Tech Buzz China

When ChatGPT launched in late 2022, the consensus inside China’s AI industry was that domestic labs were roughly two years behind on models, while export controls would leave China’s AI stack woefully incomplete.

Three and a half years later, that framing looks ... outdated.

DeepSeek and GLM forced a rethink of what was possible on the model building and also inference side. Meituan LongCat raises the same question for training.

None of this means China has caught up. Nvidia still has a massive manufacturing advantage, and US labs continue to lead in enterprise adoption, developer ecosystems and revenue.

But we do think the burden of proof has shifted.

The question is no longer simply whether China can build a frontier AI stack on domestic infrastructure. It’s how quickly that stack improves from here.

AI Made Double 11 Faster, Not Bigger

Alibaba, JD, Douyin and Kuaishou all put AI at the center of this year's Double 11, from AI customer service to AI-generated product images. One egg seller told the source AI cut his manual work to 20-30% of the job and tripled his output. His store's volume rose only 15-20% in the October 20-26 week.

Three Squirrels (三只松鼠) is the sharper case. It deployed DeepSeek in-house and runs AI across every SKU. In 2025 revenue fell 4.08% to RMB 10.19 billion (~$1.4B), while online retail grew 8.6%. The tools sit almost entirely on the supply side, making goods cheaper and faster, and the efficiency every merchant now gets cancels out as an advantage. We doubt the next Double 11 gets bigger until someone gives shoppers a reason to buy more, rather than a faster way to make the goods.

h/t @TMTPostGlobal

Anthropic's updated S-1 adds third-quarter financials and returns the company to a loss. The quarterly numbers sharpen a question the filing already raised: whether revenue can grow fast enough to absorb a compute obligation structure that is largely fixed.

1. Third-quarter revenue was $16.96 billion, up 46.7% sequentially. Compute and infrastructure costs rose 77%, reaching 58% of revenue. The company recorded a $567 million GAAP operating loss. Adjusted operating margin fell from 19% to 15%.

2. Through September, cumulative purchase and lease commitments reached $895.4 billion. About $69.8 billion falls due within twelve months, and roughly half of the total comes due before 2030. Most agreements are take-or-pay, meaning Anthropic owes payment regardless of actual usage.

3. If commitment payments were held to half of revenue, Anthropic would need about $140 billion in revenue over the next twelve months. That is the number implied by the filing, not a forecast. At a 25x multiple, roughly $58 billion in 2026 revenue supports a valuation closer to $1.5 trillion than the original $2 trillion.

4. The pricing pressure is explicit. Anthropic acknowledged in its prospectus that discounts on new model versions and contract concessions partially offset usage-driven revenue gains, with the impact estimated at about $600 million for the quarter. Over 80% of revenue came from models released in the prior six months.

5. Model switching is getting cheaper, and corporate buyers are exercising that option. This is consistent with fintech Ramp's observation that enterprises are rotating between AI providers more frequently. Every reduction in migration cost weakens the pricing power of individual model vendors.

6. Open-weight models are becoming a meaningful supply channel. On OpenRouter and Together AI, token consumption from open-source models is approaching Google Gemini API volumes under comparable measurement. That puts a price cap on closed frontier models rather than leaving them a niche alternative.

7. The harder question sits beyond Anthropic. Model vendors need enterprises to keep increasing compute and API spend. But if the promised productivity gains arrive through labor-cost reduction, they may not create enough new purchasing power to sustain that spend.

DeepSeek has opened its path from Nvidia to Huawei Ascend, releasing programming tools, compute libraries, and communication libraries mapped one-to-one to its Nvidia stack. The release comes as DeepSeek pushes a second funding round and plans at least 160,000 Ascend 950 DT chips in an Inner Mongolia datacenter.

1. The core is TileLang, a domain-specific language for high-performance kernels. DeepSeek validated it on Nvidia first, and it now carries most operators in DeepSeek V4 training.

2. Every TileLang operator used in DeepSeek training now has a corresponding high-performance implementation on Ascend. The team's FlashMLA sparse-attention kernel on Ascend 950 reaches 410 TFLOPS, about 95% of hardware peak, on that specific operator.

3. The migration keeps the upper-level compute expression, interface, and research workflow intact, then lets different backends handle hardware-specific work. The Ascend version wraps Ascend C and can build without CUDA installed. It does not emulate Nvidia instructions.

4. "One-to-one mapping" refers to DeepSeek's own six open-source components, not the full CUDA ecosystem. These tools provide CUDA-free implementations for covered tasks but cannot run arbitrary CUDA programs.

5. Training is already in the code, alongside inference. The TileLang Ascend examples include backward kernels for attention gradients, and DeepEP-Ascend shows gradient return paths for expert communication.

6. The work extends beyond single cards. DeepSeek and Huawei are jointly advancing a 128-card Ascend 950 supernode plan, and the two have published communication tests at up to EP128 scale.

7. Several things remain unproven: end-to-end convergence, full-model 128-card training, long pre-training stability, and failure recovery. Public communication performance still depends on non-public proof-of-concept driver firmware, with a recommended commercial version expected mid-October.

Full article in comments below.

Silicon Valley's AI giants need a bigger application than coding to justify their trillion-dollar capex. Drug discovery is where they are placing the bet.

1. The market signal is real. Tirzepatide and semaglutide, the two GLP-1 drugs, produced roughly $71 billion in combined 2025 revenue. That is more than double OpenAI and Anthropic's combined end-2025 annualized revenue run rate.

2. October marked a step change. Frontier AI companies announced investments, acquisitions, self-built wet labs, and industrial partnerships nearly every week. The goal has shifted from improving molecular screening to owning the full research loop.

3. Nvidia calls this Lab in the Loop. The model proposes a hypothesis, designs an experiment, the wet lab executes it, and the data feeds back into the next round of model decisions.

4. Kevin Weil's departure to start Warp Sciences is the clearest signal. The former OpenAI science lead is building an AI-automated drug discovery lab in the US. He explicitly aims to pull work back from Chinese contract research organizations like WuXi AppTec.

5. The data is the prize. Automated labs produce compounds, and they also produce proprietary experimental data, including negative results. That data is the feedback mechanism that lets models keep improving.

6. China's lead in CRO and CDMO execution is real, built on years of outsourced research and skilled lab engineers. Lab in the Loop is designed to dilute that advantage by moving both execution and learning back to Silicon Valley.

7. Google's VERID framework is one serious attempt at the missing infrastructure. It acts as a digital twin of a wet lab, tracking equipment layouts, reagent identity, and liquid volumes in real time. That lets AI agents simulate experiments before execution.

8. Closed loops are already running. OpenAI and Ginkgo Bioworks ran six design-test-analyze cycles, testing over 36,000 reaction combinations and cutting cell-free protein synthesis costs by about 40%. Anthropic has built its own Bay Area wet lab and validated a new CRISPR-like enzyme system.

9. The barrier is forming. Anthropic's September protein design competition with Adaptyv explicitly excluded participants from China and other regions. That signals frontier AI and wet-lab infrastructure are becoming linked in ways that may not stay open.

For China, the pattern is familiar. Its AI companies have followed Silicon Valley's validated paths before, but arriving a few steps late can mean missing the window when compounding advantages accumulate.

Full article in comments below.