Samsung HBM4E Samples: AI Memory Race Shifts in 2026

Samantha: Welcome to The State of Tech, The European Edition, Friday May twenty-nine, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Samsung ships the world's first next-generation AI memory samples, Japan's megabanks deploy fresh AI models from OpenAI and Anthropic for cyber defence, a Nvidia server maker warns of new AI bottlenecks beyond memory, the wider squeeze on data centre parts pushing prices to records, plus two things you can actually try this weekend. Let's get into it.

Samsung ships the world's first HBM4E samples, racing ahead in the AI memory war.

Samantha: Samsung Electronics has started shipping samples of its twelve-layer HBM4E memory chip. That makes it the first company in the world to deliver samples of this next-generation AI memory. And the timing is striking. Samsung only began mass production of the previous version, HBM4, in February. Three months later, the successor is already going out the door.

Bob: For listeners who don't live inside chip spec sheets, HBM is the stacked memory that sits right next to AI processors and feeds them data at very high speed. It is the bottleneck for almost every AI accelerator on the market. Samsung says HBM4E hits sixteen gigabits per second and is more energy efficient, thanks to a redesigned chip and production process. The headline customers being lined up are AMD, Nvidia and Google. Mass production will follow their schedules.

Samantha: And the context matters. Samsung has been chasing SK Hynix in this category for years. Hynix has been Nvidia's preferred partner for the current generation. Getting first-mover status on HBM4E is Samsung's attempt to flip that script.

Bob: The bottom line is this: the AI build-out is not just about Nvidia and TSMC anymore. Memory has become the gating factor. If Samsung can deliver HBM4E at volume and on price, it changes the supply picture for every AI data centre planned for 2027 and beyond. And it gives buyers a second strong source, which usually means more competition and, eventually, better availability.

Samantha: For Europe this is a reminder of where the dependency really sits. The EU keeps talking about sovereign AI infrastructure, gigafactories for AI, French and German data centre projects. But the memory inside those machines comes overwhelmingly from South Korea. Samsung racing ahead of Hynix doesn't change the geography. It just shuffles which Korean supplier European cloud builders depend on. The Chips Act money targets fabs, not advanced memory, and that gap is becoming harder to ignore as AI buildouts accelerate.

Bob: And European hyperscaler customers, the likes of OVHcloud or the new Mistral compute partners, will be watching pricing closely. If Samsung undercuts Hynix to win Nvidia volume, that pressure could ripple into the cost of training runs across the continent.

Japan's three biggest banks plug new OpenAI and Anthropic models straight into their cyber defences.

Bob: Story two. MUFG Bank, Sumitomo Mitsui and Mizuho, the three megabanks that effectively run Japanese finance, are getting access to new AI models from OpenAI and Anthropic specifically to strengthen cybersecurity. According to sources cited in the reporting, OpenAI's new model has capabilities comparable to Anthropic's Claude Mythos AI, which is known for spotting system vulnerabilities. The banks plan to run both in parallel.

Samantha: The driver here is regulatory pressure. Japan's Financial Services Agency and the Bank of Japan have been urging financial institutions to upgrade defences against AI-driven cyber threats. The concern is that the same models criminals can misuse to find weaknesses can also be turned around and used by the defenders to find those weaknesses first. So the banks are essentially deploying offensive-style AI to audit their own systems.

Bob: This is one of the first concrete examples of a whole banking sector going beyond pilots and embedding frontier AI models in security operations. If it works, expect every major bank globally to follow. If it goes wrong, expect a very public incident.

Samantha: For Europe the parallel is obvious. DORA, the Digital Operational Resilience Act, came into force last year and demands exactly this kind of upgraded cyber posture from banks and insurers. The European Central Bank has been running stress tests on cyber resilience. So when ING, BNP Paribas or Deutsche Bank look at the Japanese model, they are looking at a possible blueprint, both technically and in terms of how regulators react. One open question: which provider European banks pick. OpenAI and Anthropic are both US-based, and EU data residency rules make those contracts complicated.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

Nvidia server maker Wiwynn warns memory is just the start of the AI hardware crunch.

Bob: Wiwynn, a major Taiwanese builder of Nvidia servers, is warning that the AI infrastructure squeeze goes well beyond memory chips. Chair Emily Hong says shortages are emerging across the board, networking chips, power components, cooling, the lot. And she expects demand for data centre hardware to stay elevated for the next three to five years, driven by Meta, Microsoft and the rest of the hyperscaler spending wave.

Samantha: The specific timeline is what caught my eye. Wiwynn doesn't see meaningful price relief until late 2027 or even 2028. That is a long runway of expensive hardware. And it lines up with what we heard earlier this week from other Asian suppliers about networking switches and optical components being just as constrained as the headline GPUs.

Bob: It also tells you something about where margins are flowing. When a server assembler is publicly flagging shortages, they are usually also flagging that they have pricing power. Wiwynn benefits from this environment, but they are signalling to customers: lock in orders now or pay more later.

Samantha: For the broader AI story this matters because it punctures the idea that more capex automatically equals more compute. You can announce a hundred-billion-dollar data centre programme, but if you can't get the networking chips for two years, the rollout slips. That changes the timeline for everything from new model launches to enterprise AI adoption.

Bob: And Europe sits in an awkward position. The US hyperscalers have long-standing supply relationships and the deepest pockets. European players building sovereign cloud capacity, the French, German and Nordic projects, are smaller customers in a constrained market. When supply is tight, the big buyers get served first. That risks widening the compute gap between European AI infrastructure and the US, just as Brussels is trying to close it.

From networking to power, the hidden parts shortage is reshaping data centre economics.

Samantha: Story four picks up where Wiwynn's warning leaves off. Because the components Hong is flagging are the parts most people never think about. Optical transceivers that connect racks together. Power distribution units. High-voltage transformers. Custom cooling loops. None of those make headlines. All of them are now bottlenecks.

Bob: Wiwynn describes intense competition for everything from memory to networking chips, with hardware prices at record highs. And the knock-on effect is that smaller data centre operators, the ones serving regional businesses, get squeezed out. They can't match the order volumes that hyperscalers commit to. So a shortage at the top of the market becomes a capacity problem at the bottom.

Samantha: The pattern that emerges is concentration. A handful of buyers, Meta, Microsoft, Amazon, Google, plus a few Chinese players, absorb most of the constrained supply. Everyone else waits. That has implications for competition, for pricing, and for where AI workloads physically run. It also has implications for the energy grid, because the power components in short supply are exactly the ones grid operators need too.

Bob: In Europe we are already seeing this collide with grid capacity. Ireland has paused new data centre connections in parts of Dublin. The Netherlands has tightened rules in Noord-Holland. Frankfurt is running into power limits. Now layer on a hardware shortage on top of a grid shortage, and the operators trying to expand European AI capacity are fighting on two fronts. The likely outcome: more workloads pushed to Nordic countries where power is cheaper and grid headroom exists, and slower buildout everywhere else.

Samantha: To close out, two things you can actually use this weekend.

Try Claude's new project memory to stop re-explaining yourself to your AI.

Bob: First one is for anyone using Anthropic's Claude. If you haven't switched on project memory yet, it's worth ten minutes this weekend. You set up a project, drop in the documents or context you keep referring to, a CV, a style guide, your company's brand voice, ongoing meeting notes, and Claude will reference those automatically every time you open that project.

Samantha: The practical use case is the one most people miss. If you write a weekly newsletter, drop your last five issues in a project. Claude will pick up your tone, your typical structure, even your favourite turns of phrase. Then every time you ask for a draft, you skip the whole "write in my voice, here are some examples" routine. Same trick works for code with your team's conventions, or for customer emails with your brand guidelines.

Bob: It's free on the paid tiers, and the setup is genuinely two minutes. The biggest mistake people make is overloading it. Three to five tightly relevant documents per project beats twenty loosely related ones.

Samantha: The honest reason to try it: it removes the most tedious thing about working with chatbots, which is re-uploading the same context every session.

Bob: Works in Dutch, French, German, Italian, Spanish, the lot. And for European listeners worried about data, project files sit under Anthropic's standard terms, not used for training on the paid tiers. Worth checking the settings, but it's a reasonable starting point.

Proton's new Docs editor offers an end-to-end encrypted alternative to Google Docs.

Samantha: Second one, non-AI. Proton, the Swiss company behind Proton Mail, has rolled out a polished update to Proton Docs. It's a collaborative document editor, like Google Docs, but end-to-end encrypted by default. Real-time editing with colleagues, comments, the basic formatting you expect. Nothing fancy, but everything encrypted on the way in.

Bob: It's free for personal accounts up to the basic Proton Drive storage limit. Works in any modern browser, plus iOS and Android apps. If you already have a Proton Mail account, Docs sits right inside the same interface. No new login.

Samantha: The use case is straightforward. Anything you would not be comfortable having Google scan, contracts, medical notes, internal HR drafts, journalism source material, this is the alternative that doesn't force you back into emailing Word attachments around. It is not as feature-rich as Google Docs yet. No deep spreadsheet integration, no slides. But for plain collaborative writing, it works.

Bob: Proton is headquartered in Geneva, servers in Switzerland and the EU, and the company has been GDPR-aligned from day one. For European listeners who want a collaborative editor without sending every keystroke through US infrastructure, this is one of the few options that actually delivers on the privacy promise. Try it out this weekend on something small.

Samantha: Today we covered: Samsung's first HBM4E samples and the AI memory race, Japan's megabanks deploying OpenAI and Anthropic for cyber defence, Wiwynn's warning on AI hardware bottlenecks, the wider data centre components squeeze, Claude project memory, and Proton's encrypted Docs.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.