A three-decibel reduction in antenna efficiency doesn't just weaken a signal — on NB-IoT or LTE-M networks, it can force the network to command up to 128 retransmissions of the same message, multiplying energy use. It's a detail most edge AI conversations never consider, because they're too busy talking about the chip.

That's the blind spot the industry keeps stepping into. Intelligence at the edge — devices that sense, decide, and act without waiting on the cloud — is one of the defining trends in IoT today. But intelligence must travel somewhere to matter, and the radio link carrying it rarely gets the same scrutiny as the silicon making the decisions. As Dermot O'Shea, CEO and co-founder of Taoglas, puts it: “An Edge AI device that can't reliably move data is just an expensive way to make a decision that nobody hears.”

The Physics that Compute Can't Escape

Processors keep shrinking, cheapening, and sipping less power with every product cycle. Antennas don't get that gift. They're bound by the physics of the wavelength they need to resonate with, and no amount of clever firmware changes that.

O'Shea, whose company designs antennas for demanding industrial and outdoor deployments, points to real-world examples where the toughest requirements come down to environment: “If there are extremes in temperature and if it's outdoors, it's probably going to be subject to impact.” He cites metro and traffic-light systems in Toronto, where even a bird strike on an antenna enclosure in freezing conditions can crack it open, forcing far more rugged designs than most teams have budgeted.

Shrink an antenna and efficiency, range, and bandwidth all suffer. Get the impedance match wrong, and a device can lose half its transmitted power before it ever leaves the board. Wrap the same design in a metal enclosure — standard for industrial or automotive housings — and range can collapse by around 30% because this screens the antenna.

Since battery life on constrained networks degrades in steps rather than a smooth curve, a mediocre antenna doesn't cost a business a little efficiency. Rather, it costs a multiple of the energy per message, because the network compensates for a weak link by repeating transmissions rather than boosting them gently.

There's a second, quieter problem sitting on top of this: self-interference. The second harmonic of a common cellular uplink band sits almost exactly on the GPS L1 frequency, alongside Wi-Fi, Bluetooth, and switching noise from nearby digital circuitry. The field symptom is usually described simply as “GNSS is flaky sometimes.” Teams can spend months chasing it as a firmware or coverage bug before realizing it's an RF layout issue from the start.

When Position Data Can't Be Trusted

Positioning brings its own emerging concern, and it's shifting from an accuracy question to a security one. GNSS jamming is more mundane than it sounds. In trucking, the dominant source is often cheap, driver-installed devices meant to defeat fleet tracking, which end up jamming every other receiver in range as collateral damage. Spoofing is the more dangerous cousin. Rather than knocking a signal out, it feeds a device in a plausible, precise, entirely wrong position that the cloud trusts without question.

This is squarely O'Shea's territory as an antenna specialist, and the framing worth adopting industry-wide is straightforward: Wherever position data drives a decision or a payment, GNSS should be treated as an untrusted input rather than simply an accuracy problem.

Real mitigations exist at the antenna level, from shaping radiation patterns so that they favor satellites overhead rather than jammers near the horizon, to using polarization discrimination and multi-band reception, since many low-cost jammers only target a single frequency.

The Power Budget Nobody Wants to Own

If antenna performance sets the ceiling, power budgeting is where most projects break, and this is where Adam Cousin, sales and business development manager at Telit Cinterion, speaks from two decades of watching the same mistakes repeat.

“People are doing essentially today what they've been doing for the last 20, 25 years," said Cousin, “and that is switching the system on, switching the modem on, transmitting a small amount of data, and then switching it off again.”

Layering AI onto that pattern only works if it reduces what needs to be sent, not if it adds another power-hungry process competing with the radio for the same battery. Cousin warned, “If the AI system uses more power than cellular communication, then you've defeated the object of using AI in the system.”

That tension plays out clearly in large-fleet deployments — utility meters, smart-city sensors, asset trackers — where shaving a couple of euros off a bill of materials by pushing compute to the cloud can quietly add tens of milliamp-hours of daily radio activity. At scale, across tens of thousands of units, that tradeoff doesn't save money; it just moves the cost from procurement budgets to truck-rolls for battery replacements years later.

The smarter approach, borrowed from recent cloud-positioning techniques, is to send a small raw signal snapshot uplink-only and let a server resolve the computation, rather than requesting a computed result back. That’s because a round trip on a constrained network is almost always more expensive than one-way traffic.

The same logic applies to network selection. A marginal impedance match forces a power amplifier to work harder for the same effective output, draining a battery that better RF matching would have preserved.

Making the Case for Judged Adoption

None of this is an argument against edge AI. It's an argument against adopting it uncritically. Cousin is candid about the hype surrounding the space: “The problem with the hype curve of any technology is that the people who are out there evangelizing about it will make you believe this technology A exists, B is available, and C is completely infallible,” when in reality, none of that is yet true for every use case.

His advice to teams starting out is blunt: "Ask yourself why. Ask yourself what you're trying to achieve, what you're trying to fix, improve, or monetize" before committing hardware and budget to an AI-enabled redesign.

The clearest wins so far sit in industries already comfortable with technical complexity. Pavesh Govender, head of Northern Europe sales at Vodafone IoT, points to fleet management,  using AI cameras for driver-behavior monitoring and predictive maintenance, and remote patient monitoring in healthcare, as two applications furthest along. Both run largely on mains or rechargeable power rather than pure battery deployments.

Crucially, Govender noted, these sectors “are not deploying AI because it's fashionable,” but because it solves a defined operational problem. Battery-only use cases remain the harder frontier, and progress there depends as much on RF efficiency as on model design.

Connectivity Once Data Leaves the Device

Getting a signal off the device is only half the job. What happens to that data afterwards matters just as much, particularly as connected products are increasingly expected to remain in the field for a decade or more. Security for a device with a 10- to 15-year lifespan can't rely on a single shared key across a fleet, since one extracted key would compromise every unit. Per-device cryptographic identity, encrypted transport, and hardware-protected keys are becoming baseline expectations.

Supply-chain posture is following the same trajectory. Certifications like ISO 27001 and defense-grade frameworks such as CMMC Level 2 are increasingly used as a proxy for whether a connectivity partner can be trusted with sensitive data at all.

That same interconnectedness is what unlocks the actual business value of edge intelligence, and it's the point Govender returns to repeatedly. The real payoff of a device making decisions locally, he argued, is “only really felt when this is interconnected into enterprise systems.”

A smart sensor that never reports back into an operational workflow is a technical achievement without an outcome attached to it. Scaling that across borders adds its own complications. Network capabilities and regulatory environments can vary significantly between markets, making it difficult to deliver a consistent experience across geographies. Addressing those differences remains one of the most challenging aspects of global IoT deployment.

Designing the Nervous System, Not Just the Brain

The most forward-looking developments — federated learning across device fleets, swarm coordination between robots or drones, split compute between a device and a nearby edge node — all share one buried assumption: That a wireless link exists whereby it’s good enough to carry model updates and shared observations reliably, within a usable time window. Take that link away, and the entire architecture collapses back into a single disconnected device running on a stale model.

That's the shift worth internalizing, and it's one all three specialists converge on from their own vantage points. Govender stated the underlying philosophy: “I would not look at the components of an edge AI solution as discrete components, but again, as an entire system...the sum of the parts would always be greater than each component.”

O'Shea agreed, echoing that same conclusion from the antenna side: “I like that saying — it's one system.” And Cousin, closing the loop from the module and connectivity side, brought it back to discipline over hype: “Get your ducks in a row when it comes to this stuff, and don't just leap, because it could be quite a costly mistake.”

As intelligence moves from the chip to the network of devices around it, the antenna stops being a component chosen late in a design cycle and becomes the substrate that the whole system depends on. The teams that treat connectivity as a core design input from day one are the ones likely to get furthest past the pilot stage. Everyone else will keep discovering, expensively, that a brilliant decision made at the edge is worthless if nothing ever hears the result.

This article draws on insights shared during a Taoglas‑hosted panel discussion on edge AI and IoT connectivity, featuring Dermot O’Shea (Taoglas), Adam Cousin (Telit Cinterion) and Pavesh Govender (Vodafone).