Allen Salmasi

The through-line

Everything Salmasi says orbits a single conviction: intelligence has to move to where the data is created, and that forces the wireless network, the compute layer and the security layer to stop being separate things. He frames it as a set of converging structural trends, of which the most important is that "AI needs to become context aware and low latency in order to have real world applications for many of the use cases" [1]. The second is that communications and cybersecurity are now "fully intertwined, and integrated as one fabric" [1]. The consequence he keeps returning to is architectural: the hierarchical telecom design that has held for fifty years breaks, and what replaces it is a distributed, meshed, virtualized fabric where radios, compute and workload governance are orchestrated together.

He is candid that this is hard. Asked about the tangle of spectrums, vendors and infrastructures at the edge, he agrees it is "probably the soupiest mess that anyone is going to tackle anytime soon" [1], while arguing he has been working the problem for two decades rather than reacting to the current AI cycle. The continuity matters to him: the vision he describes now is the one he says began with a public safety build in the mid-2000s and was incorporated into a company in 2014 [2].

On the hyper-converged edge and where the idea came from

The origin story is operational, not theoretical. Around 2003 and 2004, partnered with Northrop Grumman, he won a New York City public safety bid against Ericsson and Motorola that he says the team did not expect to win, and built the network between 2004 and 2009 [2]. Serving first responders across 500,000 installed cameras and over two million sensors demanded a hyper-converged network from the start, because "everything had to be orchestrated" rather than manually provisioned [2]. From that experience he drew the forward bet that shaped the company: "it's just really a matter of time before machine learning, computer vision, and every application that we were applying to this public safety network is going to come to the commercial market" [2], to operators as services and to enterprises as solutions. Veea was formed in 2014 to carry that vision into commercial markets [2].

The applications he cites now are the same shape at wider scale: AI made applicable across logistics, healthcare and transportation, customized per use case with the right chip sets and ultimately the right AI factories [1].

On the coming architecture change in wireless

He describes the current design as a fifty-year inheritance running from AT&T's first analog system: switching that became the core network, base stations that became more distributed with Open RAN, radio units and DUs pushed outward [2]. His claim is that distribution now goes much further. At terahertz frequencies, "every small room is going to have some type of a radio head" [1], so radio becomes fully distributed, and that distribution only works if it is "completely in sync with a distributed compute type of capability" [1]. His architecture therefore supports a compute mesh and a microservices mesh sitting on top of the radio, forming and reforming dynamically and in real time [1].

Underneath sits full convergence of wireline and wireless at the core, to the point where "you won't be able to differentiate between optical connections, any landline connections and wireless connections" [2]. That opens a different arrangement in which compute elements that used to be rigidly defined talk to each other through AI agents back to the core network, raising RF efficiency, total throughput and quality of service location by location as frequencies climb [2]. He calls this a "Major architecture change" [2]. The deployment model he wants follows: "the way all of the network equipment should be deployed is very much like the way we deploy Wi-Fi access points" [2]. Wi-Fi fails outdoors today only because it is first come, first served and ad hoc rather than managed [2], and he sees no reason Wi-Fi and cellular cannot converge, with RF chips now covering roughly 800 megahertz up to terahertz, so a gateway or hub anywhere in a city can serve users across all ranges in a fully distributed, peer-to-peer network [2].

On NVIDIA, he says he is completely aligned with the vision Jensen articulated in Washington alongside Nokia, and describes Veea as piggybacking off NVIDIA's Open RAN work, first tested by colleagues at Vapor in Las Vegas roughly four years ago and since opened up and made open source, extending it all the way to the edge [1].

On security as one fabric, and trusting agents

Cybersecurity is not a layer in his model. It is embedded into every connection on a zero trust basis and treated as one fabric with communications [1]. His justification is practical rather than compliance-driven: "unless you can trust the AI agents that you're connecting to, you are not going to really have physical AI introduced at the edge" [1]. The same principle shows up in his enterprise products, where the promise is that a user working from a car, a home or a customer site is "connected to the network as though you're in the office", with the same level of cybersecurity carried into that wide area solution [3].

On governing AI agents and the gap between deploying and operating

As agents and automated workflows become the standard workflow at the edge, he argues the binding constraint shifts from deployment to operations. There is "a growing gap, really major gap between deploying edge infrastructure and operating it" [2]. His answer is TerraFabric, introduced days before MWC26, a control plane that manages edge workloads, makes them more secure and provides governance over work that is otherwise left to AI agents [2]. The reason governance is non-negotiable in his telling is that these agents "make decisions, they take action and they interact with physical environment" [2], which is a different risk class from software acting on software.

On the virtualized, multi-vendor network

The demonstration he points to is deliberately heterogeneous: NVIDIA Jetsons, X86 servers, Linux servers, various flavors of access points and Raspberry Pis, all running as one fully virtualized network with a single middleware managing every element [2]. That middleware orchestrates workloads and applications from the edge to the cloud, and lets new services or capabilities be introduced from any core network [2]. The commercial implication he draws is shared infrastructure: a common facility or solution that multiple operators can use to deliver their own services and solutions to end users [2].

On telcos' last chance

He sees a large opening for operators to move beyond commodity connectivity and add real value to the use cases running at the edge [1]. The physical basis is the estate they already own. Where telco aggregation points once held huge SS7 switches, the core network has shrunk into smaller racks, leaving space [1]. Operators "have the facilities, they have the cooling, they have all the wiring in these thousands of locations" [1], and the only real question is how they architect it and integrate it into existing infrastructure [1]. Pressed on whether telcos, having missed the cloud wave, could miss this one too, he agrees the risk is real [1].

On scaling the business through telco partnerships

He is explicit that the company changed direction. After nine million in revenue the prior year, "we totally pivoted to focusing on businesses that can scale very rapidly, primarily through telco partnerships" [1]. The proof point is a supply agreement announced in August of the previous year with América Móvil and its subsidiary Telcel, an operator group he describes as the largest in the world, spanning Telcel and Telmex in Mexico, Claro across Central and South America and A1 Networks in Europe, around 200 million subscribers [1]. What Veea brings to the edge of that network is fixed wireless access with cybersecurity plus AI-driven use cases and applications, including surveillance cameras and a range of IoT solutions, all bundled into a single device roughly the size of an Apple TV [1].

On virtualizing the workplace, and real estate as a use case

The same platform logic reappears in a very different market. He starts from the observation that the boundary between home and office has dissolved to the point where the location of the person on the other end of a call is unknowable and beside the point [3]. His solution completely virtualizes the business, so a professional can work from an office, a home, a car or a Sunday afternoon showing with office-grade security, video conferencing, side-channel chat and defined groups of colleagues and customers all in one place [3]. Rather than assembling standalone tools from different vendors, Veea supplies the hardware to install at home or office and smaller portable versions to set up a temporary office at a property for as long as you are there [3].

What interests him most is the applications layer on top, treated like a smartphone platform capable of "applications that no one has ever seen before" [3]. His worked example is augmented reality during a showing: an agent walking a house with an iPhone, connected to a box sitting on the kitchen table, overlaying IKEA furniture or a dining table, showing a planned renovation from the architect's drawings, or swapping appliance ranges and colors in an unfinished kitchen while discussing schools and the neighborhood, with the buyer never having to be at the property [3]. He contrasts this with the tens of thousands of dollars spent on a curated video that fails to convey what closets and pantries actually look like, a frustration he says he hit while selling his own home [3]. The platform also carries brokerage-to-associate training, learning systems and communication [3], and includes a full advertising platform aimed at enterprise use, with large wall displays, office tablets and window screens serving educational content and letting title insurers, real estate lawyers and home service professionals "advertise in a hyper local way" and open one-to-one chats with agents about upcoming closings [3]. Passersby can point a phone at a static window display's QR code and pull up a curated video or interact live with cameras inside the home [3]. He describes the same products already being delivered into retail, buildings, factories, cities and other vertical markets [3].

Takeaways

• The core premise: AI only becomes useful for real world applications once it is context aware and low latency, which is why it has to be architected at the edge rather than delivered from central data centers [1].
• Distributed radio and distributed compute must move together; at terahertz frequencies effectively every small room gets a radio head, supported by a compute mesh and microservices mesh above it [1].
• Communications and cybersecurity should be one zero trust fabric embedded in every connection, because physical AI at the edge is impossible if the agents you connect to cannot be trusted [1].
• The operational bottleneck is not deploying edge infrastructure but running it; TerraFabric is positioned as the control plane that manages, secures and governs agent-driven workloads at the edge [2].
• Network equipment should be deployable the way Wi-Fi access points are, with Wi-Fi and cellular converging across chips spanning roughly 800 MHz to terahertz in a fully distributed peer-to-peer network [2].
• Telcos already own the thousands of former aggregation sites with space, cooling and wiring needed to host edge compute; the open question is architecture and integration, not real estate [1].
• The company pivoted from nine million in revenue to scaling through telco partnerships, anchored by a supply agreement with América Móvil and Telcel covering roughly 200 million subscribers, delivering fixed wireless access, cybersecurity, cameras and IoT in an Apple TV sized device [1].
• Twenty years of proof: a New York City public safety network built between 2004 and 2009 with 500,000 cameras and over two million sensors demonstrated that everything had to be orchestrated, and predicted the commercial arrival of machine learning and computer vision [2].