DeepSeek's Nvidia Detour

China's AI Self-Sabotage?

It is a feature of top-down industrial policy that you can mandate patriotism but not performance. Beijing has been encouraging its local tech champions to use Huawei's homegrown AI chips, but as the star Chinese startup DeepSeek discovered, encouragement doesn't fix technical glitches. Last week, the Financial Times reported that DeepSeek’s next-generation model R2 was delayed after a failed attempt to train it on Huawei's Ascend processors. The stumble was so significant it forced the company back into the arms of America's Nvidia, for the critical training phase.

This is a little strange. China has been pushing its tech companies to "buy Chinese" and ditch US technology. DeepSeek was reportedly encouraged by authorities to use Huawei's chips. But the effort failed. From the FT:

Huawei sent a team of engineers to DeepSeek's office to help the company use its AI chip to develop the R2 model, according to two people. Yet despite having the team on site, DeepSeek could not conduct a successful training run on the Ascend chip, said the people.

The episode reveals a fundamental truth about the AI arms race: the most significant barrier to competing with the US is not merely about fabricating a powerful GPU. The true chokepoint is the entire, incredibly complex ecosystem of software and hardware that makes those processors work in concert—an advantage Nvidia has spent nearly two decades building.

One way to think about it is that while Huawei's chip might be getting close to Nvidia's on paper, an AI data center is not just a collection of powerful processors; it's a finely tuned system. The hardware that connects the chips and, most importantly, the software that orchestrates them are just as critical. Nvidia's deepest and most defensible moat is not its silicon, but its CUDA software platform. It is an ecosystem so vast and entrenched that competitors are still, to this day, trying to build a bridge across it. Huawei is trying with its own version, called CANN, but DeepSeek's stumble is a powerful illustration of just how wide that software and integration gap remains.

This all gets weirder. The timing of DeepSeek's failure is particularly awkward for Beijing, as it coincides with the US government allowing Nvidia to resume selling its less powerful H20 AI chips to China in exchange for a 15% cut of the revenue. For Chinese tech companies, this creates a fascinating dilemma. They are under intense political pressure from Beijing to use the less-proven Huawei ecosystem. But they also now have the option to once again buy the (nerfed) Nvidia systems that are more reliable. This pits national policy directly against commercial and technical reality.

For now, it seems, the reality is winning. And yet, there is another way to look at this. For DeepSeek and its government backers, a few months of delay might be a perfectly acceptable price to pay for progress toward technological sovereignty. The R1 model, trained on Nvidia chips, gave the US stock market a brief jump scare, to the tune of $1 trillion. An R2 model that was trained entirely on Chinese hardware, would have been something else entirely—perhaps something more like a full-blown cardiac event for investors, especially the Nvidia bulls. The delay, in this light, is not just a failure, but an investment in a much bigger future shock.

Meta's Organized Chaos

It is a little strange when a company reorganizes its most critical division four times in six months. It is stranger still when that company is Meta, and the division is its moonshot bet on artificial intelligence. The dizzying pace of change—coupled with a multi-billion-dollar hiring blitz and soaring spending forecasts—suggests a company in a frantic sprint to catch up in the AI race. The constant shuffling has left investors, and even some employees, asking a fundamental question: is this a brilliant, agile strategy, or is it simply organized chaos?

The official plan is to create a clandestine, startup-like unit focused on building artificial general intelligence (AGI), unencumbered by the bureaucracy of the parent company. But the sheer cost of this ambition is beginning to look uncomfortably familiar. As Reuters reported, the spending is set to accelerate, a direct echo of the capital-intensive Metaverse era:

Rising costs to build out data center infrastructure and employee compensation costs — as Meta has been poaching researchers with mega salaries — would push the 2026 expense growth rate above the pace in 2025, the company has said.

There is certainly evidence for the chaos theory. The constant re-orgs have reportedly "unsettled" existing AI staff, who fear being sidelined. Waning internal confidence in Meta’s own Llama models has led the company to allow some teams to use AI from other companies, suggesting a lack of a unified technical direction. And the strategy appears reactive, with Meta having reportedly explored bids for very different companies—from AI infrastructure player Scale AI to consumer search app Perplexity—suggesting a vision driven more by opportunity than a clear roadmap.

On the other hand, one could argue the chaos is a feature, not a bug. Zuckerberg may be deliberately engineering turmoil to recreate the conditions of a startup, betting that by giving a handful of elite, newly poached leaders near-limitless resources and structural freedom, they can innovate faster than the rest of the market. In this view, he is not buying a plan; he is buying talent and hoping a plan emerges from the creative destruction.

For investors, the whole thing is triggering a powerful sense of déjà vu. The massive, escalating spending—capital expenditure forecasts have soared to as much as $72 billion for the year—is uncomfortably reminiscent of the Metaverse bet, a visionary project that torched billions with little to show for it. The question now is whether Zuckerberg is building a cohesive, world-changing technology, or just the world's most expensive and chaotic science project.

The Scoreboard

• AI: SoftBank lifts stake in Nvidia as Son goes all in on AI (Nikkei Asia)
• AI: OpenAI staff looking to sell $6 billion in stock to SoftBank, others, source says (Reuters)
• Consumer: Samsung taking market share from Apple in U.S. as foldable phones gain momentum (CNBC)