Alibaba AI Stack: New Model, Chips & Servers Explained

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

Bob: And I'm Bob Russell. Today: Meta cuts eight thousand jobs to pour cash into AI, Alibaba launches a full new AI stack with its own chips, a clever new method that could shave millions off training big AI models, Samsung dodges a massive chip strike at the last minute, two of the world's biggest banks tell staff to make peace with AI, and Western Digital ships hard drives built to survive quantum hackers. Let's start with Meta.

Meta lays off eight thousand staff, roughly ten percent of its workforce, to fund AI push.

Samantha: Meta is cutting around eight thousand jobs, about ten percent of its global headcount. Mark Zuckerberg sent an internal memo framing it as a reallocation of capital toward artificial intelligence, and he told remaining staff there won't be another broad round of cuts this year.

Bob: This is the biggest reduction at Meta since the twenty-two and twenty-three layoffs. The pain falls mostly on corporate support, recruiting, and marketing. Engineering and AI teams are largely shielded, because that's where the money is going.

Samantha: And the message to the rest of Silicon Valley is loud and clear. We've now seen Meta, Google, and Microsoft all trim non-AI functions in the same quarter. The new rule seems to be: if your job isn't building or selling AI, you're vulnerable.

Bob: It also tells you something about where Meta thinks the next ten years of revenue come from. They're betting that AI-powered ads, AI assistants in WhatsApp and Instagram, and the agent economy will more than cover the cost of those data centres.

Samantha: Which is a huge bet. Meta's capital spending is now north of seventy billion a year, mostly on chips and cooling. That money has to come from somewhere, and clearly the answer is: from the headcount.

Bob: Workers in those eliminated roles aren't going to retrain into AI researchers overnight. So you get this awkward gap, where companies say AI creates jobs, but the jobs it creates need PhDs and the ones it cuts don't.

Samantha: For Europe specifically, Meta has big offices in Dublin, London, and Paris. Recruiting and marketing teams in those hubs are right in the firing line, and Dublin in particular has been here before with the twenty-two-twenty-three round.

Bob: And it lands at an awkward time politically. Brussels has been pushing Meta hard on the Digital Services Act and on AI transparency. Cutting trust and safety adjacent roles while spending more on generative AI is exactly the combination regulators worry about.

Samantha: European workers who do get laid off also have stronger protections than their American colleagues, so expect this to play out slowly, with works councils and consultations. The eight thousand number is global, but the European chunk will take months to land.

Alibaba unveils a full-stack AI upgrade with a new model, new chips, and a new server.

Bob: Over to China. Alibaba has rolled out a top-to-bottom overhaul of its AI offering. There's a new flagship language model called Qwen three point seven Max, a new training and inference chip from its in-house semiconductor arm, and a new server design built for running AI agents at scale.

Samantha: The chip part is the real story. Alibaba's T-Head subsidiary now has its own AI processor, plus its own software stack to run on it. That's Alibaba saying: we don't want to depend on Nvidia, and given the US export rules, we can't anyway.

Bob: And the model itself is pitched at what Alibaba calls agentic work. Meaning AI that doesn't just answer a question, but actually carries out tasks across software, like writing code, running tests, and finishing a long project step by step.

Samantha: It's the same pitch we hear from OpenAI and Anthropic, but Alibaba is targeting Asian and emerging market customers who want a cheaper alternative, and don't want their data sitting on American servers.

Bob: The combination matters. If you control the chip, the cloud, and the model, you can undercut Western rivals on price and lock customers in. That's the playbook Huawei is running in telecoms, and now Alibaba is running it in AI.

Samantha: For European businesses, this is a real choice. If you want generative AI that isn't routed through American hyperscalers, Qwen is now genuinely competitive. Some German and French firms are already testing the open-source versions.

Bob: But it comes with strings. Hosting sensitive data on Alibaba Cloud, or running a Chinese model on your own kit, brings supply chain and security scrutiny from Brussels. Expect the EU AI Office to take a closer look at deployments at scale.

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.

New scaling law method could save millions of dollars and weeks of time on training frontier AI models.

Bob: Researchers have published a new approach to so-called scaling laws, which are basically the formulas labs use to predict how a bigger AI model will perform before they spend a fortune training it.

Samantha: And training a frontier model is not cheap. We're talking up to a billion dollars per attempt for the biggest systems. So even shaving a few percent off that bill, by predicting more accurately whether the model is worth building, is enormous.

Bob: The trick here is borrowed from measurement science. Instead of training lots of small models to extrapolate, they use statistical techniques to get reliable forecasts with far fewer test runs. Less compute, less electricity, less risk.

Samantha: It also lowers the barrier slightly for smaller players. If you don't need to spend tens of millions just to predict whether your next model will be any good, mid-sized labs and university groups can actually compete in research again.

Bob: Although let's not get carried away. The hyperscalers will use this to train more models, not fewer. Cheaper experiments means more experiments. The total energy bill of AI isn't shrinking, the productivity per dollar is just going up.

Samantha: European AI labs have been quietly drowning in the cost gap. Mistral, Aleph Alpha, the academic consortia, they all train at a fraction of the budget of OpenAI or Anthropic. Anything that makes prediction cheaper helps them close that gap.

Bob: It also dovetails with the EU's AI Factories initiative, where Brussels is funding shared supercomputers for European startups. If those startups can stretch their compute further, the public money goes further too. Modest news on the surface, but real impact underneath.

Samsung averts massive seventy thousand worker strike with last minute wage deal.

Samantha: Samsung Electronics and its biggest union struck a tentative wage deal just hours before more than seventy thousand workers were due to walk off the job. The strike would have run until June seventh and risked serious disruption to global memory chip supply.

Bob: Memory is the unglamorous part of the chip industry, but it's everywhere. In your phone, your laptop, every AI server. Samsung is the world's biggest maker of it, and a multi-week strike would have sent prices upward right when AI data centres are buying everything they can find.

Samantha: The union has paused the action pending a ratification vote. Samsung also issued a rare apology for the prolonged uncertainty, which tells you how nervous management was about a walkout heading into the summer.

Bob: This is the second year of serious labour action at Samsung. The Korean tech giants used to have an almost union-free culture. That's clearly changing, and it's changing at exactly the moment when global demand for their products is white-hot.

Samantha: European carmakers in particular were watching this closely. Modern cars eat memory chips, and brands like Volkswagen and Stellantis still haven't fully forgotten the supply shocks of a few years ago.

Bob: For European consumers it means laptop and phone prices probably won't spike this summer. But the deeper lesson is concentration risk. One union vote in Suwon can move global tech prices. That should be making Brussels think harder about its chip resilience plans.

HSBC and Standard Chartered tell staff to embrace AI as banking jobs reshape.

Bob: Two of the world's biggest banks are now openly addressing what AI means for their staff. HSBC's chief executive Georges Elhedery told employees not to resist, and said the goal is to make people, in his words, more productive versions of themselves.

Samantha: Standard Chartered is in similar territory, after some earlier comments about replacing what they called lower-value human capital landed badly with staff. So now we get the softer message: AI will both eliminate and create jobs.

Bob: Back-office banking is where this hits hardest. Compliance checks, document review, customer service triage, fraud screening. All of it is being automated. The banks aren't denying it, they're just trying to manage how staff hear about it.

Samantha: And the wording matters. Calling people lower-value human capital is the kind of phrase that ends up in a union pamphlet very quickly. The fact that CEOs are now actively coaching their language tells you this is a sensitive moment internally.

Bob: It's also a recognition that you can't just buy the AI and fire the staff. You need people who know your processes to actually deploy these tools. So the banks are walking a line between cost cutting and keeping institutional knowledge in the building.

Samantha: European banks are in the same position but with very different rules. Works councils, collective bargaining, and strict redundancy procedures mean the same restructuring takes years, not months. That's a brake, and arguably also a buffer.

Bob: ING and BNP Paribas have both said publicly that AI will reshape their headcount over the coming years. The honest version is: fewer people in the call centre, more people building and supervising AI tools. Whether the same humans can make that jump is the real question.

Spotify rolls out verified badges for podcasts as AI-generated shows flood the platform.

Samantha: Time for two things you can actually do something with today. First, Spotify is expanding its verification system to podcasts. You'll start seeing green verified badges on shows that the platform has confidently checked out.

Bob: It's the same idea as the blue tick on social media, but aimed at a real problem. Spotify is being flooded with AI-generated podcasts, including fake shows that clone real hosts' voices. The badge is meant to help listeners tell the genuine ones apart.

Samantha: To get verified, a show has to clear a few hurdles. Real listener engagement, no policy violations, and what Spotify calls audience authenticity. Which is a polite way of saying: no bots streaming your own podcast to game the numbers.

Bob: They're also banning unauthorised AI voice clones. So if someone spins up a fake show using your voice without permission, Spotify says it will take it down. Whether that actually works at scale is a different question, but the policy is now on paper.

Samantha: For listeners, the practical thing is simple. Over the next few months, when you open Spotify and look at a podcast, glance at the show page. If there's a green check, it's been reviewed. If there isn't, especially on a brand-new show with a familiar-sounding host, be a bit sceptical.

Bob: And if you make a podcast yourself, it's worth checking Spotify's creator dashboard in the coming weeks for verification options. The rollout starts small, but they've said it will expand. Getting in early might help you stand out from the AI noise.

Western Digital ships hard drives designed to resist future quantum computer attacks.

Bob: Second find, and this one is for the slightly nerdier listener. Western Digital has launched a new line of Ultrastar hard drives built with post-quantum encryption. In plain English: drives designed to stay secure even once quantum computers can crack today's encryption.

Samantha: The threat they're addressing has a memorable name. Harvest now, decrypt later. The idea is that attackers steal encrypted data today, sit on it, and wait until quantum computers are powerful enough to crack it open in five or ten years.

Bob: The encryption methods in these new drives are based on standards approved by the US National Institute of Standards and Technology. So they're not experimental, they're the official next-generation algorithms that governments and big enterprises will be moving to anyway.

Samantha: Western Digital also flagged a second risk: tampered firmware updates. If a quantum-capable attacker can forge the signature on a firmware update, they can push malicious code straight into your storage. The new drives are designed to refuse anything that isn't properly signed.

Bob: For most listeners, you're not going to buy one of these for your home PC, they're for data centres and businesses. But if your company stores sensitive records that need to stay secret for a decade or more, medical, legal, government, this is the kind of upgrade IT teams will be quietly planning this year.

Samantha: So today we covered: Meta's eight thousand AI-driven job cuts, Alibaba's new homegrown AI stack, a smarter cheaper way to train big models, Samsung dodging a chip strike, banks telling staff to embrace AI, and two practical finds, Spotify verifying podcasts and Western Digital's quantum-proof hard drives.

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.