Top 8 HPC Stocks for DeepSeek AI-Style Scaling (2025): High-Load Inference & Hardware Investments

Which AI high-performance compute (HPC) investments may crush it if DeepSeek's scaling methods go mainstream?

Jan 31, 2025

∙ Paid

Over the last few years, large-scale AI training has dominated high-performance computing (HPC) investments, with ever-larger pre-training clusters consuming vast amounts of GPU cycles.

Yet a new shift is on the horizon: methods such as DeepSeek’s cost-efficient approach suggest that labs could accomplish near state-of-the-art reasoning with smaller base models, heavier reinforcement learning (RL), and more demanding inference loads—all at a fraction of the traditional pre-training cost.

If this strategy gains traction, it could reshuffle where and how labs spend on HPC. Instead of pouring money into colossal pre-training clusters, teams might allocate more resources to high-load inference and partial “test-time” training.

The outcome isn’t necessarily a reduction in total compute spending; rather, it’s a reconfiguration of budgets toward hardware suited for multi-step inference, networking, and memory capacity.

Preface: DeepSeek’s AI Scaling Approach May Not Become Dominant

DeepSeek’s R1 and R1-Zero approach drastically reduces the cost of building a strong reasoning AI—in some benchmarks achieving results on par with more expensive models.

This is not guaranteed to become the universal standard: large labs might still do huge pre-training expansions (e.g., OpenAI, Anthropic).

However, it provides a credible blueprint for more cost-friendly training and might inspire many labs to adopt similar architectural ideas (FP8, mixture-of-experts (MoE), multi-token prediction, and so on).

If many follow suit, model training may demand fewer GPUs, whereas inference (chain-of-thought expansions, test-time RL, partial “on-the-fly” updates) could drive the next HPC wave.

(Don’t conflate “model training may demand fewer GPUs” with “total GPU needs decrease” — there is a difference.)

What Did DeepSeek Achieve Architecturally?

DeepSeek developed R1 (and R1-Zero) with several key features:

Pure RL (R1-Zero) or Minimal SFT (R1)

R1-Zero used no supervised data pre-step, just large-scale reinforcement learning with rule-based rewards.
R1 used only a small “cold-start” dataset and then large-scale RL. Both approaches reduce the typical huge data-collection overhead.

Mixture-of-Experts (MoE) with Balanced Gating

Only ~37B parameters are “activated” out of hundreds of billions.
This cuts memory usage and GPU overhead during forward/backward passes.

FP8 Mixed Precision & Multi-Token Prediction

FP8 training drastically lowers memory and compute.
Multi-token heads further reduce the token-by-token overhead and let the model train/infer faster with less cost.

Parallel RL & Distillation

They run large-scale RL sampling, but each step is cheaper than naive giant pre-training.
Distilled smaller models further cut inference cost, broadening adoption potential.

The Result: DeepSeek claims 6× or more cost-efficiency vs. older approaches for building near–o1-level reasoners. They effectively shift HPC usage from an extremely large “one-time pre-train” stage to a more iterative, smaller RL cycle. This doesn’t necessarily shrink HPC usage globally, but it changes how HPC is allocated.

Likely Evolution & Scaling if Labs Adopt DeepSeek-Like Methods

Less “Mega” Pre-Training, More RL & Inference
- If major labs find R1’s strategy viable, they might drastically scale down multi-trillion-parameter pre-training.
- Instead, they run relatively smaller base models and rely on extended reinforcement learning or “chain-of-thought expansions” at test time.
Heavier Test-Time Workloads
- Chain-of-thought inference or partial “test-time training” can saturate HPC clusters if thousands of parallel user queries do multi-step reasoning.
- HPC usage for inference might overshadow the old training HPC usage.
Potential Test-Time Training
- Some foresee model weight updates even at inference, a more radical approach. Then HPC usage spikes again because you effectively do partial training for each user session.
- That further multiplies GPU cycles and cluster memory demands, overshadowing any cost savings from smaller base training runs.
Net HPC Patterns
- Yes, labs might buy fewer GPUs for huge base training.
- But, they might build out HPC clusters for large-scale, multi-step inference or partial RL. That still leads to big networking and memory expansions.

Impact on Cost, Equipment, HPC Infrastructure

Cost:

The total cost to achieve a top-tier reasoner is significantly lower per model training cycle. This appeals to mid-tier labs.
However, if usage scales up across thousands or millions of inference calls daily—each with chain-of-thought or partial weight updates—the overall HPC budget might remain large or even grow, but allocated to different gear (like advanced networking, memory capacity for multi-instance RL).

Equipment:

Training: Fewer monstrous “64 GPU pods,” but still some base HPC clusters.
Inference: Potentially large horizontally scaled GPU or accelerator clusters, heavy on network and memory.
If partial test-time training emerges, you need partial “training-grade” hardware in inference clusters, fueling HPC expansions again.

Hence: The HPC wave might pivot from “massive training expansions” to “massive inference expansions,” which can still drive robust hardware demand in networking, memory, and advanced chip designs.

Top 8 Hardware Investments for a DeepSeek AI-Style Scaling Paradigm

Included below is a rank-ordered list of HPC/hardware beneficiaries under widespread adoption of the DeepSeek approach—with rationale, confidence rating, and possible growth/peak timelines.

Keep in mind that some of these stocks actually sold off in response to DeepSeek.

The sell-off was mediated by several factors including: potential tariffs/sanctions from Trump (would reduce rev/profits), genuine fear of DeepSeek competition (and Chinese hardware), thinking most AI labs already have too much equipment (can gain efficiency before buying more), etc.

Stock price changes in bullish conditions can swing rapidly with breaking news and sentiment changes (i.e. FUD). DeepSeek was essentially a strong source of FUD for the entire semiconductor/HPC sector.

Also keep in mind that just because these may be good investments for DeepSeek-esque scaling, it doesn’t mean they are valued optimally to buy in right now (some might be “bad deals” in terms of financial/valuation metrics).

Will mention it again: No guarantee that DeepSeek-esque scaling becomes dominant. AI architectures evolve pretty rapidly. Something new could emerge in the blink-of-an-eye (perhaps designed by AGI itself).

Keep reading with a 7-day free trial

Subscribe to ASAP Drew to keep reading this post and get 7 days of free access to the full post archives.