Montreal, Tuesday, September 1, 2026. The first day of September is, for the technology industry, the last quiet morning before the storm. IFA Berlin — the world's oldest and largest consumer electronics show — opens its Media Days tomorrow and its show floor on Friday, September 4. Qualcomm has confirmed its Snapdragon Summit for later this month. Hot Chips 2026 has just closed, leaving behind a stack of architecture disclosures that will shape the laptops, tablets and displays sold for the next three years. And underneath all of it runs a story that touches every single price tag in this store: the global memory shortage, now in its fourth consecutive quarter of price increases.

This is Tech Science Daily, our attempt to do something slightly unusual for a retail blog — to explain the physics and engineering behind the headlines rather than simply repeating them. Today we look at three subjects with real scientific depth: the arrival of Micro RGB backlighting and what it actually does to the colour you perceive; the memory wall that has turned DRAM into the most contested commodity in computing; and the 2-nanometre gate-all-around transition that is quietly rewriting what a phone, tablet or laptop can compute on its own, without a data centre.

As always, we start with a scan of the week's ten most consequential stories, then dive into the three we think reward a closer reading.

Today's Tech Radar

1. Micro RGB: What Happens When You Stop Filtering Light and Start Emitting It

Modern large-format panels live or die on how precisely they can control light at the sub-millimetre scale. Photo: Nicolas J Leclercq / Unsplash.

Samsung confirmed that its 2026 Micro RGB television lineup will span 55, 65, 75, 85, 100 and 115 inches — a dramatic broadening from the single 115-inch model introduced in 2025. The interesting part is not the size chart. It is the sentence buried in the technical description: the panel uses sub-100µm red, green and blue LEDs that each emit light independently.

To understand why that matters, you have to understand what almost every LCD television you have ever owned actually does to light.

The subtractive problem at the heart of LCD

A conventional LED-backlit LCD is, physically speaking, a lamp with a stencil in front of it. Behind the panel sits a backlight unit (BLU) made of white LEDs. Those white LEDs are usually not white at all: they are blue gallium nitride emitters coated with a phosphor (or, in "QLED" designs, paired with a quantum-dot film) that converts part of the blue output into longer wavelengths, producing something the eye reads as white.

That broadband white light then passes through a liquid-crystal layer, which acts as a per-pixel light valve, and finally through a colour filter array — a mosaic of red, green and blue dyes. Here is the thermodynamic insult: the colour filter works by absorption. To make a red subpixel, the filter throws away the green and blue components of the white light and lets a slice of the spectrum through. Typically only around a third of the light entering the filter stack survives; in practice, once you include the polarisers, total optical efficiency from backlight to viewer is often in the single-digit percentages.

Worse, the colour you get is limited by the shape of the transmission curve of the dye and the spectral shape of the source. Broad, overlapping emission peaks produce desaturated primaries. Narrow peaks produce saturated ones. This is the entire physical reason quantum dots improved LCD colour: they replaced a wide phosphor emission with a narrow, size-tunable one.

What Micro RGB changes

Micro RGB attacks the problem at the source. Instead of a white backlight plus an absorptive filter, the backlight unit itself is composed of discrete red, green and blue LEDs, each under 100 micrometres — smaller than the width of a human hair — and each individually drivable. The colour of the light arriving at a region of the panel is now set by which emitters are on and how hard, not by what a dye is willing to let through.

Three consequences follow directly from the physics:

Wider gamut. Direct-emission LEDs have narrow spectral linewidths. Samsung states that the 2026 lineup's "Micro RGB Precision Color 100" achieves 100% of the BT.2020 colour space, verified by VDE. BT.2020 is the ultra-high-definition colour standard defined by the ITU; its primaries sit on the spectral locus of the CIE chromaticity diagram, meaning they correspond to essentially monochromatic light. No display using broadband sources and absorptive filters gets close to full BT.2020 coverage. Reaching it requires narrow-band, directly modulated primaries.

Finer local contrast. Because the emitters are individually addressable at a far higher spatial density than the few hundred or few thousand dimming zones of a conventional mini-LED set, the backlight can follow image content much more closely. The visible artefact this reduces is blooming — the halo you see around a white subtitle on a black background, which is simply light from a dimming zone that is larger than the bright object it is illuminating.

Colour-aware dimming. This is the subtle one. With a white backlight you can only modulate luminance per zone. With independent RGB emitters you can modulate chromaticity per zone. A region of the image dominated by deep red can be driven with the red emitters running hot and the green and blue nearly off, which raises the saturation ceiling in that region rather than pushing more white light through a filter that will only throw most of it away.

Where this sits between mini-LED and true micro-LED

It is worth being precise about terminology, because the marketing is genuinely confusing. In a true micro-LED display, each subpixel is an LED; there is no liquid-crystal layer at all. That is what direct-view LED video walls do, and it is why they are so expensive per square metre: you are transferring millions of individual dies onto a substrate. In Micro RGB, the LCD layer remains, but the backlight has been reconstructed from coloured micro-emitters. It is an intermediate architecture — more manufacturable than full micro-LED, and optically far better behaved than white-LED LCD.

Samsung also pairs the panel with a "Micro RGB AI Engine Pro" chipset performing frame-by-frame processing, plus 4K AI Upscaling Pro and AI Motion Enhancer Pro, and a proprietary Glare Free anti-reflection treatment. Anti-reflection matters more than most buyers expect: in a bright room, ambient light scattering off the front surface raises the black level, and no amount of backlight precision can recover contrast that has been destroyed at the glass.

Practical buying advice

Micro RGB sets will be premium products for some time. The engineering question for most buyers — and for every business buyer — is not "should I wait for Micro RGB" but "what am I actually limited by today?" In our experience the answer is usually one of three things: peak brightness in a bright room, uniformity across a very large diagonal, or duty-cycle reliability if the panel runs long hours.

If your constraint is a very large diagonal at a sane price, the LG 86-inch commercial LED display, 3840×2160 at 350 cd/m² (in stock) is the pragmatic answer. Eighty-six inches of 4K puts the pixel pitch at roughly 0.49 mm, which means that from a normal seating distance of three metres the individual pixels are comfortably below the resolving limit of a person with 20/20 vision.

If your constraint is continuous operation — a screen that runs all day in a lobby, clinic or classroom — consumer TVs are the wrong tool, because their thermal and backlight-ageing budgets assume a few hours a day. The Samsung 55" Crystal UHD Signage QBC (in stock), the Samsung 75" Professional Display QET Series (in stock) and the ViewSonic 75" 4K UHD wireless presentation display rated for 24/7 operation (in stock) are all built for that duty cycle.

And if you want to see what the direct-emission endpoint of this technology actually looks like today, that already exists in the commercial channel: the Samsung LED cabinet with 1.5 mm pixel pitch (in stock) is a true direct-view LED tile, with no LCD layer and no colour filter, assembled into walls of arbitrary size. Micro RGB is, in a sense, the consumer-scale descendant of that idea. If you are sizing a video wall and are unsure how many cabinets a given viewing distance requires, request a free quote from our team and we will work the geometry with you.

2. The Memory Wall: Why RAM Became the Most Expensive Thing in Your Computer

Every DRAM die competes for the same finite wafer starts and the same advanced packaging capacity. Photo: Laura Ockel / Unsplash.

TrendForce's latest survey, reported by Tom's Hardware, projects conventional DRAM contract prices rising 13–18% quarter-over-quarter in Q3 2026, with NAND flash up 10–15%. Those are large numbers presented as good news — because the second quarter saw increases of roughly 60%. The market is decelerating, not reversing.

The reason given in the report is worth reading carefully: the cooldown is driven by consumer electronics manufacturers being unwilling and unable to absorb further cost, not by supply improving. Demand destruction, in other words, rather than relief.

What a "memory wall" actually is

The phrase predates the current AI boom by about thirty years. It describes a structural divergence: processor throughput has historically improved far faster than the rate at which data can be moved between the processor and main memory. Compute capability roughly doubles on a short cadence; DRAM bandwidth improves on a much slower one. The gap compounds. Eventually the arithmetic units spend most of their time idle, waiting for operands.

For large language model inference this is not a subtlety — it is the dominant term. Consider what happens when a transformer model generates text. In the decode phase, the model produces one token at a time. To produce each token it must read essentially the entire weight matrix of the model from memory. The arithmetic per weight read is tiny: roughly one multiply-accumulate. The operation is therefore memory-bandwidth-bound, and to a first approximation:

tokens per second ≈ memory bandwidth ÷ model size in bytes

An 8-billion-parameter model quantised to 4 bits occupies roughly 4–5 GB. On a laptop with 135 GB/s of usable memory bandwidth, the theoretical ceiling is around 25–30 tokens per second; real systems land lower because of overheads. Double the bandwidth and you roughly double the throughput. Double the TOPS rating of the NPU and, for this phase of the workload, you change very little. This is why hands-on testing consistently finds that memory bandwidth and capacity predict local AI performance better than the NPU's headline TOPS number, and why moving from 16 GB to 32 GB of RAM often does more for real-world local inference than moving to a faster accelerator.

HBM4 and the physics of getting data off a chip

The industry's answer to the memory wall in the data centre is High Bandwidth Memory, now in its HBM4 generation and reported in mass production in 2026 from Samsung, SK hynix and Micron. Micron has confirmed 36 GB 12-high stacks delivering more than 2.8 TB/s.

The underlying trick is geometric. A conventional DDR5 DIMM communicates over a relatively narrow bus at very high frequency; the signalling has to survive centimetres of PCB trace, and every doubling of frequency costs power roughly linearly and signal integrity considerably more. HBM inverts this. DRAM dies are stacked vertically — eight, twelve or sixteen high — and connected by through-silicon vias, vertical copper columns etched straight through the silicon. The stack sits on an interposer, millimetres from the processor die, and communicates over an enormously wide, relatively slow bus.

HBM4 doubles that interface from 1,024 to 2,048 bits per stack across 32 independent channels, exceeding 2 TB/s per stack while keeping per-pin data rates around 8 Gbps. Because energy per bit transferred scales with distance and with the square of signalling voltage, moving the same data over a wide short bus at modest speed is dramatically more energy-efficient than over a narrow long bus at high speed. In an accelerator where power is the binding constraint, that efficiency is the performance.

Intel's Hot Chips 2026 disclosures illustrate how central this has become: Crescent Island, its next-generation inference GPU built on the Xe3P architecture, is specified with HBM4 and liquid cooling. Micron used the same conference to argue that memory architecture and advanced packaging — not raw arithmetic — now sit at the centre of AI system design.

Why this reaches your laptop invoice

Here is the mechanism, and it is unglamorous. HBM stacks are built from DRAM dies fabricated on the same wafers, in the same fabs, using much of the same capacity as the DDR5 and LPDDR5 that goes into laptops and phones. HBM also consumes disproportionate amounts of advanced packaging capacity, which is separately constrained. Server memory carries higher margins. When a manufacturer must choose how to allocate a fixed number of wafer starts, the allocation follows the margin.

The result, reported across CNBC, IEEE Spectrum and IDC's market analysis, is a consumer memory market that is structurally short. IDC has forecast device price increases in the 10–20% range across PCs, tablets and smartphones for 2026. TrendForce expects notebook makers to keep replenishing inventory while passing costs through, and smartphone vendors to raise handset prices to offset persistently high LPDRAM costs.

Practical buying advice: buy the memory you will need, now

There is a specific, actionable conclusion here, and it runs against normal purchasing instinct.

In an ordinary market, the rational strategy is to buy the minimum configuration and upgrade later, because component prices fall over time. In 2026 that logic is inverted. Memory prices are rising, most thin-and-light notebooks now use soldered LPDDR5X that cannot be upgraded at all, and the price delta between a 16 GB and a 32 GB configuration at purchase is far smaller than the cost of replacing the whole machine in eighteen months.

Concretely: if the machine will run virtual machines, large datasets, heavy browser workloads or any local AI model, specify 32 GB or more at the outset. In stock today, the Dell Pro 16 Plus PB16250 with an Intel Core Ultra 7 268V, 32 GB and a 512 GB SSD and the Lenovo ThinkPad T14s Gen 6 with a Snapdragon X Elite X1E-78-100, 32 GB and a 1 TB SSD both start at that level rather than treating it as an upgrade.

For workstation-class local inference, where you want to hold a large model resident, the HP Z2 G9 Workstation with a Core i7-14700, 64 GB and a 1 TB SSD (in stock) and the Dell Pro Max Tower T2 with a Core Ultra 9 285, 32 GB across two DIMMs and a 1 TB SSD (in stock) both use socketed DIMMs — which, in a rising market, at least preserves the option of adding capacity from a module you buy separately. Note the two-DIMM configuration on the Dell: populating both channels is not optional if you care about bandwidth, since a single-channel configuration halves it.

For fleet standardisation where per-seat cost matters more than peak performance, the Lenovo ThinkCentre M70q Gen 5 with a Core i5-14400T, 16 GB and a 512 GB SSD is in stock in quantity — and, being a tiny-form-factor desktop with accessible memory slots, it is one of the few remaining categories where "upgrade later" is still a real option. If you are planning a refresh cycle and want help modelling the memory-cost exposure across a fleet, request a free quote from our team.

3. Two Nanometres, Gate-All-Around, and the Quiet Arrival of Serious On-Device AI

The N2 node marks the first change in transistor geometry since FinFET arrived more than a decade ago. Photo: Vishnu Mohanan / Unsplash.

Two announcements this month bracket the same technical transition. Apple has shipped its first 2 nm Mac — the M6 Mac mini, launched in late August, built on TSMC's N2 process. And Qualcomm has confirmed its Snapdragon Summit for September 22–24, where the Snapdragon 8 Elite Gen 6 and a new "Pro" tier are expected, both reported to be built on a 2 nm process, with clocks reaching 5 GHz.

"2 nm" is, to be clear, a marketing name rather than a physical dimension. Nothing on an N2 die measures two nanometres. What the node designation signals is a package of density, performance and power characteristics — and in this case, a genuine change in the shape of the transistor itself.

From FinFET to nanosheet

A transistor is a switch. A gate electrode, separated from a silicon channel by a thin insulator, creates an electric field that either permits or blocks current between source and drain. The engineering problem, for the last twenty years, has been electrostatic control: as channels get shorter, the gate's authority over the channel weakens, and current begins to leak even when the switch is nominally off. Leakage is why a chip idles warm.

The fix, introduced commercially in 2011, was the FinFET: stand the channel up as a vertical fin and wrap the gate around three sides of it. More gate-to-channel surface area means more control. That geometry has been scaled, thinned and multiplied for over a decade, and it has run out of room.

TSMC's N2 replaces the fin with horizontally stacked silicon nanosheets — typically three to five layers, each on the order of five to ten nanometres thick — with the gate material deposited so that it surrounds each sheet on all four sides. This is the gate-all-around (GAA) architecture. The channel is now completely enclosed by its controlling electrode, which is the best electrostatic configuration available short of exotic research devices.

Two practical benefits follow. First, leakage drops sharply for a given channel length, which means the supply voltage can be lowered without the transistor becoming unreliable — and since dynamic power scales with the square of voltage, that is a large win. Second, drive strength becomes tunable by design: a designer can specify the width of the nanosheets to trade current drive against area, in a way FinFET's quantised fin count never allowed.

TSMC rates N2 at roughly 10–15% more performance at the same power, or 25–30% less power at the same performance, versus the N3E node, with about 15% higher transistor density.

Why a mobile SoC cares more about the power number than the performance number

In a phone or a fanless tablet, sustained performance is set by thermal dissipation, not by peak clock. A device that can hit 5 GHz for two seconds and then throttles is, from the user's perspective, a slower device than one that holds 3.5 GHz indefinitely. The 25–30% power reduction at equal performance therefore translates fairly directly into sustained throughput, which is what you feel when exporting video, running a long camera pipeline, or executing a local model over a document.

This is precisely why the 2 nm transition and the on-device AI story are the same story. Neural processing units are wide, low-clock, highly parallel arrays of multiply-accumulate units. Their efficiency is dominated by two things: how many operations you can perform per joule, and how much data you have to move to feed them. GAA improves the first. Wider and faster LPDDR improves the second. Neither alone is sufficient.

The honest state of local AI in 2026

It is worth being blunt, because the marketing is not. Current-generation NPUs in the 40–60 TOPS range are excellent at what they were designed for: small, sustained models running continuously at low power — background noise suppression, camera processing, live transcription, semantic search over local files, the operating system's built-in assistant features. They are not a path to running a 70-billion-parameter model on a laptop.

Two structural caveats deserve mention. First, as discussed above, the decode phase of LLM inference is bandwidth-bound, so TOPS is not the limiting number for that workload. Second, much of the popular local-inference software ecosystem still targets CPU and GPU backends rather than NPUs, so the accelerator in your machine may be idle while the model runs. This is improving, but slowly, and it is worth verifying against your actual toolchain rather than a specification sheet.

The practical guidance that falls out of this: for local AI work, prioritise memory capacity first, memory bandwidth second, and NPU TOPS third. A discrete GPU with its own high-bandwidth memory changes the class of what is possible entirely, at the cost of power and portability.

Security: the other reason on-device computation matters

There is a security argument for on-device inference that is often made badly, so let us make it carefully. Running a model locally means the input data does not leave the machine. For a law firm summarising a client file, a clinic processing patient notes, or an engineering team searching internal drawings, that is a meaningful reduction in exposure surface — not because cloud providers are careless, but because data that never transits a network cannot be intercepted in transit or retained by a third party.

It is not, however, a substitute for basic operational hygiene, and this month supplied a sharp reminder. Microsoft's August 2026 Patch Tuesday addressed 421 CVEs, among them CVE-2026-68820, a use-after-free vulnerability in afd.sys, the Ancillary Function Driver for WinSock — a kernel-mode component that sits between user-space socket calls and the network stack. It was being exploited in the wild before the patch shipped. A use-after-free is exactly what it sounds like: memory is released, a stale pointer to it survives, and an attacker who can control what gets allocated into that freed region can steer execution. In a kernel driver, successful exploitation means privilege escalation.

Separately, the Cl0p group listed more than forty organisations as victims of a campaign against PTC's Windchill and FlexPLM product-lifecycle-management platforms, with names spanning manufacturing, healthcare and finance. PLM systems are attractive targets for the same reason file-transfer appliances were before them: they are internet-reachable, they are centrally deployed, and they contain a concentrated store of exactly the intellectual property an extortion group can monetise.

The lessons are not novel, which is rather the point. Maintain a patch cadence measured in days rather than quarters for internet-facing and kernel-level components. Inventory what is actually exposed — most organisations are surprised by their own attack surface. Prefer hardware with vPro or equivalent remote-management and attestation capability when you are managing a fleet you cannot physically touch, which is why the Dell Pro 16 Plus with vPro Technology (in stock) is a sensible fleet baseline. And if you are unsure whether your current endpoint estate can even be patched within a useful window, that is a conversation worth having — request a free quote from our team and we will walk through it with you.

Choosing a device on the eve of a node transition

A reasonable question: if 2 nm silicon is arriving, should you wait?

Generally, no — and the memory market is the reason. Node transitions deliver incremental efficiency gains that compound over years. The DRAM situation is delivering double-digit price increases over months. Waiting two quarters to gain 15% efficiency while paying 15% more for the same configuration is not obviously a winning trade. Buy when the specification meets the requirement.

What the 2 nm transition genuinely changes is the shape of the market from 2027 onward, and it is worth understanding so you specify sensibly today. Among in-stock options, the Snapdragon-based Lenovo ThinkPad T14s Gen 6 with a Snapdragon X Plus X1P-42-100, 16 GB and 512 GB illustrates the Arm-on-Windows efficiency case at a mainstream price, while the Lenovo ThinkPad X1 Gen 10 convertible with a Core Ultra 5 226V (in stock) covers the x86 compatibility case in a 2-in-1 chassis.

On the mobile side, the Samsung Galaxy Z Fold7 with 512 GB, 12 GB of RAM and an 8-inch folding Dynamic AMOLED 2X panel (in stock, limited quantity) is the current expression of the "large screen, high memory, on-device AI" thesis in a pocketable device — and 12 GB of RAM in a phone exists specifically so that a model can stay resident without evicting your applications. For tablets, the Samsung Galaxy Tab S11 with a 3 nm MediaTek Dimensity 9400+ and 12 GB (in stock) sits at the performance end, the Samsung Galaxy Tab A11+ with 6 GB and 128 GB (in stock, deep inventory) at the deployment-fleet end, and the Panasonic rugged Windows 11 Pro tablet with a Core Ultra 5-135U, 16 GB and 512 GB (in stock) where the device has to survive a job site.

Reading the Week as a Whole

Three stories, one theme. Micro RGB is an admission that the classical LCD light path — generate broadband white, then throw most of it away in a filter — has been optimised to exhaustion, and that the only remaining gains come from generating the right light in the first place. HBM4 and the DRAM shortage are an admission that arithmetic is no longer the scarce resource in computing; bandwidth is. And gate-all-around is an admission that the FinFET, after fifteen years, can no longer be made shorter without leaking.

All three are cases of an industry hitting a physical limit and responding by changing the structure of the device rather than shrinking it. That is a healthier kind of progress than the alternative, but it is also slower, more capital-intensive, and — as anyone pricing a 32 GB laptop this quarter has discovered — more expensive in the short run.

Glossary of the Week

Setup at a Glance

Closing

The through-line of this week is that the easy scaling is over — in displays, in memory and in transistors alike — and that the industry's response is to redesign the device rather than shrink it. For buyers, that means specifications matter more than they did five years ago, because the trade-offs are no longer hidden behind a generational improvement that fixed everything. Memory capacity is now a decision you make once. Duty cycle determines whether a panel lasts three years or ten. Sustained power, not peak clock, determines what a thin laptop can actually finish.

If you would like help matching any of this to a specific room, fleet or budget — sizing a video wall, standardising a laptop refresh under current memory pricing, or working out whether local AI is realistic for your workload — request a free quote from our team. We would rather talk through the constraints with you than sell you a specification you did not need.

Sources & Further Reading

Samsung Newsroom, Samsung Expands Premium Micro RGB TV Lineup for 2026 with New Sizes and Advanced Features · Samsung Global Newsroom, Samsung Unveils World's First 130-Inch Micro RGB TV · Tech Guide, Samsung reveals the technology behind its 2026 Micro RGB TVs · Tom's Guide, What to expect at IFA 2026 · Stuff, IFA 2026: everything worth knowing · Tom's Hardware, Memory price surge begins to cool as consumers hit affordability limit · TrendForce, Memory Wall Bottleneck: AI Compute Sparks Memory Supercycle · IEEE Spectrum, AI Boom Fuels DRAM Shortage and Price Surge · CNBC, AI memory is sold out, causing an unprecedented surge in prices · IDC, Global Memory Shortage Crisis · ServeTheHome, Micron: Evolving Memory Architectures for AI at Hot Chips 2026 · Intel Newsroom, Intel Outlines Architectures for Agentic AI at Hot Chips 2026 · Tom's Hardware, Hot Chips 2026: Intel dives deep on Crescent Island · Computerworld, At Hot Chips '26, all eyes were on AI costs, GPUs — and the future · MacRumors, 2nm guide · DIGITIMES, Weekly news roundup: Apple debuts 2nm M6 · Android Headlines, Qualcomm Confirms Two Snapdragon 8 Elite Gen 6 Chips Are Coming · Beebom, Snapdragon Summit 2026 dates confirmed · SecurityWeek, August 2026 Patch Tuesday: Microsoft Fixes 421 CVEs, One Exploited Zero-Day · CrowdStrike, August 2026 Patch Tuesday: Updates and Analysis · SWK Technologies, Cybersecurity News Recap August 2026 · Tech Startups, Top Tech News, August 27, 2026 · StorageReview, Best Laptops for Local AI in 2026. Photos: Unsplash (free commercial license).

Product availability and inventory counts were verified at time of writing and can change without notice.