Montreal, Thursday, October 29, 2026. It has been a remarkable week for the hardware that sits on your desk and in the data centres behind it. Apple spent three days unveiling an OLED, touch-enabled MacBook Pro, an OLED iPad mini and the new M6 chip. Meanwhile, a report says OpenAI intends to buy Google's Tensor Processing Units, customs officers in Malaysia seized yet another shipment of smuggled AI servers, and Cloudflare disclosed that a subtle bug let some AI customers see fragments of other customers' data. Each of these stories hides a beautiful piece of physics or computer engineering, and each affects what you should buy next.

Today we go deep on three of them: how tandem OLED panels make light and why touch changes the laptop; why specialised AI chips such as TPUs exist and what the supply fight means for ordinary PC buyers; and how a cache designed to make AI fast became a security hazard. As always, we only report facts found in published sources (listed at the end), and where details are unconfirmed we say so.

Today's Tech Radar

Here are the ten stories we shortlisted from the last few days of news, ranked by significance and substance.

Deep Dive 1: Touchscreen Tandem OLED Arrives on the Laptop

A laptop display lit in a dark room, the environment where OLED shines. Photo: Aerps.com / Unsplash.

What Apple announced

According to MacRumors and TechRadar's live coverage, Apple's new MacBook Pro is its first Mac with an OLED display and touch support. It comes in 14-inch and 16-inch sizes with M5 Pro or M5 Max chips, memory from 24 GB to 128 GB, and storage up to 16 TB. TechRadar describes the panel as a tandem OLED with up to 2,000 nits of peak brightness. MacRumors reports a $400 premium over comparable mini-LED models, which stay on sale; TechRadar lists 14-inch pricing from $2,899 and 16-inch from $3,399, with deliveries from November 6. Apple also introduced an OLED iPad mini (8.5 inches, A20 Pro chip, IP67 rating) and M6 versions of the 14-inch MacBook Pro and iMac, with the M6 claimed to deliver up to 30 percent faster graphics and up to 25 percent faster AI workloads than M5.

How an OLED pixel makes light

Most laptop screens are LCDs: a white backlight shines through liquid-crystal shutters and colour filters. Even in mini-LED designs, which divide the backlight into many dimmable zones, light is created in one place and blocked in another. An organic light-emitting diode (OLED) works differently. Each subpixel is a thin stack of organic semiconductor layers sandwiched between electrodes. When a voltage is applied, electrons and holes are injected from opposite sides and meet in the emissive layer, where they recombine into an excited state called an exciton. When the exciton relaxes, it releases its energy as a photon. The colour depends on the molecular structure of the emitter. No backlight is needed, so a pixel that is off emits no light at all.

That is why OLED achieves what spec sheets call an effectively infinite contrast ratio: black is the absence of emission, not an imperfect attempt to block a backlight. It also explains the wide viewing angles, because light is generated at the pixel itself rather than passing through a polarising shutter that changes with angle. Response times are extremely short, which reduces motion blur and ghosting.

Why 'tandem' matters

Single-stack OLED has two classic weaknesses: limited peak brightness and a finite lifetime, because organic molecules degrade as current density rises. A tandem OLED stacks two emissive units in series, connected by a charge-generation layer. Because the same current flows through both units, each unit can run at a lower current density to deliver the same brightness, or the panel can reach a much higher brightness for the same stress. In principle that gives higher peak luminance, improved efficiency and better longevity than a single stack. A figure such as 2,000 nits of peak brightness, as reported for this MacBook Pro, is in the range that was historically reserved for LCD displays with powerful local-dimming backlights, and it matters for HDR highlights such as sunlight on a car or a specular reflection. Note that peak brightness is usually measured on a small window of the screen for a short time; sustained full-screen brightness is lower for every display technology.

Why touch on a laptop, and why now

Touch input on an OLED panel is typically integrated into the display stack as a capacitive sensing layer. A finger changes the local capacitance between a grid of electrodes, and a controller computes position from the change. Apple says touch supports gestures such as scroll, pinch, rotate and draw, and the company points to uses like manipulating 3D objects and playing virtual instruments. MacRumors reports that adding touch did not reduce battery life. Windows laptops have offered touchscreens for years, so the point is not novelty but convergence: laptops, tablets and phones now share the same display technology and interaction model, and software on all of them is expected to handle fingers, pens and keyboards.

Practical buying advice

OLED is best for creative work, video, HDR content and dim environments. Its practical caveats are the usual ones: static content displayed for very long periods can in theory cause burn-in on any OLED, and glossy panels reflect bright rooms. For office work in bright offices, a high-quality matte LCD remains a sensible, less expensive choice. Whatever panel you pick, remember that screen quality is only part of the system; memory, storage and the processor decide how long the laptop stays useful.

If you want a premium, touch-enabled Windows laptop today, the Microsoft Surface Laptop 7 13.8-inch touchscreen (Core Ultra 7, 32 GB, 512 GB) is in stock, as is the larger Surface Laptop 7 15-inch with Snapdragon X Elite. For business users who prefer a classic keyboard-first design, the Lenovo ThinkPad X1 Carbon Gen 11 (i7, 16 GB, 512 GB) and the Dell Pro 14 Plus (Core Ultra 7 265U, 32 GB, 512 GB SSD) are both stocked. For a rugged tablet that you can write on and drop, the Panasonic Toughbook 33 Mk4 (12.0-inch QHD touchscreen) is available. Not sure which one fits your workload? request a free quote from our team and we will match a device to your budget.

The wider display picture: TVs and signage

The same physics governs larger screens. Most TVs and commercial displays sold for offices, schools and shops remain LED-backlit LCD because the technology is mature, bright and economical for large sizes and long operating hours. In stock now are the LG 50PK640S0UB 50-inch Smart LED-LCD 4K TV, the LG 75PK340S0UA 75-inch 4K and the LG 86PK640S0UA 86-inch 4K. They are a strong fit for meeting rooms, waiting areas and classrooms, where brightness and reliability matter more than perfect black levels.

Deep Dive 2: Why OpenAI Wants Google's TPUs, and What It Means for Your PC

A processor chip in close-up. Photo: BoliviaInteligente / Unsplash.

The report

According to a Reuters report carried by The Star, which cites The Information, OpenAI plans to buy Google's Tensor Processing Units (TPUs). No financial terms were disclosed, Google did not immediately comment, and OpenAI declined to comment, so this should be treated as a report rather than a confirmed deal. The same article notes that Reuters reported in June 2025 that OpenAI had begun renting Google's chips, that Amazon, Microsoft and Alphabet plan to spend up to $600 billion combined this year on AI infrastructure, that Google has begun selling TPUs directly to some customers, and that OpenAI is already one of the largest buyers of Nvidia GPUs and unveiled its first custom AI chip, built with Broadcom, in June. The Neuron's daily digest independently carries the same story and stresses that volumes and confirmation from either company are not yet available.

What a TPU actually is

Training and running a neural network is dominated by one operation: multiplying huge matrices of numbers and adding the results. A general-purpose CPU is built to do many different things quickly and spends much of its silicon on branch prediction, caches and control logic. A GPU has thousands of simple arithmetic units that excel at parallel work, which is why it dominates AI today. A TPU goes one step further. Its core is typically a systolic array: a grid of multiply-accumulate units through which data flows rhythmically from neighbour to neighbour, like blood through a circulatory system. Each number fetched from memory is reused many times as it passes across the array, which sharply reduces the number of memory accesses per calculation. Since moving data costs far more energy than arithmetic, this design can deliver more calculations per watt for the workloads it is built for.

The trade-off is flexibility. GPUs have a mature software ecosystem and can be programmed for graphics, simulation and many AI models. TPUs are tightly coupled to Google's software stack and are optimised for a narrower range of tensor operations. A buyer such as OpenAI diversifying across Nvidia GPUs, TPUs and its own Broadcom-built chip is hedging against supply constraints, price and a single point of dependence.

Memory, the real bottleneck

Compute is not the only scarce resource. Modern accelerators depend on high-bandwidth memory stacked next to the processor, because a chip that cannot be fed data sits idle. This week's Tom's Hardware archive notes that Micron plans to use wafer bonding and High-NA EUV lithography for future DRAM and is developing vertical high-bandwidth memory and memory placed on top of processors. The common thread is clear: the performance of AI hardware is increasingly limited by how fast data can move, not how fast numbers can be multiplied.

Export controls and a smuggling bust

Demand this intense creates a grey market. Tom's Hardware reports that Malaysian customs seized more than $12 million (49.5 million ringgit) of advanced AI chips in 30 servers at Kuala Lumpur International Airport's free trade zone on October 15. The servers were falsely declared as ordinary electronic goods. Officials linked the network to seizures worth more than $25 million across at least five cases, including a June 5 seizure of 72 server units worth about $13 million. The chip type, final destination and recipient have not been confirmed, although the article says China is the most likely destination given U.S. export controls. The takeaway for buyers is that advanced accelerators are a strategically controlled, scarce commodity, and any hardware you purchase should come through a legitimate channel with proper documentation and warranty.

What it means for laptops and desktops

Data-centre accelerators are not what you buy for an office, but the same ideas trickle down. Modern laptops include a neural processing unit (NPU), a small matrix engine built on similar principles, to run AI features locally. The Neuron highlights Liquid AI's LFM3 family, small enough for phones and laptops, and Opera's WebAI standard for running supported models locally, noting that performance depends on hardware. Local processing can cut cloud costs and keep data on the device.

For on-device AI work, look for ample RAM (our recommendation is at least 16 GB, and 32 GB for larger local models), a fast SSD and a Copilot+ class NPU. The Surface Laptop 7 15-inch (Snapdragon X Elite, 16 GB) and the 32 GB Core Ultra 7 model above are both Copilot+ PCs and are in stock. For heavier local development, the Lenovo Legion T7 desktop (Core Ultra 9 285K, 1 TB, 64 GB) is in stock with generous memory. For general office fleets, the Lenovo ThinkPad E16 Gen 3 (Ultra 5, 16 GB) is a sensible, stocked workhorse. If you are planning a fleet purchase, request a free quote from our team for volume pricing and configuration help.

Deep Dive 3: When a Cache Leaks, the Cloud Leaks

Security starts with isolation: a padlock on a keyboard. Photo: Towfiqu barbhuiya / Unsplash.

The incident

As summarised by The Neuron, a configuration mismatch in Cloudflare's GLM-5.3 Workers AI service triggered a bug in the SGLang inference engine, so some long requests received cached data from other customers. Cloudflare identified 155 affected responses between August 27 and September 9, found no evidence of deliberate exploitation, and patched the configuration and the engine. A more detailed write-up at Tech Insider adds that each affected response involved a prompt longer than 681,280 tokens, that no other Workers AI model was affected, and that the upstream fix shipped in SGLang v0.5.20 with follow-ups in v0.5.21 and v0.5.22. Which customers were affected, and what the exposed content contained, are not confirmed in the sources we reviewed.

What is a KV cache?

A large language model generates text one token at a time. For every new token, the model's attention mechanism needs to look back at keys and values computed from every earlier token in the conversation. Recomputing them each time would be wasteful, so the engine stores them in GPU memory: the key-value (KV) cache. The cache is divided into pages, much like virtual memory in an operating system, and pages are allocated to requests and recycled when requests finish. This design is what lets one GPU serve many users at once, and it is a major reason why AI services are affordable.

The security rule is simple: a page may belong to only one request at a time, and a recycled page must never expose its previous contents. Tech Insider's account says the SGLang bug exposed KV-cache pages across requests, so a new request could read tokens from another customer's earlier request to the same model. Think of a hotel that gives the next guest a room without clearing the previous guest's belongings. Importantly, the service continued to return normal successful responses, which is why such errors are hard to detect without specific tests: a healthy status code does not prove the right context was used.

Lessons for businesses

First, multi-tenant isolation is a property you must test, not assume; Tech Insider's remediation advice includes confirming that the configured context length matches what the hardware can serve and testing oversized prompts in staging. Second, treat AI prompts as sensitive data. Do not paste credentials, customer records or confidential documents into any cloud AI service without understanding its data handling. Third, keep self-hosted inference software patched. Fourth, favour defence in depth: strong authentication, encrypted storage, network segmentation and tested backups all limit the damage when any one layer fails.

There is also encouraging news. The Neuron reports that Anthropic and Theorem used Claude to generate machine-checkable security proofs for 123 open-source libraries, with each guarantee holding only under the conditions encoded in its specification. Formal verification, which uses a proof checker to confirm that code meets a precise specification, will not eliminate bugs such as configuration mismatches, but it narrows the space where they can hide.

Practical advice for your own hardware

Local, on-premises AI is one answer for organisations that cannot send sensitive data to a third party, which is another reason on-device and on-premises models are gaining attention. Whatever your approach, endpoint hardware matters. A hardware security key supports phishing-resistant sign-in; the VeriMark Guard USB-C fingerprint key (FIDO2, WebAuthn) is in stock. Business laptops with TPM-based protections, such as the ThinkPad and Dell models above, add device-level encryption. If you would like a security review of your devices and sign-in setup, request a free quote from our team.

Glossary of the Week

Setup at a Glance

Closing Thoughts

The theme running through today's news is convergence and consolidation: displays are converging on OLED and touch, AI workloads are converging on specialised silicon and memory, and security is converging on isolation and verification. If any of this raises questions about your own setup, request a free quote from our team and we will help you choose.

Sources & Further Reading

Sources: MacRumors; TechRadar live blog; The Star / Reuters; The Neuron daily digest; Tom's Hardware (Malaysia seizure); Tom's Hardware news archive; Tech Insider (Cloudflare incident). Photos: Unsplash (free commercial license). Background explanations of OLED, TPU and KV-cache engineering are general technical knowledge.