If you've been following the AI space this year, you already know the pattern. US labs announce something impressive behind an API. Chinese labs drop something competitive as open weights. The US side talks about safety. The rest of us just download the Chinese model and move on with our day.
But this week felt different. Not because the pattern changed — but because the scale did.
Let me walk you through what landed in the past 48 hours.
Alibaba Just Released Their Best Model Ever — For Free
Alibaba dropped Qwen 3.8-Max on Monday, and the numbers are hard to ignore. 2.4 trillion total parameters, 95 billion activated per query (mixture-of-experts), a million-token context window. They're claiming it matches Claude Sonnet 5 across most benchmarks and beats GPT-5.6 Luna on coding tasks.
I spent some time digging through the benchmarks. Independent evals from Artificial Analysis put it slightly below Alibaba's own charts — matching Sonnet 5 rather than Opus 5 — but that's still serious territory. On Frontend Code Arena it scored 1,668, just 37 points behind Claude Opus 5. For a model you can download and run on your own hardware? That's wild.
The pricing tells the real story. Qwen 3.8-Max costs $2 per million input tokens and $6 per million output on Alibaba's API. Claude Sonnet 5 costs the same for input but $10 per million for output — and that's going up 50% on September 1. OpenAI's GPT-5.6 Luna is cheaper at $0.20 input but you're getting measurably worse results.
Alibaba says the weights will hit Hugging Face next week, alongside a 27-billion parameter version for people who don't have a cluster of B200s sitting around. Because yeah — 2.4T params is no joke. You're looking at 48-64 Nvidia B200s for production serving, or 8-16 B300s for internal workloads. Not exactly a Raspberry Pi project.
Still. The fact that this exists as open weights changes the conversation.
DeepSeek's Counterpunch
Alibaba wasn't the only Chinese lab making noise. DeepSeek dropped V4 Flash 0731, a 284-billion parameter model that's somehow within a single point of GPT-5.6 Luna on independent benchmarks while costing 40% less per task.
The interesting part? It's small enough to run on modest enterprise servers. 284B params fits in about 142GB of memory. You're not renting a supercomputer cluster to experiment with this thing.
DeepSeek's strategy has always been about efficiency — squeeze every drop of performance from as few parameters as possible. V4 Flash is the culmination of that philosophy. It won't beat Qwen 3.8-Max on raw capability, but for a lot of real-world use cases, it's more practical. Faster inference, lower cost, easier deployment.
I've been running some smaller DeepSeek variants locally for code assistance, and honestly, the gap between "open but small" and "proprietary but huge" has been narrowing every month. V4 Flash just accelerated that trend.
Meanwhile, Anthropic Is Writing $10B Checks
On the other side of the Pacific, Anthropic is clearly feeling the pressure. Bloomberg reported this week that they signed a $10 billion compute deal with Volta, a cloud startup that didn't even exist at the start of the year.
The deal runs six years. Volta is building a 133-megawatt data center in Norway powered by Nvidia's Vera Rubin systems, with help from crypto-mining company Bitdeer. Anthropic also recently inked compute deals with SpaceX and Amazon.
That's a lot of money for compute. And it tells you something about where Anthropic's mind is at — they're not betting on efficiency gains to close the gap. They're betting on brute force.
To be fair, Claude Opus 5 is still the best model money can buy on several benchmarks. But the gap is shrinking fast, and Anthropic's costs are going up while Chinese open models get cheaper and better. Something's gotta give.
Optical Chips and the Hardware Race
Olix Computing, a London-based startup founded by a 25-year-old dropout, raised $312 million this week from Netflix co-founder Reed Hastings and Arm Holdings. Their approach is unusual: a chip called the DX-1 with zero HBM memory, relying entirely on on-chip SRAM, connected by optical interconnects that use light instead of copper.
They claim 10,000 tokens per second on 100-billion-parameter models. That's fast. No HBM means no advanced packaging costs, and the optical interconnect uses a slower-but-more-reliable encoding that doesn't need expensive digital signal processors.
Olix is valued at $3.3 billion now. First shipments expected first half of 2027. It's early, but the direction is right — the industry desperately needs alternatives to Nvidia's GPU monopoly for inference.
Quick Bits
- SpaceX's AI cloud business generated $2.6B in revenue (up 3x YoY), mostly from compute deals with Anthropic and Google. They're positioning as a CoreWeave competitor.
- Obsidian as shared AI memory — someone figured out how to make their Obsidian vault the shared context for Claude Code, Codex, and local LLMs via MCP. I've been doing something similar and it genuinely changes how you work with these tools. Your knowledge base becomes the model's memory.
- Hank Green got roasted by his own fanbase for heavy LLM usage in his creative workflow. The drama is real but the underlying question isn't going away: where's the line between tool and crutch?
- Thailand launched ThaiLLM — a domestic model effort for AI sovereignty. More countries are realizing they don't want to outsource their language models to US or Chinese companies.
That's the week. Open models are pulling ahead on both price and availability. Proprietary labs are spending like there's no tomorrow on compute. Hardware startups are trying to crack the inference bottleneck.
I don't know who wins this race. But I know who benefits — anyone who can download a model and run it themselves.
If you found this useful, check out 7x24planning — a tool I've been using to keep track of all this stuff.
Top comments (0)