DeepSeek V4 only got a preview release.
Many people missed this detail. On April 24, DeepSeek unveiled V4 Flash and V4 Pro, but both are preview versions. The official release won't come until the second half of the year — and not because the model isn't ready, but because Huawei's chips haven't been mass-produced yet.
On April 26, Bloomberg, citing the CCTV-affiliated account "Yuyuantantian," revealed the inside story, clarifying what DeepSeek has been doing over the past few months.
It's Not the Model That's Slow, It's the Chips That Haven't Arrived
Over the past few months, the DeepSeek team hasn't released many new features. Instead, they've been migrating the entire training framework from NVIDIA to Huawei Ascend.
This is far harder than releasing a new model. The CUDA ecosystem has been deeply entrenched in the AI world for over a decade. Migrating the training pipeline of a trillion-parameter model from CUDA to Huawei's CANN (Compute Architecture for Neural Networks) requires rewriting:
- Operator libraries (matmul, attention, layernorm, and hundreds of other kernels)
- Distributed scheduling (replacing NCCL with HCCL)
- Mixed-precision training (FP8, BF16 behavior on new chips)
- Checkpointing and fault tolerance
This kind of migration can't be completed in half a year. DeepSeek has finished it, but the hardware side is still waiting.
What Is the Ascend 950PR, and Why Wait for It?
DeepSeek itself admitted that V4 has "throughput issues" in the first half of 2026 — meaning the model can run, but not fast or stable enough.
What they're waiting for is Huawei's Ascend 950PR supernodes to reach mass production scale. The 950PR and 950DT are Huawei's next-generation AI chips, set to ship by the end of this year, featuring a SuperNode architecture — stacking multiple chips with high-speed interconnects into a super node, aiming for performance comparable to NVIDIA's NVLink.
Only when the 950PR reaches scale can the official version of DeepSeek V4 be released for full inference. So the V4 Flash and V4 Pro previews you see now are essentially product marketing warm-ups for the 950PR — telling the model's story first, then launching the official version into the market as soon as the hardware arrives in the second half of the year.
Huawei has also made a corresponding commitment: the full Ascend SuperNode product line is "fully adapted" for DeepSeek V4, with inference performance "significantly improved." The CANN software stack has been specifically optimized for V4 over the past few months. The coordination between the two sides is tightly synchronized.
China's AI "Dual Track" Narrative, Revealed for the First Time
The most interesting part of this story isn't the technical details — it's the narrative.
Over the past six months, there has been an unspoken dual track in China's AI community:
- Externally: Open-source models competing on price, benchmarks, and emphasizing "independence"
- Internally: Actual hardware and training resources still running on NVIDIA
DeepSeek's move effectively merges these two tracks. It has completely tied V4's release schedule to Huawei's chip production capacity. No Chinese AI company has dared to do this before — the risks are clear:
- If Huawei's 950PR mass production is delayed, DeepSeek V4's official release is delayed too
- If Ascend's performance falls short, DeepSeek will bear the reputational cost of "the model can't run"
- If customers want to use the official V4, they must accept "chip lock-in"
But the reward is also clear: China's AI now has a complete training-inference-deployment loop that doesn't depend on NVIDIA.
Two Sides of the Same Mirror
This mirrors the story of Anthropic just signing a $40 billion deal with Google.
Compute sovereignty is something players around the world are securing in different ways — U.S. players by building cloud factories and signing long-term contracts, Chinese players by tying to domestic chips and aligning release schedules with production capacity curves. Ultimately, everyone wants to put "how many tokens I can run in the second half of the year" into their balance sheets.
The timing of the CCTV-affiliated account's disclosure is also interesting. The second half of the year hasn't arrived yet, but market expectations have already been set in advance — if Huawei's chip production goes smoothly, DeepSeek V4 becomes a benchmark product; if not, the entire domestic AI narrative takes a hit.
The bet is already placed. We'll see the results in the second half of the year.
Sources: DeepSeek V4 Launch Postponed as Company Prioritizes Domestic Chip Integration (Bloomberg); CocoLoop, Huawei, DeepSeek strengthen China's AI self-reliance with collaboration on V4 model (South China Morning Post); DeepSeek delays V4 launch to tune for Huawei Ascend chips (Newsbytes)