Eight Things the World Cup Built That Your Engineering Team Should Steal

The 2026 World Cup finished on 19 July with Spain beating Argentina 1-0 in extra time. Over five weeks it ran 104 matches across 16 venues in three countries, delivered every match in UHD for the first time, peaked at 117 Tbps of streaming traffic, and resolved offside calls to roughly 10 centimetres.

That is the largest, most scrutinised, least forgiving production deployment of the year. It ran under adversarial conditions with no maintenance window, no ability to slip a date, and an audience that notices every defect within seconds.

I have written about the individual systems elsewhere. This is the consolidated version: eight patterns that transfer directly to systems that have nothing to do with football.

1. Fuse sensors, do not downsample to the slowest one#

The offside system pairs a 500 Hz inertial sensor in the ball with roughly 50 fps camera tracking. The ball supplies the precise moment of contact; the cameras supply the geometry. Neither answers the question alone.

The transferable rule: when a fast signal meets a slow one, join them at their native rates using as-of or interval joins. Do not bucket both to the coarser grain because it makes the SQL easier. That is how you delete the resolution you paid a sensor to capture. See the connected ball as an event-time problem for the full argument.

2. Automate the evidence, keep the judgement#

The system is called semi-automated for a reason. It detects, reconstructs, and presents — a human official decides. That is not caution, it is a correct read of an asymmetric cost function on a workload composed entirely of edge cases.

Most AI implementation projects reach for full automation too early and end up rolled back. The durable version removes the reviewer’s evidence-gathering work and leaves the decision with them. Detail in the human-in-the-loop write-up.

3. Collapse fan-in before it reaches your origin#

Origin shield turned hundreds of concurrent cache-fill requests into a single origin fetch per segment per region. Over 1,140 embedded points of presence inside ISP networks served more than 80% of match traffic in some markets without touching a peering point.

Both are the same idea: put a collapsing tier between many requesters and one scarce resource. It applies to CDNs, to database connection pooling, to third-party API quotas, and to any cache in front of anything expensive. Numbers in the 117 Tbps breakdown.

4. Plan from bursts, never from averages#

Opening day: 35 Tbps. Group stage peak: 45. Nine matches above 70. Final: 117. That is not a curve you extrapolate — it is step functions with a spike, triggered by which teams keep winning.

Inside the venues the same shape appeared: over 50 TB of fan data per stadium per match, concentrated into the seconds after goals, and carrier capacity boosted three to five times to absorb it.

Your capacity number is peak concurrency times payload per client at the worst ten seconds. An average is a description of a state your system is never in.

5. Unify the operational view before you optimise any part of it#

Sixteen venues ran through a single command environment, with security, ticketing, crowd management and infrastructure telemetry in one place and digital twins predicting congestion before it formed.

Most organisations have the same problem at smaller scale and answer it with browser tabs. The questions that matter cross system boundaries, and no individual dashboard can answer them. Building the join is worth more than improving any one of the sources. More in the command centre piece.

6. Model access is not a moat#

FIFA gave a generative analysis platform to all 48 teams. Same capability, wildly different outcomes.

Whatever separated the teams that got value from it, it was ingestion, entity resolution, freshness, and whether the output reached a decision-maker in a form they would act on. The model was the commodity. This is the single most under-priced fact in enterprise AI right now, and it is only becoming more true as frontier capability keeps getting cheaper each quarter. Expanded in the analytics platform post.

7. Your labels are your ceiling#

Behind the computer vision sits a large human annotation layer — workers in Manila, Rio, Cairo and elsewhere tagging up to 3,000 actions per match.

If your ground truth is 92% self-consistent, no model exceeds 92% on that task. Measure inter-annotator agreement before you touch model architecture; it reframes “the model is underperforming” into “the task is underspecified,” which is a solvable problem. See the annotation piece.

8. Your pricing model is part of your security architecture#

Demand-based pricing plus an uncapped official resale market removed the profit ceiling on scalping, which funded the professionalisation of the bots attacking the platform, which pushed priced-out buyers toward spoofed sites — prompting an FBI advisory during the tournament.

Every component worked as specified. The composition was the defect. Anywhere scarcity meets uncapped upside, you have specified an arbitrage, and arbitrages get automated. Detail in the ticketing analysis.

What we take into client work#

Strip the football away and these eight collapse into three habits.

Preserve resolution. In time, in identity, in geometry. Every layer that aggregates early is a layer that decides, on your behalf, which questions can never be asked again. We keep the raw grain in ClickHouse and let consumers roll up, precisely so nobody has a performance excuse to pre-aggregate.

Design the handoff. Between machine and human, between edge and origin, between subsystem and command view. Almost every failure we are called in to fix lives at a boundary, not inside a component. A Hospital Management System that fails does so at the seam between the ward and the pharmacy; a School ERP fails between attendance and the person who could act on it.

Rehearse the peak. Nine matches above 70 Tbps before the final was not luck. It was a five-week ramp validating the system against progressively harder load. Your equivalent is a load test scheduled before results day, before the admissions window, before month-end — not a dashboard you watch nervously during it.

The tournament is a useful teacher because the stakes were public. Nothing in that architecture was exotic. It was ordinary engineering, done deliberately, with the boundaries taken seriously.

---

Every one of these eight patterns is available to a team of five with a normal budget. The constraint is deciding to build the seams properly. Tell us where yours are failing.