Companies / Baidu (ERNIE)
Search platform and full-stack AI
Baidu (ERNIE)
State of play as of
The Chinese company with the most complete AI stack on paper, chip to model to robotaxi, and the one whose own earnings show what assembling it costs.
Baidu is the closest thing China has to a full-stack AI company: it designs its own accelerator through Kunlunxin, trains its own models under the ERNIE (文心) name, rents compute and model serving through Baidu AI Cloud, and operates the most widely deployed Chinese robotaxi service through Apollo Go. It is also the clearest place to read what that costs. Its second quarter of 2026 put both halves in one filing, a net profit down about 68 percent and a GPU cloud line up 283 percent, and the company told the same earnings call that ERNIE would return to the top tier of the industry, which is a target rather than an achievement and reads as an admission that it left. The two pieces most often quoted as Baidu wins, the Kunlunxin listing valuation and the Apollo Go overseas map, sit in units with their own pages and their own caveats. This page tracks the parent: what its filings actually say, where the AI revenue is and is not growing, how it is reorganising around model serving, and which of the widely quoted figures are disclosures rather than reports. Every figure carries its date and its source.
Fast facts
- What it is
- Baidu (百度), founded in 2000 and best known for China’s largest search engine. Its AI work runs through the ERNIE (文心) model family, Baidu AI Cloud and its Qianfan model-as-a-service platform, the Comate coding assistant, the Kunlunxin chip unit and the Apollo Go robotaxi service.CAD reporting
- Second quarter 2026
- Revenue of 31.3 billion yuan, down 4.4 percent, and net profit of 2.32 billion yuan, down 68.33 percent, with adjusted earnings of 7.22 yuan per ADS against an expected 9.75. The shares fell 5.1 percent pre-market on the print. The same filing reports GPU cloud revenue up 283 percent year on year, a fourth consecutive triple-digit quarter, and AI at more than half of core revenue for the second quarter running.Baidu Q2 2026 results, 18 Aug 2026, via ithome and Sina Finance
- Where the AI revenue actually is
- The split matters more than the total. AI Cloud infrastructure revenue was 7.3 billion yuan and grew 50 percent, with GPU cloud inside it up 283 percent. AI applications did 2.5 billion yuan and grew 3 percent. Baidu is growing fastest in the business of renting compute to other people’s models, and barely at all in selling its own applications.LatePost, 28 Aug 2026, citing second-quarter figures
- What it says about ERNIE
- On its 19 August 2026 earnings call Baidu said ERNIE would return to the top tier of the industry and that it targeted a primary Hong Kong listing by year end. The listing happened. The ERNIE claim remains a stated target and CAD carries no benchmark result establishing that it has been met.Baidu earnings call, 19 Aug 2026, via TechNode
- Listing status
- Dual-primary. Baidu converted its Hong Kong secondary listing to a primary listing on 1 September 2026, giving it primary status in Hong Kong alongside its Nasdaq listing.TechNode, 27 Aug 2026
- The units with their own pages
- Kunlunxin, the chip arm, filed a confidential Hong Kong carve-out application on 2 January 2026 and has been reported at a listing target near 50 billion dollars, larger than Baidu’s own market value at the time; only its 21 billion yuan round of December 2025 was a price anyone paid. Apollo Go, the robotaxi arm, reported 28 cities and 350 million cumulative autonomous kilometres in the second quarter, of which 240 million were fully driverless. Both figures sets are covered in full, with their caveats, on the unit pages.The Information, 29 Jun 2026; Baidu Q2 2026 results, via Sina Finance
The state of play
The Cyberspace Administration’s phase-two note on its Qinglang campaign against chaos in AI applications, published on 2 September, named Baidu twice and at two different layers. As a platform, alongside Douyin, Kuaishou, Weibo, Tencent, Bilibili, Xiaohongshu, Zhihu, Douban and Taobao, it is named as having expanded face, voiceprint and violation sample libraries. Separately ERNIE is named as one of four model products, with ByteDance’s Doubao, Tencent’s Yuanbao and Alibaba’s Qwen, told to guard the content-review pass on original training corpora, hard-limit non-compliant output and label AI-generated synthetic content so as to prevent illegal information at source. Being on both lists is the thing to notice: an obligation on a feed can be met by sweeping it afterwards, and an obligation on a pre-training corpus has to be met before the model exists.
Baidu’s chief financial officer said AI could soon match the search business in profit and cash returns. Read it against the segment split the company reported six weeks earlier, where the growth is in AI Cloud infrastructure rather than in AI applications: the claim is about the compute-renting business reaching the economics of search, not about ERNIE doing so. It is a forward-looking statement by the company about its own business and no figure was attached to it.
Baidu broke up the division that held its model-serving, data and agent businesses, according to an exclusive in LatePost. Qianfan and the rest of the model-as-a-service stack move into the infrastructure unit under Hou Zhenyu, general agents become their own unit under Yin Dawei, and the data platform goes to the applications unit under Ruan Yu, with all three reporting to Shen Dou. The reasoning given to LatePost by someone close to the change was that model-as-a-service is essentially infrastructure, and the quarter’s numbers say the same thing: AI Cloud infrastructure revenue of 7.3 billion yuan growing 50 percent, with GPU cloud up 283 percent, against AI applications at 2.5 billion growing 3 percent. This is one outlet’s sourcing on an internal reorganisation rather than a company announcement.
Baidu converted its Hong Kong secondary listing to a primary listing on 1 September, giving it dual-primary status in Hong Kong and on Nasdaq. It arrived inside a fortnight in which Changxin Memory, DapuStor, Moore Threads and Enflame all moved on domestic or Hong Kong capital formation, and it delivers the listing plan Baidu had named on its own earnings call twelve days earlier.
Baidu told its second-quarter earnings call that ERNIE would return to the top tier of the industry and that it planned to complete a primary Hong Kong listing by the end of the year. Quarterly revenue was 31.3 billion yuan, AI businesses were more than half of it for the second quarter running, and GPU cloud revenue grew 283 percent year on year. A company saying out loud that it intends to get back to the top tier is a company confirming that it left, and the GPU cloud line is the one to hold onto: Baidu is growing fastest in the business of renting compute to other people’s models.
Baidu reported two quarters at once. Net profit came in at 2.32 billion yuan, down 68.33 percent, on revenue of 31.3 billion, down 4.4 percent, with adjusted earnings of 7.22 yuan per ADS against an expected 9.75, and the stock fell 5.1 percent pre-market. The same filing reported GPU cloud revenue up 283 percent, a fourth consecutive triple-digit quarter, AI at more than half of core revenue for the second quarter running, and a robotaxi arm at 28 cities and 350 million cumulative autonomous kilometres, of which 240 million were fully driverless. English coverage that day led on the profit line. Both things are in the same document, and reading either alone gets Baidu wrong.
Kunlunxin, Baidu’s chip unit, was reported to be targeting a Hong Kong listing at a valuation near 50 billion dollars, higher than Baidu’s own market value at the time, with investors reportedly asked to commit to buying chips worth three to seven times the value of their intended subscription. The detail that matters is the subscription condition rather than the headline number, because tying share allocation to chip purchases makes demand for the stock and demand for the product hard to tell apart. It is also a target rather than a price: the only figure anyone has paid is the 21 billion yuan round Reuters reported in December 2025, and the Hong Kong application is confidential, so none of the quoted valuations is filed.
Our reporting on Baidu (ERNIE)
- Doubao, Qwen and ERNIE Now Answer for Their Training Corpus
The Cyberspace Administration published a phase-two note on its Qinglang campaign against chaos in AI applications on 2 September, and it places obligations at three points in the pipeline rather than one. Platforms including Douyin, Weibo, Tencent and Xiaohongshu are named for expanding face, voiceprint and violation sample libraries, with 46 governance announcements published between them. Four model products, Doubao, Yuanbao, Qwen and ERNIE, are told to guard the content review pass on original training corpora and to label AI-generated synthetic content so as to prevent illegal information at source, which is an obligation on the pre-training set rather than on a feed. The third layer is distribution: Huawei, Xiaomi, OPPO and vivo are named as setting developer admission standards for AI applications and spot re-testing apps already listed, so four handset makers become the admission gate. The note landed one year and one day after the labelling Measures took effect on 1 September 2025. No published standard, audit or methodology sits behind the corpus clause, so what changed this week is scope rather than enforcement capability.
- Zhipu Grew Revenue 400% and Its Gross Margin Halved
Z.AI, the entity Zhipu lists under in Hong Kong at stock code 2513, filed interim results on 31 August 2026 covering the six months to 30 June. Revenue was 953.892 million yuan, up 399.7 percent, while gross margin fell from 50.0 percent to 26.4 percent. The move is mix rather than discounting: open-platform and API revenue was 825.176 million yuan, up 2,735.7 percent, and went from 15.2 percent of revenue to 86.5 percent, while on-premises deployment fell from 73.7 percent of 2025 revenue to 13.5 percent of this half. Token volume on the MaaS platform is up more than 40 times since the start of 2026 and the average API selling price rose about 101 percent over the same stretch, so volumes surged while prices went up. That leaves the 24.6 percent gross margin on the open-platform and API business a cost problem, not a pricing one.
- Six Percent of Moore Threads Actually Trades
Biren’s first interim report as a listed company showed 1.2355 billion yuan of half-year revenue, up 1,997.6 percent, with a net loss 76.4 percent narrower than a year earlier. The number that decides the story is the float. Moore Threads carries 250.94 billion yuan of market value on 6.43 percent of its shares actually trading, and MetaX 270.03 billion yuan on 4.63 percent. Two Shanghai listings hold 520.97 billion yuan of combined market value, about 77.4 billion dollars, priced each day by 28.6 billion yuan of tradable stock. Biren, listed in Hong Kong with just over half its shares in public hands, trades at about 71 times half-year revenue against Moore Threads at about 144.
- Changxin Stopped Asking for Money
Changxin Technology, the listed parent of CXMT, filed first-half revenue of 150.31 billion yuan, up 873.64 percent, and net profit attributable to shareholders of 77.605 billion yuan against a 2.332 billion yuan loss in the same half of 2025. Gross margin was 84.74 percent, against 13.00 percent a year earlier. The line that decides what it means is the cash flow statement: operating activities threw off 131.156 billion yuan of net cash while financing contributed 1.002 billion, so the company covered a 35.577 billion yuan investing programme nearly four times over out of what it sold. Across all of 2025 it had taken 46.121 billion yuan net from financing.
- The Paying Tier
Chinese cloud vendors are reselling competitors’ models at roughly 30 percent of list price, below their own cost, because in AI coding the customer names the model. Leiphone reported the discount ladder on 21 August from named cloud sales sources: 60 to 80 percent of list six months ago, 48 percent now common, roughly 30 percent at the low end. The same day, Approaching.AI vice-president Guan Jiawei told the South China Morning Post that domestic chips can only serve the low-quality inference tier, the one with weak demand and weak monetisation, and that high-quality tokens, especially coding, still depend on Nvidia. iFlytek said the same week that a flagship model trained entirely on domestic compute would reach the domestic first tier on code capability.
- The Collateral
A Guangzhou district launched Token贷, a Bank of China loan that underwrites AI companies on tokens burned rather than property owned, with lines to 30 million yuan on four non-property inputs. The trial tranche is about 28 million yuan, but CITIC Bank and Bank of Guangzhou have followed, and Beijing’s E-Town and Anhui have written token lending into policy. The National Data Administration puts national token consumption above 30 trillion a day as of end-June.
These link to the full issues on the newsletter. New pieces go out in the daily first.
Common questions
What is Baidu’s AI strategy?
Own every layer and sell the one that is growing. Baidu designs its own accelerator through Kunlunxin, trains the ERNIE (文心) model family, sells compute and model serving through Baidu AI Cloud and its Qianfan platform, and runs robotaxis through Apollo Go. The second-quarter 2026 segment split shows which layer is working: AI Cloud infrastructure revenue of 7.3 billion yuan growing 50 percent, with GPU cloud inside it up 283 percent, against AI applications at 2.5 billion growing 3 percent. In August 2026 it reorganised to match, moving the Qianfan model-as-a-service stack into the infrastructure unit on the reasoning, given to LatePost, that model-as-a-service is essentially infrastructure.
Is Baidu making money from AI?
It is growing AI revenue and shrinking profit at the same time, and the second quarter of 2026 is the clearest example. Revenue was 31.3 billion yuan, down 4.4 percent, and net profit 2.32 billion yuan, down 68.33 percent, while GPU cloud revenue grew 283 percent year on year and AI passed half of core revenue for the second quarter running. On 2 September 2026 Baidu’s chief financial officer said AI could soon match the search business in profit and cash returns, which is a forward-looking statement by the company with no figure attached, and it is a claim about the compute-renting business rather than about ERNIE.
Where does ERNIE stand against other Chinese models?
Baidu itself says it has work to do. On its 19 August 2026 earnings call the company said ERNIE would return to the top tier of the industry, which is a target rather than a result, and a company that says it intends to get back to the top tier is confirming that it left. CAD carries no benchmark result establishing that ERNIE has returned. What is on the record more recently is regulatory rather than competitive: the Cyberspace Administration’s 2 September 2026 Qinglang note named ERNIE as one of four model products, with Doubao, Yuanbao and Qwen, made responsible for the content-review pass on their original training corpora.
What chips does Baidu use?
Increasingly its own. Kunlunxin, Baidu’s in-house chip unit founded in 2012, makes the third-generation P800, which handles the majority of Baidu’s inference and search-retrieval work and serves the ERNIE models, reported at about 345 TFLOPS at 16-bit floating point. Baidu said in April 2025 that it had brought up a 30,000-card P800 cluster, a company claim that has not been independently verified. Kunlunxin filed a confidential Hong Kong carve-out application on 2 January 2026 and began STAR Market listing guidance on 7 May 2026; the valuations quoted for it range from a paid 21 billion yuan round to a reported target near 50 billion dollars, and none of them is a filed figure.
Why is Baidu listed in both Hong Kong and New York?
It converted its Hong Kong secondary listing to a primary listing on 1 September 2026, so it now holds dual-primary status in Hong Kong and on Nasdaq. Baidu had named the plan on its 19 August 2026 earnings call, targeting completion by year end, and delivered it inside a fortnight in which Changxin Memory, DapuStor, Moore Threads and Enflame all moved on domestic or Hong Kong capital formation.
Do Chinese developers use Baidu’s coding assistant?
The signal CAD carries is not flattering, and it is one reviewer’s opinion rather than a benchmark. A Bilibili creator ranking coding tools for a Chinese developer audience grouped Baidu’s Comate with Alibaba’s Tongyi Lingma and Zhipu’s CodeGeeX as the early domestic plug-ins and put all three in his second-lowest tier, with the note that they have made no waves since 2025 and their presence keeps shrinking against Cursor and Claude Code. In a later ranking he gave Comate the Chinese idiom for getting up early and still arriving late, meaning every capability present and every one of them average. Bilibili items carry no publish date, so treat these as recent reviewer opinion of unknown exact vintage and not as a market-share measurement.
Every dated fact above traces to a named source, and where it comes from our reporting, to the dated issue that carried it. Where a claim is self-reported or disputed, we say so.