China Update: Powerful Open-Weight Models, DeepSeek, Brain Implants

China (I): Z.AI GLM-5.2 Beats GPT-5.5 at One-Sixth of the Price

Z.ai (formerly Zhipu AI) has released GLM-5.2, an open-weights model with 753 billion parameters designed for autonomous coding and engineering tasks over long periods. The weights are available on Hugging Face under an unrestricted MIT license, along with a stable 1-million-token context window and coding plans starting at $12.60 per month. On SWE-bench Pro, the model achieves 62.1%, beating GPT-5.5 (58.6%); on FrontierSWE, it scores 74.4%, ahead of GPT-5.5 and nearly on par with Claude Opus 4.8. An architecture optimization called IndexShare shares an indexer across every four sparse-attention layers, reducing compute FLOPs per token at full context length by a factor of 2.9. The API costs $1.40 per million input tokens and $4.40 per million output tokens, significantly below US providers. The release comes in a week where the Trump administration, via an export directive, barred foreign users from accessing Anthropic Claude Fable 5, prompting Anthropic to take the affected models completely offline. → VentureBeat

Synthszr Take: In February, we wrote that China's AI offensive was continuing unabated; in May, cheap tokens were booming. Now, the calculation is changing for every engineering lead who previously reached for GPT-5.5 or Claude by reflex. A model that leads on SWE-bench Pro and costs one-sixth as much can no longer be dismissed with the argument that “proprietary is safer,” especially after Washington just demonstrated how quickly a US model can disappear by directive. This is exactly where the lock-in severity score pays off: an MIT-licensed open-weights model on your own compute has near-zero regulatory risk because no one can turn it off for you. The honest limitation remains: on Terminal-Bench 2.1, GLM-5.2's score of 81.0 still lags behind the 84 to 85 of the top models, and anyone embedding Chinese weights into a regulated codebase without an audit is acting negligently. But as an on-premise backup alongside the established Claude code stack, GLM-5.2 belongs in your own stack as of today. Anyone who takes “Reasonable Sovereignty” seriously will test this next week, not after the next AI offsite.

China (II): DeepSeek's $7 Billion Round Opens a New Front in the AI Race

Chinese AI provider DeepSeek has raised more than $7 billion in its very first funding round, valuing it at over $50 billion. Founder and CEO Liang Wenfeng is investing 20 billion yuan (around $3 billion) from his own pocket, nearly 40 percent of the total sum. Also participating are Tencent (10 billion yuan), battery manufacturer CATL (5 billion), along with JD.com, NetEase, and IDG Capital with 3 billion yuan each. The structure is unusual: external investors are not buying shares but are paying into a limited partnership led by Liang, with no voting rights and a five-year lock-up period. The only exception is China's state-owned National AI Industry Investment Fund, which receives voting rights and no lock-up period for 1 billion yuan. Since 2023, DeepSeek had been exclusively financed by Liang's hedge fund, High-Flyer, but is now changing course as Western export bans on Nvidia silicon and the compute hunger of AI agents force the need for fresh capital. For comparison: OpenAI has already raised $122 billion by 2026, Anthropic $65 billion. → Techpresso

Synthszr Take: In April, the talk was of $300 million; now it's $7 billion. Twenty times that amount in seven weeks, and the real driver is in the fine print. The narrative of a cheap open-weight model operating without commercial pressure no longer holds once agents require industrial-scale compute and frontier chips from the West are banned. With the limited partnership, Liang is building himself a money machine that leaves him in full control: Tencent, CATL, and the state fund provide oxygen but get no say. CATL's presence in the consortium is interesting—a battery maker that supplies energy and power systems for AI data centers. A domestic ecosystem of model, compute, and energy is de-fragmenting under state supervision, and the only investor with voting rights is flying Beijing's flag. Anyone who still reads DeepSeek as an underdog hasn't updated their market map.

China (III): First Brain Implant IPO and China's Health Insurers are Paying

On March 13, 2026, China's National Medical Products Administration (NMPA) approved the NEO-ONE SCI, a brain-computer interface from the Shanghai-based company Neuracle Technology (博睿康), founded in 2011 by two biomedical Ph.D. students from Tsinghua University. The device reads neural signals via electrodes placed outside the dura mater, without penetrating brain tissue; it is intended to allow paraplegics to regain their grip using a pneumatic glove. It is the world's first invasive brain-computer interface approved for commercial medical use. Three months later, on June 11, Neuracle filed for an IPO on the Shanghai Stock Exchange's STAR Market, aiming to raise the equivalent of $345 million, with 1.54 billion renminbi earmarked for BCI research alone. The prospectus, accompanied by CITIC Securities, is one of the most detailed financial disclosures seen in this industry to date. The catch is hidden in a single number: revenue in 2025 was 108 million renminbi (up 64 percent), but it came entirely from non-invasive EEG devices for clinics. China's National Healthcare Security Administration (国家医保局) has already published a guide for pricing neurological treatments, creating its own billing items for brain-computer interfaces: an “invasive BCI implantation fee,” an “invasive BCI removal fee,” and a “non-invasive BCI adjustment fee.” → Hello China Tech

Synthszr Take: While Neuralink is still navigating the FDA maze in the US, a Tsinghua spin-off founded in 2011 has become the first to receive commercial approval for an invasive implant. This is the same pattern we've seen with DeepSeek and the token economy: China industrializes a technology while the West is still debating it. The zero-revenue line for the implant would make any banker on Nasdaq nervous, but on the STAR Market, it's part of the plan, as the state is financing the path from the lab phase to production. The EEG business with 108 million renminbi is the oxygen keeping the expensive research alive, and it's this dual-track approach that makes the case interesting. Anyone who wants to spot hardware innovations early should now be looking at Shanghai stock exchange prospectuses and not just at Californian demo videos. The real race will be decided by who first moves from regulatory milestone to a repeatable procedure. The approval was the easy part.

China (IV): Alibaba is Building the AI Factory that Spans All Levels of the 'AI Stack'

Alibaba has introduced its first model series for robots, right in the middle of the industry's shift from chatbots to agents. At its core is RynnBrain, a system that allows machines to understand space, objects, and movement; in a demo by the DAMO Academy, a robot recognizes a piece of fruit and places it in a basket. Alongside it comes Qwen3.7-Max, which Alibaba claims can run autonomously for up to 35 hours without performance degradation (this figure comes from the company itself). Alibaba describes itself as the only company in China that operates all five levels of the “AI stack,” from chips and the agentic cloud to models, serving platforms, and applications. This is vertical integration as a moat: whoever owns every layer lets the profits of one layer ripple through all the others. Robotics is the most physical form of this bet; the agent moves from the screen to the warehouse and the home, similar to how an Nvidia-powered humanoid is already being tested in logistics at Siemens. Pricing, availability, and initial customers remain undisclosed. → Techpresso

Synthszr Take: The interesting number here is not the 35 hours, but the five layers. China is playing the robotics game from a position that a pure software company can hardly counter: a domestic model stack on a manufacturing base that is genuinely ahead in hardware and supply chain. In May, we wrote that Trump's export controls have strengthened rather than slowed China's AI; Alibaba is now drawing the logical conclusion and wiring chip, model, and machine into a single factory. The 35-hour claim should still be taken with a grain of salt, as the gap between a controlled demo and a reliable machine has already crushed many robotic dreams (a robot that drifts after a few hours is worthless in a warehouse). If you're serious about agents, you have to maintain them, otherwise they become orphaned—this applies to the warehouse as much as to the office. It will be interesting to see if Alibaba turns the demo into products; the vertical bet has been placed, and it's harder to copy than any chatbot.

China (V): In Image Ratings, China's AI is Overtaking the US in More and More Countries

A survey by the institute Public First of around 18,000 people in 15 countries shows a shift: in 11 of the surveyed countries, China is now considered the AI leader, not the US. In Germany, mistrust is the highest, with only 23% seeing the US ahead, compared to 46% for China. While 51% of Americans perceive themselves as leaders, the sentiment towards the technology in their own country is souring: the proportion of those who view AI positively for society fell from 39% (2024) to 31% (2026). A Gallup poll found that 7 out of 10 Americans oppose the construction of data centers in their communities. In China, by contrast, less than 10% are worried about job losses due to AI, according to a UCL study. In parallel, the performance gap between Chinese open-source models and US models has almost closed, at a significantly better cost, according to the Stanford AI Index Report. → The Deep View

Synthszr Take: In the technology market, perception is not a soft factor; it is the precursor to adoption. Back in March, we warned against China hysteria and recalled the Japan euphoria of the 80s. But the numbers are telling a different story now: China's labs are betting on efficient open-source models, while the US is locking its proprietary high-performance systems behind export barriers (Anthropic is no longer allowed to release its most powerful models to non-US citizens). When AI costs put pressure on companies and the cheaper model delivers just as well, the accounting department decides, not the flag. On top of that comes the real leverage that the US debate is sleeping on: 31% optimism for a technology you want to sell is a self-starter into irrelevance. If you don't like a technology yourself, you'll be bad at exporting it. This can't be turned around with a better benchmark, but only by people feeling a concrete benefit in their daily lives, and that's precisely where China's unique advantage currently lies.

SpaceX Acquires Cursor for $60 Billion

SpaceX is acquiring the AI coding startup Cursor (founded in 2022 as Anysphere) for $60 billion in stock, just days after the biggest IPO in history. The option has been on the table since April: either buy for $60 billion or pay a $10 billion break-up fee. The deal is intended to advance SpaceX's AI division, built around the integrated xAI, which is currently being rebuilt from the ground up after deepfake scandals and the Grok “MechaHitler” incident (all 11 co-founders of xAI are out, with Musk himself admitting it was “not built right”). Before the deal, Cursor was valued at around $29 billion and was about to raise $2 billion from a16z, Thrive, and Nvidia to reach a $50 billion valuation. SpaceX had sold investors on a total addressable market of $28 trillion, with $26 trillion related to AI. Since the IPO last Friday, the stock has jumped from $135 to over $200, adding almost a trillion in value or, in the currency of this deal, about 16 Cursors. → Techpresso

Synthszr Take: This is a failed AI lab being patched up with stock-based oxygen from an overheated IPO. xAI was practically dead after the scandals, and instead of fixing it, Musk is buying a working product with freshly printed paper. $60 billion for a company that wouldn't have even broken even with the $2 billion from its planned round: that's not a valuation, it's a bet that the market will keep playing along. In May, I wrote that the IPO was escaping the gravity of valuation models; now we see what you do with that weightlessness—namely, acquisitions that only work as long as the stock price rises. Cursor itself is a real asset (best IDE experience, up to 8 parallel agents, best-in-class completions), and that's precisely why it's a shame that it's ending up in a structure whose AI component has recently been known mainly for lawsuits. Anyone betting on coding agents in 2026 should take a close look at whose closed system their tool is now gravitating towards and keep an open-source fallback on hand. If the next trillion can be bought as easily as 16 Cursors, then the real question isn't whether the deal was expensive, but how long the flywheel will keep spinning.

ChatGPT Falls Below 50% Market Share for the First Time

Three and a half years after its launch, ChatGPT is still the most used AI assistant worldwide, but its lead is shrinking. According to Sensor Tower's State of AI Report 2026, the market share of OpenAI's chatbot fell to 46.4% at the end of May, after holding above 50% until January. The decline is driven by Google's Gemini (27.7%) and Anthropic's Claude (10.3%); Grok, Perplexity, DeepSeek, and Meta AI are each below 5%. In absolute numbers, ChatGPT remains in the lead with over 1.1 billion monthly users, followed by Gemini with 662 million and Claude with 245 million. ChatGPT was previously the fastest app in history to reach one billion monthly users. Users are now visibly switching between assistants. → Techpresso

Synthszr Take: The ChatGPT moment in November 2022 was a tear in reality; I was sitting in front of the screen myself and felt it. Now we're seeing what comes next: the first-mover bonus is decaying faster than valuations would have you believe. 46.4% sounds like a market leader, but the curve is pointing down, and Gemini is picking up the momentum because Google already owns the distribution (Android, Chrome, Workspace, the old PageRank empire). This was exactly Zuckerberg's calculus with Llama, and it's now working from the other side: as soon as model quality becomes a commodity, the winner isn't the best model, but the best contact point. OpenAI understood this in March with “Code Red” and the superapp idea, but a billion users are only a moat if they stick around. Losing twelve percentage points in four months with monthly measurable shares isn't a reach problem, it's a retention problem. The next round won't be decided by benchmarks, but by where the assistant is already built-in before the user even opens an app.

OpenAI Burns $34 Billion and Reports a $38.5 Billion Loss

Ed Zitron has seen audited financial documents from OpenAI, independently verified by the Financial Times, and the numbers are brutal. In 2025, OpenAI recorded a net loss attributable to the company of $38.53 billion, compared to $5.09 billion the previous year. On revenue of $13.07 billion, total costs were $34 billion, with research and development alone consuming $19.18 billion. SoftBank transferred $867 million, and Microsoft $303 million, while OpenAI paid $17.2 billion to Microsoft in the same year (of which $10.59 billion was for R&D, which likely means training costs). The conversion from a non-profit to a for-profit structure also introduced $41.55 billion to the bottom line through fair value changes. At year-end, OpenAI had about $50 billion in assets, with nearly half of that in cash. Zitron himself calls the situation “deeply concerning” and sees no clear path to profitability. → Techpresso

Synthszr Take: The loss increased nearly eightfold—that's the number that sticks with you. In April, we wrote that CFO Sarah Friar was warning of massive losses until 2026; now the audit is on the table, and it's worse than most had calculated. The accounting mechanics are interesting: a raw loss of $60.35 billion becomes $38.53 billion by moving $21.8 billion away as “noncontrolling members capital.” What remains is a business model where every dollar of revenue generates more than two dollars in costs, with the largest single item going back to Microsoft. The cash power, which was still considered a strength of the frontier bet in the code-crash framework, is here a combustion furnace with a $50 billion reserve and a run rate that will burn through it in a few years. Anyone betting on OpenAI as a platform standard today should have a migration path to Anthropic or open source in their risk register—not as a panic move, but as sober compute discipline. This scaling has to add up at some point, and so far, it doesn't.

Android 17 is Here, but Without Gemini Intelligence

Google is rolling out Android 17 starting today, first to Pixel devices, then to others later in the year. It includes features that were not shown at I/O 2026 or the Android show: Bubbles turns any app into a floating window, Screen Reactions records the display and selfie camera simultaneously, and the Foldable Gaming Mode on the Pixel 10 Pro Fold creates a 50/50 layout with a virtual gamepad. Security gets a boost with biometric “Mark as Lost” and granular location and contact sharing. In parallel, the June 2026 Pixel Drop is landing with Gemini Omni (video from text prompt), Lyria 3 (music by prompt), and Conversational Editing in five European markets, including Germany. Wear OS 7 also launches today, with Live Updates, cross-device control, and up to 10 percent more battery life. The real heavyweight, Gemini Intelligence, will arrive for phones and watches later this summer. → www.tomsguide.com

Synthszr Take: The most exciting feature is still missing. Bubbles and a gamepad for the foldable are nice tweaks, but what everything is leading up to, Gemini Intelligence, is being deliberately pushed back by Google. We know this pattern: first, push the OS integrated with AI, then turn every touchpoint, from the lock screen to the smartwatch, into an entry point for your own assistant. Conversational editing per language, video from a prompt, music from an image template: Google is gradually shifting creative work from the user to the models, and the smartphone is becoming the stage for it. What I like is the velocity. Four platform updates in one day, plus the monthly drop—that's applied compute discipline against the inertia of traditional release cycles. To compete here, you no longer build apps for an operating system, but experiences around an assistant that absorbs everything. By the end of summer, Android will no longer be the system you operate, but the one you talk to.

Why South Korea Loves AI

While discontent with AI is growing in the US, South Korea is emerging as perhaps the most optimistic country in the world: only 16 percent of South Koreans are more worried than excited, the lowest figure among 25 countries surveyed by Pew. In the US, it's 50 percent. A majority of Koreans use AI daily, both privately as an assistant and at work. This is supported by a national agenda: President Lee Jae-myung wants to bring the country into the “top three AI powers” alongside the US and China, has established a presidential council for AI strategy, and initiated a state-backed foundation model project. The foundation is provided by Samsung and SK Hynix, which produce the majority of the world's high-bandwidth memory chips for Nvidia hardware and are each worth over $1 trillion in 2026. As early as 2024, the parliament passed the AI Basic Act, which prioritizes speed over safety. → The Download from MIT Technology Review

Synthszr Take: South Korea's AI enthusiasm is orchestrated. The country has followed a clear recipe since the 70s: first steel and ships, then semiconductors, then broadband, then smartphones. Each stage was a state-orchestrated ascent from poverty, and each time, technology delivered. When you've seen the bet on the next technology pay off for fifty years, you don't ask about risks, you ask about the next use case. That's the real difference compared to Europe, where at the end of May we wrote about Denmark as a positive outlier: there, 42 percent of companies use AI, while here, concern dominates. Optimism isn't created by appeals, but by a history of fulfilled promises. As long as Germany invests its digitalization dividend mainly in concern-mongering, the gap with Seoul will remain as large and as growing as it is today.