Recent headlines about Chinese AI startup DeepSeek training a competitive model at a fraction of the cost of its US-based counterparts sent ripples through the stock market. One effect was a sharp dip in NVIDIA’s stock value, because some speculated that DeepSeek was proving that fewer GPUs are needed to train AI models than previously thought. But that interpretation misses the bigger picture.
While DeepSeek’s training cost was remarkably low, the company still relied exclusively on NVIDIA GPUs—specifically, relatively basic H800 accelerators—to get the job done. Rather than being a sign of NVIDIA’s decline, this moment underscores a growing trend in AI development: the search for greater efficiency in model training. And NVIDIA remains at the center of that evolution.
DeepSeek’s cost efficiency: a potential game-changer, not yet a market disruptor
DeepSeek’s efficiency stems from a combination of hardware and software optimizations. Reports indicate the company spent approximately 2.79 million GPU hours on H800 accelerators, which are a less advanced version of the H100. The total estimated training cost as claimed by the company was just $5.58 million—a figure significantly lower than what US-based models typically require. However, it’s not possible to verify this figure and some are claiming that this figure refers to the cost of the final training run only, excluding experimentation. In any case, the seeming time and cost efficiency raised eyebrows in the West, particularly given the enormous investments companies like OpenAI and Anthropic make in training their large-scale models.
But let’s put this in perspective:
- DeepSeek’s model still required NVIDIA GPUs. The training process used H800 accelerators—hardware designed by NVIDIA for AI workloads in China, albeit with some export restrictions compared to the H100 models available elsewhere.
- Training cost reductions don’t mean less demand for high-performance GPUs. AI developers will always seek ways to optimize and reduce costs, but this doesn’t eliminate the need for powerful GPUs.
- NVIDIA GPUs remain the gold standard for both AI training and inference. Even as efficiency gains improve, AI developers continue to rely on NVIDIA’s ecosystem, from TensorRT optimizations to CUDA-powered acceleration.
- Stock market reactions tend to be exaggerated. While investors initially reacted negatively to DeepSeek’s success, analysts suggest that this response overlooks the broader trends favoring sustained demand for NVIDIA hardware.
- Western market adoption is uncertain. DeepSeek’s model has yet to gain traction outside China, where regulatory concerns, data privacy issues, and geopolitical factors could limit its growth.
- The AI hardware market isn’t a zero-sum game. Increased efficiency in AI training does not translate to reduced GPU demand—it simply redistributes workloads to more strategic compute resources. Increased efficiency in training models doesn’t mean that the need for high-end hardware disappears. Instead, it shifts demand toward scalable, cloud-based GPU solutions that can adapt to evolving AI workloads. It also democratizes training: More companies can afford to train a $6M model, if this is truly the new baseline, so more companies will engage in training rather than only fine-tuning and/or inference.
- Infrastructure remains key. Enterprises with mission-critical AI applications will continue prioritizing scalable, high-performance GPU solutions, ensuring NVIDIA’s continued dominance in the space. While startups experiment with cost-effective training, enterprises with mission-critical AI workloads still rely on premium infrastructure to meet performance, latency, and security requirements.
Does the $500B Stargate Project still makes sense?
Some have questioned whether the massive $500B Stargate Project looks foolish in light of DeepSeek’s cost savings. The answer? Not necessarily. DeepSeek is a reminder that necessity is the mother of invention—faced with resource constraints, they found ways to make do. At the same time, there’s ongoing debate about whether SoftBank’s massive investments encourage innovation or, conversely, lead to complacency and reckless spending. Some argue that when resources are too abundant, teams burn cash rather than find more efficient solutions.
Still, the broader trajectory of AI computing needs remains clear. The industry is in its early stages, and model performance is far from hitting a ceiling. Every major AI breakthrough requires extensive experimentation with novel model architectures, and we are still far from AGI (Artificial General Intelligence). Each step closer will demand more training cycles, not fewer.
Despite emerging cost optimizations, high-end GPUs remain indispensable for cutting-edge AI models. These GPUs aren’t just about raw power; they enable rapid parallel processing, power-efficient computing, and deep integrations with AI frameworks. The world’s most advanced AI models will continue to rely on NVIDIA hardware for training—at least until another compute architecture disrupts the industry.
Who needs the most advanced NVIDIA GPUs?
While some AI startups optimize for cost efficiency, there are plenty of use cases where the highest-end NVIDIA GPUs remain a necessity. Here’s who needs the most powerful NVIDIA accelerators.
- Large-scale AI training: Companies training frontier models require the most powerful GPUs available. The sheer scale of their training datasets and the complexity of their architectures demand the computational power only NVIDIA’s most advanced GPUs like the GB200 and H200 can provide.
- High-performance AI inference at scale: Running AI models efficiently in production is just as critical as training them. Enterprises deploying AI-powered applications—such as real-time video analytics, autonomous vehicles, and advanced recommendation systems—need inference-optimized GPUs like the NVIDIA L40S or H100. These models help ensure that AI applications run smoothly, even at massive scale.
- Scientific and industrial applications: From drug discovery to climate modeling, high-performance GPUs power some of the most computationally intensive workloads in the world. Scientific institutions and enterprises working on complex simulations need access to high-memory, high-speed computing resources—areas where NVIDIA remains unrivaled.
- Enterprise AI and edge deployments: As AI expands beyond data centers and into edge computing, businesses looking to run AI applications in real-time environments require powerful yet efficient inference hardware. NVIDIA’s edge-optimized GPUs power everything from AI-driven cybersecurity to automated industrial robotics to next-gen smart cities.
What this means for the future of AI infrastructure
AI model training is evolving, and the industry is moving toward a hybrid approach that balances cost, speed, and performance. The rise of more efficient training techniques doesn’t eliminate the need for high-performance GPUs—it reinforces their importance.
Market dips do not reflect long-term trends. Historical patterns show that the need for specialized hardware like NVIDIA’s accelerators increases as AI capabilities advance.
Our predictions for short-term AI evolution
As a result of this recent news, we predict the following:
- Cloud-based AI computing will continue to grow, with companies leveraging platforms like Gcore’s AI cloud to access NVIDIA GPUs on demand.
- AI model training and inference will require hybrid solutions, balancing high-end compute for critical workloads with efficiency-driven optimizations.
- The demand for NVIDIA GPUs will remain strong as AI models become more sophisticated and require more computing power to stay competitive, and due to an overall growth in the number of AI projects around the world.
Drive AI innovation with Gcore and NVIDIA
DeepSeek’s cost optimizations are impressive, but they don’t spell the end for NVIDIA’s dominance in AI computing. As AI models grow larger and more complex, NVIDIA’s role will remain central to the industry.
At Gcore, we see NVIDIA’s role in AI development as stronger than ever. We provide access to NVIDIA’s cutting-edge GPUs for both training and inference workloads. Whether you’re a startup looking for an intuitive inference solution or enterprise training at scale, our AI solutions provide the power and flexibility you need.
Learn how to easily deploy DeepSeek on Gcore Everywhere Inference
