Alibaba just showed Nvidia its own playbook

Good morning 👋 Alibaba spent one Hangzhou keynote telling Nvidia it doesn't need it anymore: a new chip, a 10-trillion-parameter model roadmap, and 20GW of data centers to run it all on.

In today's issue:

• 🔭 Alibaba just showed Nvidia its own playbook
• 🧠 A power/cooling cert from NVIDIA and Grok 4.6 lands on Bedrock
• 🏛️ OpenAI tries to write its own rulebook, then names a math advisory board that includes one of its own critics
• 🛠️ Test-drive a quantized Qwen before the giant ones ship
• 🔗 A Hong Kong IPO, a new European unicorn, and BMW's cost-anomaly dashboard

🔭 THE ONE THING

⚡ Alibaba just showed Nvidia its own playbook

At Apsara in Hangzhou today, CEO Eddie Wu laid out the full stack: a new Zhenwu V900 chip claiming 3x the throughput of its predecessor and clusters scaling to 500,000 cards when it hits commercial production in Q1 2027. On the model side, Qwen 4 is training now, with Qwen 4.5 and 5 targeted at 5 to 10 trillion parameters, backed by a pledge to hit 20GW of global data-center capacity by 2032. The market bought it: Hong Kong shares gained about 5% overnight. Treat the 3x number as a vendor claim until someone independently reproduces it, but the 500,000-card cluster figure is the one that matters, because that's the first time a Chinese chip-to-agent stack has been pitched as a real Nvidia alternative instead of a stopgap. If Alibaba can actually deliver clusters at that scale, training runs that used to require a Nvidia allocation just got a second supplier.

🧠 MODELS & RELEASES

• ⚡ NVIDIA launched DSX Ready, a certification program that qualifies third-party power and cooling hardware against its own AI factory reference designs. The compute race increasingly runs through the electrical closet as much as the GPU aisle. Details
• 🤖 xAI brought Grok 4.6 to Amazon Bedrock, with a 500K token context window and four reasoning-effort tiers across both bedrock-mantle and bedrock-runtime endpoints. Builder read: one more frontier model you can call without leaving your AWS IAM and billing setup. Details

🏛️ POLICY & REGULATION

• OpenAI is pushing a US-led standards regime for frontier models: shared evaluation methods, defined triggers for human oversight, and an actual incident-reporting pipeline routed through the safety-institute network already up in the UK, Japan, Korea, and seven other countries. Read past the safety framing and it looks like a company trying to write the rulebook before recursive self-improvement stops being a hypothetical. Worth tracking who signs on. Building standards for the next phase of AI

• OpenAI also stood up a nine-mathematician advisory board (Gowers, Hairer, Witten, Vakil among them), three weeks after 25 Fields medalists signed a letter warning that labs are outrunning scientific oversight. One of the nine, Camillo De Lellis, signed that letter too. The board gets input on how OpenAI's math claims get reviewed and communicated. It has no say over how fast the research moves. Advisory Group on Mathematics and AI

🛠️ TRY THIS

Test-drive a quantized Qwen checkpoint before Alibaba's giant ones show up

Apsara talk of 5 to 10 trillion parameter Qwen models isn't running on your laptop, or anyone's. But Hugging Face just made it trivial to load the current Qwen3.5 and Qwen3.8 GGUF quants straight through Transformers, no llama.cpp conversion step required. Good week to find out if a small Qwen already covers what you need before the roadmap ships something you can't self-host.

1. Grab a Q4_K_M or Q5_K_M Qwen3.5 GGUF from Unsloth or bartowski on Hugging Face.

2. Load it with AutoModelForCausalLM.from_pretrained(model_id, gguf_file=filename). No conversion, no separate llama.cpp runtime.

3. Run your actual prompts against it, not a benchmark set. Log latency and where the answers diverge from whatever's in production now.

4. If it holds up, you've cut a dependency and a cost line. If it doesn't, you know that before the 10T-parameter version ships and someone asks why you're not on it.

Prompt: Run these 15 [your domain] prompts against the quantized Qwen3.5 checkpoint and the model currently in production. Flag any answer where they disagree on a factual claim, and log latency for each.

Worth a look

• AWS AgentCore published the pattern for moving a three-model healthcare agent off self-managed ECS/Fargate onto Bedrock AgentCore runtime, without touching the agent logic, just wrapping the entrypoint and running agentcore deploy. Keep it open for the day step 4 says your local model's fine but your infra isn't.

🔗 QUICK LINKS

• Ligent Technologies raised $727M in its Hong Kong IPO and closed up 9.8% on debut, putting the optical-transceiver maker near a $4.3B market cap. Bloomberg
• Verda closed a $189M Series B led by Emergence Capital. The Helsinki neocloud (formerly DataCrunch) now tops a $1B valuation, Europe's newest AI-infra unicorn. Bloomberg
• BMW Group built automated daily cost-anomaly alerts across 14,000 cloud accounts on its CLEA platform, using Prophet forecasting and AWS Step Functions to replace reactive dashboards. AWS
• NVIDIA showcased Egypt's AI builders hitting production scale at a Cairo event in the Grand Egyptian Museum. Regional-ecosystem flag-planting, but worth tracking who's actually deploying. NVIDIA
• Multiverse Computing frames LLM block-pruning as an Ising optimization problem, borrowing a physics trick to decide which transformer blocks are safe to cut. Hugging Face
• V7 turns a company's scattered files into source-linked memory agents can actually cite. OpenAI

See you tomorrow.

Pradeep Perugu