From GPUs to MaaS: Phancy’s CSO on Why Full-Stack Beats Pure Neocloud Compute

Much of the AI conversation is concentrated on frontier models and applications. But what happens in between—the work of making fragmented compute, model libraries, and enterprise systems function together—is just as important to the economics of adoption. In this episode, I speak with Maryann Tseng, Chief Strategy Officer of Phancy 06682.HK (formerly known as 4Paradigm), about the infrastructure middle layer that helps enterprises deploy AI more reliably and at lower cost. We unpack Phancy’s HAMi and RISE vGPU technologies, ModelHub, and its approach to improving GPU utilization and adapting models across heterogeneous hardware.

Maryann explains how Phancy works with, rather than simply competes against, major cloud providers; why it sees a different growth trajectory for public-cloud Model as a Service (MaaS) than for its more established enterprise business; and how it thinks about SMEs, international expansion, and distribution partnerships. Underlying our discussion is a broader question: will the next stage of AI adoption be driven primarily by more compute and better models, or by making much better use of the hardware and technology already available?

*apologies for the audio

The AI Proem Podcast is part of the AI Proem newsletter, which has ~13k followers globally. To learn more about China AI, the business of AI, and how AI is impacting society, please check out the ⁠newsletter here ⁠and more ⁠insightful conversations here.

Chapters

* 0:00 – Intro: Maryann Tseng’s role and what Phancy does

* 4:32 – Who Phancy’s enterprise customers are

* 6:09 – Why AI infrastructure matters, in plain terms

* 8:36 – Enterprise bottlenecks: hardware limits, cost and reliable scaling

* 10:59 – Phancy vs. traditional and new cloud providers

* 13:35 – A crowded middle layer: competition and Hami vGPU’s market share

* 15:59 – Flexibility, long enterprise relationships and trust

* 19:16 – “Token factory” vs. Models-as-a-Service (MaaS)

* 21:47 – Which industries are adopting AI fastest (entertainment, SMEs)

* 24:01 – Competing with Alibaba, Huawei and Baidu as an independent platform

* 26:16 – 12–18 month outlook: private vs. public cloud growth

* 34:43 – Industry challenges: energy, chips and efficiency

* 38:06 – A non-consensus view: efficiency over more spending

AI-generated transcript, for reference only

Grace Shao (00:01)

Maryann, thank you so much for joining us today. I’m really excited to be speaking with you. To start, tell us about your role and help us understand what Phancy does today. For many of our listeners, the company was previously known as Fourth Paradigm, and it has gone through a number of iterations over the years. Help us understand your main businesses, your core revenue drivers, and how the company has evolved.

Maryann Tseng (00:30)

I’m currently the Chief Strategy Officer at Phancy. I help our co-founders realign our longer-term strategy, identify long-term business partnerships, and translate our technology into more sustainable growth. Expanding into international markets and building awareness of Phancy internationally is also one of our key priorities this year.

Phancy was founded in 2014. Today, we provide AI infrastructure middleware, connecting heterogeneous compute environments and resources with different models to support a variety of enterprise solutions and workflows. We have two important products: HAMi vGPU and RISE vGPU. They improve GPU usability and utilization.

Another key product is ModelHub, a community with a library of more than 300,000 models and capabilities for model adaptation and optimization across heterogeneous hardware environments. We commercialize these technologies through our AI Platform business, supporting enterprise deployments in public clouds, private deployments, and MaaS, or Model as a Service, which delivers inference services through APIs.

Grace Shao (02:15)

Maryann, as we mentioned earlier, the company has gone through several iterations. You now have a few different businesses. What has really changed from the pre-IPO period to being listed on the Hong Kong stock exchange?

Maryann Tseng (02:32)

What we’ve always done, from day one, is serve enterprise customers and understand the needs of companies across different industries. When we founded the company, we realized that enterprises wanted to use more AI products internally, but finding suitable AI models and integrating them into their existing workflows and systems was a challenge. That is why we built our AI Platform: to help enterprises deploy AI within their existing workflows and improve internal efficiency.

Today, there are millions of models, but compute and hardware remain severely constrained. How can we help enterprises make use of heterogeneous chips so that these models run efficiently and effectively? With agentic AI, we’re seeing more workflows broken down into smaller, more fragmented tasks.

Our vGPU technology virtualizes GPU resources to make compute more usable and improve utilization. Together with ModelHub and its library of more than 300,000 models, we can help enterprises deploy models effectively across different heterogeneous chip environments and integrate them into their systems with higher efficiency.

Grace Shao (04:32)

Who are your main customers, then?

Maryann Tseng (04:37)

We focus on enterprise customers, including major telecom operators, financial institutions, electricity providers, and power producers. These are some of the key industries we’ve served over the past decade.

Grace Shao (05:58)

Maryann, let’s talk about what’s top of mind in the industry right now. Much of the AI conversation focuses on models and applications. But what you do could be considered the less glamorous, yet much-needed, work in between. Help us understand why solid AI infrastructure matters and what it actually includes. Break down the technicalities for us in layman’s terms. What exactly are you doing, why is it so important, and why are so many customers paying for this service?

Maryann Tseng (06:45)

If we take a step back, AI infrastructure determines whether models can be delivered as reliable, affordable services at scale. It includes GPU virtualization, heterogeneous resource management, scheduling, model adaptation, and inference optimization. Both usability and utilization matter. How can we use the available hardware to run the required workloads, and how can we make sure that hardware is being used productively?

With agents, we’re seeing workflows divided into much smaller tasks. Allocating an entire GPU to a task that only needs a fraction of its capacity creates a lot of waste. Our vGPU technology enables fine-grained slicing and concurrent execution, while ModelHub helps select and adapt models to run effectively on the underlying chips.

This middle layer is what we do. It connects models and hardware and, more importantly, helps enterprises deploy AI more efficiently. Sometimes we see utilization improve from around 30% to 65–85%. That helps companies bring down deployment costs. Even though it may sound less glamorous because it’s a middle layer, we believe it’s a critical component for mass AI deployment.

Grace Shao (08:36)

That makes a lot of sense. What I’m hearing from enterprises is that adoption speed is one challenge. There’s a lot of bureaucracy in older companies. The second is the high cost of adopting frontier models. Beyond cost, what are the biggest challenges you’re seeing among your clients in turning AI into something they can reliably scale? What other bottlenecks do they face?

Maryann Tseng (09:06)

Hardware constraints are a major challenge. How do we maximize the existing supply chain and make the best use of the resources we already have? This is something we’re actively helping enterprises solve.

Going back to ModelHub and vGPU, because hardware is constrained, we need to intelligently combine different types of chips in heterogeneous environments and break workflows into much smaller tasks, sometimes into a hundred pieces. At the same time, we need to combine th…

Get full access to AI Proem at aiproem.substack.com/subscribe