DeepSeek Just Ordered 160,000 Huawei Chips — the 'Horrible Outcome' Jensen Huang Predicted Just Got Real

DeepSeek Just Ordered 160,000 Huawei Chips — the 'Horrible Outcome' Jensen Huang Predicted Just Got Real

Let me tell you something that's been sitting with me since the Bloomberg report landed. DeepSeek — the Hangzhou lab whose API prices every AI developer on the planet watches like a hawk — is reportedly placing an order for at least 160,000 of Huawei's next-generation Ascend 950DT chips to fill a one-gigawatt data center in Ulanqab, Inner Mongolia. Not Nvidia. Huawei. And if that order gets fulfilled, it becomes one of the largest known clusters of Chinese-made AI chips on Earth.

I've been running hosting infrastructure for over a decade, and I've watched this decoupling story from the server-room floor. For years the working assumption was simple: China could build models, but it needed Western silicon to do it. Every export control debate assumed the hardware was the leash. This report is the strongest signal yet that the leash is slipping — and that the world is quietly becoming two separate compute planets.

What Bloomberg Actually Reported

Let me be precise about what's confirmed and what isn't, because that matters to people who buy and sell compute for a living. Bloomberg reported September 4, citing people familiar with the matter, that DeepSeek intends to deploy at least 160,000 Ascend 950DT accelerators at a massive data center it's building in Ulanqab, Inner Mongolia — a site reported since the summer at roughly one gigawatt of capacity, which is enough to power about 750,000 homes. At the current street price of around 111,000 yuan — roughly $16,000 per chip — the order carries a face value in the neighborhood of $2.6 billion.

Neither DeepSeek nor Huawei has confirmed the plan. No contract has been publicly disclosed. But the reporting is consistent with everything else we've seen this year: DeepSeek tuned its V4 model to run on Huawei's Ascend 950 when it launched in April, and Chinese cloud providers — Alibaba, ByteDance, Tencent — reportedly followed with orders for hundreds of thousands of the processors. Ulanqab sits about 350 kilometers northwest of Beijing, where cheap wind and solar meet an average temperature of 4.3 degrees Celsius — free cooling and cheap power, the two things every data center operator on earth is desperate for.

The Chip Itself — 'David' Against 'Goliath'

Here's where the technical story gets interesting. The Ascend 950DT, internally codenamed David — as in the underdog against Goliath — is built on Huawei's third-generation DaVinci v3 architecture and manufactured by SMIC on its N+3 process, an advanced 7-nanometer-class node. Each chip carries 144 gigabytes of Huawei's proprietary HiZQ 2.0 high-bandwidth memory with 4.0 terabytes per second of bandwidth. Huawei plans to launch it in the fourth quarter of 2026, and its Atlas 950 SuperPod links up to 8,192 of these chips into a single cluster.

But read the fine print. DeepSeek's own founder, Liang Wenfeng, gave the most candid assessment in a leaked investor call circulated in July: all tasks a GB300 can do, the Huawei supernode can do, with the same latency. The only cost, he said: four Huawei chips equal one Nvidia chip, and it's two years behind. Four for one. That's not a marketing number — that's the founder of the customer telling investors the real exchange rate. And yet DeepSeek is reportedly still buying 160,000 of them, primarily for inference — running models — not for training, where it continues to rely on Nvidia.

The Two Readings — a Self-Reliance Milestone, or a One-Gigawatt Compromise

There are two ways to read this, and like most things in this industry right now, both are true at once.

Reading one: this is the milestone. Six months ago, an official newspaper in Shenzhen celebrated the start of China's first 10,000-chip intelligent computing cluster using Huawei's Ascend series. A hundred and sixty thousand chips in a single facility is sixteen times that — the ballpark of what Western companies now build, filled with domestic silicon. Beijing has pushed state-funded data centers to use Chinese chips, told ByteDance and Alibaba to halt Nvidia testing, and watched Nvidia's share of China's AI chip market collapse from roughly 95 percent to zero. When Jensen Huang calls a DeepSeek-on-Huawei outcome "horrible" for America, that's not hyperbole — it's the sound of an ecosystem losing its toll booth.

Reading two: it's a compromise, and an expensive one. Four Huawei chips per Nvidia chip means a gigawatt of Chinese compute delivers a fraction of the work a Western gigawatt does. The Council on Foreign Relations estimated America's best chips are still about five times more powerful than Huawei's best on total processing performance — a gap it projected could widen to seventeen times by late 2027. DeepSeek tried to train on Huawei before and went back to Nvidia. This order is for inference, the revenue-generating half of the business, while the training half stays on the hardware Beijing can't yet replace. That's not capitulation — but it's not parity either.

The Secondary Bottleneck Nobody's Talking About — You Can't Order Your Way Out of a Memory Shortage

Now the part that should make every infrastructure operator sit up straight. The reason this order can't be fulfilled quickly has nothing to do with demand and everything to do with memory. Huawei's production of the 950DT is expected to be limited to the low hundreds of thousands of units this year because of shortages of advanced high-bandwidth memory — the same HiZQ stacks Huawei designed specifically to get around US export restrictions on HBM. Fulfilling DeepSeek's order alone could take more than a year.

Read that again: the chip champion of China's self-reliance push is bottlenecked on the same component that's bottlenecking every AI buildout on earth — high-bandwidth memory. Huawei is spending more than a quarter of its revenue on R&D — its first-half profit fell 36 percent — to build an AI stack without Nvidia at the center, and it still can't print enough memory to fill the orders sitting in front of it. DeepSeek has reportedly even asked Beijing to pressure Huawei to prioritize its allocation. When the customer of a state champion has to ask the state to lean on the supplier, the shortage is real, and it's structural.

What This Means for Independent Hosting Providers

If you're running an independent hosting or colocation business, this story is not a distant geopolitical curiosity. It's your pricing environment taking shape. Here's what I'd do with it.

First, start planning for two compute planets. The US-China decoupling is no longer a trade-policy argument; it's a physical reality of separate silicon supply chains. Price your services against a world where a major inference provider runs on cheaper domestic silicon with state backing — because that world is being built right now, and it will export price pressure.

Second, watch where Huawei sells next. Washington is reportedly scrambling to counter a Huawei proposal to build a 2,008-chip AI data center in Egypt with iFlytek, complete with surveillance software. The pattern is unmistakable: the same export controls that slowed China at home are becoming Huawei's sales pitch abroad. Cheap Chinese clusters are coming to emerging markets, and that is your future competition for the same customers.

Third, respect the gap between announced and delivered. A 160,000-chip order is not 160,000 chips running. It's an order that could take more than a year to fill, into a facility that may only partially open by late 2027 or early 2028. When you see megawatt headlines, ask about the memory allocation before you believe the timeline — that discipline will save you from betting capacity on vapor.

Fourth, learn the four-for-one lesson. DeepSeek is buying four less-capable chips at $16,000 each to do a job one Nvidia chip does better, because it controls the whole stack and the alternative is nothing at all. Don't assume your customers won't do similar math with commodity hardware. Sometimes total cost of ownership beats raw performance — price accordingly.

Fifth, hedge your own supply chain. The same export controls that locked Nvidia out of China can redirect GPU supply westward, but memory shortages hit everyone. Keep alternative suppliers warm, and don't let a single component become the load-bearing wall of your infrastructure plan.

The Bottom Line

Jensen Huang called DeepSeek running on Huawei chips a "horrible outcome" for America. He's right, but not for the reason the headlines suggest. The horror isn't that China built a better chip — it hasn't. The horror is that China built a second chip ecosystem that's good enough, at a scale that's now measured in hundreds of thousands of units, backed by a state willing to lean on its own suppliers to make it happen. That's not a switch that gets flipped back when the export controls loosen. Ecosystems don't work that way, ent?

For the rest of us, the lesson is simpler than the geopolitics. Demand was never the constraint on this buildout — memory, power, and permission were. And now we know that's true on both sides of the divide. Plan for scarcity, respect the timelines, and never confuse an order with a running workload. That's how you survive whichever planet you're standing on.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: Bloomberg, The Next Web, Moneycontrol, Outlook Business, TechTimes, Tech-ish.