By Song Sihang
For much of the past 18 months — from the launch of DeepSeek R1 to the recent release of V4 — founder Liang Wenfeng was widely viewed as one of China’s last committed “AI idealists”: an engineer focused more on technical breakthroughs than commercialisation.
That image has begun to fade.
Reports that DeepSeek is preparing to open itself to outside investors have fuelled speculation that the Hangzhou-based group is entering a new phase. Recent market reports say battery giant CATL plans to participate in DeepSeek’s first funding round alongside potential investors including e-commerce giants Tencent, JD.com, NetEase, along with venture capital firms IDG Capital and Monolith Management. The fundraising target is reportedly about 70 billion yuan ($10.3 billion), implying a valuation of roughly $45 billion.
At almost the same time as the fundraising rumours intensified, DeepSeek announced another dramatic move: a permanent 75% cut in the price of its V4-Pro API.
Taken together, the developments suggest DeepSeek is repositioning itself. The key issue in AI is no longer simply whose model is smartest, but who can sustain soaring levels of token consumption, agent calls and inference costs.
In that sense, DeepSeek is being pushed towards a new role: an AI infrastructure provider requiring vast amounts of computing power, electricity, data centres and industrial capital.
From AI lab to infrastructure race
One thing is beyond doubt: Liang Wenfeng is not short of money.
DeepSeek’s parent, quantitative hedge fund High-Flyer, has long been one of China’s most profitable quant funds. Public data suggest it generated average returns of about 56.55% in 2025 and managed more than 70 billion yuan in assets.
That backing explains why DeepSeek initially resembled an internal AI lab more than a conventional start-up. Before R1 became a breakout success, the company showed little urgency about fundraising or monetisation. It remained relatively small, highly engineering-driven and notably restrained in public relations.
But the AI landscape has shifted rapidly. Over the past year, groups including OpenAI, Google’s Gemini, Alibaba’s Qwen, Moonshot AI’s Kimi and MiniMax have all entered a phase of relentless iteration. Since the rise of AI agents, competition has increasingly centred not only on model quality but also on token consumption.
That is the backdrop to the V4 release.
The economics of the agent era
AI agents consume vastly more computing resources than traditional chatbots. Unlike simple chatbot interactions, agents break down tasks, call external tools, verify outputs and can run for extended periods.
As a result, a single agent task may consume tens or even hundreds of times more tokens than a conventional chatbot exchange. Once token usage begins to grow exponentially, the economics of the AI industry change with it.
Previously, the key question was whose model was more intelligent. Increasingly, companies are concluding that the real winners may be those able to support massive inference demand at lower cost and with greater stability.
That helps explain why DeepSeek’s 75% permanent price cut looks far more strategic than promotional. Pricing is no longer simply about margins. It is about attracting developers, capturing future agent traffic and controlling token gateways.
DeepSeek’s strategic shift
Liang’s search for fresh capital may ultimately signal a deeper transformation in DeepSeek’s identity.
For much of its existence, the company’s primary mission was to prove it could build world-class large language models. Whether through the low-cost training methods behind R1 or V4’s technical advances — including Hybrid Attention, MegaMoE2 and adaptation to Chinese-made AI chips — DeepSeek’s strength lay in engineering capability. But recent moves suggest its priorities are evolving.
Historically, companies willing to sustain long-term price cuts have usually been infrastructure platforms. Early Amazon Web Services and later Alibaba Cloud both sacrificed short-term profits in pursuit of larger developer ecosystems and market share. DeepSeek now appears to be entering a similar phase.
As AI agents become the industry’s next battleground, model companies are increasingly competing over who can host and sustain more agent activity. That may also explain why DeepSeek’s reported investor list spans industries ranging from batteries to ecommerce and gaming.
The company is no longer simply seeking financial backing. It appears to be assembling the industrial ecosystem needed to support the next phase of AI computing.
Source:
Huxiu Technology
