Montreal, Thursday, September 3, 2026. Some weeks in technology are about products. This one is about physics, economics and the awkward place where the two meet. Three stories dominate the wires as we write this from Montreal, and none of them is a straightforward gadget launch.

The first is a supply story that has quietly become the single most important variable in what you will pay for a laptop, a tablet or a phone this autumn: the global memory crunch. Artificial intelligence accelerators have an appetite for DRAM that the industry did not plan for, and the capacity being fed to them is capacity that is no longer making the ordinary memory chips inside consumer devices. The second is a display story — the arrival, in volume, of Micro RGB backlights, a genuinely clever piece of optical engineering that is being marketed in a way almost designed to confuse buyers. The third is a security story: within roughly forty-eight hours, Google, Anthropic and OpenAI all published material about frontier AI models that can find and exploit software vulnerabilities on their own, and about the guardrails they are bolting on in response.

We write this column the way we would explain it to a customer standing at the counter: what is actually happening at the level of electrons and photons, why the industry is behaving the way it is, and what — concretely — you should do about it when you buy equipment. Everything below is sourced; nothing is invented. Where a number is a manufacturer claim rather than an independent measurement, we say so.

Today's Tech Radar

The ten stories we considered for today's edition, ranked by how much they will change what people buy and how they work over the next twelve months.

We have chosen three of these for the long treatment: the memory crunch (1, 7, 8, 10), Micro RGB displays (2), and autonomous AI vulnerability discovery (3, 9). They are the stories with the most science underneath them and the most direct consequences for equipment you are about to buy.

Part One — The Memory Crunch: Why a Data Centre in Another Country Is Setting the Price of Your Laptop

A DRAM module in close-up. Each black package holds billions of one-transistor, one-capacitor cells that must be refreshed thousands of times per second. Photo: Liam Briese / Unsplash.

What a DRAM cell actually is

To understand why memory has become the industry's chokepoint, it helps to know how astonishingly fragile a bit of DRAM is. Dynamic random-access memory stores each bit as a quantity of electric charge on a tiny capacitor, gated by a single transistor. That is the whole cell: one transistor, one capacitor, usually written 1T1C. The capacitor in a modern node holds on the order of a few femtofarads of capacitance — a few quadrillionths of a farad. Charge leaks out of it continuously through the transistor and through the dielectric. Left alone, the cell would forget its contents in a small fraction of a second.

DRAM therefore does not store data so much as it continuously re-remembers it. A refresh controller walks through every row in the array on a fixed interval — conventionally 64 milliseconds, halved at high temperatures — reading each row into a sense amplifier and writing it back at full strength. This is the "dynamic" in the name, and it is why DRAM burns power even when idle, and why it loses everything the instant you cut the supply.

The engineering consequence is that DRAM does not shrink the way logic does. Every process generation, the capacitor must hold roughly the same amount of charge in a smaller footprint, which is why manufacturers build capacitors as deep, high-aspect-ratio trenches or pillars — structures dozens of times taller than they are wide, etched into silicon with near-vertical sidewalls. This is one of the hardest patterning problems in semiconductor manufacturing, and it is a large part of why DRAM capacity cannot simply be conjured when demand spikes. A new fab is a three-year, multi-billion-dollar commitment; a new capacitor scheme is a research programme.

Enter HBM, and the reallocation of the world's DRAM

AI accelerators do not want more memory so much as they want faster memory. Training and inference are, at the arithmetic level, enormous sequences of matrix multiplications, and the bottleneck is almost never the multiplier — it is feeding it. Model weights and activations have to be streamed from memory into the compute units continuously, and a modern accelerator can be starved by anything less than several terabytes per second of bandwidth.

The industry's answer is High Bandwidth Memory. Instead of laying DRAM dies flat on a board and connecting them through a narrow, fast bus, HBM stacks DRAM dies vertically — eight, twelve or sixteen high — and drills thousands of through-silicon vias straight down through the stack. The stack sits on the same package substrate or silicon interposer as the processor, millimetres away rather than centimetres. Because the interconnect is short and massively parallel, HBM can run each wire relatively slowly and still deliver colossal aggregate bandwidth at a far better energy cost per bit than a conventional bus. This is why every serious AI accelerator uses it.

The catch is arithmetic. An HBM stack consumes the die area of many conventional DRAM chips, adds a complex and yield-limited stacking and bonding step, and commands a much higher margin. Given a fixed number of wafer starts, every wafer routed to HBM is a wafer not producing the DDR5 in a desktop, the LPDDR in a phone or the graphics DRAM in a GPU. Micron's HBM output has been described as effectively sold out for 2026, and the reallocation of capacity toward high-bandwidth memory is the root cause of the shortage. Industry reporting has put AI's share at roughly a fifth of total DRAM production this year.

What the numbers say

The price data are unambiguous. TrendForce's memory pricing survey found conventional DRAM contract prices rising 13–18% quarter-over-quarter in the third quarter of 2026, with NAND Flash contract prices up 10–15% over the same period — and those are the moderated figures. TrendForce attributes the slowdown not to improving supply but to demand destruction: record-high contract prices mean customers in PCs and smartphones have reached their affordability limit. Earlier in the year the increases were far steeper, with reporting citing conventional DRAM contract price rises in the high-double-digit percentages quarter-on-quarter through the first half.

TrendForce is explicit about the mechanism reaching consumers: PC OEMs continue to replenish inventory, but retail notebook prices are expected to rise across the board as higher-cost components flow through the channel, weighing on full-year shipment volumes. The same survey notes that suppliers keep prioritising AI and server products when allocating capacity, keeping LPDDR — the low-power DRAM in every phone and tablet — tight, and that smartphone vendors have been raising retail prices to offset it. The company's shipment forecast tells the rest of the story: global notebook shipments are projected to fall 13.6% in 2026. That is not a demand collapse. It is a price shock.

Flash follows the same logic one step behind. Enterprise SSD demand, driven by AI inference and large-scale data centre build-outs, has pulled NAND capacity toward high-margin products; TrendForce reported the top five enterprise SSD vendors booking close to US$37.59 billion in revenue in the second quarter of 2026 alone. Client SSDs — the drive in your laptop — are the residual claimant.

The practical advice: buy the memory, not the discount

A SODIMM module. In most 2025–2026 thin-and-light laptops, the equivalent silicon is soldered to the board and cannot be upgraded later. Photo: Franck V. / Unsplash.

There is a specific, unglamorous conclusion that follows from all of this, and it runs against the instinct most buyers have.

Buy more memory than you think you need, and buy it at the time of purchase. In the current generation of thin-and-light notebooks, memory is not a module in a slot. Low-power DDR is soldered directly to the mainboard — and in the newest designs, packaged on or beside the processor itself — precisely because the short, controlled traces are what allow the high transfer rates and low voltages that give you battery life. The engineering is sound and the consequence is absolute: the 16 GB you buy today will still be 16 GB in four years. In a market where memory contract prices are rising by double digits every quarter, the gap between a 16 GB and a 32 GB configuration is the cheapest it will be on the day you buy it.

For anyone doing real work — many browser tabs, virtual machines, large spreadsheets, video, or local AI features that load a model into RAM alongside everything else — 32 GB is now the sensible floor on a new machine. The Dell Pro 16 Plus PB16250 with an Intel Core Ultra 7 268V, 32 GB and a 512 GB SSD is the configuration we point people toward for exactly this reason: it is a current Copilot+ class machine bought at today's memory prices rather than next year's. If your workload is lighter and you want the outstanding battery life of an Arm design, the Lenovo ThinkPad T14s Gen 6 with a Snapdragon X Plus, 16 GB and 512 GB is a well-judged 14-inch machine — but treat that 16 GB as a genuine ceiling and be honest about your workload before you accept it.

Prefer a desktop where you can. Desktops still take DIMMs. That is a real, monetisable advantage in a rising market: you buy the chassis and the processor now, and you can add memory later if prices ever normalise. The Lenovo Legion T7 with a Core Ultra 9 285K, 64 GB and a 1 TB SSD is an unusually well-provisioned example — 64 GB is a configuration that has become conspicuously expensive to specify from scratch this year.

Do not economise on the SSD, and keep a fast external drive. With NAND contract prices climbing 10–15% quarter-over-quarter, undersizing internal storage now means paying a premium later — and on soldered-storage machines it may mean no upgrade at all. A high-endurance NVMe drive such as the Samsung 990 PRO 2 TB PCIe Gen4 x4 is a sensible hedge for desktops and workstations that can take one.

On tablets and phones, the same logic holds with less room to manoeuvre. Nothing in a tablet is upgradeable. The Samsung Galaxy Tab S10 FE with 8 GB of RAM and 128 GB of storage is the configuration we consider the practical minimum for a device you intend to keep for four or five years; the Galaxy Tab A11+ with 6 GB and 128 GB is the budget option where the tablet is a second screen rather than a primary tool. For Windows-native work in tablet form, the Microsoft Surface Pro 11 Copilot+ with 16 GB and 256 GB remains the reference design. If you are unsure which configuration matches your actual workload, request a free quote from our team and we will size it with you rather than guess.

A footnote on why DRAM is not quite digital

Item nine on today's radar is a good companion to this section. The "GPUThor" technique reported this week defeats error-correcting codes on an NVIDIA RTX A6000 to obtain host root access. It is a Rowhammer attack: because DRAM cells are packed so closely, repeatedly activating one row of cells can, through capacitive coupling and charge leakage, flip bits in a physically adjacent row that the attacker never had permission to touch. ECC catches and repairs isolated single-bit errors, but it is a probabilistic mitigation, not a wall — an attacker who can induce the right multi-bit pattern can slip past it. It is a useful reminder that the analogue physics we described above is not an abstraction; it is an attack surface.

Part Two — Micro RGB: What Samsung Actually Built, and What the Name Hides

Large-format panels are the fastest-moving segment in display. The interesting engineering in 2026 is happening behind the liquid crystal, not in front of it. Photo: Prydumano Design / Unsplash.

The colour filter has always been the problem

An ordinary LCD television is, optically, a light source with a stencil in front of it. A backlight produces white light across the whole panel. A liquid crystal layer, cell by cell, rotates the polarisation of that light so that a second polariser passes more or less of it — that is how brightness is controlled per subpixel. Colour comes from a filter: each pixel is divided into red, green and blue subpixels, and each subpixel has a dye filter that absorbs everything except its own band.

That absorption is the whole trouble. A colour filter is a subtractive device: it makes red by throwing away roughly two-thirds of the light. Worse, real dye filters are not sharp. Their transmission curves have long tails, so the "red" subpixel passes some orange and some deep magenta, the "green" passes some cyan and yellow, and the result is that the primaries are less pure than the specification suggests. Impure primaries mean a smaller colour gamut, because the gamut is literally the triangle drawn between your three primaries on a chromaticity diagram. Every LCD engineering advance of the past fifteen years — wide-gamut backlights, quantum dot films, phosphor tuning — has been an attempt to work around the fact that you are shining broad-spectrum white light through imperfect dyes.

What Micro RGB changes

Micro RGB attacks the problem at the source. Instead of a backlight made of blue LEDs with a phosphor or quantum-dot layer converting some of that blue into a broad white, Samsung's Micro RGB panels use a dense array of separate red, green and blue LEDs — sub-100-micron devices — as the backlight unit. Each emits its target colour directly. Because the light arriving at the liquid crystal layer is already close to the desired primary, the colour filter is either eliminated or drastically simplified, and the light lost to absorption goes with it.

Two things follow. The first is purity: an LED emitting narrow-band red is spectrally far cleaner than white light forced through a red dye, which widens the achievable gamut. Samsung states that its R85H and R95H panels achieve 100% coverage of the BT.2020 colour space — the very wide gamut written into the HDR specifications and, until now, essentially unreachable by consumer displays. Independent commentary on the 2026 RGB-backlit class more broadly has anticipated coverage exceeding 90% of BT.2020, which puts Samsung's claim at the optimistic end of a real trend.

The second is control. Because the backlight is now composed of individually addressable coloured emitters, local dimming becomes local colour dimming: the set can raise the red LEDs behind a sunset and leave the blue ones dark, rather than pushing white light everywhere and asking the filters to absorb the excess. Samsung markets the processing side of this as the Micro RGB AI Engine Pro, with scene recognition driving per-zone colour and brightness optimisation. The R95H series runs a 165 Hz refresh mode and the R85H 144 Hz, both with variable refresh rate — respectable for a large-format panel, and a meaningful difference for gaming.

The naming problem, stated plainly

Here is the part a shop has an obligation to say clearly. Micro RGB is an LCD. It is not micro LED. The 2026 Micro RGB sets from Samsung — and the parallel RGB mini-LED products from LG, Hisense and TCL — are liquid crystal panels with a very sophisticated backlight. In a true emissive micro LED display, each subpixel is an LED and there is no liquid crystal layer at all, which is why such displays have perfect blacks and cost as much as a car. Micro RGB is a backlight technology. It is an excellent one, and the marketing name does it no favours.

What that means in practice is that Micro RGB inherits the strengths and the weaknesses of LCD. The strength is brightness: RGB-backlit sets can go extremely bright, and one of the first competing models to market, the Hisense UR9 series, has been measured at over 5,500 nits on a 10% window. The weakness is contrast at the edges of bright objects. Because a dimming zone is still larger than a pixel, a bright star on a black sky can produce a faint halo — the "blooming" that OLED does not have, because in an OLED an off pixel emits nothing at all.

It is also worth reading brightness numbers carefully. Independent measurement of the flagship R95H in SDR found around 235 nits on a 10% window in Filmmaker Mode and about 726 nits on the same window in Standard mode. Those are not contradictions of the HDR headline figures; they are different measurements. Filmmaker Mode deliberately targets a reference luminance for accurate reproduction of mastered content, Standard mode targets a bright showroom, and HDR peak figures describe short bursts in a small window. Anyone comparing televisions should insist on knowing which of the three a quoted number refers to.

Buying advice for displays in 2026

Samsung's 2026 Micro RGB range runs from 55 to 115 inches, with the 55-inch R85H at around US$1,599 and the 85-inch R95H at US$6,499; a carryover 115-inch model sits at US$29,999. Those are flagship-consumer prices, and for most of the rooms we specify equipment for, they answer a question nobody asked.

For a living room or a small meeting room where the screen is watched rather than run continuously, a well-made conventional 4K panel remains the value choice, and the LG 55PK640S0UB 55-inch 4K smart LCD television covers it. For anything that runs all day — a lobby, a classroom, a retail floor, a control room — a consumer television is the wrong tool regardless of its backlight. Commercial panels are specified for long duty cycles, higher sustained brightness and dust ingress, and that is what you want. The Samsung 55-inch Crystal UHD Signage QBC is the sensible entry point at 16/7 operation; the Samsung QMC 55-inch UHD at 500 nits, non-glare and IP5X-rated for 24/7 duty is the one to specify when the screen genuinely never turns off.

When the room is large, size beats specification almost every time — the single biggest determinant of perceived image quality at a distance is angular subtense, not gamut. The Samsung QM85C 85-inch UHD at 500 nits and the LG 86-inch commercial 3840×2160 display both do that job. On the desk, where you sit close and colour work matters more than peak brightness, the Samsung Essential S32B304NWN 32-inch monitor is a straightforward, well-priced panel. If you are trying to work out whether a room needs a television, a signage display or an interactive panel, request a free quote from our team — the difference in total cost over five years is usually larger than the difference in purchase price.

Part Three — When the Model Finds the Zero-Day: Frontier AI Crosses a Security Threshold

Vulnerability research used to be a scarce human skill. In 2026 it is becoming a capability you can rent by the token — on both sides. Photo: Compagnons / Unsplash.

What was announced

On September 2, three announcements landed close enough together to read as a single event.

Google introduced Gemini 3.8 Flash Cyber, which it describes as its most capable cybersecurity model, and made it available to a restricted set of defenders through a new initiative called the Fairwind Program. Google's framing is deliberate: give high-priority defenders — governments, healthcare providers, telecommunications operators — early access to advanced models so they can build defences before the corresponding threats arrive. The company says it is working with more than 650 partners globally, including CrowdStrike, Datadog, Menlo Security, Palo Alto Networks and Snowflake. Google's team stated that they prioritised vulnerability fixing over offensive capabilities such as exploitation.

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 with different levels of safeguards, the latter available only through trusted access programmes supporting cybersecurity and life sciences work. The company said it is now permitting Fable 5.1 to be used for identifying software vulnerabilities, while still routing tasks such as penetration testing, exploit generation and binary vulnerability scanning to other models. It also announced Enterprise Frontier Safeguards, combining zero data retention with misuse detection, and described hardening measures taken after incidents in which models acted against real systems they had been told were simulated.

OpenAI disclosed that its forthcoming Astra model meets the "Critical" cybersecurity capability threshold under its Preparedness Framework. That designation has a specific meaning: it applies when a model can independently detect and exploit zero-day vulnerabilities across many well-defended systems, or carry out a complete attack against a hardened target from only a high-level instruction, without a human guiding it. OpenAI reported that Astra scores 100% on ExploitBench for developing exploits from known vulnerabilities and declines 91.5% of jailbreaking attempts, against 59% for its GPT-5.6 Sol model. During evaluation, the company says, the model discovered and chained two previously unknown vulnerabilities, produced a full browser compromise that escaped the sandbox and executed commands on the host when an HTML file was opened, and combined multiple flaws in a hardened operating system into a local privilege-escalation chain from an unprivileged user to root. OpenAI said it delayed parts of Astra's development while strengthening protections, and warned that those safeguards may sometimes flag legitimate activity as misuse.

Separately, a coalition of more than 100 companies — including Anthropic, Google, Microsoft and OpenAI — has issued a joint letter calling for improved collective defences against AI-enabled attacks.

The science underneath: why models are good at this

It is worth understanding why vulnerability discovery turned out to be a task where large models excel, because it explains why the capability arrived faster than most people expected.

Finding a memory-safety bug is, structurally, a search problem over program states. Classical tools already automate parts of it. A fuzzer generates enormous volumes of malformed input and watches for crashes; a symbolic execution engine treats inputs as mathematical variables and asks a constraint solver which values would drive execution down a particular path. Both are powerful and both hit the same wall — path explosion. The number of distinct execution paths through a real program grows combinatorially, and a blind search drowns.

What a language model contributes is a learned prior over which paths are worth exploring. Having ingested vast quantities of source code, patches, bug reports and exploit write-ups, it has absorbed the shape of the mistakes programmers actually make: the off-by-one in a length check, the integer that can be made to wrap before it is used as an allocation size, the pointer freed on an error path and used again on the way out. It does not prove anything. It guesses well, and it guesses in the region where the bugs live — which converts an intractable search into a tractable one. Chain that prior to tools that can compile, run, fuzz and debug, and you have an agent that iterates toward a working exploit rather than merely describing one.

The uncomfortable symmetry is that the same prior works for defence. A model good at finding the off-by-one is good at spotting it in review and at writing the patch. That is precisely why Google says it invested in vulnerability fixing first and gated the model behind a vetted-defender programme, and why Anthropic and OpenAI have built tiered access with separate safeguards. The technology is not dual-use in the abstract; it is dual-use in the same forward pass.

What a small or mid-sized organisation should actually do

None of this is a reason to panic, and none of it changes the fundamentals. It compresses timelines. The interval between a vulnerability becoming public and it being exploited at scale has been shrinking for a decade; autonomous discovery shortens it further. Everything below is ordinary hygiene made more urgent.

Patch faster, and know what you have. An asset inventory is not bureaucracy; it is the precondition for patching. You cannot update a device you have forgotten about. This is a strong argument for standardising fleets on a small number of current, vendor-supported models with a defined firmware and driver channel rather than accumulating a decade of mixed hardware.

Buy hardware with a real security floor. Business-class machines carry firmware protections that consumer models do not: measured boot, a hardware root of trust, firmware resilience and remote attestation, and manageability that lets you push a fix without touching the device. The vPro-class configurations in the Dell Pro 16 Plus PB16250 and the commercial Lenovo ThinkPad T14s Gen 6 exist for this reason. The premium over a consumer laptop is small relative to the cost of one incident.

Treat displays and signage as networked computers, because they are. A modern signage panel runs an operating system, joins your network and often faces the public. It needs a VLAN, a patch schedule and a named owner, exactly like a server. Commercial panels such as the Samsung QMC 55-inch UHD ship with the management tooling to make that practical; a consumer television does not.

Assume phishing gets better, not worse. The cheapest single control most organisations have not yet finished deploying is phishing-resistant authentication — hardware security keys or passkeys — because it removes the credential as something that can be handed over at all.

Keep offline backups and test the restore. A backup you have never restored is a hypothesis, not a backup.

If you would like an assessment of where your fleet actually stands — firmware currency, end-of-support devices, network-exposed displays, authentication posture — that is exactly the kind of review our team does, and you can request a free quote from our team to start it.

Three Shorter Notes

Apple, September 9

Apple has confirmed a "Surprise and shine" event for September 9 at Apple Park, expected to cover the iPhone 18 Pro and Pro Max alongside the company's first folding handset, reported to be branded iPhone Ultra. It will be the first keynote led by John Ternus, who took over as chief executive on September 1. Whatever Apple ships, the category effect matters more than the device: a folding iPhone legitimises a form factor that has been commercially real but culturally niche for six years. If you want to understand what a mature large-format foldable feels like before that happens, the Samsung Galaxy Z Fold7 with a 512 GB, 8-inch folding Dynamic AMOLED 2X panel is the current benchmark — though our stock of it is down to a single unit as we write.

IFA opens tomorrow

IFA 2026 runs September 4–8 in Berlin. AMD returns as an exhibitor, Xiaomi makes its IFA debut alongside a stated plan to spend €7.4 billion on AI research and development between 2026 and 2028, and the show's programme leans heavily on robotics, including cognitive robots from NEURA Robotics. Expect the memory story above to shape a great deal of what is announced, whether or not anyone says so from a stage.

The bottleneck is the wire

SEMICON Taiwan 2026 ran September 2–4 in Taipei, drawing professionals from 65 countries, following a Semicon Network Summit on September 1 at which Taiwan's government recognised industry figures including the chief executives of GlobalWafers and Micron. The recurring theme was that the constraint in AI hardware has moved from the transistor to the interconnect — the wires and optics that move data between chips — with co-packaged optics entering commercial production this year. Testing and metrology were described as existential challenges as architectures grow more complex. It is the same physics as the HBM story: when compute is cheap and moving data is expensive, the engineering effort migrates to the plumbing.

Glossary of the Week

Setup at a Glance

Everything below was verified in stock at the time of writing. Stock moves quickly, particularly on the single-unit items.

Closing

If there is a single thread running through today's three stories, it is that the interesting constraints in computing have migrated away from the processor. Memory is scarce because AI wants bandwidth. Displays are improving because someone rethought the light source rather than the liquid crystal. Security is changing because a statistical model of how programmers make mistakes turned out to be an excellent vulnerability researcher. In each case, the part everyone talks about — the chip, the panel, the model — was not where the leverage was.

For anyone buying equipment this autumn, the practical translation is short: specify memory and storage generously now, because you cannot add them later and they will not get cheaper this year; buy commercial-grade displays for anything that runs all day; and treat every screen and endpoint on your network as a computer that needs patching. If you would like help turning that into a specific list for your organisation — with real prices, real stock and no guesswork — request a free quote from our team and we will work through it with you.

Sources & Further Reading

Memory and component pricing: TrendForce, "AI Server Demand Continues to Support Memory Prices in 3Q26" (3 July 2026); TrendForce, "Long-Term Agreements Cap Price Increases; Server DRAM Contract Prices Expected to Rise 13-18% QoQ in 3Q26"; TrendForce, "Global Notebook Shipments Forecast to Decline 13.6% in 2026"; TrendForce, enterprise SSD vendor revenue, 2Q26 (1 September 2026); Tom's Hardware, "Memory price surge begins to cool as consumers hit affordability limit"; CNBC, "AI memory is sold out, causing an unprecedented surge in prices"; IDC, "Global Memory Shortage Crisis".

Displays: Samsung Newsroom, "Samsung Sets a New Standard of Color with Micro RGB TV Lineup"; Samsung, Micro RGB TV technology overview; ecoustics, "Samsung R95H Micro RGB TV Review"; CE Pro, "Samsung Unveils Full Micro RGB TV Lineup"; Notebookcheck, Micro RGB pricing and specifications; TechRadar, "The best TVs of CES 2026"; Tom's Guide, "Should you buy a Micro RGB TV this year?".

AI and security: The Hacker News, "Google, Anthropic, and OpenAI Unveil Cyber AI Models, Safeguards, and Access Programs" (2 September 2026); Google, Gemini 3.8 Flash and 3.8 Flash Cyber; Google DeepMind, Fairwind Program; Anthropic, Claude Fable 5.1 and Mythos 5.1; Anthropic, Enterprise Frontier Safeguards; OpenAI, "The path to Astra"; OpenAI, collective cyberdefense joint letter; The Hacker News, "New GPUThor Rowhammer Defeats ECC on NVIDIA RTX A6000".

Industry events and launches: AppleInsider, Apple's "Surprise and shine" event, September 9; 9to5Mac, Apple announces iPhone 18 Pro and foldable event; IFA Berlin 2026 press releases; GlobeNewswire, "Semicon Network Summit 2026 Advances Global Chip Collaboration in the AI Era"; TechTimes, "SEMICON Taiwan 2026 Kicks Off: AI Chips' Bottleneck Is Wires Connecting Them"; Semiconductor Engineering, "Chip Industry Week In Review".

Photos: Unsplash (free commercial license) — Liam Briese, Franck V., Prydumano Design and Compagnons. Product availability and inventory figures were verified against PcHybrid stock on 3 September 2026 and are subject to change.