Does Your Hundredth Deployment Cost More or Less Than Your Tenth?

Scaling vs. Replicating#

There's a chart that Baldur draws on whiteboards, a simple graph with two lines. One line is revenue. The other is cost per incremental deployment. In a real infrastructure business, the lines diverge: revenue grows while marginal cost declines. In most AI companies, the lines track each other almost exactly. Revenue goes up. Cost goes up. Headcount goes up. The company isn't scaling. It's replicating.

"The word 'scale' has become meaningless in technology," Baldur says. "It's used to describe growth. To describe ambition. To describe fundraising. None of those are scaling. Scaling is when each incremental unit of deployment costs less than the previous one, and the operational quality improves simultaneously."

He's seen this pattern hold back companies across three continents and four decades. The signs are always the same. More cloud compute for each new customer. More annotation labour. More support engineers. A cost structure that tracks revenue linearly, dressed up with a growth narrative. What the company actually has is a labour-intensive services business with a software wrapper.

"The test is simple," he says. "Does your hundredth deployment cost more or less than your tenth? Does it require more people or fewer? Is it more reliable or less? If it is roughly the same (or worse) you are accumulating operational debt while calling it growth."

The pause that follows isn't dramatic. It's Icelandic. The silence is the sentence.

The Chart That Diverges#

At OZ, the lines diverge.

This isn't a theoretical claim. It's measurable in the deployment data. Commissioning time per venue has decreased with each installation. Model accuracy on new venues improves because the training corpus grows with every match at every site. Per-venue operational cost trends downward because Audur's playbooks eliminate the manual variance that drives cost in traditional deployments. Rushikesh's hardware enclosures carry forward every environmental insight earned across seasons. Every match at every venue feeds data back into the Venue Graph, which makes the next deployment faster, more accurate, and cheaper.

"OZ's architecture is designed so that every new OZ VI Venue deployment improves the unit economics of the entire fleet," Baldur explains. "The Venue Graph accumulates data from every match at every venue. The operational playbooks capture learnings from every installation. The AI models improve with every venue's unique lighting conditions and camera setups. The hardware carries forward every environmental insight."

This is the deployment loop. A structural property of the architecture that produces compounding returns at fleet scale. Each node enriches the playbooks, lowers marginal cost, and increases data value. It's the economic signature of infrastructure, not software.

The Switching Wall#

Baldur's background in telecoms and energy taught him something that most software companies have never experienced: the physics of real switching costs.

"People are drawn to platform narratives," he says. "And I am not skeptical of platforms. I am skeptical of platforms without switching costs. In software, switching costs are lower than most people want to admit. A customer can migrate from one cloud analytics platform to another in a quarter. It is painful, it is expensive, but it is possible. The vendor knows this. The customer knows this. The pricing power reflects it."

Physical infrastructure is categorically different.

When an OZ VI Venue node is physically installed in a stadium (sealed, calibrated, integrated into the venue's power and network infrastructure, with the Venue Graph accumulating months of match data specific to that venue) the switching cost is not a data migration. It is a physical removal. A new structural assessment. A new installation. A new calibration period. A new AI training cycle for the replacement system. And the complete loss of all accumulated venue intelligence.

"That is not a switching cost," Baldur says. "That is a switching wall. And it is a switching wall that increases over time, because every season that passes adds data to the Venue Graph, adds operational history to the playbooks, and deepens the venue operator's integration with OZ's data platform."

Baldur Stefansson

Executive Chairman

Strategy & Governance

“Anyone can raise money for AI. The question is whether your unit economics improve with every deployment or degrade.”

Atoms Plus Intelligence Plus Scar Tissue#

The advantage isn't any single layer. Baldur is precise about this.

"Hardware without intelligence is commodity infrastructure; stadiums are full of it. Intelligence without hardware is cloud software, vulnerable to latency physics and the switching cost problem I just described. And both hardware and intelligence without operational excellence produce systems that work in demos and fail in the January rain."

OZ has all three because they were built as a single integrated system from the beginning. The full-stack edge (enclosure, compute, runtime, inference, control, data output) designed as one vertical system. That was Gudjon's original architectural insight, and it's the strategic decision that makes the governance framework possible.

"You cannot govern a fragmented vendor stack the way you govern a vertically integrated infrastructure company," Baldur says. "The unit economics are different. The feedback loops are different. The compounding dynamics are different. When the Head of AI calls the Head of Hardware and they solve constraints in the same room, without a vendor intermediary, that is vertical integration as an operating principle, not a strategy slide."

The implication is structural: a well-funded team with access to every AI model on the market still cannot replicate hardware deployed at live venues, operational playbooks refined through hundreds of live matches, or a compounding Venue Graph built over seasons of real data. The test isn't about technology. It's about time under load.

Governance Is the Product#

Most companies treat governance as overhead, something to formalize later, a box to check. Baldur built it from day one. Not because he's conservative. Because in an infrastructure company, governance is the product.

"When a venue operator signs a contract with OZ, they are buying governance," he says. "They are buying published performance guarantees: the system sees and reacts faster than a human blink, with zero dependence on cloud connectivity for critical operations, and clearly defined failure categories with specified response times. They are buying liquidated damages provisions that hold us financially accountable to those guarantees. They are buying audit trails that document every operational event."

This isn't the governance of quarterly board meetings and curated presentations. At OZ, operational metrics flow continuously. Performance against guarantees, deployment economics, installation speed, AI accuracy trends: these are living dashboards, not quarterly reports.

"In a traditional model, the board meets quarterly, receives a curated presentation, asks questions filtered through management's framing, and approves the next capital tranche based on a narrative," Baldur says. "The board and the operating team inhabit different realities. In our model, the problem is already visible, already measured, already being addressed before anyone needs to schedule a meeting."

The discipline follows directly. Every resource allocation decision is tied to a deployment milestone, not a development milestone. OZ doesn't fund feature development. It funds deployment outcomes. The leadership reviews the same performance metrics that customers see. There's no translation layer where operational reality gets polished into a narrative. The numbers are the numbers.

Concentric Expansion as a Governance Strategy#

Every AI company today mentions multi-vertical potential. Baldur draws a sharp line between aspiration and evidence.

"There is a meaningful difference between a company that says 'our technology could work in other verticals' and a company that has proven its platform operates under constraints that are harder than those other verticals demand."

Live sports is the hardest real-time AI problem in commercial deployment. Twenty-two unpredictable athletes at full speed. Zero tolerance for downtime. Thousands of concurrent viewers. The system must see and react in a fraction of a second. Every match is a live stress test with no second takes.

A company that operates reliably under those constraints has already demonstrated capabilities that are directly relevant to adjacent verticals. Broadcasting is a relaxation of the real-time constraint. CCTV is a relaxation of the motion complexity constraint. Industrial monitoring relaxes the quality bar.

The governance framework gates this expansion. You do not enter a new vertical until the adjacent vertical demonstrates compounding returns. You do not allocate resources until the current vertical has validated the platform's reliability, the team's operational capacity, and the unit economics. Each ring of expansion produces the evidence that justifies the next ring.

"Concentric expansion is a governance strategy, not just a market strategy," Baldur says. "The discipline of starting where the physics is hardest means every subsequent vertical is a proven step, not a speculative leap. Evidence first, expansion second. That is the discipline."

Infrastructure Company With AI Economics#

When Baldur explains OZ to people outside the company, he returns to the same distinction, because it determines everything downstream.

"This is an infrastructure company with AI economics, not an AI company with infrastructure aspirations. The distinction matters because it determines how we allocate resources, how we scale, how we govern, and how we measure success. Infrastructure companies compound. AI software companies grow, plateau, and face margin compression as competition arrives."

OZ's node economics (node license plus usage plus data and API access) produce recurring, expanding margins. Zero-ops deployment means zero marginal operational cost per node. The deployment loop means each new venue lowers the cost basis of the entire fleet. The Venue Graph means the data asset compounds with every match, every venue, every season.

"Anyone can build AI," he says. "The tools are widely available. The question (the only question that matters for long-term value) is whether your economics improve with every deployment or degrade. At OZ, they improve. That is not a claim. It is a measurable property of the architecture, enforced by governance, and visible in every deployment we have completed."

He pauses again. In Icelandic board culture, the silence after a statement carries its own weight, not for emphasis, but because the words are finished and nothing else needs to be added.

"Discipline is not conservatism. It is the only form of ambition that survives contact with physical reality."