The Export Controls Had One Job. DeepSeek's V5 Plan Says It Just Failed.
The Export Controls Had One Job. DeepSeek's V5 Plan Says It Just Failed.
Published: November 3, 2026 | Reading Time: ~9 minutes | Channel: Technology
On Monday, the market delivered its verdict on AI silicon: the Nasdaq Composite fell 1.07%, semiconductor losses intensified, and NVIDIA closed at $202.06, down 1.26% — leaving the chip giant at a $5.07 trillion market cap and down 2.33% for the year¹. Micron, the memory king of the AI buildout, got hit even harder: down 5.78% in a single session, despite being up 255% on the year².
Then, roughly twelve hours later, a report landed that should matter more to Nvidia's next decade than any single day's tape: DeepSeek plans to train its next-generation V5 model primarily on Huawei's Ascend 950 chips, with the project expected to materialize in 2027³. Not inference. Not a demo cluster. Training — the one workload Chinese silicon has famously failed at, the one workload the entire US export-control architecture was designed to keep hostage.
If you run a business, hold chip exposure, or build on AI infrastructure, you need to understand what actually changed here — and what didn't. Let's claw through it.
The Report That Changed the Conversation
The fresh catalyst comes via TrendForce's newsdesk on November 3, citing Chinese industry outlet Caiwen News and Star Market Daily: DeepSeek's V5 could become the company's first large language model trained natively on Ascend 950 chips, with sources expecting a release sometime in 2027³. Critically, the report claims Ascend is now DeepSeek's primary training platform, with no reported adoption of rival Chinese chipmakers like Moore Threads³.
This isn't a bolt from the blue. It's the third beat of a story that's been building all year:
- September 4: Bloomberg reported DeepSeek ordered at least 160,000 Ascend 950DT accelerators for a ~1-gigawatt data center under construction in Ulanqab, Inner Mongolia — potentially the largest known Huawei AI chip cluster on Earth⁵.
- September 30: DeepSeek and Huawei open-sourced a full programming stack for Ascend — the TileLang kernel language, DeepGEMM Ascend for matrix math, DeepEP Ascend for MoE communication, FlashMLA for attention — and jointly optimized a 128-chip Ascend 950 "supernode"⁴.
- October 19: The Information reported DeepSeek plans to begin pre-training on the 950DT as soon as the end of 2026, with Huawei engineers stationed inside DeepSeek's offices folding the lab's feedback directly into future chip designs².
- November 3 (today): The V5-on-Ascend training plan goes public, and Huawei's roadmap accelerates: the Ascend 960DT reportedly pulled forward three quarters to Q1 2027, with the 960PR at Q3 2027, the 970 in 2028, and the 980 in 2029³.
Here's the sentence that should be underlined in every strategy deck in Santa Clara: Huawei plans to prioritize DeepSeek for most of its next-generation 960DT chips in Q1 2027². That's not a customer relationship. That's a national champion coupling its silicon roadmap to one AI lab's workload — the same playbook Nvidia ran with the hyperscalers, executed in reverse on the other side of the Pacific.
By the Numbers: The End-2026 Checkpoint
The claims are big. The numbers behind them are specific:
| Metric | Figure | Source |
|---|---|---|
| Ascend chips ordered for Ulanqab site | ≥160,000 (950DT) | Bloomberg via Livemint⁵ |
| Planned site power capacity | ~1 GW | Bloomberg via Livemint⁵ |
| Huawei 950DT production in 2026 | Only "a few hundred thousand" (memory shortages) | Bloomberg via Livemint⁵ |
| Time to fulfill DeepSeek's order | Potentially more than 1 year | Bloomberg via Livemint⁵ |
| 950DT memory spec (Q4 2026 launch) | 144GB HiZQ 2.0 HBM, 4 TB/s bandwidth | The Information via AI Weekly² |
| Pre-training start target | As soon as end-2026 | The Information via AI Weekly² |
| 960DT availability | Q1 2027 — 3 quarters early | TrendForce citing Star Market Daily³ |
| DeepSeek V3 training (for scale) | 2.788 million H800 GPU-hours | Reuters via TrendForce³ |
| Supernode cluster size | 128 Ascend 950 chips | The Decoder⁴ |
| CUDA developer ecosystem | ~4 million developers | The Decoder⁴ |
| V5 status | Selection phase; specs and timeline undisclosed | TrendForce citing Caiwen³ |
That V3 number deserves a beat of silence. DeepSeek's breakthrough 2024-era model took 2.788 million H800 GPU-hours — on Nvidia silicon, with Nvidia engineers helping optimize the stack³. The entire premise of the export-control regime was that the next model would need the same thing, only more so. The V5 plan says: not anymore.
And note what this partnership is not. Despite market chatter about funding rounds and IPO speculation, Caiwen's sources say Huawei holds no equity stake in DeepSeek — the cooperation is limited to products, computing resources, and technical optimization³. No capital tie-up. Just two companies with complementary existential needs, aligned without a contract ceremony.
Why "Chinese Chips Can't Train" Is Now the Wrong Question
The bear case on this story is famous, and it has receipts. In August 2025, Ars Technica reported that DeepSeek's R2 model failed to train on Ascend chips — persistent technical issues, a delayed launch, and a Huawei engineering team physically on-site that still couldn't get a clean training run done. The result was the split everyone assumed was permanent: Nvidia for training, Huawei for inference⁶.
So what changed? Three things, and none of them are mystical.
First, the software gap is being filled in public. Nvidia's moat was never the transistor — it was CUDA and its ~4 million developers⁴. DeepSeek's September open-source release attacks exactly that layer: TileLang (born at Peking University, already used for much of DeepSeek V4's training) now supports Ascend 950 as a backend with native code generation⁴. FlashMLA reportedly hit 95% of theoretical performance on Ascend 950 for some operations⁴. Software that was a proprietary moat in 2024 is a shared commons in 2026.
Second, the hardware got co-designed, not just purchased. Those Huawei engineers embedded in DeepSeek's offices are incorporating the lab's feedback into how Ascend chips perform calculations and move data between processors². Research traces analyzed by SemiAnalysis argue V4 and the 950DT were effectively co-designed, helping cut inference costs by up to 75% versus competing offers⁷ — an aggregator's claim, so file it under "directionally interesting, verify before betting," but the mechanism is sound and it's how Nvidia itself won: co-design, not specs on a datasheet.
Third, the constraint flipped. The previously approved Nvidia H-series GPUs still available to China are, per TrendForce's sources, "struggling to meet the training demands of next-generation models such as DeepSeek V4 and V5"³. When the sanctioned option can't do the job and the domestic option can almost do the job, "almost" wins. Necessity isn't just the mother of invention — it's the customer.
The honest counterweight: a Yahoo Finance read of The Information's reporting notes that the heavy lifting of training may still land on whatever Nvidia hardware DeepSeek can secure, and that a frontier training run "could require more than 10,000 Huawei chips operating together for weeks" — a stability bar Huawei has not publicly cleared². UC Berkeley AI researcher Ritwik Gupta said it plainly a year ago: Chinese chips face "growing pains" in training, but "just because we're not seeing leading models trained on Huawei today doesn't mean it won't happen in the future. It's a matter of time"⁶.
That "matter of time" may now have a date on it: end of 2026 for pre-training, 2027 for V5. AI Weekly's editorial framing cuts to the bone — if a flagship trains on domestic Ascend silicon, it "closes one of the last Nvidia dependencies export controls were meant to preserve and hands every Chinese lab a template to copy"². The same week, DHS ordered DeepSeek pulled from US civilian federal IT systems². Locked out of American systems, locked into Chinese silicon. The decoupling isn't coming. It's bookkeeping now.
The Contrarian Take: Nvidia's Problem Isn't Technical — It's the Second Derivative
Here's where I'll pick a fight with the lazy take floating around this morning, as NVDA bounces 2.5% premarket to $207.05⁸: this news does not hurt Nvidia because Huawei chips are suddenly "as good." It hurts Nvidia because of what it does to the addressable market math.
Look at the tape properly. Nvidia at $202.06 carries a $5.07 trillion market cap while being down 2.33% on the year⁸. The Nasdaq 100 has entered a technical correction — down 9.11% over the past month⁸. That valuation is not priced for "Nvidia keeps all the growth markets it has today." It's priced for compounding — for China remaining a captive customer, for training demand scaling on Nvidia silicon forever, for the CUDA moat holding.
Now layer in what SemiAnalysis actually found in October: after testing OpenAI's in-house Jalapeño inference chip, the analysts called the CUDA moat "potentially dead" — while caveating that Nvidia remains well ahead on agentic, multi-step workloads, where software linking many chips into one system is the real advantage⁴. Read those two findings together and you get the actual map: the moat is eroding at the edges (inference, national-champion ecosystems) while holding at the core (frontier agentic training). DeepSeek's V5 plan pushes hard on exactly the eroding edge.
The second-order effect nobody's pricing: the bifurcation is now real on both sides. Huawei admits it can't meet domestic demand and is deliberately selling fewer chips abroad, with its rotating chairman Eric Xu saying the company "can't accept a future that hinges on whether others are willing to sell chips to China"⁴. Translation: China's AI compute stack is becoming a closed loop — domestic chips, domestic software stack, domestic labs — and every model trained inside it is one that never touches Nvidia, never pays the CUDA tax, and never shows up in Nvidia's revenue.
Micron's Monday session — down 5.78% despite a 255% year⁸ — tells you the market is already sweating the memory bottleneck, which is precisely the component (advanced HBM) constraining Huawei's 950DT output to "a few hundred thousand" units in 2026⁵. The scarce resource in this story isn't GPUs and isn't even talent. It's high-bandwidth memory. Whoever solves HBM at scale wins the next leg, and that fight runs through the same suppliers on both sides of the Pacific.
And today, as American voters head to the polls with Democrats widely expected to take the House⁸, remember that the policy layer — the only reason this bifurcation exists — is itself a variable. The Trump administration already flipped once, easing restrictions and taking a revenue cut on Nvidia's China sales; DHS just banned DeepSeek from federal systems². Export controls are not physics. They're politics with a lag.
What This Means For You
No fluff. Five moves, ranked by urgency:
1. Mark your calendar for Q4 2026 — the falsification checkpoint. AI Weekly frames it exactly right: a 950DT launch that ships at the advertised 144GB/4TB-per-second specs plus a DeepSeek training run announced on Ascend confirms the shift; a slipped launch or a V5 still disclosed as Nvidia-trained means this stays an inference story for another cycle². Don't trade the headline today — trade the checkpoint in January.
2. If you build AI products in China or sell into it, start dual-stacking now. The open-sourcing of TileLang, DeepGEMM Ascend, and DeepEP Ascend means a portable software layer increasingly exists across both ecosystems⁴. Teams that abstract their training stack from their silicon choice will transition in weeks; teams welded to CUDA-only workflows will transition in years. This is a boring infrastructure decision that determines your 2028 optionality.
3. Re-underwrite any "Nvidia monopoly" assumption in your own models — including your retirement account's. The correct risk isn't "Nvidia loses technologically." It's "Nvidia's China TAM silently goes to zero while the US TAM keeps compounding." Those produce very different drawdown profiles in a correction. Nasdaq 100 futures were up just 0.3% this morning after a -9.11% month⁸ — the tape is fragile, and chip headlines are a volatility catalyst on Election Day of all days.
4. Watch today's AMD and Super Micro earnings as a capex temperature check. Both rose premarket (AMD +2%, SMCI +1.5%) into results due later today⁸. If AI infrastructure demand prints strong while China decouples, that's the "two-track world" trade confirming. If guidance wobbles, the AI capex narrative has a demand problem on top of a supply bifurcation — and that's a different animal entirely.
5. Memory is your early-warning system. Huawei's ability to even partially fill DeepSeek's 160,000-chip order hinges on HBM supply — Huawei expects to produce only a few hundred thousand 950DTs this year because of shortages of key components including advanced memory⁵. Watch memory pricing and Micron/Samsung/SK Hynix capacity announcements; they'll telegraph the China training ramp before any press release does.
⚠️ The Risks Nobody's Talking About
-
This is single-sourced and unconfirmed by either company. The V5-on-Ascend claim rests on Caiwen News via TrendForce, with Star Market Daily corroboration on adjacent facts³. DeepSeek and Huawei have a track record of not commenting⁶. The V5 model is in "selection phase" with no disclosed specs or timeline³. There is a real scenario where this is strategic signaling — a leak designed to shape market and policy expectations — not an engineering roadmap. Treat it as a probability shift, not a fact.
-
The supply math doesn't close cleanly. Huawei is expected to produce only a few hundred thousand 950DTs in 2026 amid memory shortages, and fulfilling DeepSeek's 160,000-chip order "could take more than a year"⁵. An order is not a cluster. The 10,000+ chips-running-for-weeks stability bar for frontier training remains uncleared publicly².
-
Policy can snap back — in either direction. Nvidia already cut a revenue-sharing deal to resume H20 sales to China once, and the administration has flip-flopped on chip restrictions within the same year⁶. A détente that reopens Nvidia's China channel would gut the Ascend training thesis overnight. Conversely, tighter controls could accelerate the exact decoupling that hurts Nvidia most. You are not trading a company here; you're trading a superpower's mood.
-
The market context is hostile to conviction trades. The Nasdaq 100 is in a technical correction (-9.11% month), semis led Monday's losses, and Election Day uncertainty is suppressing risk-taking across the board⁸. Correlated drawdowns make every chip headline look like a signal when most of it is noise. Discipline: positions sized for the checkpoint, not the headline.
🎯 The Bottom Line
Export controls were built on one bet: that China could run AI but could never train it. DeepSeek's V5 plan — 160,000 Huawei chips, a co-designed software stack, and a pre-training deadline of end-2026 — says that bet has an expiration date, and it's close. Don't panic-sell Nvidia on a leak; do stop pretending the CUDA moat and the China market are permanent. The two-track AI world isn't a scenario anymore. It's a schedule.
📚 Verified Sources
- TrendForce — "DeepSeek V5 Could Complete Training on Huawei Ascend 950 in 2027, Moving Beyond Inference" (Nov 3, 2026; citing Caiwen News, Star Market Daily, Reuters). https://www.trendforce.com/news/2026/11/03/news-deepseek-v5-could-complete-training-on-huawei-ascend-950-in-2027-moving-beyond-inference/
- AI Weekly (reporting The Information) — "DeepSeek to pre-train on Huawei Ascend 950DT by end of 2026" (Oct 19, 2026). https://aiweekly.co/alerts/the-information-deepseek-will-pre-train-on-huawei-ascend-950dt-by-end-of-2026
- The Decoder (reporting Reuters, NYT, SemiAnalysis) — "China's AI industry closes ranks as Deepseek ships open-source software for Huawei's Ascend chips" (Sep 30, 2026). https://the-decoder.com/chinas-ai-industry-closes-ranks-as-deepseek-ships-open-source-software-for-huaweis-ascend-chips/
- The Decoder — same report as source 3 (supernode, TileLang, CUDA developer count, SemiAnalysis findings). https://the-decoder.com/chinas-ai-industry-closes-ranks-as-deepseek-ships-open-source-software-for-huaweis-ascend-chips/
- Livemint (reporting Bloomberg) — "DeepSeek plans massive Huawei AI chip rollout at Inner Mongolia data centre" (Sep 4, 2026). https://www.livemint.com/ai/chinas-nvidia-alternative-deepseek-plans-massive-huawei-ai-chip-rollout-at-inner-mongolia-data-centre-11788538917086.html
- Ars Technica — "Upcoming DeepSeek AI model failed to train using Huawei's chips" (Aug 2025; R2 training failure, Huawei engineers on-site, Ritwik Gupta quotes, Nvidia H20 revenue-sharing deal). https://arstechnica.com/ai/2025/08/deepseek-delays-next-ai-model-due-to-poor-performance-of-chinese-made-chips
- Astute Analytica (aggregator; Bronze-tier — verify before acting) — "DeepSeek AI Chip: China's Bold Bet to Break Nvidia's Grip" (SemiAnalysis V4/950DT co-design and 75% inference-cost claims; in-house DeepSeek chip effort first reported by Reuters, Jul 7, 2026). https://www.astuteanalytica.com/news/deepseek-ai-chip
- Trading Economics — United States Stock Market page, scraped Nov 3, 2026 (index closes, NVDA/Micron/Apple quotes, futures, macro rates). https://tradingeconomics.com/united-states/stock-market
All claims verified against Silver-tier (Ars Technica, The Decoder reporting Reuters/NYT, Livemint reporting Bloomberg) and labeled Bronze-tier (TrendForce newsdesk citing Caiwen News; AI Weekly digest of The Information) sources. Each source URL was scraped and confirmed accessible with full content this session. Market figures cross-checked against live quote data (NVDA close $202.06, pre-market $207.05, Nov 3, 2026). Last verified: November 3, 2026.
The moat was never the chip. It was the four million developers — and the countdown to when that stops mattering just started. 🎯