The five business uses of virtual reality that stuck

In the fourth quarter of 2025, Meta’s Reality Labs division earned $955 million and lost $6,021 million. Revenue in that quarter was lower than the $1,083 million it took a year earlier. For the full year the segment turned over $2,207 million against a loss from operations of $19,193 million, and the company’s own results announcement of 28 January 2026 told investors to expect Reality Labs operating losses in 2026 to remain similar to 2025 levels.

Those four numbers are the most useful thing published about virtual reality in the last decade. They say that the largest, best-funded, most patient attempt to build a consumer market for headsets has produced revenue that is not growing, at a cost that is not falling, and that the company running it has told the market to stop expecting the loss to close.

Meanwhile, in a set of applications that attract almost no coverage, headsets went into routine use and stayed there. What follows is five of them, chosen because in each case somebody with a budget signed a contract, a regulator issued an approval, or a trial reported a result. They have one property in common, and it is worth naming before the list rather than after it: in every case the thing being replaced is expensive, dangerous, or impossible to schedule.

1. Flight training that counts towards a licence

The decisive moment for virtual reality in aviation was not a product launch. It was a qualification.

On 26 April 2021 the European Union Aviation Safety Agency qualified a virtual reality based flight simulation training device, a Robinson R22 Beta II trainer built by the company then called VRM Switzerland, at Flight and Navigation Procedures Trainer level II. EASA’s stated reasoning was blunt: around 20% of rotorcraft accidents occur during training flights, and the manoeuvres that most need rehearsing are the ones that most need height perception and a wide field of view. Autorotation. Hovering. Slope landing.

That approval changed the category the technology sits in. Before it, a headset in a flight school was a marketing asset. After it, time spent in one could be credited, which means it appears in a training organisation’s regulatory paperwork and in its cost model. A simulator that reduces the hours a student must fly in a real helicopter has an economic case that survives contact with a finance director, because rotorcraft flying hours are among the most expensive units of training anyone buys.

Note what did the work here. Not resolution, not tracking latency, not content. A regulator agreeing that a class of device could stand in for something it previously could not.

2. Army training bought by the fifteen-year contract

On 10 July 2026 the Ministry of Defence awarded a £2 billion contract for the Army Collective Training Service to Omnia Training, a consortium of five UK companies led by Raytheon UK. The service runs for fifteen years. The department says up to 60,000 soldiers a year will be trained through the platform, that it will integrate simulation, live systems and analytics, and that it will improve readiness for formations from teams of 100 soldiers up to 50,000.

Set the marketing language aside and read the contract shape instead. Fifteen years is not a pilot. It is an estate decision, made by a buyer that has to justify it to the National Audit Office and to a parliamentary committee, about a training method it expects to be running in 2041.

One qualification belongs inside the item rather than in a footnote, because it changes what can honestly be claimed for it. The announcement never uses the words virtual reality, and it never mentions a headset. What it describes is a Combat Laboratory using artificial intelligence, advanced analytics and virtual environments. Whether head-mounted displays form part of the delivered service is not established anywhere in the public record, and anyone stating what share of that training will be worn rather than watched on a screen is guessing. What the contract does establish is a fifteen-year commitment to synthetic training, which is the category virtual reality is bought inside and the budget line it has to compete for.

Defence has been the most consistent buyer of synthetic training for forty years, for reasons that have nothing to do with novelty. Live collective training requires land, ammunition, fuel, movement orders and weather. Every one of those is a constraint that a synthetic environment removes, and armies were doing this arithmetic long before consumer headsets existed.

3. Isolating an electrical supply you cannot practise on

The most instructive British example is also the least glamorous. Sellafield’s engineering teams need to be able to isolate electrical equipment correctly, on a site where the consequence of getting it wrong is not a training incident. You cannot rehearse on live plant, and building a physical mock-up of every configuration is not affordable.

The site’s emerging technology team, working with the supply chain firm V360 Energy, built a virtual reality electrical isolations simulator that runs at the Engineering Centre of Excellence in Cleator Moor. Sellafield Ltd’s account of the project, published on 22 November 2024, says the simulator lets users practise and refine isolation procedures for multiple pieces of equipment in a controlled environment, and reports that the training now takes one day where it previously took several weeks.

That last figure deserves a caution, because it is the sort of number this desk usually declines to print. It appears in an awards citation published by the operator, not in an independent evaluation, and no methodology accompanies it. It is the operator’s own claim about its own programme. The reason to quote it anyway is that the operator is a public body and the claim is therefore on the record and answerable, which is more than can be said for the equivalent figure in a vendor case study. Treat it as a stated result, not a measured one.

4. Procedural skill in surgery

Here there is peer-reviewed evidence, and it points the same way while being messier than its advocates suggest.

A systematic review and meta-analysis published in the Journal of Medical Internet Research in 2025 pooled 23 randomised controlled trials covering 1,091 participants in orthopaedic education. It reported standardised mean differences favouring virtual reality of 1.08 for knowledge scores, 1.44 for clinical operation scores and 1.75 for surgical design scores, along with odds ratios above four for learner interest and satisfaction.

Effect sizes that large, that consistently, in a field this heterogeneous, should make a careful reader slightly suspicious rather than entirely persuaded. The authors themselves flag notable heterogeneity across virtual reality platforms and call for multicentre, double-blind, large-sample trials. Learner satisfaction is also the outcome most vulnerable to novelty: people enjoy the new thing, and enjoyment is not skill.

What survives the scepticism is the direction and the durability. Surgical simulation has been studied for two decades, the studies keep finding a benefit for procedural competence acquired before a trainee reaches a patient, and no serious body now argues that the first attempt at a technique should happen on a person. That is why this use stuck and the enterprise meeting room did not. The counterfactual is a human being.

5. When the headset is the treatment, not the training

The fifth case is different in kind, and it is the one most likely to matter to an NHS commissioner.

gameChange is an automated virtual reality therapy for people with psychosis who avoid everyday situations. It was funded through the National Institute for Health and Care Research invention for innovation programme and tested in a randomised controlled trial of 346 patients across nine NHS trusts in England, reported in Lancet Psychiatry in 2022. At the six-week primary endpoint the therapy group showed significant reductions against usual care alone in agoraphobic avoidance (adjusted mean difference 0.47 on the Oxford Agoraphobic Avoidance Scale, p=0.026) and in distress (4.33, p=0.014). Both effect sizes were small, which the authors attribute in part to floor effects.

Two things about that trial are routinely reported wrongly, and the second-hand version is the one in circulation. The University of Oxford’s project page says the benefits were maintained at the six-month follow-up, and that those who gained most included patients with the most psychiatric symptoms. The trial paper does not support either statement as written. Differences between the groups on the primary measure were not significant at 26 weeks. Maintenance at 26 weeks appears only in a post-hoc analysis of patients with high or severe agoraphobic avoidance at baseline. And the moderation analysis found no evidence that outcomes varied with negative verbal auditory hallucinations (p=0.68), hopelessness (p=0.17) or appearance concerns (p=0.36). Severity of agoraphobia moderated the result. General psychiatric severity did not.

That distinction is the whole commissioning question. A therapy that helps the patients who cannot leave the house unaccompanied is a targeted service with an identifiable cohort and a validated screening measure. A therapy that helps everyone a bit is a procurement.

The mechanism worth understanding is not the graphics. It is that the therapy is delivered by a virtual coach rather than a clinician, which means the constraint it relieves is the supply of trained therapists. In a service where the waiting list is the problem, a treatment that works without consuming a therapist hour is a different economic object from one that works better than the alternative.

What is not yet public is how much of this has reached routine practice. A trial result is not a deployment, and no national figure exists for how many patients have received virtual reality therapy in the NHS outside research settings. Anyone quoting one has made it up. Our wider coverage of the sector sits under healthcare technology.

What failed, and the reason it failed

The uses that did not stick were the ones where the alternative was fine.

Virtual meetings failed because a video call is adequate and a headset is not free. Virtual showrooms failed because a photograph is adequate. Virtual training for tasks that can be practised on the real equipment failed because the real equipment was already there. In each case the technology worked and the comparison did not favour it, which is the most common way a capable technology loses.

The consumer story failed for a related reason and at extraordinary expense. Meta’s segment reporting shows Reality Labs revenue of $2,207 million in 2025 against $2,146 million in 2024, a business that has stopped growing while the loss attached to it widened from $17,729 million to $19,193 million. Apple has published no unit sales for Vision Pro at all, and any number circulating for it comes from a supply chain estimate rather than the company.

There is a second-order effect worth naming. Every one of the five survivors is a training or clinical application, which means the content is the product and the headset is a peripheral. That inverts the assumption most buyers started with in 2016, when the device was the purchase and the content was assumed to follow.

It also explains why the surviving suppliers are unglamorous. The firms doing well here are simulation houses, training contractors and clinical research spinouts, not consumer hardware brands. They sell a procedure rendered accurately, validated against a standard, and maintained when the standard changes. None of that is a technology business in the sense a 2016 investor meant.

The test to apply before buying one

Three questions, in order, and a no to all three is a no.

Does a regulator, an insurer or a professional body treat time in the simulator as equivalent to time in the real thing? If yes, the business case is arithmetic rather than argument, because you are buying a substitution at a known price.

Is the alternative genuinely unavailable? A live nuclear isolation, a live patient, an occupied battlefield, a helicopter in autorotation. Scarcity of the real thing is what pays for the fake one.

And can the organisation state what the headsets will cost to keep running in year three, including content maintenance when the underlying procedure changes? The capital purchase is the cheap part, which is the same lesson learned expensively in what automation costs to run.

The pattern across all five surviving uses is that virtual reality became useful precisely where it stopped being interesting. It is now a piece of training and clinical infrastructure bought by people who will never describe it as immersive, and that is the condition an emerging technology reaches when it stops emerging. The rest of that argument, and the record of which predictions about it held, sits in emerging technology and in our marking of the 2019 forecasts.