Apple Vision Pro M5 for Manufacturing: Spatial Computing on the Factory Floor (2026)

A headset that costs $3,499, weighs roughly three quarters of a kilogram and runs about two and a half hours per charge should not make sense on a factory floor. Yet Apple Vision Pro M5 manufacturing pilots keep surviving procurement review, because the device does something no ruggedised monocular or phone can: it renders a full-resolution, spatially anchored digital twin at human scale, with eye and hand input, and lets a second person stand in the same coordinate space. The honest question is not whether the hardware is impressive. It is which factory jobs justify its limits.

That question is sharper now than a year ago. The M5 refresh shipped in October 2025, visionOS 26 opened parts of the enterprise API set to every developer, HoloLens 2 is on a clock that ends in 2027, and Apple’s own roadmap signals have turned toward glasses. A buyer in November 2026 is choosing a platform whose future is unclear while its present is the best in class.

This rewrite replaces the stale numbers of the first version with sourced facts, and separates what is verified from what is rumour. You leave with a reference architecture, the exact enterprise entitlements and what each unlocks, a streaming pattern for heavy OpenUSD twins, a comparison against the alternatives, and a rollout checklist.

What this covers: verified M5 specs, the enterprise API surface, a reference architecture, Omniverse streaming, factory use cases, the competitive field, failure modes and a pilot plan.

Context and Background

Industrial augmented reality has had three eras. The first was the tethered-tablet era, where a phone or iPad showed a work instruction over a camera feed. The second was the optical see-through era led by Microsoft HoloLens 2, which put hands-free overlays on a wearable but with a narrow field of view and modest compute. The third, now arriving, is the video pass-through era: the headset captures the world with cameras, composites digital content in a GPU pipeline, and shows the result on micro-OLED displays. Pass-through gives opaque, correctly occluded, high-contrast graphics at the cost of weight, latency and the fact that the user sees the plant through a screen.

Microsoft told partners in October 2024 that HoloLens 2 sales were ending with a last-time buy, with security updates and critical fixes promised through 31 December 2027 and no successor announced (Road to VR). That single decision pushed industrial buyers to look at pass-through devices. Apple had already been courting them: it said in May 2024 that more than half of the Fortune 100 had bought at least one Vision Pro, naming KLM, SAP, Porsche, Lowe’s and EnBW among users of custom software (UploadVR). Treat that as a purchase-count claim, not proof of production deployment. Buying one unit for an innovation lab is a different commitment from equipping a shift.

The software side matters as much as the optics. A digital twin is only valuable if the headset can show live state, not just geometry. That means the twin pipeline discussed in our OPC UA historian to digital twin ingestion architecture has to reach the headset, either as streamed pixels or as compact scene updates. It also means the twin needs to be light enough for the device to render, which is the problem addressed in large-scale digital twin visualization with 3D Tiles and LOD.

The rest of this article takes the position that Vision Pro is best treated as a high-fidelity, low-duty-cycle review and collaboration surface for the twin, not as an all-shift wearable. That framing predicts both where it wins and where pilots stall.

What Is Actually Verified About the M5 Vision Pro

Apple announced the M5 Vision Pro on 15 October 2025 and it reached stores on 22 October 2025. The US starting price stayed at $3,499 for 256 GB, with 512 GB and 1 TB options, and the device launched with visionOS 26 (Road to VR). In Canada the entry price was $4,999 CAD; in the UK and euro area the M5 started at £3,199 and about €3,700, below the M2 model’s launch prices (MacRumors comparison).

The silicon changes are real but incremental. The M5 is a 3 nm part with a 10-core CPU and 10-core GPU, hardware ray tracing and mesh shading, and a 16-core Neural Engine that Apple says speeds AI tasks by up to 50 percent. The displays render about 10 percent more pixels and run up to 120 Hz, against 100 Hz before. MacRumors lists memory bandwidth rising from 100 GB/s to 153 GB/s and adds AV1 decode and LPDDR5X memory. The cameras, sensors and panels themselves are unchanged, and the software is identical to the M2 model.

Battery life is the number that governs factory scheduling: Apple quotes up to 2.5 hours of general use or 3 hours of video playback, and the device can run while plugged in. Weight is where reviewers disagree on framing. MacRumors lists the M5 at 750 to 800 g with the new counterweighted Dual Knit Band, versus 600 to 650 g for the M2 with its original bands. The extra mass is the band’s counterweight, intended to shift the centre of gravity rearward. Apple’s own published figure should be checked on its specification page before any ergonomic sign-off, because weight varies with band and light-seal size.

Two claims in circulating material, including the first version of this post, should not be repeated. A “10 to 12 millisecond passthrough latency” figure and a “15 percent weight reduction” for the M5 do not appear in any source I could verify, and the weight direction is in fact the opposite. Apple quoted 12 ms for the original R1 sensor-to-display pipeline at launch; I did not find a re-measured number for M5. If your safety case depends on latency, measure it on your own hardware.

Reference Architecture: Where the Headset Sits in a Factory Twin

A direct answer first: in a factory twin architecture the Vision Pro is a thin, high-fidelity client. Live state arrives from the OT network through a gateway and a twin service; heavy geometry either streams as pixels from a GPU server or loads as pre-optimised USDZ assets on the device; the headset adds spatial anchoring, eye and hand interaction, and optional camera input that flows back to the twin.

Figure 1: Reference architecture. The headset consumes twin state and geometry and returns camera, anchor and annotation data.

The diagram shows four zones. At the bottom is the OT zone with PLCs, OPC UA servers and sensors. A gateway in the DMZ translates and filters this data into the twin service, which holds state, asset hierarchy and the OpenUSD scene. A render tier, optional and GPU-backed, produces streamed frames for scenes too heavy for the device. The headset sits at the top, with native visionOS apps for on-device content and a streaming client for server-rendered content. Management systems, identity and the Apple Business Manager distribution channel sit to the side because they gate everything else.

Design principle one: keep OT traffic out of the headset

The headset should never speak OPC UA or Modbus to a controller. A headset is a consumer-grade endpoint that roams, sleeps, drops Wi-Fi and gets handed between people. Controllers need stable, authenticated, least-privilege sessions. Put a twin service between them, expose read-only state over a protocol the headset handles well, such as HTTPS with server-sent events, WebSockets or MQTT over TLS, and write back only through a separately authorised command path with human confirmation. This is the same segmentation argument made in the Purdue-model literature, applied to a new endpoint class.

A practical pattern is to subscribe the twin service to the OPC UA address space, project values onto scene prims by an ID mapping, and publish deltas at a modest rate. Operators do not read a pressure gauge at 1 kHz; two to five updates per second to the headset is ample, and it keeps battery and radio use low. The maturity concepts in our digital twin maturity model are useful for deciding how much state is worth pushing: a headset overlay of status colours is a level-two twin, while closed-loop what-if simulation in the headset needs far more.

Design principle two: decide on-device versus streamed rendering per asset

Apple’s RealityKit renders USDZ content natively and well. A single machine, a cell or a piping run of moderate polygon count belongs on the device. A full plant with millions of triangles, CAD-derived assemblies or physics-driven simulation belongs on a server. Mixed designs are common: native UI, annotations and anchors on the device, heavy geometry streamed behind them. We return to the streaming case later.

Design principle three: anchor to the plant, not to the room

ARKit world tracking and world anchors let content persist at physical locations, but the persistence is local to the device and its sensing of the room. Factories change: pallets move, fixtures are repositioned, lighting varies. Anchoring digital content to a machine requires a reference that survives that churn. Object tracking, which uses a trained reference object built from the machine’s USDZ model, is the more robust option for equipment. Fixed QR or barcode markers, readable through the barcode-detection entitlement, give a deterministic re-localisation fallback at shift start.

The visionOS Enterprise API Surface

Apple groups its enterprise capabilities into two categories, enhanced sensor access and platform control, gated by entitlements and a license file (Apple developer documentation).

The sensor-access category provides:

main camera access, the forward-facing camera feed;
camera region access, a feed from a chosen spatial region of the wearer’s view;
shared coordinate space, aligning several devices in the same physical space;
screen capture including passthrough, a composite of what the wearer sees;
barcode detection, with decoding and spatial positions.

The platform-control category provides:

protected content, which blocks screenshots, recordings, AirPlay and SharePlay of a view;
object-tracking parameter adjustment;
an app compute category, which allows more CPU and GPU power at the cost of heat and battery;
window body follow, so windows move with the user;
visual fidelity monitoring of fit and field-of-view coverage.

The licensing mechanics matter operationally. Only the Account Holder of an Apple Developer Program or Developer Enterprise Program membership can request an enterprise API license, an individual enrolment cannot, and apps must be for business use only. Distribution is private, as proprietary in-house apps or custom apps through Apple Business Manager. A Development Only license works at team level for test devices, while a distribution license applies to a single App ID. The license file carries an expiry date and must be renewed; WWDC25 described over-the-air renewals and a VisionEntitlementServices framework for checking license status in code.

Apple’s WWDC25 session on spatial business apps (session 223) added several changes. UVC camera access and Neural Engine access became available to all developers without an enterprise license. Object tracking training can now be scripted with a command-line tool, which matters for pipelines that regenerate reference objects when CAD changes. CameraFrameProvider can deliver left, right or both camera streams for stereo processing, in either Shared Space or an Immersive Space. SharedCoordinateSpaceProvider lets an app align participants’ coordinate systems over any transport it chooses, while SharePlay covers the out-of-the-box case. The contentCaptureProtected() SwiftUI modifier hides sensitive views from captures and mirrored displays while the wearer still sees them (WWDC25 session 223).

The camera-region guidance is worth quoting in spirit because it reveals the cost model. Apple advised keeping camera-region anchors to about one sixth of the visible area or less to manage memory and processing, and noted that stabilisation gives a higher frame rate than contrast enhancement. For a technician reading a small gauge or a serial plate, a stabilised, magnified window over that region is the practical feature; whole-view machine-vision analytics is not.

An earlier tranche from visionOS 2 in 2024 is still the conceptual core: forward camera access for anomaly detection, passthrough recording for remote assistance, QR scanning for warehouse verification, a Neural Engine path for machine learning, multi-object tracking for tools and parts, and a higher-compute mode that costs battery and fan noise (Computerworld).

Figure 2: Mapping enterprise entitlements to factory tasks, from sensor access through to platform control.

Figure 2 maps each family of entitlement to the factory task that justifies requesting it. The guiding discipline, which Apple’s own session repeats, is to request only the entitlements an app needs: each one lets the device operate outside its default configuration and may affect other apps. In a regulated plant, every extra entitlement is also an extra line in the security review.

Streaming Heavy Twins: OpenUSD, Omniverse and CloudXR

The M5 GPU is capable by headset standards, but a full factory twin built from CAD, point clouds and simulation output will not fit. The established answer is hybrid rendering. NVIDIA’s Omniverse spatial streaming documentation describes a server running an Omniverse Kit application that renders and manipulates an OpenUSD stage, with native Swift and SwiftUI clients on Vision Pro and iPad Pro. Head pose and JSON events travel up from the client; scene logic on the server, written with ActionGraph and Python, reacts by changing prim variants or user locations. Custom UI built natively on the device sends commands that the server processes with low latency (NVIDIA Omniverse spatial streaming docs).

For production, the documented route packages USD data as a zip and publishes it to NVIDIA’s Graphics Delivery Network, which offers an XR machine configuration for streaming to Vision Pro users without a local workstation. The documentation mentions Omniverse Kit 109 and CloudXR Framework 1.0, and it also warns that the page is preserved for information and may be outdated, pointing to a newer NVIDIA spatial documentation site. That warning is the right instinct for any integrator: pin versions, test on your exact build, and expect the client SDK to move.

I could not verify end-to-end latency, bitrate or frame-rate figures for this path from primary sources, so none are quoted here. Streamed XR is sensitive to network jitter more than to raw bandwidth, because head-pose-to-photon delay governs comfort. Plan on a dedicated, wired-backhauled Wi-Fi 6 or 6E cell with quality-of-service for the headset VLAN, test with your own geometry, and set a rendering budget in the server application rather than hoping the network absorbs overload.

Figure 3: Hybrid rendering. Native UI and anchors stay on the headset while heavy OpenUSD geometry renders on a GPU server and streams back.

Figure 3 shows the control loop. The headset sends pose and events; the Kit application updates the stage from those events and from live twin state; frames return through the streaming layer; the headset composites them with native SwiftUI controls. If the stream drops, the native layer should fall back gracefully to a cached low-polygon USDZ model so the worker is never staring at a frozen frame.

Why OpenUSD is the right interchange

OpenUSD gives one scene description that spans CAD import, simulation, robotics and visualisation. The same asset path that feeds a robot simulator can feed the headset. Our comparison of URDF, SDF, MJCF and OpenUSD explains why composition arcs and variants make USD suited to configuration-driven scenes, and the lightweight CAD format comparison covers the conversion step from PLM sources. For the device-native path, USDZ is the package format RealityKit loads, so the pipeline normally ends in a decimated, texture-compressed USDZ per machine.

A related development on the content side is the Siemens-NVIDIA ecosystem work around industrial twins; our explainer on Visual Components FactoryLens and NVIDIA shows how factory layout simulation is meant to feed immersive review.

Factory Use Cases, Ranked by Fit

Not all uses are equal. Rank them by how well they match the device’s strengths, which are display quality, spatial anchoring, collaborative presence and short-session intensity, against its weaknesses, which are weight, battery and the opaque view of the plant.

Strong fit: design review and layout validation

Walking a planned line at full scale before concrete is poured is the clearest win. A review team shares a coordinate space, sees a full-scale cell with clearances, reach envelopes and maintenance access, and annotates it. Sessions are short, seated or slow-walking, and happen in a meeting space, not on a hot floor. Battery and weight are non-issues, and the cost of an error found here is orders of magnitude below an error found at commissioning. This is the use where Vision Pro’s display quality earns its price.

Strong fit: training and procedural rehearsal

Training in a headset avoids tying up a production machine and avoids safety exposure. A twin with scripted fault conditions lets a trainee rehearse an isolation, lock-out or changeover. Siemens, according to a 2024 CIO.de report, has field technicians overlay a virtual copy of the device to be installed and repeat the installation steps on site; the same report describes Porsche showing unbuilt Taycan configurations as 3D objects in the room and SAP using Personas around a shared data visualisation. The article offered no quantified savings, so treat these as qualitative evidence of fit, not ROI (CIO.de).

Conditional fit: remote expert assistance

A field technician streams what they see to an expert, who annotates. The enterprise passthrough-capture entitlement and camera access make this possible, and spatial annotation anchored on the machine is genuinely better than a video call. The conditions are that the technician can tolerate 90 minutes of wear with the headset, that connectivity is good, and that the plant permits camera capture at all. Many plants prohibit cameras, and protected-content features cut both ways: they stop sensitive views leaking, but they also complicate recording for audit.

Conditional fit: maintenance overlays

Showing live sensor state and step-by-step instructions on a machine is the headline use case, and the one that most often underdelivers. Object tracking can lock instructions to a machine, and the 2024 visionOS 2 APIs were built with this in mind. The limits are practical: workers in gloves, face shields or hearing protection; noisy and reflective environments degrading tracking; and the need for the headset to be safe near moving equipment. A tablet or a single-eyed display is often better for a 20-minute task.

Weak fit: continuous shop-floor wear

A full shift in a pass-through headset is not realistic with a 2.5-hour quoted battery, a heavy front, and no safety certification for hazardous areas that I could verify. Do not buy for this use. If it is your core requirement, look at the alternatives below.

Figure 4: Decision flow for matching a factory task to a headset class.

The Competitive Field in Late 2026

Microsoft HoloLens 2

Discontinued, with updates through 2027, as described above. It remains deployed in many plants and retains advantages: optical see-through, no pass-through blindness, and a head-mounted form that sits slightly lighter. It has no future roadmap, so new purchases are a bridge decision at best.

Samsung Galaxy XR and Android XR

Samsung launched Galaxy XR on 21 October 2025 at $1,799 to $1,800, using Qualcomm’s Snapdragon XR2+ Gen 2 with 16 GB of memory, a micro-OLED display of about 29 million pixels at up to 90 Hz, a separate battery pack with up to 2.5 hours of use, and Android XR, a platform co-developed by Samsung, Google and Qualcomm (Android Central). It is roughly half the Vision Pro’s price. I did not verify its enterprise management, industrial app ecosystem or sustained-use weight, so it should be evaluated rather than assumed.

Meta Quest line

Meta’s headsets are the cheapest way to run pass-through mixed reality and have enterprise management offerings. I did not verify current Quest enterprise licensing, device-management features or pricing for this article, so treat any Quest assumptions as unverified. Cost is the draw; visual fidelity, text sharpness and the Apple-style hand-and-eye interaction model are where Vision Pro leads.

Rugged monoculars and smart glasses

For hands-free procedures with a small display and long battery, single-eye wearables remain the mainstream industrial tool. Apple’s own direction matters here: Bloomberg reported in October 2025 that Apple paused a cheaper, lighter Vision Air headset to accelerate smart glasses, while saying the pause did not affect the M5 upgrade (UploadVR). A pre-WWDC 2026 preview citing Mark Gurman, which is unconfirmed, said Vision Pro 2 is at least two years away (vr.org). The implication, if both reports hold, is that the current M5 device will be Apple’s headset for some time, which lowers obsolescence risk but also means no cheaper, lighter model soon.

Decision matrix

Requirement	Vision Pro M5	Galaxy XR	HoloLens 2	Rugged monocular
Review at full scale	Excellent	Good, unverified for enterprise	Fair	Poor
Collaborative presence	Strong, Personas and shared space	Unverified	Fair	None
Full-shift wear	Poor	Poor to fair, unverified	Fair	Strong
Optical safety, see-through	Pass-through only	Pass-through	Optical see-through	Optical
Vendor roadmap	Uncertain, glasses focus	New platform	Ending	Mature
Unit price	$3,499 and up	about $1,800	Discontinued	Varies

visionOS 27 and What the Roadmap Means

Apple presented visionOS 27 at WWDC on 8 June 2026. Coverage describes a 3D Siri that responds to gaze, redesigned Control Center, eye-controlled notifications, curved windows, panoramas turned into spatial scenes and faster Wi-Fi. The same coverage notes that Apple offered few new Vision Pro updates and reports that Apple is reportedly halting work on the headset (Tom’s Guide). I did not find confirmation of specific new enterprise APIs in visionOS 27 from Apple’s documentation, and I could not confirm the public release date, so do not assume any 27-specific enterprise feature in a plan. A pre-event article speculated about MDM, guest-session and protected-content improvements; that was speculation then and remains unverified here.

Device management deserves a warning. When MDM support arrived in visionOS 1.1, management required account-driven enrolment with Managed Apple IDs, and automated device enrolment and supervision were reported as unsupported at that time (Ivanti). That is two years old. Current support for Automated Device Enrollment, supervised mode, single-app lock and shared-device use should be confirmed against Apple’s current deployment documentation and your MDM vendor before you plan a fleet.

Trade-offs, Gotchas, and What Goes Wrong

Battery and thermal budgeting

Take Apple’s 2.5-hour figure as an upper bound for general use, not a plan. The enterprise compute category explicitly trades battery and fan noise for performance, and continuous camera processing and streaming add radio and sensor load. As an illustrative calculation, not a measurement: if a maintenance task needs 25 minutes of active headset time and a typical inspection round needs four such tasks, the device is in use for 100 minutes, which leaves little margin on a 150-minute budget once you add donning, calibration and breaks. Design sessions to be shorter than the battery by at least a third, provide wired charging stations at the point of use, and measure on your own workload.

Weight and ergonomics

At roughly 750 to 800 g by MacRumors’ listing with the counterweighted band, the device is heavy compared with industrial wearables. Neck strain is a documented complaint of head-mounted displays generally, and it grows with session length and head-down tasks. Run an occupational-health review, and pilot with workers of different sizes, because fit depends on the light seal and band sizing. The visual-fidelity entitlement, which monitors fit and field-of-view coverage, can help detect poor fit automatically.

Safety and situational awareness

Pass-through video means the worker is looking at a camera image of the plant. If the pipeline stalls or the battery dies, vision is degraded. Ban use near moving machinery, forklifts, hot surfaces and edges unless a documented risk assessment says otherwise. Apple’s own enterprise guidance asks developers to make sure users can work safely around real equipment. I found no verified certification of the Vision Pro for hazardous-area or intrinsically safe use, so assume none.

Tracking drift and relocalisation

World anchors are device-local and depend on the visual environment. Repetitive structures, glossy floors and changing layouts degrade localisation. Use object tracking against the machine itself plus marker re-anchoring, and expose a one-gesture “re-align” command so a worker never has to restart the app.

Data governance

Camera access plus passthrough capture means a headset can record trade secrets, people and screens. Use protected content for sensitive views, log which entitlements each app holds, and write a retention policy for captured video. Where unions or works councils are involved, consult them before piloting; head-mounted cameras are a labour-relations topic as much as a technical one.

Total cost and vendor risk

The headset price is the smallest line. Add developer time, USD optimisation, an MDM and an Apple Business Manager setup, a GPU render tier if streaming, charging infrastructure, spare headsets and light seals, and training. Vendor risk is real: Apple’s reported pivot towards glasses means feature investment in the headset may slow. Mitigate by building the content layer on OpenUSD and open protocols, so your twin survives a change of headset.

Anti-patterns

The recurring failures are consistent. Teams stream a whole plant when a single cell was needed. They put live control writes in the headset. They anchor content to a room rather than a machine. They pilot with enthusiasts, then roll out to sceptics. And they measure “wow” instead of task time, error rate or rework.

Practical Recommendations

Start from a task, not a device. Pick one job where a spatial, full-fidelity view changes a decision: a layout review, a changeover rehearsal, a maintenance procedure that currently needs a travelling expert. Define a baseline in task time, error rate and rework before the pilot starts, and agree the success threshold with the line owner in advance.

Buy for the use you can verify. Five to ten headsets for design review and training is a defensible pilot. Fifty for shift wear is a bet against the battery. Keep a rugged monocular or tablet in the comparison for any task over 30 minutes.

Build the content layer so it outlives the headset. Keep the twin in OpenUSD, expose state through a twin service, and treat the visionOS app as a client. If Apple’s roadmap slows, or Android XR matures in your sector, you swap the client and keep the twin. The agentic digital twins pattern is a natural next layer, where an AI agent pre-selects which alarms and states a headset overlay should surface.

Checklist for a 90-day pilot:

Choose one use case and one location, with a named process owner.
Capture baseline metrics before any headset is worn.
Confirm current MDM, supervision and Apple Business Manager support with your vendor.
Request only the enterprise entitlements the app needs, from the Account Holder, and calendar the license renewal.
Decimate assets to USDZ per machine and set a polygon budget.
If streaming, test on a dedicated wireless cell and define a cached fallback.
Complete an occupational-health and safety risk assessment, including session length limits.
Agree a camera and recording policy with security, legal and worker representatives.
Measure battery, thermal behaviour and tracking drift on the actual floor, not in a lab.
Review at 30, 60 and 90 days against the baseline, and be willing to stop.

Frequently Asked Questions

Is the Apple Vision Pro M5 approved for industrial hazardous areas?

I could not verify any certification of the Vision Pro for explosive-atmosphere or intrinsically safe use, so plan as if there is none. It is a consumer-derived pass-through headset. Use it in meeting rooms, training areas and controlled sections of the plant, and check with your safety officer and the local hazardous-area regulations before any wear near process equipment or moving machinery. Rugged purpose-built wearables exist for those zones.

How much does the Vision Pro M5 cost and when did it ship?

The M5 model was announced on 15 October 2025 and went on sale on 22 October 2025. In the United States it starts at $3,499 for 256 GB, with 512 GB and 1 TB options. The Dual Knit Band is included, and the band is sold separately for $100 in the US. Regional prices differ, for example $4,999 CAD in Canada. Enterprise volume pricing was not verified.

What enterprise APIs does visionOS 26 offer for manufacturing?

Apple offers main camera access, camera-region access, shared coordinate space, passthrough capture, barcode detection, protected content, object-tracking parameter control, a higher compute category, window follow and visual-fidelity monitoring. These require an enterprise API license requested by the Account Holder, plus entitlements and a renewable license file. Since visionOS 26, UVC camera and Neural Engine access no longer need an enterprise license.

Can Vision Pro show a live digital twin from Omniverse?

Yes. NVIDIA documents spatial streaming of OpenUSD stages from an Omniverse Kit application to Vision Pro, with a native Swift and SwiftUI client that sends head pose and events to the server. Production deployments can use NVIDIA’s Graphics Delivery Network. The documentation page itself warns it may be outdated and points to a newer site, so verify current SDK versions and requirements, and test latency on your own network.

Is HoloLens 2 still supported?

Microsoft discontinued HoloLens 2 in October 2024 and said security updates and fixes for critical regressions would continue through 31 December 2027, with warranty and developer support unchanged. No successor was announced. Existing deployments can keep running, but new projects should not depend on the platform, and teams should plan a migration path using portable content such as OpenUSD and glTF.

Should a factory wait for a lighter Vision Pro?

Probably not, if the use case is design review or training, because those are strong fits today. Reports from Bloomberg say Apple paused a cheaper, lighter headset in favour of smart glasses, and an unconfirmed preview says a Vision Pro 2 is at least two years off. For all-shift wear, the answer is to choose a different device class now rather than wait on an unannounced product.