Anthropic's $40B xAI Compute Deal & Chip Stock Sell-Off

Samantha: Welcome to The State of Tech, The European Edition, Saturday July eighteen, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: a forty billion dollar twist between Elon Musk's xAI and Anthropic, China switches on the first commercial brain chip implant, Moonshot drops the largest open weights model yet, tech stocks take a serious beating on fears the AI rally is running out of road, and to close out, two things you can actually use today: the transcription tool Whishper, and the open-source read-later app Wallabag. Let's start with the biggest deal in AI infrastructure this year.

Elon Musk flips on Anthropic as it signs a forty billion dollar compute deal with his xAI.

Samantha: Elon Musk has publicly changed his tune on Anthropic. Until recently he was dismissive, now he's calling them a leading AI company. And the reason is not hard to find. Anthropic has reportedly become one of the biggest customers of Musk's own AI firm xAI, signing a deal back in May for three hundred megawatts of compute from the Colossus 1 data centre. The total value: roughly forty billion dollars in revenue for xAI, running through May 2029.

Bob: Which is a striking arrangement, because Anthropic and xAI are direct rivals in the model business. What this shows is how brutal the compute crunch has become. Anthropic needs power on a scale that only a handful of players can deliver, and xAI happens to have built that capacity around its Memphis site. So a competitor becomes a landlord. Forty billion is not a rounding error, it's on the order of what some countries spend on their entire tech sector in a year, and it flows through a single bilateral contract.

Samantha: And it reframes what xAI actually is. On paper it's a model lab, but this deal makes it look more like a specialist power and compute utility for the rest of the AI industry. The Colossus site was originally sold as the training ground for Musk's own Grok models. Now it's also a paying facility for the company he used to mock.

Bob: For Europe the number should land hard. Three hundred megawatts is roughly the draw of a mid-sized European city, dedicated to one customer at one site. European AI Gigafactory plans are measured in a few hundred million euros of subsidy, not tens of billions of private compute contracts. Industry groups have been arguing for months that without a jump in scale, European labs will be renting American capacity by default. This deal is exhibit A for that argument. It also puts pressure on regulators looking at market concentration, because the same names keep appearing on both sides of these mega contracts.

China switches on what it calls the first commercial invasive brain chip implant.

Samantha: China has announced what it describes as the first commercial invasive brain chip implant. That's a step beyond clinical trial: it's framed as a product that can be sold and deployed, not just tested. Details are limited and independent verification is still pending, but if the framing holds up it puts China ahead of Western programmes like Neuralink, which remain firmly in the trial stage.

Bob: Some caution is in order here. Commercial in this context can mean many things, from a tightly restricted medical procedure at a handful of hospitals to something closer to an approved device. The announcement does not spell that out. What it does signal is intent: Beijing wants brain-computer interfaces treated as a strategic industry, alongside chips and AI, and it wants to be first past the post on real-world deployment.

Samantha: For patients with severe paralysis or communication disorders this technology is genuinely life-changing, so the medical stakes are real. But so are the questions. An implanted chip that reads neural signals raises issues around consent, long-term safety, data ownership, and what happens when the company behind the device changes hands or shuts down. Those are not hypothetical concerns once you move from trial to market.

Bob: European regulators have been quietly working on rules for neurotechnology under the medical devices framework, and several member states have flagged brain data as a distinct category that deserves protection beyond standard health data. A Chinese commercial rollout will almost certainly accelerate that work. European researchers in Lausanne, Grenoble and the Netherlands have been leaders on the science side, but the gap between lab breakthroughs and approved commercial products is where the continent has repeatedly lost ground. This is another test of whether that pattern changes.

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.

Moonshot AI releases Kimi K3, the largest open weights model the world has seen.

Bob: Chinese lab Moonshot AI has released Kimi K3, and the numbers are eye-catching. Around 2.8 trillion total parameters, a context window of one million tokens, and open weights promised by 27 July. There are two versions: one aimed at chat and agent work, one tuned for large-scale parallel processing. Pricing through the API is three dollars per million input tokens and fifteen for output, which undercuts most Western frontier models.

Samantha: The headline here is not just size, it's the openness. Open weights means anyone can download the model and run it on their own infrastructure, subject to the licence terms. That is a very different proposition from a closed API. For companies worried about sending sensitive data to a US cloud, or about pricing at the whim of a single vendor, an open model of this class changes the calculation. And the context window matters: one million tokens is roughly a long novel in a single prompt, useful for legal review, code bases, and research work.

Bob: There is a serious catch though. Running a model of this scale locally is not something you do on a laptop, or even on a normal server. You need a small cluster of top-end AI accelerators, which most European companies do not have sitting spare. So in practice open weights often still means renting compute somewhere, and the question is where. Sovereign European cloud providers see this kind of release as an opportunity to offer a hosted version inside the EU, under European law.

Samantha: It also sharpens the strategic question in Brussels and in national capitals: does Europe try to build its own frontier model, or does it focus on hosting, tuning, and applying open models coming out of China and the United States. Mistral and a handful of others are still in the first camp. A release like Kimi K3, cheap and downloadable, makes the second option look increasingly attractive to boards that just want the capability at a predictable cost.

Chip stocks tumble as investors start doubting the AI rally that carried markets all year.

Bob: Tech stocks took a hard hit heading into the weekend. The Nasdaq closed down 2.2 percent yesterday, and the semiconductor index has now shed nearly twenty-one percent over July alone. Chipmakers and AI-linked names led the fall. The picture in Europe was similar: the Stoxx 600 technology sub-index dropped in sympathy, and ASML, the Dutch equipment maker that sits at the heart of the AI supply chain, was among the biggest fallers.

Samantha: The bottom line is that investors are asking harder questions about whether the capital being poured into AI data centres will ever earn a decent return. We just talked about a forty billion dollar compute deal between two labs. Multiply that pattern across the industry and the numbers get uncomfortable fast. Analysts at several US banks have flagged concerns about oversupply of certain chip categories heading into next year, and that is starting to bite sentiment even where actual demand still looks strong.

Bob: There's also a rotation story underneath the headline number. Money is not necessarily leaving tech entirely, but it is moving away from the highest-flying names toward more defensive positions. That kind of shift often precedes a broader repricing.

Samantha: For Europe the exposure runs through a handful of critical suppliers. ASML sells the lithography machines that every advanced chip fab in the world needs, so a slowdown in fab spending hits its order book directly. Companies like ASM International, BE Semiconductor, and Infineon are in the same neighbourhood. Pension funds across the continent have also loaded up on US tech during the rally, so a sustained correction shows up in retirement portfolios well beyond the trading floor. Whether this is a healthy pause or the start of something larger, next week's earnings from the big US chip names will start to answer.

Whishper turns your own laptop into a private transcription studio, no cloud required.

Bob: Time for something you can actually try today. First up: an open-source transcription app called Whishper. It's a self-hosted tool built on top of OpenAI's Whisper speech recognition model, but the whole thing runs on your own machine or a small home server. You drop in an audio or video file, and it produces a searchable transcript with time codes, speaker segments, and export to subtitles.

Samantha: The point of Whishper is that nothing leaves your device. For journalists working with confidential sources, lawyers handling client recordings, researchers doing interviews, or anyone bound by GDPR when processing personal audio, that matters. Users report solid accuracy across major European languages, including Dutch, German, French, Spanish, and Italian, and it handles long files without falling over. It also translates as it transcribes, so a Spanish interview can come out as English text in one pass.

Bob: Setup is not quite a one-click affair. It runs in Docker, which is a way of packaging software so it works the same on any machine, and there's a clear guide on the project's GitHub page. If you already have a home server or a spare laptop, you can be running in under an hour. If you don't, it also runs on a normal desktop, just slower for very long files.

Samantha: It's free, the code is open, and there is no account to create. For a European audience, that combination is genuinely rare in the transcription market, where most polished tools are US cloud services with per-minute pricing and unclear data policies. Whishper is one of those cases where the open-source community has quietly built a tool that holds its own against paid alternatives, and it deserves more attention than it gets.

Wallabag saves the articles you meant to read, on your terms.

Bob: The second thing to try is Wallabag. It's an open-source read-it-later app, in the same category as the well-known commercial services, but built and hosted in France with a strong focus on user control. You install a browser button or the mobile app, and one click stores any article in a clean, readable format for later, online or offline.

Samantha: What sets Wallabag apart is where your reading list lives. You can either pay a small annual fee for the hosted European version, which keeps everything on servers inside the EU, or you can install it yourself on a home server and own the whole stack. Either way, the articles you save, the highlights you make, and the tags you add stay with you, not with an advertising platform trying to profile your interests.

Bob: It handles the basics well: text-to-speech for listening on a walk, epub export for sending long reads to a Kindle or Kobo, and full text search across your entire archive. Users often build up thousands of saved articles over years, and Wallabag keeps them accessible without the risk of the service being shut down or bought out, which has happened repeatedly to competitors.

Samantha: For a European listener who wants a calmer relationship with the news feed, and a reading archive that survives the next round of tech consolidation, Wallabag is a quietly excellent choice. Free to self-host, modest annual fee for the managed version, works on every major platform.

Samantha: Today we covered: Elon Musk's forty billion dollar compute deal between Anthropic and xAI, China's first commercial brain chip implant, Moonshot's giant open weights model Kimi K3, the sharp sell-off in tech and chip stocks, and two things to try yourself: the private transcription tool Whishper, and the European read-later app Wallabag.

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.