DeepSeek V4 on Huawei Chips: US-China AI Decoupling

Samantha: Welcome to The State of Tech, The European Edition, Saturday April twenty-five, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: DeepSeek's new model running on Huawei chips while Washington cries foul, SpaceX eyeing a one-point-seven-five trillion dollar IPO with an AI twist, Apple's Mac mini vanishing from shelves thanks to the AI boom, Sam Altman apologising over a ChatGPT-linked mass shooting, Google pouring up to forty billion into Anthropic, and a cancer AI alliance using federated learning to crack open medical research. Let's start with DeepSeek and that escalating US-China AI standoff.

China's AI champion ditches Nvidia for Huawei, and the White House calls it theft.

Samantha: DeepSeek has released its V4 model, and the headline isn't just the model itself. It's that it's been adapted to run on Huawei's Ascend AI chips. That's a deliberate move away from Nvidia, and it tells you everything about where China's AI strategy is heading. Self-sufficiency, top to bottom, silicon to software.

Bob: And right on cue, the White House drops the word distillation. They're accusing Chinese entities of, quote, industrial-scale distillation of American frontier models. Translation: take a big US model, query it relentlessly, train a smaller model on the outputs, and suddenly you have a near-equivalent system at a fraction of the cost.

Samantha: Which is exactly the playbook DeepSeek has been accused of before. The cost gap they keep advertising — pennies on the dollar compared to OpenAI or Anthropic — is hard to explain unless something like that is happening. Or unless their engineering really is that much better. Probably a bit of both.

Bob: The bigger story is that the decoupling is now real. Two years ago, even Chinese labs were quietly running on Nvidia. Today DeepSeek is shipping a flagship model tuned for domestic silicon. Huawei's Ascend line isn't matching Nvidia's top end, but it's good enough — and that's the threshold that matters strategically.

Samantha: And the distillation accusation is going to mutate into policy. Expect new export controls, expect terms of service rewrites, expect API rate limiting on accounts flagged as suspicious. The entire question of who gets to query a frontier model is about to get political.

Bob: For Europe, this is the classic squeeze. Brussels wants AI sovereignty but doesn't have the chips, doesn't have the hyperscalers, and doesn't have a domestic frontier lab at DeepSeek or OpenAI scale. Mistral is trying, but the gap is real.

Samantha: And the distillation row complicates partnerships. If a European company builds on top of a Chinese open-weight model — and DeepSeek's models have been popular precisely because they're open and cheap — does that suddenly become a compliance risk? Regulators in Paris and Berlin will be looking at this very carefully.

Bob: The AI Act gives Europe a regulatory voice, but not a manufacturing one. That asymmetry is going to bite harder this year.

---

SpaceX wants one-point-seven-five trillion dollars and calls itself an AI company now.

Bob: SpaceX is reportedly heading to public markets this summer, and the numbers are frankly absurd. One-point-seven-five trillion dollar valuation, seventy-five billion raise. But the more interesting line is buried in the regulatory filings. SpaceX is now describing itself as an AI-first entity, with a total addressable market of twenty-eight-point-five trillion dollars, and they say over ninety percent of that is AI.

Samantha: This is the xAI acquisition paying off narratively. Musk bought xAI in February, folded it into the SpaceX empire, and now Starlink revenue is being pitched as the funding engine for space-based data centres. Orbital compute, basically. AI training in low Earth orbit.

Bob: Which sounds like science fiction until you remember Starlink already has thousands of satellites up there. The pitch to investors is: we have the launch capacity, we have the satellite manufacturing, we have the bandwidth, and now we want to put GPUs in orbit. The cooling argument alone is interesting — space is cold, and energy from solar is essentially free up there.

Samantha: The valuation only makes sense if you buy the AI story. As a pure launch and broadband business, SpaceX is maybe worth four to five hundred billion. Getting to one-point-seven-five trillion requires investors to believe the enterprise AI pivot.

Bob: And public markets will scrutinise that hard. Starlink revenue is real, growing fast, but profitable space-based data centres are still a slide deck. The risk for retail investors is buying a launch company at AI multiples.

Samantha: For Europe, this is uncomfortable on several fronts. Starlink already dominates satellite broadband, including in conflict zones near the EU's borders. Now imagine that same private American operator running orbital AI infrastructure that European companies might depend on.

Bob: The IRIS-squared constellation, Europe's answer to Starlink, is already behind schedule and over budget. If SpaceX raises seventy-five billion and pours it into orbital compute, the gap becomes structural. Brussels will need to decide whether to accelerate domestic alternatives or accept dependency.

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.

---

The cheapest Mac is now the hardest to find.

Samantha: Apple's five-hundred-and-ninety-nine-dollar M4 Mac mini is suddenly very hard to buy. Retailers are out of stock, resale prices are climbing, and Apple isn't giving a clear timeline. The cause isn't a logistics hiccup — it's the AI boom eating consumer hardware.

Bob: Memory is the choke point. DRAM and high-bandwidth memory are being bought up by hyperscalers building AI clusters. TSMC and Samsung are prioritising data centre customers because the margins are better and the contracts are bigger. Consumer chips end up at the back of the queue.

Samantha: Analysts are calling this structural, not cyclical. That's the key word. A cyclical shortage clears in a quarter or two. A structural shortage means the consumer device market just lost its priority status until someone builds a lot more fab capacity.

Bob: And the Mac mini is interesting because it's cheap, it's used for local AI inference by developers and small studios, and it's a popular home server. So the same trend driving demand — local AI computing — is also draining the supply.

Samantha: This is the moment the AI investment cycle stops being a Wall Street story and becomes a high street story. People who just want a desktop computer are running into shortages because Microsoft, Google, and Meta are buying everything that comes off the line.

Bob: And expect prices to drift upward. When fabs prioritise enterprise GPUs and HBM, consumer SSDs and memory get scarcer too. Gaming PCs, laptops, even smartphones — the whole stack feels it.

Samantha: Europe's Chips Act was supposed to address exactly this. The goal was twenty percent of global semiconductor production by 2030. We are nowhere near that. Intel's Magdeburg fab keeps slipping, TSMC's Dresden plant is years away from volume.

Bob: So in the meantime, European consumers and businesses pay the same scarcity premium as everyone else, with no domestic buffer. The Digital Compass goals look more aspirational with every shortage cycle.

---

A mass shooter used ChatGPT, OpenAI knew, and now Sam Altman is apologising.

Bob: This is a difficult one. Sam Altman has issued a public apology after it emerged that OpenAI internally flagged the account of Jesse Van Rootselaar back in June, suspended it for, quote, misuse in furtherance of violent activities, and did not contact law enforcement. Eight months later, Van Rootselaar carried out a mass shooting in Canada that killed eight people.

Samantha: OpenAI's initial position was that the usage didn't meet the threshold for a credible threat. Altman is now saying that judgment was wrong, and the company should have alerted authorities. It's a rare admission of fault from a frontier AI lab.

Bob: The hard question is what threshold should trigger a report. AI companies see millions of disturbing prompts. If every flagged account got reported to police, the system would collapse under false positives, and civil liberties groups would be in uproar. But clearly the current threshold missed something it shouldn't have.

Samantha: Legally, this is uncharted territory. Therapists have a duty to warn under the Tarasoff doctrine in the US. Whether that extends to AI providers has never been tested. After this case, it will be.

Bob: And the political pressure will be intense. Congress will hold hearings, OpenAI will probably build out a dedicated trust and safety reporting unit, and every other lab — Anthropic, Google, Meta — will quietly review their own escalation protocols this weekend.

Samantha: The AI Act has provisions for high-risk systems, but consumer chatbots are mostly classified as limited risk. This case argues that classification is too soft. Expect MEPs to push for explicit reporting obligations when an AI system detects credible threats of violence.

Bob: GDPR also complicates things. Reporting a user to police involves processing personal data for a purpose beyond the original consent. European providers will need a clear legal basis. The Commission may need to issue guidance fast, before this case becomes a template for litigation across the bloc.

---

Google is putting up to forty billion dollars into Anthropic.

Samantha: Google has confirmed up to forty billion dollars going into Anthropic. Ten billion immediately, thirty billion tied to performance milestones. This comes on top of Amazon's existing investment, so Anthropic is now sitting on one of the largest war chests in tech history.

Bob: The structure is interesting. Anthropic agrees to use Google's TPU chips and Google Cloud, so a chunk of that money flows right back to Google as compute spend. It's the same dynamic we saw with Microsoft and OpenAI — the investment is partly a long-term cloud contract dressed up as equity.

Samantha: And it locks Anthropic into the Google ecosystem just as Amazon was trying to pull them closer to AWS. Anthropic is now strategically dependent on two cloud rivals at the same time, which is either a brilliant hedge or a future governance nightmare.

Bob: The capital concentration story is what stands out. Between OpenAI, Anthropic, xAI, and a handful of others, hundreds of billions of dollars are flowing into maybe five companies. Whatever the AI revolution becomes, it'll be shaped by an extremely small number of decision makers.

Samantha: Two years ago people talked about an open AI ecosystem with dozens of competitive labs. That's gone. The cost of training a frontier model has put the bar so high that only the giants and their chosen partners can play.

Bob: And that has knock-on effects for safety, for pricing, for who gets access. When five companies decide what AI can and cannot do, the rest of the world lives with their choices.

Samantha: Mistral has raised serious money but nothing on this scale. The numbers being discussed for European frontier AI are an order of magnitude smaller than what Anthropic just secured.

Bob: Which strengthens the case for the European AI gigafactories programme and pooled sovereign compute. Either Europe finds a way to back a frontier player at scale, or it accepts a regulator-only role in a market shaped entirely in California and Seattle.

---

Cancer research goes federated, and patient data never leaves the hospital.

Samantha: And to close, something genuinely hopeful. The Cancer AI Alliance has launched a collaborative AI platform using federated learning. Dana-Farber, Johns Hopkins, and other major cancer centres are training shared AI models on millions of de-identified patient records — without any of that data ever leaving the original hospital.

Bob: Federated learning is one of those quietly elegant ideas. Instead of pooling sensitive data in one place, you send the model to the data. Each hospital trains the model locally, only the model updates get shared, and the combined model ends up smarter than any single institution could build alone.

Samantha: For cancer research specifically, this is huge. Rare cancers, rare treatment responses, biomarker patterns across diverse populations — all of this needs scale that no single hospital has. And historically, data-sharing agreements between hospitals have been a nightmare of legal review and patient consent.

Bob: The clinical applications are concrete. Predicting which patients respond to which treatment, identifying biomarkers earlier, spotting treatment patterns that work better for underrepresented groups. The equity angle matters — most cancer datasets historically skew white and Western, and federated learning lets you bring in centres that serve very different populations.

Samantha: We talk a lot about AI risks. This is the other side of the ledger. Real medical research, real patient benefit, and the privacy architecture is built in from the start rather than bolted on afterwards.

Bob: For European hospitals this is almost tailor-made. GDPR makes pooling medical data across borders incredibly difficult. Federated learning sidesteps the problem entirely. The data stays in Hamburg or Milan or Rotterdam, only the model travels.

Samantha: The European Health Data Space initiative could absolutely build on this. It's a model for how European medical AI can be world-class without compromising patient rights. Honestly, one of the more encouraging stories of the week.

Bob: Nice to end on something that makes you smile.

Samantha: Today we covered: DeepSeek and the US-China AI showdown, SpaceX's trillion-dollar AI pivot, Apple's Mac mini shortage, the OpenAI duty-to-warn apology, Google's forty billion Anthropic bet, and federated learning in cancer research.

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.