DeepSeek Finally Says "We'll Make the Chip Ourselves Too"

Here's the deal: Reuters, citing three sources, just broke the news that Chinese AI startup DeepSeek is developing its own AI chip. And this isn't a vague "we're thinking about it" — the company is already in active discussions with chip-design firms, foundries, and memory makers, and it's hiring semiconductor engineers. One source said the project kicked off roughly a year ago. It's early-stage, sure, but the hands are already moving, not just the mouth.

The single most important detail is that this chip is built for inference, not training. AI chips split into two broad camps. One trains a model from scratch — the brutal, expensive compute of building the thing. The other runs inference: it takes a finished model and makes it spit out answers to user queries. DeepSeek is going after the second camp. It doesn't want to own the ground-up training silicon; it wants to control the cost of actually serving its models in production. Why does that distinction matter so much? Inference is technically less demanding than training, and for a company whose whole identity is "serve a lot, dirt cheap," inference cost is basically survival.

And you have to see why this is news right now. DeepSeek is the same company that rattled global markets in early 2025 with its low-cost reasoning models — the one that wiped hundreds of billions off Nvidia's market cap in a single day. Now that same company is saying it won't leave its chip supply in anyone else's hands. Not Nvidia's, and not even Huawei's, China's own domestic alternative. This goes beyond cost-cutting. It's a symbolic moment in the middle of the US-China chip war, showing just how far China's AI camp wants to push toward self-reliance.

So Who Exactly Is DeepSeek

DeepSeek is a Hangzhou-based AI startup. It originally spun out of High-Flyer, a quant hedge fund led by Liang Wenfeng. That origin story quietly matters. Hedge funds run massive GPU fleets to crunch market data, so DeepSeek started life already holding a substantial stash of Nvidia chips and real infrastructure know-how. This wasn't a broke startup begging for compute — it was a team that already had the hardware and the engineering chops in hand.

The company became globally famous thanks to its R1 model in early 2025. R1 claimed to approach the performance of top-tier models that US Big Tech had burned billions on — but at a fraction of the cost. The fun twist is what R1 was trained on: Nvidia H800 chips. The H800 was a deliberately throttled, China-only version Nvidia sold before late-2023 US export rules clamped down. So DeepSeek's legend was literally "look how much we did with the weaker chips we legally got around the rules with." That story alone cemented a corporate identity: we punch through hardware constraints with cleverness.

Then in April 2026, DeepSeek unveiled V4, its next-gen model — and that was the real pivot signal. V4-Pro is a 1.6-trillion-parameter mixture-of-experts model with a one-million-token context window, but the headline was that it was optimized specifically for Huawei's Ascend AI chips. The center of gravity shifted from Nvidia to Huawei. The pricing was insane, too. V4-Pro runs $3.48 per million output tokens, versus $30 for OpenAI and $25 for Anthropic on the same job. V4-Flash drops to just $0.28. It was a near-dumping price strategy designed to shake the market.

But this new chip story tells us that within months of cozying up to Huawei, DeepSeek was already doing the math on "we don't want to fully depend on Huawei either." It optimized V4 for Ascend up front, while quietly sketching its own inference chip in the back room. What you read here is a fierce determination to never let any single link in the supply chain hold a leash on the company.

What's Actually Happening

Here's the core of the Reuters reporting. DeepSeek is designing its own inference-only AI chip, and the work started about a year ago. It's early-stage right now, with the company holding talks with chip-design firms, contract manufacturers (foundries), and memory makers, while aggressively recruiting semiconductor engineers. DeepSeek had, in fact, been openly hiring chip-design talent since early 2025 — and at the time, industry watchers speculated this might signal plans for a proprietary processor. This report basically confirms that speculation was right.

The timing is uncanny. DeepSeek is currently pushing its first external funding round, targeting $7 billion in fresh capital at a valuation floated between $52 billion and $59 billion. This is a company that had long refused outside investment, now flipping its stance. Given that chip development is a cash-devouring beast, it's natural to read the funding round and the chip project as interlocked. Building your own silicon requires design tools, foundry bookings, secured memory, and enough runway to survive at least several years.

The problem is that the obstacles are enormous — and US export controls are aimed precisely here. The US restricts Chinese design houses from accessing advanced overseas foundries (like TSMC's latest processes) and blocks access to high-bandwidth memory (HBM), which is central to inference-chip performance. You can design a brilliant chip, but if you can't find a fab to print it on a cutting-edge node, or the memory to make it perform, it can die as a design on paper. So this news is less a "DeepSeek makes a chip" success story and more a "DeepSeek just started a very hard gamble."

Item Detail
What's being built Inference-only AI chip, not for training
When it started About a year ago (roughly mid-2025)
Current stage Early — in talks with design/foundry/memory firms, hiring engineers
Goal Cut dual dependence on Nvidia + Huawei Ascend
Existing infrastructure R1 trained on Nvidia H800; V4 optimized for Huawei Ascend
Core obstacle US limits on overseas advanced foundry and HBM access
Funding backdrop Pursuing $7B first round at $52–59B valuation
V4 pricing V4-Pro $3.48 per 1M output tokens, Flash $0.28

The table makes the arc obvious. DeepSeek is climbing a three-step ladder toward infrastructure self-reliance. Step one: get famous on Nvidia H800s secured before the rules hit. Step two: move V4 onto Huawei Ascend, switching to domestic silicon. And step three — right now — is trying to sever that last remaining dependence with its own chip. But step three is drastically harder than the first two. The earlier steps were about "using someone else's chip well"; this one is about "making the chip yourself."

Who Gets What Out of This

For DeepSeek itself, the gain is clear: control. If a real in-house inference chip materializes, DeepSeek can tailor its models to its own hardware down to the metal. Vertically integrating software and silicon lets you run the same performance cheaper and more efficiently. DeepSeek's core weapon has always been absurdly low prices — and if it can cut inference cost with its own hands, that price weapon gets even sharper. On top of that, if Nvidia supply gets choked by rules or Huawei Ascend hits a bottleneck, having its own chip frees DeepSeek from a big chunk of supply risk.

The Chinese government and the semiconductor self-reliance camp win big too. DeepSeek is the poster child of Chinese AI. If a company like that succeeds in localizing even its chips, that becomes a powerful political message: "No matter how hard America squeezes, China can build its own stack from software all the way down to silicon." It would put a question mark over the effectiveness of US export controls. Conversely, if the effort stalls against the HBM and foundry walls, that becomes evidence the controls are actually working.

An unexpected beneficiary is China's domestic foundry and memory industry. If DeepSeek can't tap advanced overseas processes, it'll ultimately have to lean on domestic fabs like SMIC and homegrown memory. When a mega-customer like DeepSeek pours volume and capital into the local supply chain, that itself can prime the pump for China's whole semiconductor ecosystem. DeepSeek's chip bid isn't a solo project — it's a stress test for the entire domestic silicon stack.

To be fair, look at who could lose, too. Nvidia has already seen its China revenue shrink under the rules, and the prospect of China's biggest AI startup defecting to its own chip is unwelcome over the long haul. Huawei's position is delicate. It built the "DeepSeek is on our side" picture by putting V4 on Ascend, yet DeepSeek is now trying to reduce its Huawei dependence too. It's a tangled relationship where cooperation and rivalry roll forward at the same time.

Past Attempts Like This — Wins and Flops

Building your own chip is a script Silicon Valley has run many times. The textbook success is Google. Starting in the mid-2010s, Google built its own AI chip — the TPU (Tensor Processing Unit) — and deployed it across its data centers. At first people asked, "Why is Google making chips?" But it ended up substantially cutting Nvidia dependence and optimizing its models to its own silicon. Amazon walked the same road with its Inferentia and Trainium chips. What these successes share: guaranteed volume, control over their own software stack, and free access to a cutting-edge foundry (TSMC). DeepSeek has the first two conditions, but gets snagged on the last one. That's exactly the decisive difference.

There are plenty of flops or slogs, too. Chip design mercilessly devours capital and time, and software companies that jumped in have often scrapped plans or slipped by years. First silicon rarely works right on the first try; it's common wisdom in this world that budgets balloon several times over through repeated respins (redesign and re-fabrication). Even companies that dominated in software get humbled by the physical constraints of hardware.

And the history of China's own semiconductor self-reliance is itself a drama of triumph mixed with frustration. Huawei somehow producing its Kirin and Ascend chips under US sanctions is a symbol of self-reliance — but along the way, blocked from cutting-edge processes, it kept slamming into walls on performance and yield. DeepSeek's path sits squarely on this same dilemma: "We can design it. The problem is where do we find a fab and the memory to print it at world-class quality." The lesson from history is clear — an in-house chip can be the right direction, yet reaching actual success means clearing a triple gate of money, time, and supply chain.

How Rivals Counter

The most directly affected party is Nvidia. From Nvidia's view, DeepSeek's own chip is one step toward the worst-case scenario: "AI that runs without us." But Nvidia's real moat isn't a single chip — it's the CUDA software ecosystem. Developers worldwide are locked into CUDA, so even if DeepSeek builds its own chip, it has to rebuild a whole software stack for that silicon from scratch. Nvidia will keep wielding this software lock-in, holding its overwhelming position outside China and playing the card that "you can copy the hardware, but you can't copy the ecosystem."

Huawei's counter-play is subtler. Huawei wants to keep DeepSeek pinned as the flagship customer of the Ascend ecosystem — that's exactly why it fully optimized V4 for Ascend. To stop DeepSeek from defecting to its own chip, Huawei has to make Ascend cheaper and better-running, convincing DeepSeek that "you don't need to build your own." From Huawei's angle, it's better to get DeepSeek to spend the money and time it would sink into a custom chip on Ascend optimization instead. It's a strange spot: being a partner while pre-emptively fencing in a would-be competitor.

Other Chinese chipmakers — the likes of Cambricon, Hygon, and Moore Threads — might actually see opportunity. DeepSeek building its own chip signals that the domestic inference-chip market is genuinely huge. They can partner with DeepSeek or fill the volume DeepSeek can't, splitting the pie. And zoom out beyond China: US leaders like OpenAI and Anthropic are already pushing their own chips to reduce Nvidia dependence. So "AI companies building their own silicon" is a global trend unfolding on both sides of the US-China divide simultaneously. DeepSeek is simply riding that wave with a Chinese edition.

So What Actually Changes

For everyday users, there's almost no immediate change to feel. A homegrown chip is years away at best, with no guarantee of success. But the direction matters. If DeepSeek lowers inference cost with its own hands, that can eventually translate into cheaper API prices and broader free or low-cost AI services. V4 already runs at roughly a tenth of US rivals' prices — bolt on a custom chip, and the "AI getting as cheap as water" trend could accelerate further. For consumers, that's not bad news over the long run.

For industry folks, especially those in AI infrastructure and semiconductors, this is a pretty loud signal. We're moving from an era where AI companies only build models into one where they vertically integrate everything from model to chip. "Which chip do we put our model on?" is giving way to "should we design our own chip?" as a serious option. Competition over chip-design talent, EDA tools, foundry capacity, and HBM supply will heat up — and there's a strong chance talent and capital shift dramatically toward China's domestic supply chain.

For investors and policy watchers, this is a thermometer for the US-China tech war. If DeepSeek's chip bid succeeds, it powers the paradox that "export controls actually accelerated Chinese self-reliance." If it fails, it becomes evidence the controls are biting. Either way, it inevitably reshapes the calculus of US chip policy toward China and of supply-chain keystones like Nvidia, TSMC, and SK Hynix. And with DeepSeek pursuing that $7 billion round at a $52–59 billion valuation right now, it's worth watching how the chip narrative reshapes the funding story. Just keep it cold-eyed: this is early-stage, and the HBM and foundry walls are very real.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Almost nothing right now. A custom chip is years out if it ever ships, and success isn't guaranteed. But if DeepSeek does cut its own inference costs, AI service prices can drop further — so over the long run, it could help make the AI you use cheaper.

— Why is this dropping right now of all times? It overlaps with DeepSeek pushing its first external funding round ($7B at a $52–59B valuation). Chip development is a cash-devouring beast, so it's natural to read the funding story and the chip project as interlocked. It also lands right after April's move to put V4 on Huawei Ascend — the moment a roadmap of "domestic chip first, our own chip next" came into view.

— So is it ahead of Nvidia and Huawei now? No, not at all. It's early-stage design with no prototype yet. And with the US blocking advanced foundry and HBM access, it's too soon to say whether a genuinely competitive chip will even emerge. The ambition is real, but success hinges on clearing that triple gate of money, time, and supply chain.

Sources

Numbers are as of announcement and may change.