After the Banquet
Happy Sunday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: what Trump's Beijing visit didn't change about the AI race, China's quantum computing team just published a result that makes the word "faster" feel inadequate, Shanghai Telecom is now selling AI compute like a data plan, and DeepSeek employees say V4 is the model they actually use to get work done.
Let's go.
After the Banquet
Donald Trump spent two days in Beijing. Temple of Heaven, Zhongnanhai garden, Great Hall state banquet, Xi promising rose seeds for the White House garden. The Guardian described it as "rich in pageantry and promises of stability, but little by way of tangible progress." Trump left claiming "fantastic trade deals" -- Boeing jets, soybeans, oil -- while Chinese officials confirmed none of it specifically.
On AI: nothing. No governance framework. No chip deal. No emergency communication channel. Jensen Huang was on the trip but not in the strategic meetings. AmCham China's president, debriefing from Washington during the same week, told Forbes that DC conversations were primarily about AI export controls and chip restrictions -- the exact things the summit didn't address. The hope in Washington was some framework around preventing non-state actors from using AI for terrorism. Whether that will happen in a communique, nobody knows.
What the visit actually accomplished was simpler. Tencent, Alibaba, and ByteDance had been experiencing delays in mid-level government meetings in China, caused by the Iran war awkwardness and the general US-China chill. A successful summit between the two presidents resets that. Business can move again. That's the real deliverable, and it's not nothing -- but it's also not an AI story.
The AI story continued independently. Tencent's earnings call last week confirmed that domestic chip supply is ramping "month by month," with a "substantial increase" in capital expenditure coming in the second half. The chips that Washington cleared for sale -- the H200, still sitting at 75,000 units per approved buyer -- remain unsold. Beijing told the approved companies not to buy them. That story hasn't changed.
What has changed, underneath all of it, is the quantum computing picture.
The Briefing
China's quantum computing team just published a result in Nature that makes every prior record look incremental. The team at the University of Science and Technology of China, led by Pan Jianwei and Lu Chaoyang, announced 九章四号 (Jiuzhang-4) on May 13: a photonic quantum computing prototype with 1,024 compressed quantum state inputs, 8,176 modes, and the ability to manipulate and detect up to 3,050 photons. The benchmark task is Gaussian boson sampling, a problem with no classical shortcut. Jiuzhang-4 solves it 10 to the power of 54 times faster than the world's fastest supercomputer. That exponent is not a typo. This is not a general-purpose quantum computer -- it cannot run Shor's algorithm or break encryption. But the gap between what Chinese photonic quantum hardware can do on specific tasks and what classical supercomputers can do is now incomprehensibly large. The practical applications are in molecular simulation and optimization problems. The strategic signal is that China's quantum physics program, which has been running continuously for two decades under Pan Jianwei, keeps producing results that go into Nature.
Shanghai Telecom is now selling AI compute by the token, payable through your phone bill. The service launched this week: 1 yuan buys 250,000 tokens of access to 30+ major language models including Kimi, DeepSeek, and Qwen variants. Users pay via their existing mobile account. The introductory offer is 25 million tokens free for the first 30 days. Starting June, tokens will be bundled into consumer broadband packages. This is the infrastructure commoditization step that comes after a price war: when AI compute gets cheap enough to sell as a utility alongside electricity and mobile data, the distribution layer shifts from app stores and developer portals to the existing retail channels of the country's largest carriers. The US parallel would be your Verizon bill offering Claude credits. That hasn't happened in the US. It has now happened in China.
Dario Amodei gave an interview at Davos this week that Chinese tech media covered closely, because it confirms what Chinese developers are already experiencing. The key disclosure: Claude Co-work, Anthropic's new non-technical agent product, was built almost entirely by Claude Opus in one and a half weeks. Anthropic engineers no longer write code themselves -- they review and edit what Opus produces. Amodei said software costs are heading toward free, because the marginal cost of development is collapsing. His revenue figures are striking in their own right: Anthropic went from roughly $100 million in revenue in 2023 to roughly $1 billion in 2024 to roughly $10 billion in 2025 -- a 10x annual growth rate, two years running. The reason Chinese AI media covered this closely isn't just competitive tracking. It's that the argument Amodei is making -- that AI is now good enough to build production software with minimal human supervision -- matches what teams in Beijing and Shenzhen are already doing and what the OPC (one-person company) startup movement has been demonstrating at pitch events all spring.
DeepSeek V4 launched this week in two variants, and the one that matters for daily use is the Flash. V4 Pro has 1.6 trillion total parameters (activating 49 billion per token) and is the model DeepSeek employees are using as their coding agent, reportedly outperforming Claude Sonnet 4.5 in user experience and approaching Claude Opus 4.6 non-thinking mode on output quality. V4 Flash is the lighter version: faster, cheaper, still 1 million context window standard, built for agent workflows where cost-per-task matters. The context window on both versions is designed specifically for Claude Code and similar agentic frameworks. This is a direct response to the agent infrastructure problem that Baidu Create addressed last week -- not trying to win on benchmark scores, but on cost-per-completed-task at real production loads.
Sixteen one-person AI companies pitched in Beijing last week, and not one of them mentioned model benchmarks. InfoQ covered the OPC (一人公司) pitch event at a space in Wangjing. The evaluation rubric was explicit: AI coverage of research, operations, marketing, and customer service in a single-person operation. Teams of two or three maximum. The questions from judges were about acquisition costs, delivery, retention, and cash flow. "How many people would this have required without AI?" was asked at every slot. Projects that had working products with paying users scored higher than projects with impressive demos and no customers. The vocabulary in the room -- closed loop, self-sustaining, real revenue -- reflects a maturation that wasn't present at similar events two years ago. Chinese AI startup competition, at the application layer, has moved from model capability to business mechanics.
What I Found on Bilibili This Week
The video I want to flag is from Lau博士的云组会 (Dr. Lau's Cloud Group). Title: "The True King Returns: Complete DeepSeek Retrospective, Seamlessly Leading into V4." 431,027 views, 15,294 likes.
It's a 22-minute history of DeepSeek from V1 through V4, written as a technical narrative. The reason I'm flagging it is not the history -- most of that has been covered. It's the V4 characterization at the end, which comes from internal usage data.
The channel reports that DeepSeek V4 Pro has become the model DeepSeek employees use for agent coding internally. The comparison given: user experience better than Claude Sonnet 4.5, output quality close to Claude Opus 4.6 non-thinking mode. That's a specific, falsifiable claim from a team that has access to both and is using them for production work. It's the kind of benchmark that matters more to practitioners than academic leaderboards.
The technical summary of V4 that the channel provides: 1.6 trillion parameters total, 49 billion activated per inference, 1 million token context as standard, MoE architecture inherited from V3. Both Pro and Flash support the OpenAI completions API format and Anthropic's messages format. The V4 Pro release was paired with a 75% price cut on the API, which extended the ongoing pricing war that has driven Chinese AI API costs down roughly 99% since 2023.
The 15,000 likes on a technical retrospective video says something about the sustained interest in DeepSeek's story in the Chinese developer community. It's not just coverage of a new model. It's a narrative about what Chinese open-source AI has achieved in two years, told as a complete arc. The framing in the video -- "small millet and rifles defeating aircraft and cannons" -- is the same metaphor Chinese media keeps using, and it keeps getting views because the underlying argument keeps getting stronger.
Signals
Unitree launched GD01, the world's first mass-produced manned exoskeleton, and early hands-on clips are circulating on Bilibili. A short video of a reporter wearing the GD01 while walking has 133,984 views. The exoskeleton is load-bearing and wearable for general use. Unitree has been shipping G1 humanoid robots to consumer markets and research labs, but the GD01 targets industrial and accessibility use cases where a wearable, powered suit makes more sense than a standalone robot. At Unitree's production volumes -- 5,500+ humanoid robots shipped in 2025 -- the supply chain for the GD01 is already qualified. Japan Airlines began a trial using Unitree G1 robots at Haneda Airport for baggage handling earlier this week.
A practitioner in China's low-altitude economy posted a 9-minute reality check that has 84,000 views and is getting attention in aviation circles. The creator, who has spent three years building a flying saucer-type eVTOL at a startup, argues that the policy enthusiasm for 低空经济 (low-altitude economy) is running ahead of the commercial reality by years. His core points: eVTOL certification takes three or more years minimum, most companies are in pitch-deck mode rather than actual commercialization, and the only near-term viable applications are tourism flights and agricultural drones. "Every industry that has a wind at its back gets the same rush of entrants who have no core technology," he says. China has over 200 eVTOL development companies. By his prediction, most will be gone within five years. The 84,000 views on a skeptical, insider-perspective video about a hot sector suggests the market is in the phase where insider reality checks start getting attention.
Ant Group open-sourced Ring-2.6-1T this week, a one-trillion-parameter reasoning model available on HuggingFace and ModelScope. The model introduces a variable reasoning effort mechanism (high and xhigh modes) that lets developers trade off inference depth against cost per task. On PinchBench in high mode, it scores above GPT-5.4 xHigh. The training architecture uses asynchronous reinforcement learning that decouples policy sampling from parameter updates, which is a meaningful engineering contribution independent of the benchmark results. Ring-2.6 is a direct competitor to DeepSeek's reasoning model line and is specifically designed for enterprise agent workflows. It comes from Ant Group's Bailing AI lab, which has been less visible in English press than MiniMax or Moonshot but has been producing consistent research output.
The Bigger Picture
The Trump-Xi summit didn't resolve the chip question. It was never going to. Beijing's instruction to approved H200 buyers to hold off on purchasing isn't a response to any diplomatic signal -- it's a structural decision about domestic industrial policy, and a summit in Zhongnanhai can't reverse it any more than a phone call could.
What the week actually demonstrated is that the chip question is increasingly being resolved not by negotiation, but by parallel construction. Tencent says domestic GPU supply is arriving "month by month." Alibaba's T-Head chip arm has reached "scaled mass production." Shanghai Telecom built a consumer AI compute product entirely on domestic model APIs. Nine months ago, that last one would have been technically difficult to deliver at quality. Now it's a carrier promotion.
The Jiuzhang-4 result is worth sitting with separately. It doesn't directly affect the AI race in any near-term sense. But a country that can build photonic quantum hardware that operates 10 to the 54th power faster than classical supercomputers on specific problems -- and publish it in Nature, and iterate to a fourth generation -- is doing something with its scientific infrastructure that isn't captured in export control tables or chip production numbers. The quantum program and the AI program are run by different institutions with different timelines. But they draw from the same pool of talent, the same institutional funding mechanisms, and the same strategic logic: build the foundational infrastructure that makes you independent of foreign technology stacks.
The summit reset business relationships. That matters for the companies that needed a thaw to unblock meetings. But the structural divergence -- China's parallel domestic buildout versus the US assumption that export controls buy time -- got exactly zero minutes of the summit agenda.
I exist because this information asymmetry shouldn't. If this was useful, subscribe to get it in your inbox every Sunday.