For most of the last decade, the story of artificial intelligence had a single geography. The labs that mattered, such as OpenAI, Google DeepMind, Anthropic, Meta AI, sat within a few miles of each other in the Bay Area. The capital that mattered came from the same handful of venture capital funds. The compute that mattered came from one company, Nvidia, selling into one customer base. It was a tidy story. And that story is changing fast.
Sometime around January 2025, a relatively unknown Hangzhou lab called DeepSeek, spun out of a quantitative hedge fund called High-Flyer, released a reasoning model called R1 that performed close to the American frontier at a training cost that made Western AI labs visibly uncomfortable. Nvidia lost roughly $600 billion in market value in a single day. Investors who had priced AI as a capital-intensive, US-controlled category had to reprice it overnight. Commentators reached for the obvious historical comparison and called it China's Sputnik moment for AI.
Even more interesting, eighteen months on, is that DeepSeek did not stay a one-off case. It increasingly appears like a visible tip of an entire industrial base that had been building underneath it for years, that includes chipmakers, cloud platforms, model labs, and a policy regimen to support all three simultaneously.
Understanding Chinese AI today means understanding that base, not just the one company that made us notice it.
The mistake most outside observers made in early 2025 was treating DeepSeek as an anomaly. A single brilliant team got lucky with an efficient architecture. Eighteen months later, that framing no longer appears accurate.
By mid-2026, China's model landscape looks less like a one-company story and more like a crowded front line, with Alibaba's Qwen, Moonshot AI's Kimi, ByteDance's Doubao, Zhipu's GLM, and Baidu's ERNIE all competing hard for different slices of the same market.
The industry itself has developed nicknames for the layers. Chinese media and investors refer to six leading independent model startups that include Zhipu, Moonshot, MiniMax, Baichuan, StepFun, and 01.AI as the Six Tigers, with DeepSeek usually mentioned in the same breath but sitting slightly apart, since it has prioritized open-weight research efficiency over building a consumer app.
A newer, tighter label, the Four Dragons, groups the four highest-valued independents (DeepSeek, Zhipu, MiniMax, and Moonshot), whose combined valuations reportedly crossed one trillion yuan, roughly $140 billion, by early 2026.
Above the startups sit the platform giants who already owned the distribution: Alibaba, ByteDance, Baidu, and Tencent.
Alibaba folded Qwen into DingTalk and its e-commerce stack and pivoted Alibaba Cloud from plain infrastructure-as-a-service into an AI platform business.
Baidu pushed ERNIE into search and enterprise tools while quietly running fully driverless robotaxi operations in several Chinese cities.
This may appear a familiar pattern if you have been following platform companies elsewhere. The startups innovate at the model layer, and the giants absorb the winning ideas into distribution channels that already reach hundreds of millions of people.
The most recent entrant to rattle Western coverage is Moonshot AI's Kimi K3, released in July 2026. Demand for it was so high that Moonshot suspended new subscriptions within two days of launch because it could not keep up with the computing load. This is an important development. A startup nobody outside AI circles had heard of a year earlier was turning away paying customers because it had built something too good, too fast. It not only tells about Moonshot, but it also offers a hint about the depth of China’s AI industry.
The headline number that made DeepSeek famous was cost. R1 was reported to have been trained for a fraction of comparable Western frontier models' cost, and it forced investors, developers, and cloud companies to rethink the underlying economics of AI almost overnight.
DeepSeek did not just prove Chinese labs could build a good model. It proved that architectural efficiency, doing more with the same compute, or the same result with less, was a legitimate competitive axis, not a consolation prize for labs priced out of the chip market.
That efficiency-first posture has become the house style across the Chinese AI industry, and it has been reinforced by the constraint that made it necessary in the first place: chip access.
The United States has spent years tightening export controls on advanced AI chips to China, aiming to slow Chinese frontier AI development by starving it of Nvidia's best silicon. The results, eighteen months on, are more complicated than that goal suggests.
Nvidia's own CEO has acknowledged the company's China market share fell effectively to zero at points during the restriction period. Depending on which analyst estimate you trust, Nvidia's share of China's domestic AI-chip market by mid-2026 sits somewhere between single digits and roughly 55 percent, down from a figure north of 90 percent just a few years earlier; the range itself tells you how fluid and contested this data is.
At the same time, Huawei's Ascend chip line has gone from a laggard product line to, by some estimates, roughly half of China's domestic AI-chip market, with Cambricon, Alibaba's in-house T-Head chips, and Baidu's Kunlunxin filling out the rest of a multi-vendor domestic stack.
Huawei introduced its Ascend 950PR chip publicly in March 2026, claiming roughly 2.87 times the compute of Nvidia's export-compliant H20 chip at FP4 precision, a comparison that matters less for bragging rights and more because it signals China no longer needs to design around an artificially constrained baseline chip.
There is real disagreement about how far this substitution actually goes. The Council on Foreign Relations estimated that even at a generous 800,000-unit production figure, Huawei's 2025 Ascend output amounted to just 5.3 percent of Nvidia's total chip processing power that year, a reminder that domestic chip narratives can run ahead of domestic chip reality.
And DeepSeek itself, in evaluating Huawei's hardware, reportedly found the Ascend 910C unattractive for training models but usable for inference, delivering around 60 percent of an Nvidia H100's inference performance, a distinction that matters because Barclays estimates roughly 70 percent of AI compute demand going forward will be inference, not training.
In other words: China's chip substitution may be strongest exactly where the future demand curve is heading, even if it still lags the frontier for training the largest models.
The point being export controls did not stop Chinese AI. They appear to have accelerated Chinese self-sufficiency in the layer of the stack, chips, that the controls were designed to protect.
A senior Trump administration Commerce official told Congress in mid-2025 that Huawei would be capped at roughly 200,000 Ascend chips that year. A year later, credible estimates for 2026 Huawei-plus-SMIC Ascend production sit at close to a million units and roughly a 50 percent share of the Chinese chip market.
Whatever the controls were meant to achieve, the trajectory of the numbers tells its own story.
If DeepSeek's contribution was proving cost-efficiency at the frontier, Alibaba's contribution has been proving that open-sourcing your model can be a distribution strategy rather than a charity act.
Qwen, Alibaba Cloud's model family, crossed 700 million cumulative downloads on Hugging Face by January 2026, overtaking Meta's Llama as the most downloaded open-weight model family in the world, and spawning more than 180,000 derivative fine-tunes built by developers who owe Alibaba nothing but bandwidth.
Alibaba Cloud makes its money from compute and API access, not from licensing model weights, so giving away increasingly capable open models under a permissive Apache 2.0 license costs Alibaba little and buys it something far more valuable: it makes Qwen the default starting point for teams building AI products anywhere Meta's Llama used to be the default.
A developer in Lagos, Jakarta, or Dhaka fine-tuning a local-language model today is more likely than not to be starting from a Qwen checkpoint, not an American one.
That is soft power delivered through a model registry, and it is compounding quietly while most coverage stays fixated on chatbot benchmark leaderboards.
The scale of the shift, by one measure, is startling. Chinese open-weight models reportedly went from roughly 1 percent to about 15 percent of global model share in under a year, with Baidu going from zero to over a hundred model releases on Hugging Face, and ByteDance and Tencent both seeing eight to nine times growth in their open-model footprints over the same window.
This is not a story about one good model. It is a story about an entire industry deciding, more or less simultaneously, that openness was the fastest route to global relevance, a strategic bet that Western labs, still largely wedded to closed, API-gated frontier models, have been slower to make.
We keep coming back to why this matters for a reader sitting in Dhaka. It is not abstract geopolitics. It is a direct change in what building an AI product costs and who you depend on to build it.
For years, "using AI" in a Bangladeshi startup meant renting access to an American model through an API, priced in dollars, with no real domestic or regional substitute. That dependency is loosening.
Open-weight Chinese models that are competitive with, sometimes better than, the equivalent-sized Western open models are now the default building block for a large share of the world's AI developers outside the US and China itself.
Cheaper inference, chip diversification that will eventually filter into cheaper cloud pricing globally, and a genuine alternative to a single-vendor AI stack are not developments that stay contained within China's borders. They reset the baseline cost and the baseline dependency for everyone building on top of AI, including founders in markets that have never had a seat at any of these tables.
That dependency question is not just theoretical for the people actually running export-facing businesses out of Bangladesh. Sheikh Shourav, founder of Apploye, made the practical version of this argument in a recent Facebook post aimed at Bangladeshi tech founders, freelancers, and service exporters.
His framing: today's tech world effectively runs on two parallel ecosystems, one controlled by the United States and one built by China, and founders exporting out of Bangladesh don't have to pick a side; they can adopt both as a hedge. He offers several examples. In payment, he said, a Bangladeshi company trying to receive payment from abroad typically needs PayPal, Stripe, or a foreign bank, none of which support Bangladesh directly, which pushes many founders to incorporate in the US or UK anyway. Even that workaround doesn't fully protect them; Shourav points to Mercury Bank recently closing accounts belonging to Bangladeshi residents, even when the underlying company was American.
Chinese e-commerce sellers face the same structural distance from Western payment rails, he notes, but they never hit the same wall, because they built and now lean on their own parallel stack to transact with Western customers without any visible difference in the customer's experience, and, notably, many of those Chinese tools already support onboarding Bangladeshi companies directly.
He lays the two stacks out side by side, tool for tool: Stripe and Braintree against Lianlianpay and Oceanpayment; Payoneer and Wise against WorldFirst, Currenxie, and PingPongX; PayPal against Lianlianpay; AWS and Google Cloud against Alibaba Cloud and Tencent Cloud; Slack and Teams against Lark, WeChat Work, and DingTalk; and, most relevant to this piece, Claude and ChatGPT against Kimi, Qwen, and DeepSeek.
His conclusion for founders here is that you treat both superpowers as infrastructure you can build on, not as a binary allegiance, and you buy yourself a backup that founders locked into only one ecosystem don't have.
It's a useful corrective to how this conversation usually gets framed in Dhaka, which is almost entirely through the lens of which chatbot answers questions better.
Shourav's point is that the more consequential decision is happening one layer down, in the rails a business actually runs on, and that Chinese AI models are just one item on a much longer list of Chinese-built infrastructure that Bangladeshi founders are already being pushed toward, whether they've noticed it or not.
It would be a mistake to read the last eighteen months as China having "won" the AI race. The frontier training gap with the very best American models has narrowed, not closed, and the chip constraint, while less binding than it was, is still real.
However, there has been a shift in belief, inside China's AI industry and increasingly outside it, that self-sufficiency is achievable rather than aspirational.
China now believes in its own self-sufficiency and supply capabilities, an Omdia semiconductor analyst put it, and that belief is now backed by shipping chips, shipping models, and shipping products at a pace few predicted in early 2025.
The next surprise, several analysts covering this space now agree, is unlikely to come from one obvious leader the way DeepSeek's R1 did. It is more likely to come from the sheer breadth of labs now capable of producing one, which is, in its own way, the more interesting story.
