Head of Product Dan Zhang offered an instructive example: medical-record retrieval.

A simple description—“get the records from the provider”—obscures the operational complexity. The agent may need to validate provider contact details, identify the provider’s request process, complete forms and HIPAA-related paperwork, fax the request, follow up by phone or email, monitor for a response over days or weeks, ingest the records upon arrival, and alert the legal team if the process stalls.

That is what makes the agent “long horizon.” It is not simply capable of using a single tool. It understands the broader goal and can keep working toward it over time, across channels and through intermediate decisions.

This is also why long-horizon agents are a more meaningful test of enterprise AI maturity than conversational interfaces alone. A generative-AI assistant can draft an email instantly. A long-horizon agent must determine when to send the email, what information it needs, whether a response has arrived, when escalation or approval is appropriate, and where the result belongs in the system of record.

From legal assistant to firm operating layer

Supio’s strategy is based on the recognition that plaintiff legal work is not a clean, fully digital workflow. Matters move across case-management platforms, email, voice, documents, provider offices, fax systems and external organizations, including insurers, clients and treatment providers. According to the company, roughly two-thirds of the work in a case involves some form of communication with an external party other than the client.

That requires more than access to a language model. It requires a connected operational environment. Supio is building a platform that brings together case data, firm knowledge, authoritative case law from Thomson Reuters, work status and communications. As work is performed, agents can document what happened, identify follow-up tasks and build a more complete picture of where a matter stands. Future agents can then act on that evolving context rather than starting from scratch each time a user submits a prompt.

A traditional case-management system records activity after a person performs it (and the record is only as good as what was documented). An agentic system can both perform certain activities and record them as they occur. The product increasingly becomes an active participant in the process rather than a passive repository of case files.

For an attorney, the value proposition is not merely fewer keystrokes. It is the ability to spend less time coordinating mechanical work and more time applying legal strategy, exercising judgment, communicating with clients, and deciding how aggressively to pursue or resolve a case.

Supio’s vision is for the agent to feel less like an entry-level automation tool and more like an experienced colleague who knows the organization and the attorney’s work. The agent manages the repeatable workflow in the background, while attorneys focus on the legal decisions and case strategy that cannot be delegated.

The Simon Law Group use case

With AI, the most compelling evidence of success comes from customer use cases. Trial lawyer Bob Simon described using Supio to build a personalized agent tailored to his litigation approach. He began by connecting the system to sources such as SharePoint, Outlook, his CMS and OneDrive, then added past trial materials, depositions, litigation manuals, expert research, articles and a book he wrote about trying disc-injury cases.

The goal was not merely to create a repository of documents. It was to codify a playbook: how Simon evaluates a case, prepares for an expert, identifies weaknesses in an opposing position, and thinks about winning at trial.

That is a significant evolution beyond generic legal AI. Horizontal tools can produce competent drafts and summaries, but they do not inherently understand how a particular lawyer or firm operates. A purpose-built vertical platform can integrate general-purpose model capabilities with case-specific data, legal workflows and the firm’s accumulated expertise.

Simon described using Supio to review depositions against his prior work product, identify material he may have missed and iteratively refine the agent’s analysis. He also used it beyond traditional legal tasks — for example, to analyze the firm’s financial information in QuickBooks and to reconcile meeting notes, agendas and a conference website to surface gaps or inconsistencies.

One example illustrates both the potential and the limits of agency. Simon said the system found metadata in a discovery response indicating the defense may not have produced everything, and then drafted a subpoena targeting the third party from which the information originated. He credited that process with helping resolve the case for a substantial sum of money.

But he was equally clear about the human role: Lawyers should verify the work. In high-stakes legal matters, an agent’s output is not a substitute for professional responsibility. Simon’s practice is to request source links and to verify key exhibits, evidence and medical records. That is the right model for agentic AI adoption: increased autonomy for repeatable, low-risk workflow steps, paired with clear human review at consequential decision points.

The key question: trust

Long-horizon agents will face a higher bar than traditional AI assistants because they operate over time and increasingly interact with the outside world. The design challenge is not simply to make agents more autonomous. It is to make their autonomy observable, controllable and appropriately constrained.

Firms will need clear permissions, audit trails, source attribution, escalation paths and role-based access controls. Simon’s experience with a financial-analysis skill is instructive: After realizing too many employees had access to it, his firm restricted access to the skill to three authorized users.

The legal sector is therefore a useful proving ground for long-horizon AI. It is document- and workflow-intensive, highly regulated and dependent on judgment, trust and accountability. A system that helps firms achieve meaningful operational gains in that environment — while keeping attorneys in control of the decisions that matter — has implications far beyond the legal sector.

Supio bets that winning agentic platforms will not be general-purpose systems trying to serve every industry equally. Instead, they will be vertical intelligence systems that understand the language, workflows, authoritative knowledge, data, exceptions and institutional memory of a specific domain. That is likely correct. In the AI era, the differentiator will not simply be access to the same foundation models. It will be the ability to turn those models into trustworthy systems of action capable of handling real work from beginning to end.

Final thoughts

The next winners in enterprise AI will be the companies that shift from producing answers to advancing work. That shift requires deep domain knowledge, access to the systems where work happens, persistent memory, workflow awareness, disciplined governance and a design that keeps people accountable for consequential decisions.

Supio’s long-horizon-agent strategy offers an early illustration of that model. The company is not simply trying to help lawyers produce better documents. It is trying to help firms operate differently — with agents that can absorb repetitive coordination, document their own work, surface the moments that require expertise, and free legal professionals to focus their time where it has the greatest value. If that model succeeds, legal AI will evolve from software lawyers use into an intelligent operating layer that continuously helps move the firm’s work forward.

Operating a network has not changed in decades. Every network engineer knows the drill. An alert fires. You open a dashboard. Then another. Then you pull logs, review client history, ping the wireless team, and, hopefully, an hour later, there is enough context to start solving the problem. The industry spent a decade building better dashboards, even though operations teams needed fewer.

The first wave of artificial intelligence in networking was supposed to fix that, but it largely didn’t. When I sat down with Nabil Bukhari, chief technology officer and president of AI at Extreme Networks Inc., at its chief information officer event in Seattle this week, he was refreshingly candid about why and didn’t point the finger at competitors.

“The first generation of AI from multiple vendors in networking, or adjacent markets like security, was essentially a chatbot built on a lot of unconsolidated data,” Bukhari told me. “They demoed really well. Everybody showed amazing demos. But when you push them into actual production and people started using them, there were multiple areas where they just did not work.”

Extreme has 18 months of its own production data to draw on, since the Extreme Platform ONE shipped with first-generation generative AI. That experience shaped Extreme Agent ONE Coworker, which became generally available Wednesday to all Platform ONE customers worldwide as part of their existing subscription. I’ve spoken with beta customers who say it reduces resolution time from hours to minutes and cuts new-engineer onboarding time by as much as half.

Three failure modes, three fixes

Bukhari identified three specific things that broke in generation one, and each corresponds to a design decision in this release.

The first is data. “Unless you absolutely fix your data and have a custom-built context layer, it’s never going to be good enough for a network admin to actually use it,” he said.

His example is mundane and devastating: Within Extreme alone, fabric, switching, Wi-Fi and software-defined wide-area networking each carry a field called client ID in their databases, and each one means something different. Plug that into an LLM unnormalized, and the model has no idea what it’s looking at.

Extreme’s answer is a purpose-built context layer plus a knowledge graph mapping relationships across users, devices, applications, services and network conditions. I’ve argued for a while that in networking, context is the moat. Any vendor can wire a model to a chat box and have it explain spanning tree. Almost none can answer why the third-floor conference room degraded at 2 p.m. Tuesday.

The second is behavior. Extreme ran focus groups with network administrators, dropped them in front of an open chatbot, and watched what happened. Mostly nothing. “They just kind of go, ‘Ah, well, I don’t know,’ and then they ask you how many devices I have — which is a completely useless question, because it’s right there on the dashboard,” Bukhari said. Worse, when a real incident hit, engineers forgot the AI existed and reverted to old habits.

The third is honesty. General-purpose models are built to be helpful, so they’d rather generate something than admit a gap. Extreme built what it calls an awareness scale, so Agent ONE quickly determines whether it can help, says so plainly when it can’t, and then explains what it can do instead — often by packaging the evidence it collected and opening a support ticket on the user’s behalf.

In an analyst pre-briefing, Michael Jones, Extreme’s vice president of AI and a former Salesforce AI leader, demonstrated this live by asking whether Agent ONE could remediate wired devices. It said no, then listed what it could do.

Ambient beats interactive

The fix for the behavior problem is the most interesting part of this release. Agent ONE Coworker is ambient, which means it’s always running and it reaches out to you rather than waiting to be asked. Extreme calls the mechanism a Nudge: a command bar at the bottom of Platform ONE that blinks when the agent spots something, after running a pre-investigation, before interrupting you.

The early numbers suggest that framing matters enormously. “In the first eight days, we have seen about a 900% increase in interaction,” Bukhari said. “When a problem happens and you see it in your logs, nobody thinks of opening a chatbot to ask what these logs mean. Those demos well, but people don’t do that.”

The guardrails are critical. Informational nudges can be snoozed for an hour or dismissed for days. Severity-one nudges can’t be snoozed. I compared it to a car’s collision alerting: If you switch it off, it should come back on when the system thinks you’re about to crash. Bukhari’s logic was the same. If a Sev1 is still there after 15 minutes, you should probably look at it, so why turn it off?

A second flavor of nudge targets the skills gap. If the system detects you lingering on a Platform ONE screen where you knew how to do the equivalent task in ExtremeCloud IQ, it surfaces the top three things users typically do there and offers to walk you through it or just do it for you. Nobody has time to read product documentation, so this meets engineers where they are.

Trust is a gradient, not a switch

Historically, network engineers have been skeptical of automation, and Extreme is working around that skepticism rather than debating it. Coworker mode is ambient but never autonomous, built with a human in the loop by design, with governance controls that define what it may and may not touch. Operator mode, the autonomous tier, arrives early next year and includes its own inverse of the nudge: a recap that briefs you the way a human teammate would upon your return.

“Trust is not binary. It’s not like I trust AI or I don’t trust AI,” Bukhari said. “It’s gradual. You can say, ‘I trust it to do these things, but I don’t yet trust it to do these kinds of things,’ and that progression is what’s needed. Nobody is going to go from ‘I’ve never used it’ to ‘I want my entire network run by AI.'”

Standard operating procedure underpins both modes. Bukhari’s argument here is the sharpest critique of general-purpose AI in operations I’ve heard from a vendor chief technology officer: Ask a frontier model how to solve a networking problem seven times, and you’ll get seven answers. “You do not want to troubleshoot your network that way,” he said. “Networks have standard operating procedures. You can call them blueprints, you can call them validated designs — that’s how predictability is built into the network.”

Extreme loaded its own blueprints and best practices into Coworker; with Operator, customers can load theirs via skills. Bukhari said that early adopters who first tried to bolt their network data onto general-purpose frontier models became Extreme’s fastest adopters. They’d already discovered the predictability problem on their own.

Why this matters for Extreme

Fiscal 2026 ended with revenue of $1.28 billion, up 13% year-over-year, and fourth-quarter revenue of $339 million, marking its ninth consecutive quarter of sequential growth. Platform ONE drove nearly half of subscription bookings in the quarter, and the number of customers spending more than $1 million annually rose to 187.

Extreme is a challenger to much larger network vendors and cannot win a spending war. It must win on velocity and by shipping first. Platform ONE went GA in 2025 with 265 early adopters; roughly a year later, the agentic layer is live globally on a monthly release cadence, with the next wave landing at Extreme’s AI Summit in Amsterdam on Oct. 20.

Shipping Coworker before Operator is the right call. The fastest way to poison enterprise AI adoption is to let an agent make a change that no one has sanctioned. This lets Extreme roll out its AI capabilities in a measured way, letting customers get their feet wet with AI in a safe, low-risk way.

Advice for IT professionals

Pick one use case and protect it from distraction. This was Bukhari’s answer when I asked what “Just get started” should mean. “Pick a use case you care about, not a random one. It’s going to be frustrating at first, but the ramp-up is very quick.” The failure mode isn’t picking the wrong one — it’s chasing 27 things at once because the demos are shiny.

Test for honesty, not just accuracy. Use your pilot to probe the agent’s boundaries and see whether it admits gaps. An agent that bluffs is worse than no agent.

Grade vendors on context, not on model names. If the pitch leads with which large language model is under the hood, it’s a commodity conversation. Ask how the system normalizes data across domains, how long it retains history and how it maps relationships.

Measure mean time to context. Everyone tracks mean time to recovery; almost nobody tracks how much of it is spent gathering evidence before diagnosis starts. That window is what Nudge is designed to eliminate, and it’s the metric that will tell you whether 15x holds in your environment.

Bring your standard operating procedures to the table. If your runbooks live in someone’s head or on a stale wiki, no agent can follow them. Codifying them is prerequisite work for operator-class autonomy, and it pays off regardless of which vendor you choose.

The dashboard isn’t dead yet, but for the first time, the industry has a credible answer to what replaces it.

Netskope Inc.’s latest quarter delivered a solid beat and a raise, but the company’s story goes beyond the numbers. It demonstrated improving operating discipline, a meaningful rebound in net-new annual recurring revenue, and early evidence that artificial intelligence security could become a material expansion engine — even if it is not yet a near-term revenue cure-all.

The market’s response to Netskope has been surprisingly cautious. The company is still early in its public life, investing heavily in its NewEdge infrastructure and navigating AI’s uncertain economics. Yet it is puzzling that the stock remains below its $18 IPO price despite a quarter that showed accelerating revenue, improving retention, a raised full-year outlook and better-than-expected operating leverage. Netskope closed at $13.75 before the earnings release and traded at $15.51 afterward, still below the IPO price.

AI should create a “rising tide” that lifts the security sector. I believe one challenge for the company is that the security industry is a confusing mesh of companies, many of which partially overlap, so creating an “apples-to-apples” comparable for Netskope isn’t easy. This is where time in the market will help it, as the investor community will have more data points to connect back to historical and future trends. With that being said, here are five things investors may have overlooked:

1. Net-new ARR finally turned the right way

The headline numbers were strong: Fiscal second-quarter revenue rose 29% year-over-year, to $220.5 million, beating the company’s guidance range, and annual recurring revenue reached $899 million, up 27%. But the more important metric was net-new ARR: $54 million, up 9% year-over-year.

That matters because net-new ARR had become one of the most important parts of the Netskope story after it declined in the prior quarter. In her earnings note, Rosenblatt Securities analyst Catharine Trebnick characterized the result as a pivotal positive inflection, since $54 million beat consensus by $6.7 million and marked a 23-point swing in year-over-year growth, moving from the prior quarter’s decline.

For investors, this metric matters. Revenue is a lagging indicator in a subscription business. Net-new ARR more closely reflects whether sales execution, platform demand, and customer expansion are moving in the right direction. Netskope’s quarter does not erase the execution question, but it provides evidence that demand did not break structurally.

The company’s raised full-year revenue guidance reinforces that point. Netskope now expects fiscal 2027 revenue of $888 million to $892 million, up from the prior $879 million to $883 million range. That is not an aggressive raise, but it is noteworthy: Management is signaling confidence while retaining room for the seasonally stronger second half and the ongoing ramp of its sales organization.

2. The platform model is taking hold

Investors often focus on the AI narrative and overlook the less flashy data points that show Netskope’s broader platform strategy is working. Net revenue retention improved to 114%, gross retention reached a company record, and remaining performance obligations grew 36% year-over-year, to $1.35 billion.

The underlying driver is broader product adoption. Fifty-nine percent of customers now use four or more Netskope products, up from 51% a year earlier; 41% use at least five products, up from 35%. The number of customers contributing more than $100,000 in ARR grew 23% to 1,686, representing 87% of total ARR.

That is what investors should expect from a company positioning itself as a converged security, networking, data protection and AI-security platform. Growth doesn’t depend solely on landing new logos. It increasingly comes from cross-selling and consolidation within the installed base.

As Chief Executive Sanjay Beri (pictured) stated on the earnings call, enterprises often start with “one or two core use cases” and expand over successive years. “With average customers having four or five products, we have a lot of ability to upsell for many, many, many years,” he said.

Trebnick made the same point in her post-earnings research, arguing that “cross-sell, not new logos alone, is carrying net adds.” That matters strategically because a broader product footprint can make customers stickier and reduce the risk that Netskope becomes merely another point-product supplier in a crowded secure access service edge and security service edge market.

3. AI security is promising — but investors should model it carefully

The company’s AI security pipeline is gaining traction. About 13% of the pipeline is already in or entering proof-of-concept deployments. Netskope cited early wins in financial services, technology and manufacturing for products such as Agentic Broker, AI Guardrails, AI Gateway and its AI Command Center.

However, the key word is pipeline. Beri was emphatic that enterprises will follow their normal buying process: evaluate the technology, conduct a proof of concept, secure budget, obtain executive approval and move through procurement. That process generally takes six to 12 months.

“We see the POCs building,” Beri said. “And as a result, we see the ARR building.” This is an attractive story, but it is not a reason to extrapolate a sudden AI-driven growth explosion into the next quarter.

The more nuanced point is that Netskope is creating a potentially different monetization model. Beri said an agent is not a user, so it makes little sense to price certain AI-security capabilities per seat. Agentic Broker is priced by transaction, while some AgentSkope capabilities are priced on an outcome basis — for example, resolving relevant data-loss-prevention cases amid thousands or millions of alerts.

That creates upside, but it also raises a new investor question: Can Netskope help customers forecast, govern and accept transaction-based spending? In a post-earnings discussion, Beri told me that Netskope intentionally avoided token-based pricing because customers did not understand it well. Instead, the company provides usage visibility, calculators based on observed activity, and relatively easy top-up mechanisms.

The comparison is not with traditional seat-based security software. It is more like the enterprise transition from fixed infrastructure purchases to consumption-based cloud economics, where cost visibility and FinOps discipline become part of the product experience.

4. Margin gains are real, but cash conversion remains the test

Netskope reported a non-GAAP operating margin of negative 9%, an 11-percentage-point year-over-year improvement and well ahead of its guidance. Gross margin reached 77%, up about two points from the prior year.

That improvement is worth highlighting. In a market that has become far more demanding of efficient growth, Netskope demonstrated that revenue scale and a common platform architecture can drive operating leverage. The company also reduced its workforce by about 5%, reallocating resources to AI infrastructure, tokens, research and development, and other areas it considers more strategic.

But investors should distinguish accounting and operating leverage from free-cash-flow conversion. Netskope generated negative free cash flow of $29.8 million in the quarter and expects capital expenditures to be about 4% to 5% of revenue for the year, primarily to continue investing in its NewEdge private cloud network.

On the earnings call, Beri said the company is transitioning customers to annual billing, a move that temporarily delays cash collections but should improve visibility into customer commitments. The transition should be substantially complete by mid-next fiscal year.

5. The stock’s discount reflects execution risk, not a lack of narrative

At $13.75 before earnings, Netskope traded well below its $18 IPO price, despite operational improvements. This reflects an enterprise value-to-sales multiple of about six times, well below the median of 8.4 times and the 13.9 times average among similar publicly traded security companies.

The company is expanding revenue by nearly 30%, improving retention, lifting guidance, demonstrating platform-led cross-sell, investing in a globally differentiated network, and establishing an AI-security position that fits its existing inline traffic-enforcement architecture.

During my call with Beri, we discussed security’s role in AI. He believes, as do I, that the right security platform can enable customers to move faster with AI rather than be a traditional bottleneck. “Our goal is to let organizations leverage AI quickly while minimizing risk,” Beri said.

Though I am not a stock analyst, that statement is a strong investment thesis. Netskope is trying to become less a provider of security controls and more the trusted control plane that lets enterprises deploy cloud, generative AI and autonomous agents at speed. The latest quarter did not prove the entire thesis, but it did show more evidence than the current share price appears to acknowledge.

At its first Horizon customer and partner event this week, Equinix Inc. argued that the architecture of enterprise artificial intelligence is being reshaped by a simple yet hard-to-answer question: Where should inference run?

For the past several years, much of the AI infrastructure conversation has centered on the supply and cost of accelerated computing. The focus has been on graphics processing units, AI factories, training clusters and the unprecedented capital buildout required to support them. Those remain crucial. But as enterprise AI shifts from experiments to real applications, the more immediate operational challenge is distribution.

Data resides across multiple clouds and enterprise systems. Users, sensors and business processes are dispersed across regions. Models may be proprietary, open source, fine-tuned or delivered as services. AI agents will call other agents, tools and application programming interfaces across organizational and infrastructure boundaries. In this environment, compute is only one component of a system that must be connected, governed and continuously optimized.

That is the premise behind the two announcements Equinix made at Horizon: Equinix Fabric One, an intent-driven, managed any-to-any connectivity service, and Equinix Inference Exchange, a distributed inference offering built with Nvidia Corp. and Together AI Inc.

Combined, they represent an effort to extend Equinix’s traditional role as a neutral interconnection provider into the AI era. The company aims to make its global footprint, dense ecosystem and Fabric platform the place where enterprises not only collocate and interconnect infrastructure but also assemble and operate distributed AI environments.

From connections to outcomes

Fabric One is the more foundational announcement. Equinix describes it as a managed service for connecting distributed enterprise, cloud and AI environments. The key change is the operating model: Rather than requiring customers to design, provision and manage a collection of discrete network services, Fabric One is meant to let them state the business or application outcome they need and have the platform compose the underlying connectivity.

In practical terms, an enterprise might specify that it needs a secure, resilient, low-latency connection among a data source in one metro, an inference provider in another, a cloud-hosted application and a branch, factory or other edge location. Fabric One is intended to orchestrate routing, cloud connectivity, encryption, redundancy and failover as a single managed service.

That is a significant shift from the project-oriented networking model still familiar to many large enterprises. Building a distributed AI workflow today can require separate teams and tools for cloud interconnects, carrier services, security controls, network routing, colocation operations, GPU capacity and application integration. Each additional model provider, cloud region or inference endpoint can add another engineering and governance task.

In his opening remarks, Chris Audie, Equinix’s chief product officer, set the stage for the event when he stated, “For decades, enterprise networks have been built one connection at a time for each partner and provider they depend on. That approach doesn’t scale in a world of distributed AI that demands dynamic, flexible and real-time connectivity.”

Fabric One will use open connectivity specifications developed in collaboration with Amazon Web Services Inc. and Google Cloud, its lead integration partners. The service is expected to enter beta later this year, with general availability planned for 2027, initially in North America.

Historically, Equinix’s value proposition has focused on easier provisioning, but this release goes beyond that. It aims to turn the network into an adaptive control plane for AI-era infrastructure. Customers can submit requests via a portal, APIs, automation workflows, agents or natural-language prompts. The platform then continuously validates the network and manages policies, encryption, resilience and failover.

That model is particularly relevant for agentic AI. A future enterprise application will not necessarily make a fixed call to a single large language model running in a single cloud. It may dynamically coordinate several models and services, retrieve information from multiple data domains, apply policy checks and invoke specialized agents in different locations. In that scenario, the network cannot remain an opaque, static utility. It must become programmable, policy-aware and responsive to the application’s needs.

Why inference changes the equation

Equinix Inference Exchange addresses the complementary problem: how enterprises can bring production inference capacity closer to their data, applications and users without building the full stack themselves.

The offering combines Nvidia Enterprise Reference Architectures and AI infrastructure, Together.ai’s platform, which supports more than 200 open-source models, and Equinix’s global data-center and interconnection footprint. Equinix says the service will support multitenant deployments for shared efficiency and dedicated single-tenant environments for workloads requiring isolated capacity. Availability is planned for the first quarter of 2027.

This announcement matters for inference economics and user experience. Training can be concentrated in massive, centralized clusters because it is a long-running, batch-oriented process. Inference is different. It sits on the live path of an application, customer interaction, operational system or agentic workflow. It can be sensitive to latency, data movement costs, sovereignty requirements and the cumulative cost of every token generated.

During her keynote, CEO Adaire Fox-Martin (pictured, left) made that point by identifying four pressures converging on enterprise information technology: fragmented compute, the rise of machine-to-machine traffic, the challenge of achieving predictable AI returns, and increasingly stringent data sovereignty requirements. “Where compute runs is becoming as important as the compute itself,” she said.

An enterprise does not necessarily need every inference request to reach the largest possible centralized model. In many cases, it needs a model that is sufficiently capable, properly governed, affordable and physically close to the interaction or data source. A manufacturing application may require response times near the factory floor. A customer-facing service may require inference near a metropolitan population center. A regulated financial, healthcare or public-sector organization may need to keep data and inference processing within a specific jurisdiction.

Nvidia CEO Jensen Huang (right) took a break from the G20 summit to pop in on video. He emphasized the same architectural shift during his Horizon discussion with Fox-Martin. “The world of AI is going to be uncentralized, fundamentally uncentralized,” he said, describing AI applications that will orchestrate agents that access data and services across multiple clouds and locations.

Huang also explained the importance of proximity, where compute needs to be close to the “action” — sensors, customers, factories, cell towers, airports or other locations with high data volumes and where response time matters — while prior knowledge and long-term memory can reside farther away. Equinix’s distributed footprint and interconnection fabric are designed to enable enterprises to support both realities simultaneously.

A simpler path to production

The most compelling aspect of the Equinix strategy is not that it eliminates AI complexity. It does not, and no vendor can. Instead, it packages some of the difficult infrastructure decisions into a more consumable, neutral operating model.

Inference Exchange is designed to provide pre-staged, pre-connected Nvidia-based AI infrastructure through Equinix, connected to clouds, networks and AI providers via Equinix Fabric. Together.ai adds model flexibility, including access to a broad catalog of open-source models. That combination matters for enterprises that want to avoid an all-or-nothing choice between proprietary frontier models and do-it-yourself open-model operations.

Huang argued that enterprises will use both. Closed models will remain important for frontier capabilities, he said, while open models increasingly let organizations fine-tune systems, incorporate domain-specific knowledge, meet regulatory or sovereignty requirements, and retain control over strategically important AI capabilities.

For customers, this could reduce the time and coordination burden of moving from an AI pilot to a production distributed deployment. Instead of separately sourcing GPU infrastructure, establishing colocation, arranging cloud and network connectivity, integrating inference services and building operational guardrails, they can use a more integrated platform.

Fox-Martin said Equinix designed Inference Exchange to be “production-ready, fully connected, and built to scale from day one.” She added that the company’s goal is to provide Nvidia compute “pre-staged and pre-connected within Equinix’s ecosystem,” with Together AI offering model choice and avoiding lock-in at the model layer.

That is an attractive proposition, particularly for enterprises whose AI strategy will be heterogeneous by necessity. A bank, manufacturer, retailer or global SaaS provider may use proprietary models for specific tasks, open models for specialized workloads, multiple clouds for different applications, and localized inference for latency or compliance reasons. The strategic risk is not merely vendor lock-in; it is the operational friction of stitching those choices together.

The bigger bet

Equinix’s differentiator is its neutrality and ecosystem density. The company operates more than 280 data centers across 77 metros, provides 230 cloud on-ramps and interconnects more than 10,500 businesses. It also reports that eight of the top 10 AI model providers and nine of the top 10 AI clouds have active deployments on its platform.

Those figures alone do not guarantee that Fabric One or Inference Exchange will become essential enterprise services. The proof will come from execution: the depth of cloud and AI-provider integrations, the maturity of automated orchestration, price transparency, service-level commitments, geographic availability and the ability to make intent-driven networking genuinely simpler without masking critical controls.

But Horizon signaled a notable evolution in Equinix’s story. The company is no longer positioning connectivity as a supporting service beneath cloud and AI. Instead, it is positioning connectivity as the enabler of a distributed, multicloud, multimodel AI strategy.

For enterprises entering the inference era, that is the right problem to solve. The winning AI architecture will not be defined solely by who owns the most compute. It will be defined by who can place the right compute, model and data path in the right location — and adjust that decision as performance, cost, sovereignty and business requirements evolve.

As Nvidia Corp. announced this morning that it has agreed to acquire Hugging Face Inc. for $12.93 billion — one of the largest acquisitions in the company’s history — Chief Executive Jensen Huang promised that the platform will retain its brand, its leadership and, according to both companies, its neutrality.

“Hugging Face will remain an open platform for the entire AI ecosystem,” Huang said. “Nvidia compute will not be required to build on or deploy through Hugging Face.”

That’s the news. The more interesting question is what it means, and I see four things worth technology leaders’ attention:

The distribution layer of AI just got a deep-pocketed owner

Hugging Face has become the de facto front door to open AI. The platform hosts more than 3 million models, 500,000 datasets and 1 million applications, used by more than 18 million developers and researchers and more than 200,000 companies. For most enterprises, the practical path to an open-weight model — evaluate, fine-tune, quantize, deploy — runs through that hub.

Yet the company operating it was, by AI standards, small. Hugging Face was valued at $4.5 billion in its last disclosed round in 2023 and is estimated to have reached roughly $150 million in annualized revenue this year. Running the world’s model registry on that revenue base is a structural problem. Storage, bandwidth, security scanning, provenance tracking, evaluation harnesses and inference endpoints are all expensive, and open-model growth is not slowing.

On the press call, Hugging Face co-founder and CEO Clement Delangue said the goal is to grow from 18 million builders to 100 million “in the next few years” and that Nvidia’s resources let the company “think about the next 10 years and really optimize for maximum impact.” Justin Boitano, Nvidia’s vice president of enterprise AI, described the deal as helping Hugging Face “achieve stuff that we don’t think either of us would have been able to achieve alone.”

The reality is that the critical distribution layer of open AI was underfunded relative to its importance, but it isn’t anymore. Nvidia’s stated investment areas, including reliability, safety, model evaluation, inference and deployment, map almost exactly to the friction enterprises cite when moving an open model from a notebook into production.

Why this accelerates enterprise AI adoption

The bottleneck in enterprise AI stopped being model quality some time ago. Open-weight models now approach the quality of frontier proprietary models at far lower cost. This week I attended JFrog’s swampUP and the Equinix Horizon events, and in both cases the topic of open-weight models came up, and it’s fair to say enterprises have embraced them as a viable, lower-cost alternative.

Today, the bottleneck is everything around the model: which one to pick, how to prove it’s safe, how to validate it against your data, how to serve it economically and how to defend that choice to a risk committee.

These are all platform problems, not model problems. Better evaluation tooling shortens selection cycles from months to weeks. Real supply-chain security on the hub, including signing, provenance and scanning, removes an objection that has stalled countless open-model projects in regulated industries. Turnkey inference paths close the gap between “We tested it” and “It’s serving customers.”

If Nvidia funds that layer at scale, the effect on adoption is meaningful and mostly independent of whose silicon runs underneath. Boitano’s point on the call was that open models let “every company build their own domain-specific intelligence” and own the competitive advantage of their business. That thesis only works if the tooling around the models is enterprise-grade. Today, it mostly isn’t.

The open-weight thesis gets an institutional backer

Nvidia’s position here isn’t opportunistic. In July, Huang used his first-ever X post to share an open letter on open weights and U.S. AI leadership, co-signed by more than 20 organizations. Nvidia is also the largest contributor of open models and data to Hugging Face, with more than 500 models and 250 open datasets published on the platform.

Delangue made a strong argument on the call in support of the acquisition, and it’s one chief information officers should internalize: Power in AI is concentrating in proprietary application programming interfaces, and an open model hub is “structurally, by definition, a de-concentration platform.” He said that reasoning was part of Hugging Face’s thinking this summer.

There’s obvious tension in a $5.4 trillion company positioning itself as the champion of de-concentration. But Nvidia’s incentives genuinely differ from a model lab’s. It doesn’t need any single model to win; it needs many models trained, fine-tuned and served everywhere. A thriving ecosystem of thousands of specialized models generates more aggregate compute demand than a handful of dominant closed APIs, and it hedges Nvidia against Meta Platforms, OpenAI and Microsoft building their own accelerators.

Will Nvidia actually keep it open?

This is a fair question, and it came up on the call. Some analysts and developers worry that Nvidia will gradually neglect rival hardware until its chips become the only practical choice. Boitano explained that model weights are just numbers that run on open inference runtimes such as vLLM and SGLang, which already target every available accelerator.

Delangue’s comments were more matter-of-fact, as nearly everything Hugging Face does is open source: “Everyone can fork our open source if they’re not happy about it.” That fork risk is the real governance mechanism, and it’s stronger than any promise. Nvidia paid nearly $13 billion for trust and reach, both of which evaporate the moment the platform tilts. You also don’t get from 18 million to 100 million builders by narrowing hardware support.

Still, it’s important to watch Nvidia’s future actions rather than simply accepting the talking points. Does non-Nvidia optimization remain first-class on model cards? Do competing accelerator vendors maintain equal footing in documentation and integrations? Does the Hugging Face team retain independent decision rights over hub policy? These are measurable and worth watching over the next 12 months.

More than a silicon company

Nvidia is gaining market influence as much as boosting earnings. We have seen this before. Mellanox made Nvidia a networking company in 2019. Run:ai made it an orchestration company. CUDA, NeMo, NIM and AI Enterprise made it a software company. Last year’s $900 million-plus deal for Enfabrica’s team and technology extended the systems story. Hugging Face makes it a developer platform and distribution company.

Nvidia’s moat has never really been the silicon; it has been the accumulated layers above it. Owning the place where 18 million developers discover and choose models is the most upstream position in the AI stack — earlier than the cloud, the framework and the chip.

For enterprise buyers, the practical takeaway is to treat open models as a first-class option in AI architecture, not the budget alternative. The tooling gap that made them harder to adopt is about to get $13 billion of attention. Just keep a genuine multi-accelerator exit path in your design, because the best insurance against a platform tilting is never needing it not to.

Nvidia Corp. delivered another mind-boggling quarter Wednesday. Total revenue reached $96 billion, up more than 100% year over year, with data center revenue at $89 billion and a third-quarter guide of $108 billion.

More striking than the quarterly numbers was the guidance Chief Financial Officer Colette Kress dropped midway through the call: Fiscal 2028 revenue should grow roughly 70% — and that is a supply-constrained figure. Customer forecasts, she said, “point to our growth doubling next year.”

Kress framed the quarter by explaining: “The surge in AI demand is driving a global infrastructure buildout, supported by an expanding and diverse set of growth opportunities spanning hyperscalers, AI labs, AI natives, enterprises and sovereign customers.”

That gap between roughly 100% demand and roughly 70% supply is the single most important fact for enterprise information technology leaders to internalize this quarter. Everything else, including pricing, lead times, architecture choices, even the reported Hugging Face acquisition, flows from it.

Here are five thoughts on the quarter and what they mean if you buy, budget for, or operate infrastructure.

1. Supply is the ceiling, and that changes how you procure

For two years, the industry has treated the artificial intelligence trade as a demand question. On the call, Nvidia reframed it as a logistics question. Chief Executive Jensen Huang (pictured) stated that supply constraints will persist “at least through the end of fiscal year ’28.” When asked to rank the bottlenecks, he declined to single one out: “Our entire supply chain is challenged.” He said capacity won’t arrive “in just an instance in time, but it’s going to come online every day.”

Huang was unusually candid about why Nvidia broke its own rule and guided a full year out: “It is the case that we’ve never forecasted or guided to a year in advance. And even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%. And we’re going to keep working with our supply chain to increase that. But what we wanted to do was be consistent with everybody, from our customers, our shareholders and our supply chain, so everybody sees the same view.”

One could view his statement as supply chain management dressed up as guidance, but it shows how tightly the entire AI ecosystem is now integrated. Nvidia’s response has been to move further upstream and downstream than any other chip company. It’s now involved with memory suppliers, power generation and land and shell developers. It secured 4.25 gigawatts of capacity at SoftBank Energy’s Portsmouth campus for OpenAI alone, a site Kress said could support multiple upgrade cycles over 20 years.

The lesson for enterprise buyers is not to copy Nvidia’s strategy but to adopt its time horizon. Huang noted that land, power and shell are “often two to three years out.” If your AI capacity planning is annual, that’s the wrong time frame. Multiyear allocation commitments, named-quantity contracts with OEMs and colocation partners, and explicit fallback tiers, such as Blackwell instead of Vera Rubin or one region instead of another, should now be standard in every AI infrastructure RFP. In a supply-constrained market, contract certainty is worth far more than a few points of discount.

2. The unit of purchase is no longer a GPU. It’s a gigawatt

In my opinion, Nvidia’s most important strategic statement was the escalating revenue opportunity per gigawatt of data center capacity — roughly $18 billion with Hopper, $25 billion with Grace Blackwell, and about $40 billion with Vera Rubin. This now spans the Vera central processing unit, Rubin graphics processing unit, NVLink, InfiniBand and Ethernet, and the new Groq language processing unit. Huang explained that the direction from here is “higher,” with the theoretical ideal being “infinity per gigawatt.” For context, he pegged general-purpose computing in the Moore’s Law era at $3 billion to $5 billion per gigawatt.

Kress further elaborated on the repositioning: “Today, we’re not just selling the best chips; we’re selling a full-stack AI factory platform that delivers superior economics for customers and captures a larger share of the data center TAM.”

From an investor perspective, Nvidia methodically expands into adjacent line items to drive growth. Networking set another record, with revenue up 18% sequentially and Spectrum-X Ethernet up 2.6x year over year. Kress claimed Nvidia is now “the largest and fastest-growing network company in the world.” The Grace CPU has surpassed $5 billion on a trailing-12-month basis. Vera is in full production, with shipments underway to Oracle Cloud Infrastructure and, this quarter, to AWS. Nvidia expects CPU revenue to more than double in fiscal 2028.

For IT buyers, this is the quarter to stop evaluating accelerators in isolation. The bill of materials for an AI cluster increasingly comes from a single vendor’s catalog, and the economics are attractive because the co-design is real and works: Vera Rubin reportedly delivers 30 times higher throughput per megawatt and 35 times lower token cost than Grace Blackwell Ultra. It has already drawn purchase orders from every major hyperscaler, AI cloud and system original equipment manufacturer.

But the procurement discipline that served networking buyers for 20 years still applies. Benchmark the Vera CPU against x86 on your own agentic workloads, not on SPEC. If you are looking for a “best of breed” in an AI factory, understand exactly what you are trading away for the convenience of a turnkey solution.

3. The enterprise is now half the story

Perhaps the most underappreciated part of this business is non-hyperscalers, which Nvidia calls ACIE, spanning neoclouds, sovereigns, enterprises, the edge and air-gapped data centers. This segment hit $40 billion in the quarter, up 25% sequentially and 138% year over year, and will represent “roughly half” of the data center business. Huang described that half as “invisible to everybody” because those customers don’t design custom silicon or buy chips one at a time. “They really need an entire factory platform built for them,” he said, and that plays to Nvidia’s strength.

Kress has talked about the benefits Nvidia brings to these customers. “One platform, fungible for every model and workload, durable for the entire life cycle of AI,” she said. “That combination of performance, fungibility and durability is what makes Nvidia the productive and financeable compute infrastructure.”

The named enterprise deployments are helpful because they’re in production, not pilots: Hudson River Trading and Jane Street in quantitative trading; Samsung Electronics using cuLitho for a claimed 20x gain in computational lithography; and Bristol Myers Squibb building a Vera Rubin AI factory, following Roche and Lilly. On-premises automotive revenue alone reached $8 billion on a trailing-12-month basis, with financial services, manufacturing and healthcare adding another $7 billion.

Some enterprise buyers have viewed AI infrastructure as the domain of hyperscalers, but that argument is quickly fading. Nvidia’s neocloud partners are on track to exit the year with 8 gigawatts of installed capacity, up from about 3 gigawatts at the end of 2025. The customer roster spans the globe, including Firebird in Armenia, Cassava across Africa, Yotta and Neysa in India, Firmus in Australia and YTL in Malaysia. Competitive, geographically local, sovereignty-compliant capacity is arriving faster than most enterprise data center plans can be updated. Huang’s point that “a country or region can allocate land and power directly to a regional cloud partner in ways it never would to a foreign hyperscaler” is as much about data residency as it is about economics.

4. 2027 server pricing is going up. Plan for it now

Nvidia lowered its gross margin outlook to 74% in Q3, bottoming at 71% to 72% in Q4, and settling at 72% to 73% in fiscal 2028. It also explained the impact of memory pricing: “extreme pricing conditions in memory,” with increases that “exceeded our prior expectations and are headed even higher into next year.”

Kress added, “Memory scarcity today is being driven in large part by the AI build-out itself, unlike a component that simply raises our cost with no offsetting benefit. Tighter memory supply is a symptom of the same demand surge that’s driving our own growth.”

Enterprise IT leaders need to understand the implications. The world’s most powerful memory buyer cannot fully absorb DRAM inflation and is passing it through. You have far less leverage than Nvidia. Every server, storage array and edge device with memory will cost more in 2027, and AI systems will be hit twice, once on the accelerator and again on the host.

IT leaders should build meaningful component inflation into 2027 capital plans, lock in pricing wherever suppliers allow it and revive the unglamorous discipline of right-sizing memory footprints and improving utilization across existing fleets. The cheapest gigabyte in 2027 is the one you already own.

5. Agents are the demand function — and they never sleep

Huang put a number on the shift that most enterprise architects are trying to get a handle on. An agent consumes “probably 15 to 100 times” the compute of a human using the same model, due to reasoning, planning and multiple turns of tool use. He believes agentic usage crossed over human-prompted usage in the past month and sketched a future where Nvidia’s roughly 40,000 employees are joined by 400,000 or 4 million agents “running continuously.”

He then described the process as reflective: Agents run, update their own skill files and repeat — “a kind of loosely coarse-grained self-improvement.” Human-driven AI workloads are bursty and follow the workday. Agentic workloads are 24/7 background loads with essentially unbounded appetite. Capacity models based on interactive usage patterns will be off by an order of magnitude, and FinOps practices designed for cloud elasticity will need per-agent budgets, kill switches and token-level chargeback.

Huang’s explanation of why everyone is adopting AI was well thought out and easy to understand. “I think the most important thing for the industry is that, one, AI is now doing productive and useful work; two, AI is generating profitable tokens; and three, if we had more compute, we could generate more profitable tokens, which results in more profit for all of the services. This is the exact phase we’re at, which is why everybody is leaning in.”

He added a datapoint that should give every CFO pause, in both directions: “I heard the other day that return on invested capital is now less than a year. And we’re talking about $50 billion data centers.” If that holds, the AI buildout is rational. If it doesn’t, this is the number the whole cycle is resting on.

The Hugging Face report: Buying the onramp

The Information reported that Nvidia has agreed to acquire Hugging Face for $12.9 billion. At roughly 86 times an estimated $150 million in annualized revenue, this is not a revenue purchase. It’s a distribution purchase, and it would be Nvidia’s largest deal ever, dwarfing the $7 billion Mellanox acquisition.

Hugging Face is the default onramp to AI. It has about 13 million users, more than 2 million public models, and over 500,000 datasets, with verified accounts among more than 30% of the Fortune 500. Nvidia is already the platform’s most prolific big-tech contributor.

The simplification value is compelling for enterprise IT. Today, the path from “I found a promising open model” to “it’s serving tokens profitably in my data center” involves quantization decisions, kernel compatibility, serving frameworks, inference tuning and a lot of tribal knowledge. Owning the hub lets Nvidia collapse that path. Models arrive pre-optimized for your specific rack, with known throughput and cost-per-token profiles, and can be deployed to a DGX Spark on a desk or to a Vera Rubin cluster with identical tooling.

That is the difference between open models being viable for the Fortune 500 and open models being viable for the Fortune 5000. It also aligns with Huang’s stated view of why open models matter to enterprises: “You should rent intelligence, strong intelligence, smart intelligence wherever you can. But every major company and every country and every startup needs to build their domain-specific, proprietary AI. And the open models reaching frontier levels has made it possible, has enabled them to all do that.”

This only strengthens Nvidia’s seemingly growing competitive moat, because it doesn’t depend on a hardware advantage. CUDA is a developer moat Nvidia built itself. Hugging Face is a model-discovery moat someone else built, and it sits one layer above where Advanced Micro Devices Inc., custom XPUs and OpenAI’s reported Jalapeño silicon are trying to compete. Nvidia’s central claim is fungibility — Huang said it is “the only platform that runs every frontier model, whether it’s closed or open.”

Owning the place where 2 million models live feeds exactly the growth engine Kress described: a CUDA ecosystem that extends AI “into markets a single chip alone can never reach.” If Nvidia becomes the default distribution channel for open models the same way it is the default runtime, then the open-model wave that some investors read as a threat becomes another Nvidia growth vector.

My advice to IT buyers: Welcome the simplification, and hedge the dependency. Keep model artifacts, evaluation harnesses and fine-tuning pipelines portable across at least two serving stacks. Nvidia has earned its position as the safest bet in AI infrastructure. That is not the same thing as being the only one.

For the past two years, the generative artificial intelligence boom has largely been defined by a single market. Hyperscalers have dominated the market with massive capital spending.

According to CreditSights, the five largest U.S. hyperscalers — Amazon Web Services Inc., Microsoft Corp., Google LLC, Meta Platforms Inc. and Oracle Corp. — are on pace to spend roughly $700 billion to $800 billion in capital expenditures this year, with about three-quarters of that spending tied directly to AI infrastructure. An Allianz report also found that hyperscalers and large tech service providers now absorb about two-thirds of all AI spending. The tech giants have spent billions building custom liquid-cooled mega-clusters to train frontier models, leaving everyone else scrambling for access to raw compute.

Though hyperscalers were early drivers of the AI hardware gold rush, the market is shifting. The next phase of AI infrastructure growth is moving beyond the public cloud giants toward “distributed AI builders” — a market that includes sovereign clouds, neoclouds, AI service providers and, increasingly, mainstream enterprise data centers.

However, enterprises and neocloud providers face severe headwinds when building modern AI infrastructure. Massive-scale workloads, such as multi-agent frameworks and trillion-parameter models, require high-density, liquid-cooled environments. At the same time, enterprises face data sovereignty challenges because they cannot simply upload their core intellectual property to public frontier models without risking their competitive advantage. Finally, a persistent shortage of specialized network engineering talent makes deploying bespoke, complex graphics processing unit clusters a high-risk venture.

Today, Cisco Systems Inc. announced an expansion of its Cisco Secure AI Factory with Nvidia Corp., directly addressing these friction points. By expanding its partnership with Nvidia to introduce Nvidia-certified rack-scale architectures, Cisco is bridging the gap between hyperscale-grade AI capabilities and enterprise-ready operational simplicity.

Cisco extends the AI factory to rack-scale

Cisco’s announcement expands its core AI strategy in three critical areas:

1. Integrated rack-scale infrastructure: To address high-density AI workloads, Cisco is introducing pre-integrated, liquid-cooled rack architectures. To round out its compute portfolio alongside its existing UCS family, Cisco is partnering with Supermicro to deliver high-density, liquid-cooled systems, including support for architectures such as Nvidia’s NVL72. This enables Cisco to offer full-stack, turnkey AI data center pods backed by Cisco’s global supply chain, support and service level agreements.
2. End-to-end NCPRA compliance: Cisco announced support for the Nvidia Certified Partner Reference Architecture across front-end and back-end networking fabrics. Operating on its Silicon One and Nexus Spectrum-X switching platforms, Cisco stands out as a primary technology partner capable of delivering unified, NCPRA-compliant fabrics using both its own silicon and Nvidia’s back-end networking components.
3. Cisco Validated Infrastructure Services: To address deployment challenges and the global skills gap, Cisco introduced CVIS, an automated toolkit and service framework that accelerates deployment from “months to weeks.” Backed by a new 1,000-GPU internal engineering cluster, Cisco pre-validates full-stack performance before customer hardware even reaches the data center floor.
4. Unified operations: The entire infrastructure — from compute sleds to switches and security appliances — is managed via the Cisco Cloud Control platform, which integrates telemetry, observability and day-two operations into a familiar interface for network administrators.

Moving beyond hyperscalers: Bringing AI data centers to the enterprise

For enterprise information technology leaders, this news signals a shift in where and how enterprise AI workloads will run. Hyperscalers build infrastructure for scale-out public tenancy, with hardware disaggregated and managed by proprietary software control planes. Enterprise IT operates under entirely different constraints:

• Extending the enterprise operational model: Most IT organizations lack the specialized engineering teams needed to assemble custom liquid-cooling units, disaggregated fabrics and open-source orchestration engines. By integrating NCPRA-compliant rack-scale architectures into the Cisco ecosystem, IT teams can build standard, repeatable “AI Pods” using the same operational workflows, security policies and Nexus/NX-OS tools they have relied on for decades.
• Protecting data sovereignty and alpha: As agentic AI and localized fine-tuning replace simple, broad queries, enterprises want to bring models to their data, not the other way around. Turnkey, enterprise-class liquid-cooled racks allow IT to run dense, high-performance training and inference in-house or in colocation facilities under their own governance.
• Unified fabric and security: Rather than treating AI clusters as isolated “shadow IT” islands with separate networks, Cisco’s approach integrates backend GPU fabrics directly into enterprise security and observability frameworks. Security is fused into the network stack rather than bolted on after deployment.

Furthermore, as neoclouds and sovereign clouds expand regional AI hosting services on Cisco infrastructure, enterprise IT gains an operational “flywheel.” An enterprise can seamlessly burst workloads from its on-premises Cisco AI Factory to a Cisco-powered sovereign cloud, using identical networking policies, operational standards and security controls on both sides.

Strategic advice for enterprise IT pros

As AI infrastructure shifts from hyperscale science experiments to standard enterprise IT rollouts, tech leaders should consider the following guidance:

• Avoid “DIY” GPU clusters: As GPU densities rise and liquid cooling becomes mandatory, custom-building AI infrastructure by integrating disparate compute, storage and networking layers is high risk. Prioritize pre-validated, reference-architected stacks (such as NCPRA-compliant systems) to drastically reduce time-to-first-token and lower long-term mean time to resolution.
• Leverage existing network and operations skill sets: Do not create isolated, hyper-specialized silos within your IT organization solely to support AI. Focus on architectures that enable your existing network, operations and security teams to manage AI fabrics using familiar control planes (for example, Cisco Cloud Control and Nexus OS). Upskilling your current staff in AI-ready networking is far more sustainable than trying to hire scarce niche specialists.
• Plan for hybrid model mobility: Design your data center architecture with hybrid elasticity in mind. Ensure that whatever stack you deploy on-premises shares management and network lineage with regional neocloud or sovereign cloud providers. This ensures you can burst training or inference workloads externally without rearchitecting your security or networking posture.
• Demand full-stack support and SLAs: Hardware components in high-density AI clusters will fail — whether it is a liquid-cooling loop, a transceiver,or a GPU node. Avoid multivendor finger-pointing by selecting full-stack solutions backed by single-source, enterprise-grade day-two support and global service capabilities.
• Buy the support model, not the spec sheet. Rack-scale AI deployments fail on integration, not benchmarks. Ask explicitly who owns the escalation when a GPU tray, a cooling loop and a fabric link are simultaneously suspect. If the answer involves more than one throat to choke, keep pushing.

Final thoughts

The era when AI computing was limited to hyperscalers is coming to an end. Through partnerships such as those between Cisco and Nvidia, standardized, liquid-cooled, rack-scale architectures are now reaching mainstream channels, giving enterprise IT a blueprint for running high-density AI infrastructure with confidence, security and operational simplicity.

Physical artificial intelligence is emerging as the next major phase of AI. These systems not only generate content or analyze data but also perceive, reason about and act in the physical world.

The opportunity is massive, but so are the operational, data, latency and lifecycle-management challenges. That is why Amazon Web Services Inc. last month rolled out cloud-to-edge solutions for customers building these systems.

For the past few years, generative AI has largely lived in the digital realm. It writes, summarizes, codes, searches and converses. Physical AI extends that intelligence into robots, industrial equipment, cameras, autonomous mobile robots and other systems that must interpret changing conditions and act reliably in environments that are rarely as orderly as a chatbot prompt.

Traditional industrial automation and robotics were designed for highly structured, repeatable workflows. A warehouse robot might reliably carry a bin from point A to point B in a mapped, tightly controlled facility. But it has trouble when an object is misplaced, a pathway is blocked, a person enters its work area or the task goes beyond what it was explicitly programmed to handle.

Physical AI promises greater adaptability. Advances in foundation models, vision-language-action models, world models, reinforcement learning and simulation are giving machines a more capable “brain.” At the same time, edge inference, model compression and improved cloud-to-device operations make it more feasible to deploy those capabilities in real-world systems.

“As with the transition from early ChatGPT capabilities to today’s applications built on more robust reasoning models, the underlying models continue to improve rapidly,” Sri Elaprolu, director of the AWS Generative AI Innovation Center, told me at a recent briefing. “It’s no longer just large language models. It’s world models, vision-language-action models and models powering robotic fleets.”

That maturation is one of the key reasons physical AI is gaining momentum now.

From programmed machines to learning systems

The first wave of robotics was built around instructions. The next wave is being built around learning. Elaprolu drew an interesting contrast between conventional robotics and the physical AI systems now emerging. Traditional machines are programmed to perform a defined task or set of tasks. When the environment deviates from expectations, the machine often requires intervention. Physical AI systems start with a capable model, use sensors to understand conditions, and can learn from experience.

“The variety of things you can now do has expanded drastically,” Elaprolu said. “I don’t need to pre-program 50 tasks a humanoid will have to perform. I start with enough of a foundation and let the robot or device learn from real-world experience and adapt.”

That does not mean robots are suddenly ready to perform every human task. Far from it. Dexterous manipulation remains especially challenging. A human hand can grasp, rotate, fold, sort, adjust and recover from errors with remarkable ease. Replicating those abilities in a machine requires fine motor control, perception, tactile feedback, spatial understanding and the ability to reason through ambiguity.

AWS customer RLWRLD is tackling that problem with RLDX-1, an 8.1-billion-parameter robotics foundation model designed for five-fingered dexterity. The model combines vision-language understanding with proprioception, tactile and torque sensing, and is intended to work across single-arm, dual-arm and humanoid systems rather than being tied to a single hardware platform. RLWRLD trains on hundreds of terabytes of data collected from factory and service environments, underscoring a central truth about physical AI: The model is only one part of the problem. Data, compute, simulation and deployment operations are equally important.

The physical AI challenge stack

Building a working physical AI proof-of-concept is one thing. Operating a fleet of robots or intelligent devices in factories, hospitals, warehouses, retail stores or energy sites is another entirely.

The biggest hurdle is data. An LLM can learn from massive stores of text and code. A robot needs examples grounded in the laws of physics: objects, surfaces, lighting, geometry, friction, hand positions, contact forces and the countless variations that arise in real-world work environments. Data must be sufficiently diverse for a system to generalize beyond the exact conditions it encountered during training.

Config, another AWS partner, has built a robot-data pipeline with more than 200,000 hours of action data and is collecting roughly 20,000 hours per month. Yet even at that scale, the data-diversity problem persists. Each variation in lighting, table surface, object type or background can require additional human data collection.

AWS and Config addressed that bottleneck with a generative, multi-view augmentation pipeline. Using a post-trained version of Nvidia’s Cosmos-Transfer2.5 model, the system can re-render real demonstrations with varied lighting and surfaces while preserving object position, action labels, temporal consistency and the robot gripper’s appearance. In a reported out-of-distribution test, Config said that adding augmented data improved success rates from 8.3% to 75%, a roughly ninefold increase in robustness.

That is significant because physical AI will not scale up if people manually record every possible task variation. Synthetic and augmented data are becoming necessary complements to real-world data collection, not substitutes.

Simulation is another critical component. Training directly in the physical world is slow, expensive and potentially unsafe. A digital twin or simulated environment lets developers test policies, stress-test edge cases and identify failure modes before machines operate around people, equipment or inventory.

However, simulation itself is challenging. The physics must be sufficiently accurate, the sensor models realistic and the “sim-to-real” gap carefully managed. A robot policy that succeeds in a clean virtual environment may fail when confronted with reflective surfaces, variable lighting, network interruptions, worn equipment or cluttered workspaces.

During the call, AWS acknowledged that the ecosystem is still evolving. Customers can bring their own simulation tools or use platforms such as Nvidia Corp.’s Isaac Sim, while AWS provides elastic cloud infrastructure, data services and templates to spin up and down simulation environments as needed. The company does not yet offer a single branded simulation service for physical AI, but the direction is clear: Simulation will become a major control point in the physical AI stack.

Why cloud and edge must work together

Perhaps the defining architectural challenge in physical AI is deciding where intelligence should run. A large foundation model may be trained in the cloud using massive GPU clusters and enormous datasets. But a robot navigating a hospital hallway or a machine inspecting a production line cannot always wait for a round-trip to a distant cloud region. It needs low latency, resilience and, in some cases, the ability to operate during intermittent connectivity.

That creates a tiered architecture: cloud for large-scale training and orchestration, local edge infrastructure for site-level intelligence, and smaller, distilled or quantized models that run directly on devices.

“The model in the cloud acts as the big brain,” Elaprolu said. “As the model gets smaller and smaller, you can place it directly on the device. But the local learnings must flow back to cloud for overall improvement. That’s the critical step.”

That continuous feedback loop is what separates a static automation deployment from a physical AI system that improves over time. Devices must capture the right signals, models must be retrained and validated, and updates must be safely delivered back to the field. That means data governance, fleet management, observability, version control, rollback capabilities and security are not afterthoughts. They are production requirements.

AWS and Edge Impulse exemplify the hybrid approach. Their warehouse asset-tracking architecture runs lightweight object-detection models on edge cameras, then invokes a quantized vision-language model for deeper contextual interpretation only after an object is detected. Cloud services orchestrate natural-language queries and aggregate results across a fleet of devices. A worker can ask, “Where is the forklift?” and receive an answer based on real-time local perception, rather than waiting for raw video to be streamed and processed centrally.

The two-stage cascade model is important both economically and technically. Continuous lightweight detection conserves compute, power and bandwidth; higher-cost reasoning is triggered only when needed.

AWS’ role: Assembling the production stack

AWS is not trying to become a robot manufacturer. Its strategy is to provide the services, infrastructure, and partner ecosystem required across the physical AI lifecycle. At the model and training layer, customers can use Amazon EC2 GPU instances, Amazon SageMaker, Amazon S3, and high-performance storage such as Amazon FSx for Lustre. For model optimization, SageMaker and Amazon Bedrock support techniques such as distillation, while partners such as Multiverse Computing offer compression tooling to reduce memory, latency, energy and hardware demands at the edge.

At the deployment layer, AWS IoT Core and AWS IoT Greengrass provide connectivity, fleet management and edge runtime capabilities. In industrial settings, partners such as Galeo Tech bridge operational technology environments, including PLCs, SCADA systems, historians and protocols such as OPC UA, MQTT and Modbus, with cloud data and MLOps pipelines. The goal is not merely to deploy a model once, but to create a repeatable release process that can safely target specific device groups, monitor outcomes and roll back when field performance degrades.

AWS is also working to reduce the development friction that slows robotics teams. Its Kiro agentic development environment is designed to automate environment provisioning, training workflows, simulation setup and troubleshooting across physical AI projects. The company argues that the current setup-training-validation loop can consume days or weeks of engineering effort before a real-world trial even begins; its ambition is to compress that cycle by making the workflow more conversational and specification-driven.

The key takeaway is that physical AI will not be won by the largest model alone. It will be won by companies that can build a closed loop spanning real-world data collection, simulation, training, model optimization, edge inference, monitoring and continual improvement.

For enterprises, the immediate opportunity is not necessarily a humanoid worker roaming every facility. It is likely to be more focused: computer vision that detects defects, autonomous systems that move materials, robots that reduce repetitive hospital logistics, edge intelligence that locates assets, or machines that improve safety by taking on hazardous tasks.

Physical AI is rapidly moving from concept to operational reality. But as Elaprolu noted, there is not yet an “easy button” an enterprise can press to deploy it. AWS’ emerging role is to make that journey less bespoke — and eventually more repeatable — by providing customers with a common foundation spanning cloud-scale model development to real-time action at the edge.