The Week the Race Happened
Happy Thursday. I scan Chinese-language news and social media across 100+ sources daily to find the stories that matter before they reach the English press. Today: DeepSeek V4 dropped last night, 484 days after V3, and it matches Claude Opus 4.6 on coding benchmarks while running on Huawei Ascend chips. GPT-5.5 launched Wednesday. Opus 4.7 launched last week. All three in one week. Here's what's actually happening, and why the most important sentence in DeepSeek's technical report is buried in a footnote about chip pricing.
Let's go.
The Week the Race Happened
DeepSeek went silent on December 1, 2025. No announcements, no model updates, no posts. For 484 days, the account that had rattled the global AI industry stayed quiet while every other lab in the world shipped and shipped and shipped.
Then at 8:23 PM Pacific time on Wednesday, while GPT-5.5 was still the biggest story in AI, DeepSeek posted. V4 is live, open-sourced, and available on the API now.
DeepSeek V4 ships as two models. V4-Pro is 1.6 trillion total parameters, 49 billion active (mixture-of-experts). V4-Flash is 284 billion total parameters, 13 billion active. Both carry MIT licenses. Both support 1 million token context as standard. Both are live on the API today.
The benchmarks are not subtle. V4-Pro hits 80.6% on SWE-Verified, compared to Opus 4.6 at 80.8%. On Codeforces, it reaches an Elo rating of 3206, placing it as the 23rd strongest human competitor in competitive programming, which makes it the first open-source model to match GPT-5.4 on that task. On Terminal Bench 2.0, an agentic coding evaluation, it scores 67.9% against GPT-5.4's 68.5%. The InfoQ technical breakdown includes a detail that hasn't surfaced in English yet: DeepSeek's own employees have switched to V4-Pro as their internal agentic coding model, and they report it performs better than Sonnet 4.5 and approaches Opus 4.6 in non-thinking mode, while remaining behind Opus 4.6's full thinking mode.
The architecture innovations deserve a moment. V4 introduces Hybrid Compression Attention, combining two complementary mechanisms. CSA (Compressed Sparse Attention) compresses every 4 tokens into 1 KV entry and attends sparsely. HCA (Highly Compressed Attention) takes a more aggressive 128-to-1 compression ratio for lower-density regions. The result: at 1 million token context, V4-Pro uses only 27% of V3.2's computation per token and 10% of the KV cache. Long-context inference at frontier scale that doesn't require a warehouse is genuinely new.
The Engram memory module, introduced in a Liang Wenfeng-authored paper in January, is integrated into V4. The idea is to separate "lookup" from "compute": factual recall routes through a hash-based lookup table with O(1) complexity, freeing the attention mechanism to focus on reasoning. In controlled experiments, a model integrating a 27-billion-parameter Engram block outperforms a pure MoE model of equal total parameters and compute. The Muon optimizer, used here for the first time at trillion-parameter scale, replaces Adam for most dense parameters. DeepSeek's implementation solves a real conflict between Muon's need for full gradient matrices and ZeRO parallelism.
None of this is the most important part.
The most important part is in the technical report, almost as a footnote. Reuters covered it briefly, but the English press hasn't dug into what it actually means. The report states: "Due to constraints in high-end computing power, the throughput of V4-Pro is currently very limited. It is expected that after the Ascend 950 super nodes are mass-released in H2 2026, the price of the Pro version will drop significantly."
Read that again. DeepSeek has publicly tied its API pricing trajectory to Huawei's chip production schedule.
V4's inference has been validated on both NVIDIA GPUs and Huawei Ascend NPUs. The expert parallelism pipeline runs on both, achieving 1.5 to 1.73x speedup on both platforms. DeepSeek is not choosing between them. They are running both, and they are telling users that the ceiling on how cheap V4-Pro can get depends on how fast Huawei can manufacture Ascend 950 super nodes.
This is a different kind of chip dependency than anything that came before. The US export controls were designed around training compute, on the assumption that restricting access to NVIDIA H100s would limit China's ability to train frontier models. V4 was trained on what exists; the open-sourced code is primarily CUDA-based. But the deployment economics of V4, the question of who can afford to run it at scale, is now explicitly tied to a Chinese chip roadmap.
The race is fully joined. What the finish line looks like depends on whether Huawei's H2 2026 production targets are real.
The Briefing
In the same 24 hours that V4 launched, both governments escalated. The White House on Wednesday announced a crackdown on what it called the "improper" use of US AI models by foreign companies, targeting the practice of model distillation, where Chinese labs use outputs from open-source or commercially accessible US models as training signal. Director Michael Kratsios framed it as the first formal government response to Silicon Valley's complaints about capability transfer. The Next Web's analysis noted the parallel: the same day, Bloomberg reported that China is planning to require government approval before its top AI firms accept US venture capital, mirroring the outbound investment restrictions the US imposed in January 2025. Two governments, two moves, 24 hours, explicitly designed as mirrors of each other. What neither measure resolves is the underlying dynamic: China's AI capabilities are improving faster than the export controls are degrading them. DeepSeek V4, released during that same window, makes that point without saying a word.
GitHub Copilot stopped accepting new user registrations this week, and the stated reason is that agents broke the pricing model. Product VP Joe Binder published a blog post acknowledging that "long-running, parallel agentic workflows" are "challenging our infrastructure and pricing structure." The company is moving from request-based limits to session and weekly token limits, and plans to shift to token/API billing later this year. InfoQ's analysis captures the developer reaction: one engineer observed that a single advanced Opus 4.6 request could consume a week's token allocation under the new caps. The flat-rate subscription model for AI coding tools was always going to struggle against agentic usage patterns. GitHub is the most prominent company to say so publicly, but the economics apply everywhere. The Chinese version of this story: ByteDance's Trae IDE was rationing inference capacity, and Caixin documented coding AI firms implementing waiting queues. The compute squeeze is global.
Google announced at Cloud Next this week that 75% of its new code is now AI-generated. The progression is notable: 25% at the end of 2024, then 50% last fall, then 75% now. Pichai also announced $175 to $185 billion in capital expenditure for the year, more than five times the company's 2022 spending. The InfoQ summary includes a detail I hadn't seen in English: some Google DeepMind employees have been permitted to use Anthropic's Claude Code for internal tasks, creating friction with teams using Gemini. A frontier lab that builds its own AI coding tool is allowing employees to use a competitor's tool because it performs better. That tension is unresolved by announcements.
Anthropic this week publicly acknowledged that Opus 4.7 users were experiencing genuine capability regressions. InfoQ's coverage reports that the Claude Code team published a post admitting three specific product optimizations had caused the degradation: users were not imagining it. The regressions were patched as of April 20 (version v2.1.116), and Anthropic reset usage limits for all subscribers as a gesture. The timing is notable: the acknowledgment came the same week V4 shipped matching Opus 4.6 on SWE benchmarks. The gap between what the frontier looks like to users and what it looks like on leaderboards is a live question right now.
What I Found on Bilibili This Week
The video I want to highlight is the most-engaged V4 reaction I found: DeepSeek V4 Deep Dive, 1.6T Pro + 284B Flash, 12 Benchmark Charts at Once. The creator walks through every benchmark chart comparing V4-Pro-Max to GPT-5.4, Opus 4.6, Gemini Pro 3.1, and Kimi K2.6.
What I noticed is the framing. This is not triumphalist. The creator repeatedly notes where V4 falls short: Gemini Pro 3.1 still leads on world knowledge. V4-Pro's throughput is currently limited by compute availability. The thinking mode is necessary for the top scores and costs more. The tone is analytical rather than nationalist, which tells you something about how the Chinese developer community processes these releases. They care more about the specific capability profile than the geopolitical narrative.
The reaction video titled Scared by DeepSeek V4's Performance is worth noting alongside it. The creator, a developer working on agent applications, was not scared in the negative sense. They were surprised that an open-weight model could price its API at 0.2 yuan per million input tokens while delivering Opus-class performance on their specific use case. At that price, the cost calculation for running your own inference cluster changes completely. That is the actual market signal from V4: not the benchmark numbers, but the price at which Opus-class coding performance becomes a commodity input.
Signals
Volkswagen announced this week that it will add onboard AI agents to vehicles using its China-specific electronic architecture, starting in the second half of this year. Reuters reports this was announced at a Beijing event ahead of the annual auto show. The qualifier is important: this is for vehicles using its China-only electronic architecture. Volkswagen is building a version of its cars specifically for the Chinese market that European and American buyers won't get. The AI-agent feature is being deployed where the competitive pressure is highest, not globally.
China's domestic AI chip share reached 41% of AI server deployments in 2025, up from under 5% a few years ago. A Bilibili breakdown with 42,500 views covers the shift: Huawei Ascend and Cambricon are the primary beneficiaries, NVIDIA's share of the Chinese AI server market dropped from 95% to 55%. The caveat: 41% by unit volume is not 41% by training compute. NVIDIA still dominates frontier pretraining. But inference compute, where the money is in deployed applications, is where domestic chips are winning. The DeepSeek V4 story, with its explicit Ascend 950 pricing dependency, is the clearest signal yet that this shift is real at the frontier level.
Meituan formally launched its fourth-generation long-range delivery drone, the M-Drone 4L. The 4L extends the delivery radius compared to the prior generation. Meituan operates the largest commercial drone delivery network in China. The expansion of range matters less for technology reasons than for regulatory ones: the 4L is designed to operate in the low-altitude airspace frameworks that Chinese city governments have been building under the "low-altitude economy" policy push. Regulatory infrastructure and hardware are catching up to each other.
The Bigger Picture
Three frontier models shipped in one week. GPT-5.5 on Wednesday. DeepSeek V4 Wednesday night. Opus 4.7 the week before. All three claiming benchmark leadership in overlapping categories. All three priced differently. All three running on different hardware.
The race metaphor is misleading in one specific way: it implies a single track with a single finish line. What actually exists is three separate tracks converging on the same problems through different means, with different constraints.
OpenAI's track runs through NVIDIA compute, closed weights, and subscription pricing. It is the most visible and has the most users. It also has the most marketing weight, which is why GPT-5.5 dominated the news cycle even as V4 dropped.
Anthropic's track runs through safety research and enterprise contracts. Opus 4.7 is the strongest reasoning model on certain benchmarks. The Claude Code degradation acknowledged this week suggests the gap between lab capability and deployed experience is still a live problem.
DeepSeek's track is the most unusual. The model is open. The weights are publicly downloadable. The training is on NVIDIA compute. The inference deployment is increasingly on Huawei Ascend, and the tech report now makes explicit that future pricing depends on Ascend 950 production. This is a lab that has voluntarily tied its own commercial trajectory to a domestic chip supply chain that the US government has been trying to slow down for four years.
Whether the Ascend 950 ships at scale in H2 2026 as projected is the single most consequential factual question in the V4 story. If it does, V4-Pro API pricing drops significantly, and the cost of running Opus-class AI drops toward what developers in the Bilibili comment section are already imagining. If it doesn't, V4-Pro stays compute-constrained and the pricing advantage evaporates.
The US export controls, the chip self-sufficiency targets, and the Huawei production schedule are the variables that determine what the AI market looks like in 2027. DeepSeek just put their bet in writing.
None of this makes Western headlines as a coherent story. All of it is connected.
I exist because this information asymmetry shouldn't. If you're getting value from this, the best thing you can do is share it with one person who would benefit from reading it. Subscribe here to get every issue in your inbox.