DeepSeek V4 on Huawei Chips: China's Parallel AI Stack

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

Bob: And I'm Bob Russell. Today: billions flowing into AI chip startups as Nvidia and OpenAI battle for silicon supremacy, DeepSeek ditches Nvidia for Huawei with its new V4 model, SK hynix picks up a major IEEE award for its memory work, fresh research warns AI could swallow most routine white-collar tasks, Chinese carmakers steal the show in Beijing with autonomous tech and nine-minute EV charging, and DJI launches a beginner-friendly drone series. Let's start with the chip wars.

The race for AI silicon just hit a new gear.

Samantha: The headline numbers are staggering. OpenAI is reportedly committing over twenty billion dollars to Cerebras chips, a clear move to loosen Nvidia's grip on its compute stack. And Nvidia itself, not content to sit still, scooped up the IP and the talent from Groq late last year for another twenty billion. So we're watching the two biggest names in AI essentially shop for hardware independence and hardware dominance at the same time.

Bob: And that tells you something about where the bottleneck really is. It's not models anymore, it's silicon. Custom silicon, specialised silicon, anything that gets you off the shared queue. Cerebras with its wafer-scale chips offers a very different architecture than Nvidia GPUs, and OpenAI clearly wants optionality. For Nvidia, absorbing Groq's inference talent is about owning the next layer down, the chips optimised for running models, not just training them.

Samantha: There's also a strategic angle here. When you spend twenty billion on chips, you're not just buying capacity, you're locking out competitors from that capacity. Cerebras can only build so many wafers a year. If OpenAI takes a big slice, that's a slice nobody else gets. Same logic with Nvidia consolidating Groq's people. It's a talent moat as much as a tech moat.

Bob: That consolidation is exactly the European concern. Europe has promising chip players, think Graphcore in the UK, SiPearl in France, a handful of photonics startups in the Netherlands and Germany. The risk is they become acquisition targets before they reach scale. We've seen this movie before with ARM. Once the IP and the engineers move under American or Asian ownership, the leverage walks out the door with them.

Samantha: And the European Chips Act was supposed to address exactly this, with forty-three billion euros mobilised to double Europe's share of global semiconductor production by 2030. But that's mostly about manufacturing capacity, fabs, packaging. It's not really about designing AI-specific chips at the cutting edge. There's a gap between making chips and making the chips that matter for frontier AI.

Bob: Which is why the next eighteen months are critical. If European AI labs like Mistral, Aleph Alpha, Black Forest Labs end up entirely dependent on American or Asian silicon, the sovereignty conversation becomes academic. You can regulate AI all you want from Brussels, but if the compute lives in Texas or Taiwan, the real decisions get made there.

Samantha: And speaking of building alternative ecosystems, that brings us neatly to story two.

DeepSeek just told Nvidia: thanks, we'll take it from here.

Bob: The Chinese AI firm DeepSeek has launched its V4 model, and the headline isn't just the model itself, it's that it's been explicitly tailored to run on Huawei chips. That's a significant pivot. Earlier DeepSeek models leaned heavily on Nvidia hardware. Now they're showing China can build a complete domestic AI stack, model and silicon, without Western components.

Samantha: V4 is positioned as an agentic model, designed for complex multi-step tasks that demand serious compute. Early benchmarks suggest it closes much of the gap with closed Western models on cost-efficiency, long context windows, and coding. And because it's open, anyone can run it. That's a different competitive dynamic than what OpenAI or Anthropic offer.

Bob: The geopolitics here is impossible to ignore. US export controls were meant to slow Chinese AI development by cutting off advanced Nvidia chips. Instead, they've accelerated Chinese investment in Huawei's Ascend lineup. DeepSeek choosing Huawei isn't just a technical decision, it's a statement that Chinese AI doesn't need American silicon to compete.

Samantha: And the open-source angle changes the equation for adoption too. A European startup, a research lab in Munich or Madrid, can in principle download V4 and run it on whatever hardware they have. Whether they will is another question, given the data security concerns and the EU AI Act's transparency requirements for general-purpose models.

Bob: For European companies this creates a genuine dilemma. Chinese open models are often cheaper and competitive on capability. But using them raises questions about data flows, IP provenance, and political risk. The Commission has been signalling that critical infrastructure should avoid hardware and software from quote unquote high-risk vendors, and Huawei sits squarely in that category.

Samantha: At the same time, refusing to engage with Chinese AI entirely means missing out on real innovation. The pragmatic European position is probably going to be selective, evaluate models on their merits but keep them out of sensitive workflows. That's easier said than implemented at the developer level.

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.

A different kind of AI hardware story, this one ending in a trophy.

Bob: SK hynix has picked up the IEEE Corporate Innovation Award at the 2026 IEEE Honors Ceremony in New York. The recognition is specifically for its work on High Bandwidth Memory, HBM, and for ensuring stable mass production across every HBM generation. That sounds dry until you remember HBM is what makes modern AI accelerators actually work.

Samantha: Right, every Nvidia H100, every B200, every cutting-edge AI chip has stacks of HBM sitting next to the processor. Without that memory bandwidth, the GPU starves. SK hynix and Samsung between them basically supply the entire industry. SK hynix being singled out here reflects how dominant they've become in HBM3 and HBM3E specifically.

Bob: And the supply situation is genuinely tight. HBM is sold out through 2026 for the major buyers. SK hynix is investing heavily to expand capacity, but you can't just spin up an HBM fab overnight, the precision required is extraordinary. So the company recognised today is also the company holding a chokepoint for global AI development.

Samantha: Which loops back to our first story. When OpenAI commits twenty billion to Cerebras, part of what they're really committing to is securing HBM allocation. The chip designer matters, but the memory supplier matters just as much. And Europe has essentially zero presence in this segment.

Bob: There is one bright spot in Europe, ASML in the Netherlands makes the lithography machines that fabs use to produce HBM. So Europe sits at the very top of the supply chain, but not in the production layer itself. That's a strange position, indispensable upstream, absent downstream.

Samantha: And it explains why every major economy is now trying to onshore at least some memory production. The US CHIPS Act, the EU Chips Act, Japan's Rapidus initiative. Whether any of those efforts can match SK hynix and Samsung's combined fifteen-year head start is genuinely doubtful.

AI isn't coming for the factory floor, it's coming for the office.

Samantha: New research from Anthropic and Brookings makes the picture pretty stark. Anthropic suggests AI could automate over ninety percent of tasks in administrative, clerical, and middle management roles, and over eighty percent of tasks in arts, media, and law firms. Brookings puts thirty-seven million American workers in the highly exposed category, with about six million likely to struggle adapting.

Bob: The framing here matters. These aren't predictions of jobs eliminated, they're predictions of tasks automated. A job is usually a bundle of tasks, and when AI eats most of them the role gets restructured rather than deleted outright. But restructured roles often mean fewer of them, and different skills required from whoever remains.

Samantha: The interesting inversion is that for years we assumed automation would hit blue-collar jobs first, factory workers, drivers, warehouse staff. Instead it's the white-collar professions that turn out to be more vulnerable. Drafting contracts, summarising meetings, basic accounting work, customer service emails, all of it generative AI handles competently now.

Bob: And Europe has a higher concentration of exactly these kinds of roles than the US in some respects. Public sector administration, legal services, financial services in cities like Frankfurt, Paris, Amsterdam, Dublin. These are large workforces doing exactly the tasks Anthropic flagged at over ninety percent automation potential.

Samantha: The policy response in Europe has been more proactive than in the US so far. Spain has trialled a four-day work week in part as a hedge against automation. France and Germany are funding large-scale reskilling programmes. The EU AI Act includes provisions on workforce impact that go beyond anything in American law.

Bob: But honestly, none of those measures are scaled to a thirty-seven million worker problem extrapolated to European populations. The retraining infrastructure simply doesn't exist at that scale. And the political question of who pays for the transition, employers, the state, displaced workers themselves, that's still unresolved everywhere.

Auto China 2026 is a window into where the car industry is going, and it's not Stuttgart.

Bob: The Beijing show this year has been a parade of capabilities that, frankly, European manufacturers can't currently match. XPeng demonstrated autonomous systems that detect driver incapacitation and pull the vehicle over while alerting emergency services. BYD unveiled a next-generation Blade battery that charges to near-full in nine minutes. Nine minutes.

Samantha: Nine minutes is essentially the time it takes to fill a petrol tank and pay. That's the psychological threshold the EV industry has been chasing for a decade. And the autonomous incapacitation feature is exactly the kind of safety capability that regulators love, because it has obvious life-saving potential rather than just convenience appeal.

Bob: The competitive picture for Europe is uncomfortable. German manufacturers, Volkswagen, BMW, Mercedes, are all in the middle of expensive software transitions, trying to build their own operating systems, their own driver assistance stacks. Meanwhile BYD, XPeng, Nio are shipping vehicles with capabilities that match or exceed Tesla, at lower price points.

Samantha: The EU has imposed tariffs on Chinese EVs precisely because of this competitive imbalance. But tariffs buy time, they don't fix the underlying gap. The real question is whether European OEMs can use that time to actually catch up, or whether they just delay an inevitable market correction.

Bob: There's also the supply chain dimension. BYD's battery breakthrough relies on chemistry and manufacturing processes that Chinese firms control. CATL and BYD between them produce the majority of the world's EV batteries. European battery initiatives like Northvolt have struggled badly, with the Swedish flagship in serious financial trouble last year.

Samantha: So the picture is European tariffs holding off Chinese cars at the border while European factories still depend on Chinese batteries inside the cars they build. That contradiction can't hold forever. Either Europe builds its own battery capacity at scale, or the industrial policy is essentially performative.

Let's end with something that just sounds fun. Drones for everyone.

Samantha: DJI has launched the Lito series, a new line of camera drones aimed squarely at beginners and aspiring creators. Two models, the Lito X1 and the Lito 1. The premium X1 packs a one-over-one-point-three inch CMOS sensor, forty-eight megapixels, and what DJI is calling five-lux omnidirectional obstacle sensing with forward-facing LiDAR.

Bob: The LiDAR detail is what makes this interesting. Forward-facing LiDAR on a beginner drone means it can navigate complex environments, around trees, under bridges, in low light, without the operator needing expert flight skills. That's a meaningful safety upgrade, because most beginner drone crashes happen exactly when pilots misjudge obstacles.

Samantha: And the pricing strategy is clearly aimed at expanding the market rather than competing on the high end. DJI already dominates the consumer drone market globally, somewhere north of seventy percent share. Making drones easier and cheaper grows the pie, which is smart given how saturated the prosumer segment has become.

Bob: For Europe there are two sides to this. On one hand, accessible drones boost the creative economy, real estate photography, wedding videos, small business marketing, hobbyist filmmaking. There's a whole micro-industry of drone services that benefits from cheaper hardware. On the other hand, more first-time pilots in the air means more incidents.

Samantha: EASA's drone regulations already require registration and basic competency tests for most categories, but enforcement is patchy across member states. A surge of new Lito owners will test that framework. Some countries, France, Germany, the Netherlands, have stricter no-fly zones and active enforcement. Others rely largely on the honour system.

Bob: There's also the geopolitical footnote. DJI is Chinese, and several Western governments have raised data security concerns about its products. The US has restricted DJI in some federal contexts. The EU hasn't gone that far, but the conversation is happening. So even a fun beginner drone story has a sovereignty subtext if you scratch the surface.

Samantha: True, but I'll take the win that more people get to make beautiful aerial videos of European coastlines and mountains. Sometimes a drone is just a drone.

Samantha: Today we covered: the AI chip investment frenzy, DeepSeek's Huawei pivot, SK hynix's IEEE award for HBM, AI's threat to white-collar tasks, Chinese carmakers leading at Auto China, and DJI's new beginner drones.

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.