PRODUCTS

Chinese lab uses unique pitch to win AI support

It was just last year that DeepSeek emerged as China's first serious contender in the global AI race. Since then, a wave of Chinese labs has quietly raised the stakes, churning out bigger, cheaper, more efficient models that are giving US rivals a run for their money. Alibaba just did it again.

On Sunday, Alibaba launched Qwen3.8-Max, a 2.4 trillion-parameter model that the company is calling not only its biggest model, but also its most capable in the Qwen family. The model touts upgrades across coding, agentic use cases, long-horizon tasks, research, and other complex tasks. 

The company's promotional video focuses on Qwen3.8's ability to act as an "always on work-mate," showing examples of it doing hours worth of work while the worker can do more enjoyable tasks. 

For instance, one vignette shows Qwen3.8-Max designing chips nonstop for 12 hours while the engineer goes fishing. Another shows a biology professor playing tennis while the model verifies a protein. This approach, which contrasts with US labs by emphasizing the lifestyle users can reclaim, is garnering attention on socials. It's anchored on the idea that the model sets a new bar for coding and coworking, a claim the benchmarks are meant to support.

Qwen3.8-Max performed comparable to, or even higher than, leading models in a series of benchmarks Alibaba published. That included Anthropic's Fable 5, a model so powerful that the US temporarily banned it, and Kimi K3, a model with 2.8 trillion parameters that made headlines last month for its claim that it could outperform Anthropic, OpenAI, Google, and others.

Since benchmarks aren't always the truest indicators of performance in real-world tasks, more notable is its performance on crowdsourced, independent benchmarks such as the Code Arena for front-end web development tasks, where it placed fourth following Claude Opus 5, Kimi K3 Max, and Claude Opus 5 High. 

Another noteworthy aspect of the release is the models' cost, which, based on current pricing across the frontier field, Qwen3.8-Max's rates ($2 input / $6 output per million tokens) land on the cheaper end (by comparison, Fable costs $10 per million tokens for input and $50 for output). This is likely to further the price war that is already heating up in the US, with Chinese labs wielding both benchmark performance and token economics as powerful competitive forces. Alibaba plans to release the open weights of Qwen3.8-Max next week, another differentiator from the leading US labs.

Alibaba's latest release not only adds to the onslaught of highly capable, efficient, and inexpensive models coming out of China, but also highlights a different approach than the US labs on multiple fronts. The most obvious parallels are the same strengths seen in the releases of Kimi K3 and the DeepSeek models: cheaper, open-weight models that don't compromise on performance. The more interesting angle, however, is the focus on giving people back their time rather than simply increasing productivity, which is what US labs have focused on until now. While this framing may strike some as tone-deaf, given the very real concerns about AI displacing jobs by doing people's work for them, it's an interesting approach that might actually appeal to the emotions of people in the broader public, who tend to be very skeptical of AI. If the US adopted more of a Chinese marketing approach, I'd be curious to see whether it would bring more people on board with using the technology.

TOGETHER WITH CRUSOE

Here’s How To Customize Top LLMs For Your Business

It’s called Serverless Fine-Tuning, and it’s Crusoe’s proprietary method for taking the top LLMs in the AI game today (like those from DeepSeek, OpenAI, Meta, and Google) and fitting them to your exact business needs. 

Simply pick a base model, load your dataset, tune your settings exactly how you like them, and boom – you’re ready to deploy all in one place. Plus, there’s no infrastructure wrangling or surprise bills to hold you back. Try customizing a top LLM with Crusoe’s Serverless Fine-Tuning right here.

MARKETS

Why AI's money is moving down the stack

AI investing might not be all about powerful models and chips anymore. 

The blast radius of rapid AI innovation is starting to extend throughout industries, digging further and further into markets like energy and materials as the industry's so-called "five-layer cake" continues to bake. Now, investors are starting to put their money where their mouth is. 

On Monday, two early-growth stage startups announced massive funding rounds from some of AI's most prominent investment firms, including Sequoia, a16z and Khosla Ventures. The catch: neither of these startups work directly on AI models or chips. 

• Valar Atomics, a company that is building small-modular nuclear reactors, announced a $1 billion Series B funding round, with Bloomberg reporting the valuation to be $6 billion. The company has had previous partnerships with Nvidia, in which it's developing a waterless 30 megawatt AI factory.

• Mariana Minerals, a company that mines critical minerals for "modern energy, AI, and defense," raised a $310 million Series B funding round, though the valuation wasn't disclosed. In addition to supporting AI, the company's software, MarianaOS, uses AI and machine learning to cut mining execution timelines in half. 

"Critical minerals are the materials that decide whether America builds its own future or keeps depending on China to build it instead," Vinod Khosla, founder of Khosla Ventures, said in Mariana Minerals' announcement. 

These investments are the latest sign that the market is shifting their dollars deeper into the AI stack. Nuclear companies have seen significant traction in recent months, including Antares, which raised $470 million, and Commonwealth Fusion Systems, which raised $1 billion, both to support the buildout of small nuclear reactors. And on the minerals side, companies like Terra AI and EnergyX are also attracting investor attention.

It comes as no surprise that investors are pointing their funds towards these endeavors. AI has created a compute crisis that drills all the way down the stack, and existing infrastructure is not ready to meet the demand. Along with data centers creating a stark demand for energy, the crunch for compute has led to a hastened buildout of AI data centers, which require critical minerals for chips and other parts of the infrastructure. 

"For the last few years, AI has been defined by models and apps. What's changing is our understanding of where value will be created. Investors increasingly recognize that AI is fundamentally an infrastructure story," Chad Seiler, industry leader for KPMG's Technology, Media, and Telecommunications group, told The Deep View. "The companies that supply the compute, energy, networking, cooling, and industrial capacity required to scale AI may ultimately be just as important as the companies building the models themselves."

There is potentially another reason that investing in the foundational elements of AI seems sound to investors: The further down the stack you go, the more stable the investment. Things like energy and critical mineral mining are likely to have longevity and utility beyond AI. Additionally, crisis is the birthplace of invention, and AI is most certainly creating crises. That means that the industry is looking to quickly solve the energy and materials supply chain problems that AI has wrought. And as keen investors know, the best place to put your money is with the people that are solving core problems. All of this to say, even if AI doesn't deliver on its grandiose promises of complete societal transformation, the innovations that emerge as a result of trying to get there could offer immense benefits of their own.

TOGETHER WITH IBM

What comes after tokenmaxxing?

AI adoption once centered on tokenmaxxing, with organizations driving more usage across development teams. But many are discovering that activity alone does not guarantee results.

Enter valuemaxxing, a shift toward measuring AI outcomes, including fast delivery, code quality, reduced rework and business impact.

Discover why maximizing value is becoming more important than maximizing usage.

Read the full article

HARDWARE

How AI’s next power shift starts on your desk

What happens when powerful AI no longer has to live in the cloud?

In this episode of The Deep View Conversations, we talked with Mark Papermaster, CTO of AMD, about why the next major shift in AI could happen on the device sitting on your desk.

Papermaster explains how computers could soon run sophisticated models and teams of private AI agents locally, offering greater speed, security and control without recurring token costs. He also makes the case that AI will be even more transformative than the smartphone because it will be embedded across nearly every device, industry and aspect of daily life.

The conversation also covers:

• Why open ecosystems matter in the AI era

• How AI is accelerating science, agriculture and industry

• The growing energy demands of AI

• How leaders can reinvent workflows with agents

• Why local AI could reduce cloud dependence and vendor lock-in

• Papermaster’s lessons from four decades in technology

If you’re interested in less lock-in, open ecosystems, and how enterprises can run AI more efficiently and privately, this conversation offers a look at what a more distributed and secure AI future could look like.

📺 Watch on YouTube

🎧 Listen in your favorite podcast player

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm

LINKS

• Visa to acquire AI fraud platform BioCatch for 2.4 billion dollars

• Dario Amodei warns new Anthropic talent prioritizing pay over mission

• EU gains AI Act powers to fine and order changes at AI firms

• Actualyze AI emerges from stealth with $7 million seed round

• Agent security platform Zenity raises $125 million in Series C funding 

• Amazon becomes fifth company to reach $5 trillion market cap

• nexos.ai: Access 200+ AI models in one platform with automatic model routing. Get nexos.ai with 75% off. (sponsored)

• GenOffice: Genspark open-sourced its open-source AI Office for PC and Mac

• Motion: Now available in Claude, an AI agent for motion design 

• Higgsfield AI: Offering unlimited Seedance for 11 days

• Perplexity Computer: Brain now pre-fills a wiki of your projects, people, and preferences into every session

• New York Life: 2027 Technology, Data, AI & Ventures Summer Internship Program - Data Engineer Intern

• Goodfin: AI Engineer 

• Symmetry AI: Applied AI Engineer

• Capgemini: Gen AI / Agentic AI Developer

GAMES

A QUICK POLL BEFORE YOU GO