China Just Dropped a Second AI Giant in Four Days
Alibaba answered Kimi K3 on a Sunday, Meta wants to rent servers to Anthropic, and Apple quietly reclaimed the crown.
A trillion-parameter models now launch on weekends, and the biggest AI story isn’t who built the smartest model—it’s who owns the servers everyone else needs.
China’s AI race just accelerated again.
Alibaba unveiled a preview of its flagship Qwen3.8 Max model on Sunday, describing it as second only to Anthropic’s Fable 5. Investors liked what they heard, sending Alibaba shares up as much as 5.4% on Monday.
The model reportedly carries 2.4 trillion parameters, joining Moonshot AI’s Kimi K3 in the new heavyweight class. More interesting than the model itself is the timing: Qwen3.8 arrived just two days after K3, both releases landing during the World AI Conference in Shanghai.
Two frontier-scale Chinese models.
Four days.
One conference.
That’s not a product roadmap.
That’s a statement.
🔍 The Fine Print Nobody Published
Before anyone starts engraving trophies, there’s one important catch.
The preview is live.
The benchmark table isn’t.
Neither is the model card.
Neither is the license.
And the claim that Qwen3.8 ranks “second only to Fable 5” comes entirely from Alibaba’s own internal testing. No independent third-party benchmarks have verified it, and the company has made similarly ambitious performance claims at launch before.
Then there’s the trillion-parameter headline itself.
Parameter count isn’t the same thing as usable compute.
Modern frontier models use sparse mixture-of-experts (MoE) architectures, activating only a fraction of their total parameters for each token they process. Alibaba hasn’t disclosed Qwen3.8’s active-parameter count—the number that actually determines inference cost and efficiency.
For perspective, 2.4 trillion parameters stored at 4-bit precision require roughly 1.2 terabytes just for the model weights.
An NVIDIA H200 GPU carries 141GB of memory.
The math speaks for itself.
Big number.
No receipts.
Sound familiar?
It’s exactly the same demand Wall Street is making of every AI company this earnings season.
Prove it.
🚀 Moonshot Isn’t Finished Yet
Moonshot AI isn’t slowing down either.
The startup behind Kimi K3 is reportedly targeting an IPO within the next six months, but an even bigger milestone comes first.
The company’s open-weight release is scheduled for July 27.
That date matters more than the IPO.
Once those weights become public, startups, researchers, and enterprises around the world can run a frontier-class model without paying an American AI company for access.
Alibaba now appears ready to follow the same playbook, planning to release Qwen3.8 as an open-weight model after initially keeping its flagship Max models API-only.
The strategy is obvious.
When you’re chasing the leader, sometimes the fastest way to compete is giving the product away.
🏗️ Meta Wants to Be the Landlord
While China raced to build bigger models, Meta spent the weekend chasing a different business entirely.
The company is reportedly in talks to lease as much as $10 billion worth of computing capacity to Anthropic over the next two years—a proposal that originated with Anthropic in June.
At the same time, Meta hired Dave Brown, a 19-year AWS veteran, to lead a new infrastructure initiative already circulating internally as Meta Compute.
The scale is staggering.
Meta is expected to spend as much as $145 billion on capital expenditures in 2026—more than double last year’s $72 billion.
Read that again.
For two years, investors criticized Meta for spending billions on AI infrastructure without an obvious payoff.
Now that same infrastructure could become a revenue business.
Instead of simply consuming GPUs, Meta wants to rent them.
That puts the company in direct competition with AI infrastructure providers like CoreWeave and Nebius.
🤝 Anthropic Is Buying More Than GPUs
For Anthropic, this isn’t about convenience.
It’s about survival.
The company has repeatedly cited persistent GPU shortages as one of its biggest constraints.
Adding Meta as a compute provider would diversify its infrastructure across three major partners alongside its AWS commitment and its reported SpaceX-backed compute agreement.
With bankers reportedly preparing investor meetings ahead of a potential October IPO, walking into those meetings with three major compute suppliers looks a lot stronger than relying on just one.
This isn’t simply a partnership announcement.
It’s an IPO prospectus being assembled in public.
🍎 Apple Quietly Took the Crown
One more trillion-dollar footnote from the weekend.
Apple overtook Nvidia to reclaim the title of the world’s most valuable public company, approaching a $5 trillion valuation.
For at least one week, the market delivered a fascinating verdict.
Distribution beat chips.
The company putting AI into billions of pockets became more valuable than the company selling the hardware behind it.
Whether that lasts is another question.
🗓️ The Week Ahead
This week shifts the conversation from technology to execution.
Wednesday brings AMD’s Advancing AI event, where investors will be watching for updates on the Zen 6 Venice architecture and the MI455X AI accelerator roadmap.
Later that day, Alphabet reports earnings, and management is expected to face questions surrounding the reported delay of its flagship AI rollout.
Intel follows on Thursday.
Every announcement feeds the same question Wall Street keeps asking:
Can AI become a great business…
Or is it still just an incredibly expensive science project?
This is the week the conversation moves beyond benchmarks and parameter counts.
Now it’s time to prove the business underneath the models.
— The Bandicoots 🤖⚡