Every week, the Esper team sends an edge technology topic straight to your inbox. Here’s what we’ve been talking about this month: 

The Drive-Thru is the New Data Center

Written by Cam Summerson

The AI battles you never see 

A few weeks ago, I talked about the upcoming gigawatt datacenters that Meta, Google, xAI, and Microsoft are working on. Today, I want to talk about what’s on the other end of that wire: the hyper-local AI processing units that run the models these gigawatt data centers train. 

If the data center is the brain, these nodes are the nerve endings. And right now, those nerve endings are getting massive upgrades on the QSR front. Drive-thrus are ripe for fresh technology, with brands like Wendy’s and McDonald’s leading the charge. 

Why? Because accuracy trumps all and latency kills sales. You can’t round-trip a voice order from a drive-thru in Detroit to a data center in Memphis in under 200 milliseconds. Suddenly, “no tomato” gets lost in the cloud, and your customer is pissed off because their triple meaty melt, plain and dry is covered in veg. 

The solution isn’t “move to faster 5G,” either. It’s to go on-prem with the language models. QSRs can no longer afford to be food service companies — they have to be tech companies with distributed edge computing solutions. The restaurants that win in 2026 will be the ones that inference right next to the fryer. 

Would you like AI with that?

The two biggest players in the space right now are Wendy’s and McDonald’s, and both have big expansion plans in 2026. The fun part is that they have the same cloud partner (Google), but they’re taking totally different approaches. 

McDonald’s is going fully in house and putting server clusters in restaurants. AI processing nodes in thousands of stores, just to process data. And it’s not just about real-time voice transcription. They’re using smart “Accuracy Scales” to weigh delivery bags and catch missing items before they leave the window. And that’s really just a portion of what they’re doing with AI. 

Wendy’s, on the other hand, is going all in on SLMs (small language models) running on bare metal. They’re optimizing software across their “FreshAI” platform and deploying across more and more stores. Where McD is moving to on-prem and leaning into more and more AI, Wendy’s is working the software angle with some pretty amazing results — they report an increase in profit margins and higher average sales because AI never forgets to upsell. 

A drive-thru readiness check 

Just because AI is taking over the drive-thru doesn’t mean every shop out there is already equipped to make that happen. Here are some considerations you can run right now to approach your next restaurant upgrade. 

1. Thermal Considerations: NPUs generate a lot more heat than a “dumb” display. Can your current enclosures handle an extra 15-20 degrees (F) of internal heat generation? In hot areas like Texas or Arizona, the same internal fans displays use right now probably won’t cut it during the lunch rush when it’s triple digits outside. 
2. Vibration: AI accelerators are beefy little guys. They require heavier mounting compared to SSDs or built-in GPUs. Pair that with the vibration drive-thru displays/speakers need to deal with (win, trucks, etc.), and you’ve got yourself a potential recipe for disaster. 
3. Power: It might go without saying, but you have to plan for the extra power draw. The last thing you need is for a brownout to reboot your menu board. 

There are plenty of other considerations here — space, cost, what you’ll do with all that data, etc. also come into play here. 

Can AI help fix the ice cream machine?

The McDonald’s broken ice cream machine is a meme for a reason. But using AI, McDonald’s is enabling in-house employees to do more in real-time using its “Ask Pickles” internal LLM. If an on-prem agent can help a 19-year-old crew member run through an ice cream machine repair sequence without calling a $500/hour technician, it’s clear that the edge isn’t just about selling more fries or nixing the tomatoes — it’s about operational uptime. 

The Myth of ‘Fully Managed’: Where Vendors Stop Owning the Problem

Written by Maddie Gainza

Most enterprise tech is sold as managed: Managed OS. Managed hardware, managed cloud, managed security agents. But the moment something breaks, “managed” suddenly means distributed responsibility. And distributed responsibility is just a polite way of saying nobody owns the outcome.

The ISV wrote the app. The OEM shipped the hardware. The cloud provider hosts the backend. The MDM enforces policies.

And when an update bricks devices in the field? Everyone points sideways. This is the dirty secret of modern infrastructure: most outages aren’t caused by bad technology. They’re caused by unclear ownership at the seams.

The blame triangle in action

We’ve all seen this movie: 

• An OEM firmware update causes devices to reboot intermittently.
• The app vendor says, “It’s not our code.”
• The cloud provider says, “Our systems are green.”
• The MDM vendor says, “The update came from the OEM.”

Meanwhile, your frontline devices are down, tickets are piling up, and the only people actually fixing anything are your internal IT team, at 2:17 AM, with no rollback button and three vendors on hold. “In these multivendor environments, maintaining SLA compliance is challenging” at best and can be monumentally damaging at worst (TMCnet.com). Differing guarantees and unclear ownership can result in service loss, financial penalties, and reputational harm (TMCnet.com). 

This isn’t a one-off failure. It’s structural. The industry is optimized for speed to market and modular ecosystems. What it didn’t optimize for was clear, enforceable accountability once systems are live.

The vendor accountability gut check

Before your next renewal, rollout, or “managed” purchase, ask these questions. If you can’t answer them instantly, you already have a risk gap.

1. Who owns rollback? If a vendor ships a bad update, who can actually reverse it without manual intervention?
2. Who defines “healthy”? Is device health measured by uptime, policy compliance, app behavior, or all three? And who sets the thresholds?
3. Who sees the blast radius first? When something goes wrong, which system shows you which devices are impacted, not just that something is broken?
4. Who acts without a ticket? Can remediation happen automatically, or does every fix start with a human escalation?
5. Where does the SLA actually end? Read the fine print. Most SLAs stop at availability of the service, not functionality of your fleet.

If ownership isn’t explicit, it will default to you.

“Fully managed” is a sales term. Operations need enforcement.

The strongest teams don’t buy ecosystems because they look unified on a slide. They buy systems that enforce shared responsibility in production. That means:

• One control plane that sees the full fleet
• Enforced baselines, not best-effort configs
• Rollback and staged change as non-negotiables
• Partners aligned around outcomes, not tools

Because when vendors stop owning the problem, the business still expects you to deliver uptime. And the only way to do that at scale is to demand more than “managed.” You have to demand accountability that survives failure.

If your vendors can ship a change but can’t undo it, they aren’t managing anything. They’re just delivering risk faster. Before the next update, ask the uncomfortable question: When this breaks, who actually owns the fix? If the answer isn’t clear, the risk already is.

Mind the Skills Gap

Written by Ali Clawson

The edge computing market is on track to nearly double by 2031 (from $228B to $424B), but the single biggest drag on that growth isn't hardware costs or connectivity — it's the humans trying to run it all. 

Market research from Mordor Intelligence highlights the single biggest non-security drag on edge adoption: the “skills gap” in deploying and managing heterogeneous edge stacks. In plain terms, there aren’t enough operators who can actually run this stuff once it's in the field.

The gap shows up in four edge-native capability failures. Let’s break them down.

From the field: The activation bottleneck

The industry is at an inflection point. While AI and edge adoption are top priorities for 2026, the human element is where things get tricky.

• Fleet orchestration at scale: Pushing a software update or security patch across thousands of devices sounds straightforward until those devices aren't identical. Different OSes, hardware generations, failure modes — you name it — mean more blind spots at scale.

Mid-sized manufacturers are already straining under this complexity. The fleet sizes planned for the next wave of deployments will surface brittle processes along the way. 

• IT/OT convergence: Edge deployments live in the physical world: factory floors, utility infrastructure, remote industrial sites. IT teams understand networks and cloud infrastructure. But OT runs on industrial protocols and hardware designed to stay in the field for 10–15 years, not a standard three-year IT refresh cycle.

The software teams don’t always understand the constraints of industrial hardware. The hardware teams don’t always operate in modern software workflows. That gap shows up in both hiring and in how incidents get handled.

• Network variability: Cloud and data center work assumes reliable connectivity. Edge doesn't offer that. Designing for that reality — eventual consistency, offline queues, safe state enforcement — is an operational discipline most teams haven’t built yet. And with on-prem deployments still dominating edge infrastructure, this isn’t an edge case. It’s the default.
• Protocol diversity: 5G, Wi-Fi, LoRaWAN — often within the same deployment. Standards fragmentation is a known drag on the broader edge market, and the ops burden of supporting multiple communication protocols simultaneously falls squarely on the people managing the stack.

Taken together, these aren't gaps that a generalist hire closes. And with the average hiring and onboarding cycle running over eight months, organizations that rely on finding the right person externally are already behind.

Run the skills audit you've been avoiding

If your team feels stretched, you’re probably managing devices one by one instead of orchestrating them as a system. Run this mental checklist:

• Map the skill silos: Do you have one person who "owns" the kiosks and another who "owns" the handhelds? If so, you have a silo risk on top of a skill gap. 
• The “golden image” test:  Could you deploy a critical security patch to every device in your fleet, regardless of manufacturer, without a technician physically touching anything? If the answer is no, your management stack is too vendor-dependent.
• Prioritize edge-native literacy: Rather than training everyone on every device type, focus on containerization and zero-trust security. These horizontal skills transfer across almost any heterogeneous stack.

This week's action item: Pick one "orphan" device type in your fleet this week and document its update process. If it takes more than 10 minutes to explain to a new hire, it’s time for an abstraction layer that hides that complexity.

The skills gap in edge management is real, but it's also fixable and faster than most teams expect. The hard part isn't training. It's deciding to treat it as a priority before a deployment forces your hand.

Read more: The Edge Visibility Gap: Why IT Teams Go Blind When Devices Scale Up

Looking ahead to a 6G world 

Written by Kelsey Milligan

Better isn’t faster. It’s coordinated

Network upgrades have always been about speed. From 3G to 4G to 5G, we moved from side roads to highways to express lanes.

But 6G will land differently. It’s being designed for AI-first experiences, promising seamless connections from the cloud to the edge. 6G is not just about speed. It’s about coordination — connecting intelligence across data centers, regional infrastructure, edge servers, and end-user devices.

As faster networks plus smoother compute raise performance ceilings, operational expectations start climbing. They demand seamless rollouts, deeper visibility, and continuous deployments. That makes 6G less of an incremental upgrade and more of an architectural shift, and one that could take shape before the end of the decade.

From the field: Developments in 6G 

Chipmaker Qualcomm has positioned itself at the center of early 6G research. To its CEO, Cristiano Amon, 6G is more than bandwidth. It’s a connective layer between cloud AI and device AI, or “the network of intelligence connecting the edge and the cloud” as Amon puts it, built to coordinate distributed systems, not just move data faster.

That vision implies a redesign of compute for a 6G world. Chips and processors are already evolving toward it: A future where AI agents increasingly act on behalf of applications at the edge, learning continuously, working across apps, responding in real time. All of it dependent on fast, efficient silicon and infrastructure built for coordination.

Today’s architecture, tomorrow’s network

6G promises a network of intelligence, but intelligent networks don’t fix fragmented fleets. If your devices, telemetry, and data systems aren’t unified today, faster coordination tomorrow won’t help. Stay adaptable by pressure testing your readiness:

Unify before you amplify:  If you can’t see and update your fleet as one system, more edge intelligence will magnify fragmentation. Break device-management and security silos. Align tools, processes, and teams around a single operating model with DevOps-style workflows.

Turn telemetry into infrastructure: If the future is seamless intelligence between device and cloud, devices must be continuously observable, with clean, reliable data flowing upstream. That means collecting real-time telemetry across your fleet: CPU usage, connectivity status, performance health, all visible in a unified dashboard.

Automate like you’re already distributed: In a 6G world, devices will run more logic locally and sync more often. If updates take weeks, you’ll lose coordination at scale. Ask yourself: Are you shipping updates, configuration changes, and models through automated pipelines or manual rollouts?

Remember that no matter what network you use, edge AI demands faster edge updates. Our CEO wrote the guide: Mastering Software Deployments at the Edge. Read it for a practical primer on building the operational backbone that distributed AI will depend on.