DeepSeek Halts Its $1.5B Fundraise After Founder's Candid Investor Remarks Go Viral

DeepSeek Halts Its $1.5B Fundraise After Founder's Candid Investor Remarks Go Viral

On July 25, DeepSeek verbally notified prospective investors that it was suspending its second financing round — a deal that had been targeting a pre-money valuation of at least 480 billion yuan (approximately $71 billion) and sought to raise a minimum of 10 billion yuan ($1.5 billion) in fresh capital. The trigger was not a market downturn or a regulatory intervention. It was a leaked transcript.

According to Bloomberg↗, Fortune↗, and Reuters↗, unverified records of a four-hour private investor briefing held by founder Liang Wenfeng during the company's first financing round circulated widely on Chinese social media platforms, including Weibo. The transcripts, first reported by the financial outlet Yicai, contained candid assessments of China's AI hardware constraints, the country's dependence on Nvidia chips, and the compute gap separating Chinese labs from their American counterparts. Liang was reportedly frustrated by the leak, and the company moved swiftly to pause the deal process.

The episode is more than a corporate embarrassment. It is a window into the structural tensions that define DeepSeek's position in the global AI race: a lab that has achieved world-class model quality on constrained infrastructure, now navigating the pressures of public capital markets while trying to preserve the candour that made it exceptional.

What Liang Actually Said

The leaked remarks, whose authenticity has not been independently verified by major news organisations, paint a picture of a founder who is unusually direct about the limits of China's AI ecosystem — and unusually optimistic about the trajectory.

On the compute gap, Liang reportedly identified computing resources — not talent — as the primary bottleneck separating Chinese AI from the American frontier. He described DeepSeek as operating with compute resources equivalent to roughly 20,000 Nvidia H100 GPUs, approximately one-twentieth of what top US labs can deploy. That constraint, he argued, is why Chinese labs have been forced to prioritise algorithmic efficiency over brute-force scaling — a necessity that has, paradoxically, produced some of the most technically interesting model architectures in the field.

On domestic hardware, Liang's comments were characteristically precise. He acknowledged that Huawei's Ascend 950 series is approximately two years behind Nvidia's hardware in raw performance, and that it currently takes four Huawei Ascend 950 GPUs to match the output of a single Nvidia GB300. Yet he expressed strong optimism about the trajectory, reportedly stating that Nvidia is "digging its own grave" by creating a market environment that incentivises the development of domestic alternatives. He pointed to TileLang — DeepSeek's open-source, Python-based domain-specific language for GPU kernel development — as evidence that the software moat around Nvidia's CUDA ecosystem is eroding faster than most observers appreciate.

"The gap is in compute, not in people. And compute gaps close faster than talent gaps." — Liang Wenfeng, per leaked transcript (authenticity unverified)

On strategy, Liang reiterated a philosophy that has defined DeepSeek since its founding: the company does not aim to maximise commercial profits or compete as a "super app" in the manner of ByteDance or Tencent. Instead, it prioritises open-source distribution and "reasonable" profit margins — recovering hardware costs within ten months, for instance — to accelerate AGI research and foster a broader ecosystem. As TechNode reported↗, this philosophy has been consistent across the company's public communications, but hearing it stated so plainly in a private investor context apparently created friction with the narrative of a company seeking a $71 billion valuation.

The Fundraising Context

To understand why the leak caused such disruption, it helps to understand the deal's timeline and scale.

DeepSeek completed its maiden financing round in June 2026, raising approximately $7 billion↗ at a valuation of between $52 billion and $60 billion. The round was historic by any measure — the largest first external financing for a Chinese AI lab — and included backing from:

• Tencent Holdings, reportedly considering a 10 billion yuan commitment
• CATL (Contemporary Amperex Technology), the world's largest EV battery maker, with approximately 5 billion yuan
• NetEase and JD.com, two of China's largest internet platforms
• The National Artificial Intelligence Industry Investment Fund, a state-backed vehicle that signals Beijing's strategic interest in the company
• IDG Capital, Loyal Valley Capital, and several other institutional investors

The second round, now paused, was intended to capitalise on unquenched investor demand and fund the next phase of infrastructure buildout: proprietary data centres, custom AI inference chips, and expanded agentic AI capabilities. The target valuation of $71–74 billion represented a roughly 40% step-up from the June close in under two months — an aggressive ask that required a clean narrative.

The leaked transcript complicated that narrative. Investors who had been told a story of relentless technical progress were now reading candid admissions about hardware constraints and a two-year performance gap with Nvidia. The company's decision to pause rather than push through reflects a calculation that the deal, under these conditions, was not worth closing.

The pause does not mean the round is dead. Sources familiar with the matter describe negotiations as "fluid," and the company has not ruled out resuming the process once the media cycle moves on.

DeepSeek V4: The Technical Foundation

The fundraising drama should not obscure what DeepSeek has actually built. The DeepSeek V4 model card↗ documents a genuinely impressive technical achievement, and the V4-Pro weights on Hugging Face↗ have been downloaded extensively by developers worldwide since the April 2026 launch.

Architecture and Specifications

The V4 family consists of two variants:

• DeepSeek V4-Pro: 1.6 trillion total parameters, 49 billion active per token, using a Mixture-of-Experts (MoE) architecture with 61 transformer layers. Achieves 80.6% on SWE-bench Verified and 93.5% on LiveCodeBench.
• DeepSeek V4-Flash: 284 billion total parameters, 13 billion active per token, designed for high-throughput, low-latency applications with 43 layers.

Both models feature a 1-million-token context window and are released under the MIT license, making them freely usable for commercial applications without geographic restrictions.

Key Architectural Innovations

The V4 series introduces several technical advances that distinguish it from the V3 generation:

• Hybrid Attention (CSA + HCA): Interleaves Compressed Sparse Attention and Heavily Compressed Attention, reducing single-token inference FLOPs for V4-Pro to 27% of V3.2 and KV cache usage to 10% — a dramatic efficiency gain that makes the model practical to serve at scale.
• Manifold-Constrained Hyper-Connections (mHC): Replaces standard residual connections with a mechanism that constrains weight updates to the Birkhoff polytope, preventing gradient explosion during the 32-trillion-token pre-training phase.
• Muon Optimizer: Replaces AdamW with a Momentum + Orthogonalization optimizer that achieves faster convergence and improved stability.

The model also supports three reasoning modes — Non-think, Think High, and Think Max — allowing developers to trade compute cost against output quality depending on the task.

The Huawei Ascend Integration

One of the most strategically significant aspects of V4 is its adaptation for Huawei Ascend hardware. As Reuters reported↗ at launch, V4 is the first frontier-class Chinese model developed with deep hardware co-design for the Ascend 950 series. A Huawei-led research group subsequently reported to Tom's Hardware↗ that it had completed full-parameter post-training for V4-Pro on a cluster of 1,000 Ascend 910C chips — a significant proof point for domestic silicon's training-class capabilities.

The efficiency gains from this co-design are substantial: reports indicate a 75% reduction in AI inference costs compared to running equivalent workloads on Nvidia hardware, driven by custom kernels that interact directly with the Ascend architecture rather than going through generic frameworks.

TileLang: The Software Layer That Changes the Equation

Liang's comments about the erosion of Nvidia's CUDA moat point to a project that deserves more attention than it has received outside specialist circles: TileLang.

Available on GitHub↗ under an open-source license, TileLang is a Python-based domain-specific language for writing high-performance GPU, CPU, and accelerator kernels. Built on Apache TVM, it allows developers to generate optimised kernels — including GEMM operations, FlashAttention, and FlashMLA — across multiple hardware backends:

• Nvidia GPUs: H100, A100, V100, RTX 4090/3090, with support for Auto TMA and WGMMA
• AMD GPUs: MI250 and MI300X, with performance parity on specific workloads
• Huawei Ascend NPUs: AscendC/NPU IR backends
• Apple Metal and WebGPU

In April 2026, DeepSeek released TileKernels↗, an open-source library of production-grade GPU kernels built entirely in TileLang, covering MoE routing, FP8/FP4 quantisation, and Manifold HyperConnection kernels. The practical implication is that a developer can now write a kernel once in TileLang and deploy it across Nvidia, AMD, and Huawei hardware — reducing the software-level lock-in that has historically made Nvidia's CUDA ecosystem so sticky.

This is not a theoretical capability. It is the infrastructure that makes DeepSeek's domestic hardware strategy credible, and it is why Liang's comment about Nvidia "digging its own grave" is less hyperbolic than it might appear.

The IPO Rush and What It Means

The fundraising pause occurs against a backdrop that Fortune described↗ as "the Great Chinese AI IPO Rush." Multiple Chinese AI labs are simultaneously accelerating plans to access public capital markets:

• Moonshot AI is preparing for a Hong Kong Stock Exchange listing within six months, targeting a valuation of up to $50 billion following the Kimi K3 launch
• MiniMax and Z.ai (Zhipu AI) debuted on the Hong Kong exchange in early January 2026
• DeepSeek is targeting a Shanghai STAR Market listing, with internal goals to file as early as late 2026 and a potential debut in Q2 2027

The choice of venue is itself a strategic signal. Moonshot's Hong Kong listing reflects its international investor base and offshore VIE structure (now being dismantled). DeepSeek's preference for the STAR Market — China's Nasdaq-style exchange for technology companies — aligns it explicitly as a domestic national champion, a positioning that typically secures state support and preferential treatment in government procurement.

Why the Leak Matters Beyond DeepSeek

The viral spread of Liang's comments is not simply a corporate communications failure. It exposes a structural tension that every Chinese AI lab navigating public markets will face: the gap between the narrative required to justify frontier valuations and the operational reality of building world-class AI under hardware constraints.

Investors in a $71 billion round need to believe that the company can sustain its technical edge. Liang's candid acknowledgment that domestic chips are two years behind Nvidia, and that DeepSeek operates with one-twentieth the compute of leading US labs, is precisely the kind of information that complicates that belief — even if the underlying argument (that algorithmic efficiency can compensate for hardware deficits) is well-supported by the company's track record.

The US policy environment adds another layer of complexity. As TechCrunch reported↗, the US government is actively evaluating restrictions on Chinese open-weight models, with Treasury Secretary Scott Bessent signalling potential sanctions over allegations of "industrial-scale distillation." A coalition of over 25 companies — including Nvidia, Microsoft, Meta, and Hugging Face — signed an open letter opposing broad restrictions, but the regulatory uncertainty is real and will be a factor in any IPO prospectus.

Practical Takeaways for Developers

Despite the fundraising drama, DeepSeek's models remain fully available and unchanged:

• DeepSeek V4-Pro and V4-Flash are accessible via the DeepSeek API↗, with V4-Pro priced at competitive rates that undercut most Western frontier models
• Both models are available for self-hosting under the MIT license — no geographic restrictions, no usage fees for on-premises deployment
• The Hugging Face model page↗ provides weights, model cards, and integration guides
• TileLang and TileKernels are open-source and actively maintained, providing a path to hardware-portable kernel development

The fundraising pause has no operational impact on the API or the open-weight releases. DeepSeek completed a $7 billion round in June; it is not capital-constrained in the near term. The pause is a deal-process decision, not a sign of financial distress.

What Comes Next

DeepSeek's IPO preparations are continuing in parallel with the fundraising pause. The company is reportedly consulting with banking and accounting advisors, with an internal target to submit its STAR Market filing by end of 2026. Whether the second private round resumes — and at what valuation — will depend on how quickly the media cycle moves on and whether Liang chooses to address the leaked comments publicly.

The deeper question is whether the transparency that defines DeepSeek's research culture — the detailed technical reports, the open-weight releases, the candid investor briefings — is compatible with the opacity that public markets typically demand. Liang's leaked remarks were not reckless; they were honest. The problem is that honesty about hardware constraints and a two-year performance gap with Nvidia does not fit neatly into a $71 billion valuation story.

DeepSeek has built something genuinely remarkable: frontier-class models on constrained infrastructure, distributed freely to the world, with the technical documentation to prove every claim. The fundraising pause is a reminder that the transition from research lab to public company involves a different kind of constraint — one that no amount of algorithmic efficiency can optimise away.