Montreal, Monday, August 24, 2026. Three very different stories dominated the technology wires as this week opened, and although they look unrelated at first glance, they are all arguments about the same thing: where the light, the arithmetic and the trust in a computing system should physically live.

The first is a display story. Samsung has spent 2026 rolling out a television backlight architecture called Micro RGB, in which the white light behind an LCD panel is abandoned entirely and replaced by millions of independently driven red, green and blue emitters smaller than a human hair. The second is a silicon story. Between TSMC's Arizona expansion, AMD's Helios rack, a former Intel chief executive telling an AI conference that "GPUs suck," and Samsung's confirmed August 27 Galaxy event, the industry is visibly renegotiating which computations happen in a data centre and which happen in the device on your desk. The third is a security story, and it is the least glamorous and most urgent: Microsoft's August 2026 Patch Tuesday closed 421 vulnerabilities, one of which attackers were already using to take complete control of Windows machines.

This edition of Tech Science Daily takes the ten most consequential stories of the past week, then goes deep on the three that carry real scientific content and real consequences for the equipment you buy. We explain the physics and the engineering in plain language, and we point to hardware we actually have on the shelf in Montreal — verified in stock at the time of writing.

Today's Tech Radar

Stories 1, 3 and the cluster around 4–6 and 2 are the ones with genuine scientific depth. We take them in turn.

Deep Dive 1 — Micro RGB: why televisions are abandoning white light

Modern LCD televisions live or die by the quality of the light behind the panel. Photo: Nicolas J Leclercq / Unsplash.

The problem every LCD has to solve

A liquid crystal display does not make light. It edits light. Behind the visible panel sits a backlight, and in front of that backlight sits a sandwich of polarisers, a liquid crystal layer and a colour filter array. Each pixel is divided into three sub-pixels covered with red, green and blue filters. Applying a voltage twists the liquid crystal molecules, which rotates the polarisation of the light passing through and determines how much of it escapes the second polariser. The colour filter then throws away everything that is not the wanted wavelength band.

That last sentence contains the fundamental inefficiency of the technology. A colour filter is a subtractive device: it works by absorbing photons. If you shine broadband white light at a red filter, roughly two-thirds of the energy is converted to heat rather than image. Every improvement in LCD picture quality over the past fifteen years has essentially been an attempt to make the light arriving at that filter less wasteful and more precisely controlled.

Three generations of backlight, briefly

Generation one — white LED with phosphor. Conventional LED televisions use blue LEDs coated in a yellow phosphor. The blue photons excite the phosphor, which re-emits a broad yellow band; the mixture reads as white. It is cheap and reliable, but the spectrum is broad and lumpy. Broad emission peaks mean the red filter passes some orange, the green filter passes some yellow, and the resulting colour gamut is narrow. Contrast is limited by how finely the backlight can be dimmed — often only in a handful of edge-lit zones.

Generation two — Mini LED with quantum dots. Here the backlight is a dense array of very small blue LEDs, and a quantum-dot film sits between the LEDs and the panel. Quantum dots are semiconductor nanocrystals whose emission wavelength is set by their physical diameter, because of quantum confinement: squeeze an electron-hole pair into a smaller box and its energy levels move further apart, shifting the emitted photon towards the blue. Tune the dot size and you get a narrow, tall emission peak exactly where you want it. Narrow peaks pass through colour filters with far less loss and far less cross-contamination, which is why quantum-dot LCDs cover much wider gamuts. Because there are now hundreds or thousands of individually addressable LEDs, the set can also perform local dimming: darkening the backlight zones behind dark parts of the image to deepen blacks.

Generation three — Micro RGB. Samsung's 2026 approach removes the white-light intermediary altogether. Instead of blue LEDs plus phosphor or blue LEDs plus quantum dots, the backlight is built from sub-100-micrometre red, green and blue LEDs that each emit their own colour independently. Samsung states that the 2026 lineup spans 55, 65, 75, 85, 100 and 115 inches, and that its "Micro RGB Precision Color 100" implementation is VDE-certified at 100% of the BT.2020 colour gamut.

Why direct RGB emission changes the engineering

Three consequences follow, and they are worth understanding before you spend money.

Colour volume, not just colour gamut. Gamut describes which colours a display can reproduce; colour volume describes which colours it can reproduce at a given brightness. Most displays lose saturation as they get brighter, because they reach peak luminance by adding white. When the backlight itself can be driven as red, green and blue independently, the set can push a saturated red to high luminance without diluting it. This is the technical reason a bright, deeply saturated sunset looks different on these panels rather than merely brighter.

Colour-aware local dimming. A conventional local-dimming zone has one degree of freedom: brightness. An RGB zone has three. That means the backlight can be dimmed and tinted per zone, matching the chroma of the region it sits behind. Because the light arriving at the colour filter is already close to the target colour, less of it is thrown away, and the halo of stray light around bright objects on dark backgrounds — blooming — is both dimmer and better colour-matched to its surroundings, which makes it far less visible to the eye.

BT.2020 is a genuinely hard target. BT.2020 is the ITU colour space defined for ultra-high-definition broadcasting. Its primaries are monochromatic points on the spectral locus, meaning they sit at the physical boundary of what light can do. Practically no consumer display fully covers it; most premium sets cover a large fraction of the smaller DCI-P3 space. A VDE certification at 100% of BT.2020 is a meaningful engineering claim precisely because the target is so demanding.

Where Micro RGB does not win

Honesty matters more than enthusiasm here. Micro RGB is still a backlit LCD. The liquid crystal layer never blocks light perfectly, so there is always residual leakage, and a self-emissive technology such as OLED — where each sub-pixel generates its own light and can be switched fully off — still has a structural advantage in absolute black level and in off-axis viewing. Industry analysis through 2026 continues to place OLED and QD-OLED at the top of the premium consumer market on picture quality and availability, with true MicroLED remaining a niche product because of manufacturing cost. Micro RGB is best understood as LCD closing much of that gap while keeping LCD's advantages: higher sustained full-screen brightness, no burn-in risk, and far better economics at very large sizes.

Practical buying advice

That last point is the one that matters for most real installations. If you are lighting a bright room, running a screen for eight hours a day with static elements such as menus, dashboards or signage, or you simply need a very large diagonal, a well-engineered LCD remains the correct engineering choice regardless of how good OLED looks in a dark showroom. Burn-in is a real constraint on organic emitters under static content, and sustained full-field brightness is where backlit panels are strongest.

For large-format work we currently have the LG 86-inch commercial-grade 4K display, 3840 × 2160 at 350 cd/m² (in stock). It is a commercial panel rather than a consumer television, which matters more than the specification sheet suggests: commercial displays are rated for extended duty cycles, ship with landscape and portrait orientation support, and are built around the assumption that they will be on all day. For collaborative rooms and classrooms, the 98-inch ViewBoard 4K Ultra HD interactive flat panel (in stock) adds touch input across a diagonal large enough to be legible from the back of a room.

One dimensioning rule worth applying: at 4K resolution, the pixel structure becomes invisible to a viewer with normal acuity at roughly 1.5 times the screen height. For an 86-inch 16:9 panel, screen height is about 107 cm, which puts the "resolution-limited" seating distance at roughly 1.6 metres, and comfortable viewing at two to three metres. Rooms deeper than that are not getting the benefit of the extra pixels; they are getting the benefit of the extra size. Both are legitimate reasons to buy, but they are different reasons, and knowing which one applies to your room prevents overspending. If you would like us to size a display against your actual room dimensions and lighting, you can request a free quote from our team.

Deep Dive 2 — The great relocation: where AI arithmetic actually happens

Modern processors increasingly dedicate silicon area to specialised matrix engines rather than general-purpose cores. Photo: BoliviaInteligente / Unsplash.

Reading five separate news items as one story

Take the week's chip headlines together. TSMC expands Arizona to $265 billion and ramps 2 nm. AMD launches Helios, integrating Epyc 9006 CPUs, Instinct MI455X GPUs and Pensando networking into one rack. Pat Gelsinger stands on a stage at Ai4 2026 and says the GPU-plus-HBM stack is power-hungry and inefficient. AMD buys Taalas, Etched raises around $700 million, and frontier labs keep designing their own accelerators. Broadcom explores more than $60 billion in debt to finance AI silicon. Copper supply becomes a construction bottleneck.

Every one of those is a statement that the general-purpose graphics processor, running in a data centre, on grid power, is an expensive way to perform inference. And the logical response to that observation is not only better data-centre chips. It is also moving as much inference as possible out of the data centre and onto the device in front of the user. That is precisely what Samsung's August 27 Galaxy event is about — bringing the Galaxy S26 family's camera and AI experiences to a broader audience — and it is what the Copilot+ PC category has been about since it launched.

What an NPU actually is

A neural processing unit is not a small GPU. It is a different trade in the same three-way space of flexibility, throughput and energy per operation.

A CPU core is built for latency on unpredictable, branch-heavy code. It spends enormous silicon area on branch prediction, out-of-order execution and large caches so that a single instruction stream finishes as fast as possible. That machinery is close to useless for neural network inference, which is overwhelmingly dense matrix multiplication with a completely predictable access pattern.

A GPU exploits that predictability with thousands of simple lanes running in lockstep. But it remains programmable enough to run graphics, physics and scientific code, and that programmability costs energy: instruction fetch and decode, register file traffic, and memory movement dominate the power budget rather than the arithmetic itself.

An NPU narrows the target further. It implements matrix multiply-accumulate directly in hardware, typically as a systolic array in which data flows through a grid of multiply-accumulate cells and each partial sum is passed to its neighbour rather than written back to a register file. Because the physical energy cost of moving a byte across a chip vastly exceeds the cost of the multiplication itself — often by one to two orders of magnitude — eliminating that movement is where the efficiency comes from. NPUs also run at reduced numerical precision, commonly INT8 or 8-bit floating point instead of FP32. Halving the bit width roughly quarters the area of a multiplier and cuts the memory bandwidth needed to feed it, and modern quantisation techniques keep accuracy loss small for inference workloads.

The net result: an NPU can deliver tens of trillions of operations per second within a laptop's thermal envelope, running continuously without the fan spinning up. Microsoft's Copilot+ PC specification sets a floor of 40 TOPS for exactly this reason — it is the point at which useful models run locally and continuously rather than in bursts.

Why local inference is a privacy and cost story too

There are four independent reasons to prefer on-device inference, and only one of them is speed.

Latency. A round trip to a cloud endpoint costs tens to hundreds of milliseconds before the model even starts. For live captioning, translation, background blur or noise suppression, that budget is simply not available.

Privacy. Audio that never leaves the machine cannot be intercepted, retained or subpoenaed. For anyone handling client files, medical information or legal work, this is not a preference but a compliance requirement.

Cost. Cloud inference is metered. Device inference is a capital expense you have already paid.

Availability. A local model works on a plane, in a basement, and during an outage.

Foldables, and why the Galaxy Z Fold7 is a good illustration

The same relocation logic applies to phones, and this week's Galaxy event is the consumer-facing edge of it. It is worth pausing on the engineering of a modern foldable, because it is one of the more impressive materials-science achievements in consumer electronics.

Folding a display requires solving a problem in mechanics. When any layered structure bends, the material on the outside of the curve is stretched and the material on the inside is compressed; somewhere between them lies a neutral axis where strain is zero. The strain experienced by any layer scales roughly with its distance from that neutral axis divided by the bend radius. Glass fails in tension at a fraction of a percent of strain, so a foldable panel must be built from ultra-thin flexible glass or polymer layers, laminated so that the fragile emissive layer sits as close to the neutral axis as possible, over a hinge geometry that keeps the local bend radius large rather than creasing the panel sharply. This is why foldable hinges are multi-link mechanisms rather than simple pivots: they let the panel form a teardrop shape when closed instead of a tight fold.

We currently have the Samsung Galaxy Z Fold7, 512 GB, in Blue Shadow (in stock): an 8-inch flexible Dynamic AMOLED 2X panel at 2184 × 1968, an octa-core Oryon-based processor with two cores at 4.47 GHz and six at 3.53 GHz, 12 GB of RAM, Android 16 and 5G. The 12 GB figure is the one to notice for AI purposes. Running a quantised language model locally requires the weights to be resident in memory, and memory capacity — not raw compute — is usually the binding constraint on which models a device can host.

If a phone is not the right form factor, the same on-device AI logic applies to tablets. The Samsung Galaxy Tab S10+, 12.4-inch WQXGA+, Dimensity 9300+ on a 4 nm process, 12 GB RAM and 256 GB storage (in stock) is the capable end of that range, while the Samsung Galaxy Tab A9+, 11-inch WUXGA, Snapdragon 695 5G, 4 GB RAM (in stock) is the sensible choice where the job is browsing, documents, video and light field work rather than local model inference.

Copilot+ laptops: what to actually look for

On the PC side, two silicon philosophies are on our shelves right now, and they suit different buyers.

The Arm route is represented by the Dell Latitude 5455, a 14-inch Copilot+ PC with the Qualcomm Snapdragon X Plus X1P-42-100, 16 GB of RAM and a 512 GB SSD (in stock). Arm designs generally deliver the best performance per watt in this class, which shows up as battery life measured in working days rather than working hours. The trade-off is that a minority of legacy x86 applications and specialised drivers still run through emulation or not at all — worth checking against your specific software list before committing a fleet.

The x86 route is represented by the Dell Pro 16 Plus PB16250, a 16-inch Copilot+ PC with Intel Core Ultra 7 268V, vPro, 32 GB of RAM and a 512 GB SSD (in stock), and by the more portable Dell Pro 13 Premium PA13250 with Core Ultra 5 236V and vPro (in stock). You get full native compatibility with everything Windows has ever run, plus vPro, which — as the next section explains — is not a marketing checkbox but a genuine security capability.

Our practical guidance on memory: 16 GB is a floor, not a target, on any machine expected to run local AI features alongside a browser and an office suite. If the budget stretches to 32 GB, take it. Memory is the one specification you cannot upgrade on most modern thin-and-light designs, because the RAM is packaged on the processor module itself.

Deep Dive 3 — 421 patches, one zero-day, and the physics of a use-after-free

Kernel-level privilege escalation flaws turn ordinary user access into total system control. Photo: FlyD / Unsplash.

What Microsoft shipped on August 11

Microsoft's August 2026 Patch Tuesday resolved 421 CVEs — 236 in Windows, 98 in Office, 98 in Office 2016, 30 in SharePoint Server, 26 in developer tools, 17 in Azure, 7 in Exchange Server, 1 in Defender and 6 elsewhere, plus two non-Microsoft CVEs in the TPM 2.0 reference implementation. That is an unusually large release even by recent standards.

One flaw stands apart. CVE-2026-68820 is a use-after-free in the Ancillary Function Driver for WinSock, afd.sys — the kernel-mode driver that underpins the Windows Sockets API and therefore sits beneath essentially all network communication on the system. Microsoft states that threat actors were already exploiting it: "A locally authenticated attacker could run a specially crafted application on an affected system to trigger a race condition. Successful exploitation could allow the attacker to gain SYSTEM privileges. User interaction is not required." CISA added it to the Known Exploited Vulnerabilities catalog on August 11.

Tenable's Satnam Narang noted that this is the fourth afd.sys zero-day exploited in the wild since 2022, following CVE-2025-32709, CVE-2025-21418 and CVE-2024-38193 — the last of which was reportedly exploited by North Korean actors linked to the Lazarus group. He observed that, based on historical tradecraft against this driver, nation-state involvement is plausible.

The mechanism, explained without jargon

A use-after-free is a memory-safety bug, and the metaphor that works best is a hotel room key.

When a program needs memory, it asks the allocator for a block — checking into a room. When it is finished, it frees the block — checks out. The allocator is then entitled to give that same physical memory to the next request. A use-after-free occurs when the program keeps a pointer to the freed block and later uses it: the guest kept a copy of the key and walks back in after a new guest has moved in. Whatever the new occupant put in the room is now being read, or overwritten, by code that believes it owns the space.

In user-space applications this typically causes a crash. In kernel space it is catastrophic, because kernel code runs with complete authority over the machine. If an attacker can control what data lands in the reused block — a technique called heap grooming, in which the attacker makes carefully sized allocations to steer the allocator's reuse decisions — then the stale pointer can be made to reference an attacker-controlled structure. If that structure contains a function pointer, the kernel can be tricked into calling attacker-chosen code with kernel privileges.

Why "race condition" is the operative phrase

Microsoft's wording specifies a race condition, and this is the part that explains why such bugs survive code review.

A network driver such as afd.sys is heavily concurrent by nature: many threads on many processor cores handle sockets simultaneously. Suppose thread A checks that a socket object is valid, and thread B closes and frees that same object in the microseconds between A's check and A's use of it. Neither thread is individually wrong. Reading either function in isolation, the logic looks correct. The defect exists only in the interleaving — a specific, narrow window measured in microseconds where the two operations overlap in exactly the wrong order.

Attackers exploit these windows by widening them artificially: pinning threads to specific cores, flushing caches to make the target operation slower, or generating memory pressure so the scheduler behaves predictably. A window that occurs naturally once in millions of executions can be made to occur reliably by an attacker who controls the surrounding conditions. This is why race conditions are so difficult to find with conventional testing — the bug is invisible in almost every run.

It is also why the industry's long-term answer is architectural rather than procedural: memory-safe languages such as Rust enforce, at compile time, that a reference cannot outlive the object it points to, and that mutable access cannot be shared across threads without synchronisation. Rewriting decades of kernel code is a multi-year project, so in the meantime, patching remains the control that actually works.

Two more items from the same release worth flagging

CVE-2026-62832 is an improper link resolution flaw in the Windows User Profile Service. An authenticated attacker holding credentials for another local account can load a second user's registry hive and thereby access or modify that user's data and obtain administrator privileges. Microsoft flagged it as publicly disclosed and considers exploitation likely.

Remote code execution bugs were also patched in Windows DNS Server (CVE-2026-62878), the Windows Deployment Services TFTP server (CVE-2026-62893), Microsoft QUIC (CVE-2026-62815) and Microsoft HPC Pack (CVE-2026-59124), alongside an elevation-of-privilege flaw in Exchange Server (CVE-2026-62911). If you run any of those services, they belong at the front of your queue, because they are reachable across the network rather than requiring a local foothold.

What to actually do

Five concrete measures, in order of return on effort:

1. Patch, and verify that patching happened. "Automatic updates are on" is not the same as "the update installed." Machines that sleep, machines that are low on disk space, and machines with pending reboots routinely fall behind. Check compliance rather than assuming it.

2. Take away local administrator rights. CVE-2026-68820 is a privilege-escalation flaw, which means it is the second step of an attack. The attacker must already be executing code as some user. Standard-user accounts do not prevent that step, but they materially raise the cost of the whole chain, and they block a large class of simpler attacks outright.

3. Prioritise network-reachable services. A DNS or TFTP server exposed to a hostile network is a fundamentally different risk from a local escalation bug on a laptop. Sequence accordingly.

4. Use hardware-rooted protections. A discrete TPM plus virtualisation-based security means credentials live in memory the kernel itself cannot casually read, which limits what a kernel compromise yields. Intel vPro adds out-of-band remote remediation — you can reach a machine that will not boot without physically touching it, which is the difference between a two-hour incident and a two-day one across a distributed team. Both the Dell Pro 16 Plus PB16250 (in stock) and the HP EliteBook 840 G11 with Core Ultra 5 125U and vPro (in stock) carry vPro.

5. Watch the authentication layer as well. This week also brought reports of a phishing toolkit that abuses passkeys to maintain access after a password reset. Passkeys are a genuine improvement over passwords — they are phishing-resistant by design because the private key never leaves the device and the browser will not release a credential to the wrong origin — but any enrolled credential is a persistence mechanism if an attacker gets one registered. Auditing which authenticators are enrolled on privileged accounts belongs on the same checklist as patching.

If you are not certain where your fleet stands on any of these five points, that is a normal position to be in, and it is a solvable one. You can request a free quote from our team for a device-security review, and we will tell you plainly which of your machines are exposed and what it would cost to close the gap.

Glossary of the Week

Setup at a Glance

The takeaway

The unifying thread across all three deep dives is locality. Micro RGB moves colour generation from a filter that discards photons to an emitter that produces the right ones in the first place. NPUs move inference from a metered data centre on the far side of a network to silicon a few centimetres from the user. And the afd.sys zero-day is a reminder that the most dangerous code is the code closest to the hardware, where a single stale pointer is the difference between a crash and a total compromise.

None of this means you should replace working equipment. It means that when you do replace it, the questions worth asking have shifted. Not "how many nits?" but "how is the light generated?" Not "how many gigahertz?" but "how much memory, and is there an NPU?" Not "does it have antivirus?" but "can I reach it remotely when it will not boot, and does the user run as administrator?"

If you would like help translating any of that into a specific purchase or a fleet plan, we are in Montreal and we are happy to go through it with you — no obligation. Just request a free quote from our team and tell us what you are trying to accomplish.

Sources & Further Reading

Microsoft Patch Tuesday and CVE-2026-68820: SecurityWeek, "August 2026 Patch Tuesday: Microsoft Fixes 421 CVEs, One Exploited Zero-Day" and Zero Day Initiative, "The August 2026 Security Update Review", with the vendor release notes at Microsoft MSRC. Samsung's August event: Samsung Global Newsroom, "[Invitation] Galaxy Event August 2026" and 9to5Google. Micro RGB technology and 2026 lineup: Samsung, "Samsung Expands Premium Micro RGB Lineup for 2026", Samsung, "What is Micro RGB TV?" and Samsung US Newsroom. Display market context: Analytics Insight, "Quantum Dots vs OLED vs MicroLED" and ecoustics, 2026 TV Shootout results. Data-centre silicon, TSMC Arizona, AMD Helios and the Gelsinger keynote: DataCenterKnowledge, "Data Center Hardware Highlights: August 2026". Custom AI chips: HPCwire, "What's Hot Now: Custom AI Chips". Broadcom financing and general daily roundup: Tech Startups, August 21, 2026. AI and authentication security: SecurityWeek on encrypted prompts bypassing AI guardrails and SecurityWeek on the passkey phishing toolkit. Photos: Unsplash (free commercial license).

Product availability and stock levels were verified against our Shopify inventory on August 24, 2026, and can change without notice. This article explains publicly reported technical information and is not a substitute for a security assessment of your own environment.