#DeepSeek V4 Locked Out Nvidia and AMD — What HR Teams Using AI Hiring Tools Need to Know

TL;DR (Direct Answer): DeepSeek withheld pre-release access to its upcoming V4 model — a 1-trillion parameter system with a 1-million-token context window — from Nvidia and AMD while giving Huawei a multi-week optimization head start. A senior Trump administration official alleges V4 was trained on Nvidia Blackwell chips inside China in violation of US export controls. DeepSeek's previous models triggered the largest single-day stock market wipeout in history ($589 billion from Nvidia's market cap in one day) and have been downloaded 75 million times on Hugging Face. For HR teams, the practical questions are: which AI models power your screening and assessment tools, and what are the compliance implications of each answer?

#What DeepSeek V4 Actually Is

On February 25, 2026, Reuters reported that DeepSeek withheld pre-release access to its V4 flagship model from Nvidia, AMD, and other US chipmakers — while giving Huawei and other Chinese chip manufacturers a multi-week head start to optimize their hardware.

V4 is a significant architectural step up from anything DeepSeek has released before:

Three new architectural innovations are reported: Manifold-Constrained Hyper-Connections (mHC), Engram conditional memory, and DeepSeek Sparse Attention. Internal benchmarks claim 80%+ SWE-bench performance at 10x to 40x lower inference cost than Western competitors. DeepSeek plans to open-source the weights under Apache 2.0 — maintaining the approach that drove 75 million Hugging Face downloads for earlier models.

The model reportedly runs on dual RTX 4090 GPUs. For HR software vendors evaluating self-hosted AI options, that accessibility matters: a frontier-class model running on two consumer GPUs means any mid-size company with basic hardware can run powerful AI screening workloads without cloud API costs.

#Why Locking Out Nvidia Matters for Your HR Tools

Pre-release model access is an industry norm with a practical purpose. Chipmakers need time to tune drivers, compilers, and software stacks so a model runs efficiently on their hardware from day one. Without that access, hardware runs the model with generic settings instead of tuned ones. The performance gap can be 20 to 40% or more.

DeepSeek previously worked closely with Nvidia's technical teams on its earlier models. That relationship is now explicitly severed for V4.

By giving Huawei a weeks-long head start, DeepSeek ensures V4 will run best on Chinese hardware at launch. Nvidia and AMD engineers start reverse-engineering optimizations that Huawei has already locked in.

For HR software vendors who decide to self-host DeepSeek V4 — attractive because of V4's claimed 10x to 40x lower inference cost — hardware choice determines whether they get optimized or degraded performance at launch. That performance gap flows directly into your tool's screening accuracy and response quality.

"By prioritizing Huawei's Ascend chips for optimization, DeepSeek is accelerating the development of a parallel software ecosystem that reduces long-term dependency on US technology." — The China Academy

This is the deeper issue for HR tech buyers. DeepSeek's models are open-source with 75 million downloads. When the most widely-downloaded open-source AI models run best on specific hardware, that hardware gains share globally — including among HR software vendors in Asia, the Middle East, and emerging markets who are evaluating self-hosted options.

#The Blackwell Training Allegation

A senior Trump administration official told Reuters that DeepSeek's V4 model was trained on Nvidia Blackwell chips inside mainland China — a direct violation of US export controls. Nvidia's Blackwell architecture is explicitly restricted from export to China. Using it for AI training inside China violates current US law regardless of how the hardware was obtained.

The same official alleged that DeepSeek may attempt to scrub technical indicators from V4 that would reveal Blackwell training origins — then claim the model ran on Huawei hardware instead. AI models contain artifacts from their training environment: numerical precision patterns, memory access optimizations, kernel signatures that experienced engineers can identify. These traces can be obfuscated with enough effort, but intelligence agencies have methods for detecting hardware provenance.

For HR teams in regulated industries — healthcare, finance, government contractors — this compliance uncertainty is a due-diligence item for vendor evaluation, not an immediate operational risk. If a future regulatory determination classifies DeepSeek V4 as a product of prohibited export violations, organizations using HR tools built on V4 may face secondary compliance questions.

#DeepSeek's Track Record: Why This Matters

DeepSeek has established a consistent pattern of delivering frontier AI capability at costs that shocked the market.

DeepSeek-V3 (late 2024): 671-billion parameter model trained for approximately $5.5 million. At a time when Western frontier training runs cost $100 million or more, this was not incremental improvement — it was an order-of-magnitude disruption to training cost assumptions.

DeepSeek-R1 (January 2025): The release that triggered a 17% single-day drop in Nvidia's stock — erasing $589 billion in market value, the largest single-day market cap loss in stock market history. The Nasdaq fell 3.1%. The S&P 500 dropped 1.5%. The panic was not about one model. It was about the implication that frontier AI was achievable at a fraction of the cost Western investment theses had assumed.

If V4 delivers on its stated specs — 1 trillion parameters, 1 million token context, 10x to 40x lower inference cost — it becomes a compelling self-hosted option for high-volume HR screening workloads. The pattern says to take the claims seriously.

#Two Parallel AI Ecosystems and the HR Software Caught Between Them

DeepSeek's V4 lockout is one visible moment in a long-running structural split between US and Chinese AI ecosystems.

The four-year pattern of escalating export controls tells the same story every time: US tightens restrictions, China develops workarounds or domestic alternatives, the cycle repeats. The result is not Chinese AI falling behind. It is two parallel ecosystems forming.

One is built on Nvidia GPUs, CUDA, PyTorch, and Western API providers — OpenAI, Anthropic, Google. The other is built on Huawei Ascend chips and Chinese model providers — DeepSeek, Kimi, Baidu ERNIE.

HR software vendors are not neutral in this split. Every vendor using OpenAI, Anthropic, or Google APIs operates in the Western ecosystem. Every vendor self-hosting DeepSeek models draws from the Chinese ecosystem.

For HR buyers in multinational organizations, this creates real complexity. A vendor using OpenAI APIs has different latency and cost profiles for US operations versus Asia-Pacific. A vendor using self-hosted DeepSeek may offer better performance and lower cost in Asian markets — but with different compliance, data residency, and geopolitical exposure profiles.

#A Practical Framework for Evaluating Model Sourcing Risk

Add model sourcing to your vendor evaluation process in 2026. Here is how.

#Step 1: Ask vendors which models power their AI features

"We use AI for screening" is not sufficient information. Ask specifically: OpenAI GPT-5.x, Anthropic Claude, Google Gemini, self-hosted Llama, self-hosted DeepSeek? The answer determines your compliance profile, data handling requirements, and exposure to regulatory changes.

#Step 2: Understand data residency for each provider

When candidate data is processed by an AI screening model, it transits through that provider's infrastructure. OpenAI, Anthropic, and Google all have data processing agreements with clear jurisdiction and residency terms. Self-hosted open-source models keep data within your infrastructure. Chinese-origin model APIs have different data jurisdiction implications.

#Step 3: Check for single-model-provider dependency

Vendors locked into a single AI API provider carry concentration risk. Ask vendors whether they can route workloads to alternative model providers.

#FAQ

Should HR teams avoid DeepSeek-based tools entirely?
Not necessarily. The compliance uncertainty around V4's training origins is real but unresolved. For most private-sector HR use cases in Western markets, the current risk is low. For government contractors, defense-adjacent organizations, or companies with strict data localization requirements, additional scrutiny is warranted. Ask vendors specifically which DeepSeek version they use, whether data is processed through DeepSeek's API or a self-hosted deployment, and whether they have a compliance assessment for their model sourcing.

What is the 1-million-token context window useful for in HR?
It holds approximately 750,000 words in a single prompt — roughly five full-length novels. For HR applications this means processing an entire candidate portfolio (resume, LinkedIn, portfolio, work samples, interview transcripts) in a single pass rather than chunking. It also enables analyzing a full year of candidate pipeline data in one request.

Does DeepSeek V4's 10x lower inference cost matter for our HR software budget?
It matters more for vendors than buyers directly. If your HR tool vendor self-hosts DeepSeek V4 for screening workloads, their unit economics improve by 10x — which either improves their margins or flows through to your pricing. As a buyer, ask vendors whether their AI infrastructure costs have fallen and whether that is reflected in pricing.

Is the export control allegation about DeepSeek V4 a real compliance risk?
Currently, no US regulation prohibits US companies from using AI tools built on DeepSeek models. The export control allegation targets how DeepSeek obtained its training hardware, not how US companies use the resulting model. However, this could change if regulatory determinations or sanctions follow. Monitor the situation and build vendor contract language that gives you flexibility if model sourcing restrictions emerge.

Which HR tool categories are most likely to adopt DeepSeek V4?
High-volume resume screening vendors are the most likely early adopters — the cost economics are most compelling for workloads processing thousands of applications at low per-candidate budgets. Assessment and skills evaluation vendors are a second wave.

What should HR teams do right now based on this news?
Three actions. First, ask your current AI hiring tool vendors which models they use and add model sourcing to vendor questionnaires. Second, for any vendor using DeepSeek models, ask about their data processing model (API vs self-hosted) and compliance assessment. Third, watch regulatory developments over the next six months.