DeepSeek’s decision to design its own AI chip places the Chinese startup alongside Amazon, Google, OpenAI, Huawei, Alibaba and Baidu in a growing race to own every layer of the AI stack, from silicon and compute infrastructure to the models themselves. The result is a structural shift that could redefine who builds and benefits from artificial intelligence. As frontier labs increasingly control the chips, cloud infrastructure and models that power AI, developers risk becoming locked into proprietary ecosystems with fewer alternatives and higher switching costs. DeepSeek’s chip project suggests that the world’s biggest AI powers, despite geopolitical rivalry, are converging on the same conclusion: the future does not only belong to the company with the best model, but also the one that owns the entire stack, raising fresh questions about whether the era of open, hardware-agnostic AI is drawing to a close.
Chinese artificial intelligence company DeepSeek is designing its own chip for the first time. The chip is built for inference, the stage of computing in which an already trained model answers a user’s question, rather than for training new models from scratch. The company has been quietly hiring chip engineers and holding talks with design, foundry and memory partners for roughly a year.
The timing is notable. DeepSeek, the company that rattled Silicon Valley and Washington in early 2025 with a cut price reasoning model, has spent most of its existence avoiding outside investors and staying narrowly focused on model research rather than commercialization.
That is changing on two fronts at once. The company is simultaneously raising $7 billion at a valuation of $52 billion to $59 billion, its first outside capital, and moving into silicon design to cut its dependence on Nvidia and Huawei chips. Both moves point toward the same goal: control over its own destiny at every layer of the stack, from capital to compute.
Jack Collier, chief growth and marketing officer at io.net, tells Impact Newswire that the DeepSeek news is not an isolated data point. It is confirmation of a pattern he has been tracking across both sides of the Pacific. “The fact that DeepSeek is now building its own chips tells you everything you need to know about where the AI industry is heading, and it should concern anyone who cares about open access to AI,” he says.
“We’re watching a pattern emerge. Amazon is building custom chips. Google is building custom chips. Now DeepSeek. The US and China are taking very different approaches to AI development, but they’re arriving at the same destination: fully vertically integrated tech stacks where the companies building the models also control the silicon underneath them. Sometimes that works well, sometimes it doesn’t, but either way you’re locked in completely.
Nobody’s stopping to ask what happens when every major AI player on both sides of the world owns the full stack from chip to model. The compute shortage is real, but the answer shouldn’t be walled gardens, whether they’re built in San Francisco or Hangzhou. It should be open, decentralized compute infrastructure that any builder, anywhere, can access without being trapped inside someone else’s ecosystem.”
— Jack Collier, Chief Growth & Marketing Officer, io.net
The pattern, in numbers
Collier’s claim holds up against the record. Amazon’s custom silicon business, which includes its Trainium AI accelerators, has crossed a $20 billion annual revenue run rate and is growing at a triple digit pace, chief executive Andy Jassy told investors in April. Amazon now holds over $225 billion in Trainium revenue commitments, including up to five gigawatts of capacity reserved by Anthropic and roughly two gigawatts by OpenAI. Google, meanwhile, has pushed its own Tensor Processing Units into the millions of units a year and has begun selling TPUs to outside customers for the first time, mirroring Amazon’s move.
The custom silicon trend is no longer confined to cloud landlords. OpenAI unveiled its own chip, code named Jalapeño and built with Broadcom, last month. Anthropic has been weighing its own chips as well. In China, Alibaba and Baidu are already fielding in house chips, and Huawei supplies roughly half of the country’s $50 billion domestic AI chip market despite lagging Nvidia’s most advanced hardware. DeepSeek’s move puts it on the same path as nearly every frontier lab on either side of the export control line.
| Why compute is scarce enough to justify owning the chip • Global AI compute supply is falling 15 to 30 percent short of demand in 2026, a gap expected to last into 2027, per J.P. Morgan and industry estimates. • Orders for Nvidia GPUs total roughly $1 trillion through 2027, double the prior year, with lead times stretching toward twelve months. • All three major high bandwidth memory suppliers are sold out for 2026, and U.S. data center vacancy sits at just 1 percent. • Amazon’s Trainium revenue commitments now exceed $225 billion, including up to 5 gigawatts reserved by Anthropic and 2 gigawatts by OpenAI. • Google’s TPU shipment target for 2026 was cut from 4 million to roughly 3 million units due to packaging and memory constraints. |
The scramble for chips traces back to a physical bottleneck. J.P. Morgan research puts total orders for Nvidia GPUs at $1 trillion through 2027, double what they were a year earlier, with lead times across GPUs and custom silicon stretching toward a year. All three major memory suppliers are sold out of high bandwidth memory for 2026, and data center vacancy sits at just 1 percent for a second straight year. Apollo’s chief economist Torsten Slok has described 2026 as the year compute itself, not the model or the algorithm, became the scarcest resource in the industry, with GPUs, memory, and power all constrained at once.
The case for the walled garden
Not every analyst reads vertical integration as a warning sign. For companies burning through billions on inference, owning the chip is simply good arithmetic. Google’s TPUs are estimated to cut cloud costs by 20 to 30 percent internally, and Amazon says its Trainium2 chips deliver roughly 30 percent better price performance than comparable graphics processors, with Trainium3 improving on that by another 30 to 40 percent.
A Foreign Affairs Forum analysis of the custom chip race argues that well funded labs will keep chasing these efficiency gains regardless of the concentration it creates, leaving smaller startups dependent on off the shelf hardware or cloud rentals.
There are also real doubts about whether DeepSeek’s chip ambitions will amount to much outside China’s borders. “Nvidia is at zero in China and staying there. DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading edge manufacturing,” Richard Windsor, an analyst at Radio Free Mobile, says, noting the project does not threaten Nvidia’s position. U.S. export controls already bar Chinese firms from the most advanced overseas foundries and have cut off access to high bandwidth memory, meaning DeepSeek’s chip, even if it succeeds, is likely to stay a domestic tool rather than a global one.
An open alternative, still small
Collier’s proposed fix, decentralized compute assembled from idle GPUs scattered across data centers, crypto mining rigs and consumer machines, exists today but remains a rounding error next to the hyperscalers. Networks in this category claim cost savings of 60 to 90 percent against AWS and Azure pricing for certain workloads, and the broader decentralized AI market is forecast to grow from $9 billion in 2024 to $22 billion by 2035, according to Research and Markets. io.net says its network can reach more than 100,000 GPU devices, and rival projects such as Prime Intellect and the Bittensor subnet Templar have run distributed training experiments on models with tens of billions of parameters.
Those figures are real, but so is the gap between them and the sums hyperscalers are spending. Amazon’s chip revenue commitments alone, at $225 billion, dwarf the entire projected decentralized compute market a decade from now. The open compute movement is a genuine hedge against lock in, not yet a replacement for it.
What is actually at stake
What DeepSeek’s chip project ultimately confirms is less about any single company and more about a structural convergence. American labs are racing to own silicon so they are never again caught short by a supply crunch or an export ban. Chinese labs are doing the same for the opposite reason, because export bans have already been imposed on them. Both paths lead to the same place: fewer, larger, more self contained stacks in which the company that trains the model also owns the chip beneath it, the data center around it and, increasingly, the power plant behind that.
Whether that convergence is a genuine threat to open access, as Collier argues, or simply the rational endpoint of an industry with a real, physical compute shortage, as the vertical integration camp argues, may depend less on ideology than on who ends up building the next generation of frontier models, and how many of them there ultimately are left to build.
Stay ahead of the stories shaping our world. Subscribe to Impact Newswire for timely, curated insights on global tech, business, and innovation all in one place.
Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.
Got a story to share? Pitch it to us at info@impactnews-wire.com and reach the right audience worldwide
Discover more from Impact AI News
Subscribe to get the latest posts sent to your email.
