Snitz examines Chinese AI vs America's big 4 (Claude, ChatGPT, Grok, Gemini)!

I've asked all our AI models to evaluate if China's primary AI engine (Deep Seek) is: 1. as effective as our top platforms, 2. less costly to run, 3. uses less electrical power. The latter is very important as many politicos in this country are critical of the Data Center construction boon and it's impact on our power grid.

All our AI models agree that Deep Seek produces results close to their best models. There are some application where it lags (Claude).

Where the "close to Claude/GPT/Gemini" claim holds up and where it doesn't

• On many coding and math benchmarks, DeepSeek's models score competitively with — and on some specific benchmarks even ahead of — leading Western models.

• On more open-ended reasoning, multimodal tasks, agentic/tool-use workflows, and non-English writing quality, it tends to still lag the top-tier frontier models somewhat.

• There's also a real hallucination/reliability gap reported by several reviewers, plus data-residency concerns (data processed on servers in China), which matters a lot for regulated industries or sensitive business use.

They also agree it produces results at a far lower cost. How much lower? Like 10-25 times lower cost!

Not so fast! How about the power required to run Deep Seek vs other US models? It turns out that Deep Seek is super power hungry. Depending on the type of query it requires between 6-8 times more electricity to compute the answer. So if you think US data centers require too much energy and are "hogging the grid", the problem is 8 times worse in China!

Is it ok to allow China to gain dominance in this area and overtake the US? Sure if you want our future to be dependent on the folks who brought you COVID and have tanked the EU economy. Why worry?

Below some excerpts from Claude and Gemini's analysis:

Claude

DeepSeek's models (originally V3/R1, now the V4 family) really have delivered performance in the same ballpark as the frontier US labs at a fraction of the price. The "why" is a mix of real architectural innovation and some China-specific circumstances.

The technical reasons it's cheaper

1. Mixture-of-Experts (MoE) architecture — Instead of activating the entire model for every token, DeepSeek uses a sparse MoE design where only a small subset of "expert" sub-networks fire on any given token. You get the capacity of a huge model (hundreds of billions of parameters) while only paying the compute cost of a much smaller one at inference time.

2. Multi-head Latent Attention (MLA) — A compressed attention mechanism that shrinks the memory needed to store the "key-value cache" during inference. This directly cuts serving costs and lets them run more requests per GPU.

3. FP8 (low-precision) training — DeepSeek trained in 8-bit floating point rather than the more common 16/32-bit formats, cutting training compute and memory substantially without a big accuracy hit.

4. Reinforcement-learning-driven reasoning (the R1 approach) — Rather than relying purely on massive human-labeled fine-tuning data, DeepSeek used large-scale RL to teach the model to reason step-by-step, which is a more compute-efficient way to get strong reasoning performance than some alternatives.

5. Engineering efficiency under constraint — Because of U.S. export controls on advanced Nvidia chips, DeepSeek's team had to be unusually resourceful with a more limited compute budget, which forced a lot of these low-level optimizations.

The "how are they able to charge so little" side

Beyond the technical efficiency, pricing is also a strategic/subsidized choice — Chinese AI labs (and DeepSeek specifically) have been willing to price near or below cost to gain adoption and market share, similar to a land-grab dynamic. So part of the price gap reflects genuine compute efficiency, and part reflects a deliberate pricing strategy rather than pure cost-to-serve.

Where the "close to Claude/GPT/Gemini" claim holds up and where it doesn't

• On many coding and math benchmarks, DeepSeek's models score competitively with — and on some specific benchmarks even ahead of — leading Western models.

• On more open-ended reasoning, multimodal tasks, agentic/tool-use workflows, and non-English writing quality, it tends to still lag the top-tier frontier models somewhat.

• There's also a real hallucination/reliability gap reported by several reviewers, plus data-residency concerns (data processed on servers in China), which matters a lot for regulated industries or sensitive business use.

One caveat on the numbers: pricing and benchmark tables for DeepSeek and its competitors are shifting monthly right now — multiple sources I checked disagree on exact current figures — so if you want precise up-to-date pricing for a specific decision, I'd verify against DeepSeek's own site directly rather than trusting aggregator comparisons.

Want to be notified when Claude responds?

Gemini

The answer to this question highlights a fascinating paradox in AI: DeepSeek’s models are hyper-efficient in terms of financial cost, but their actual electricity usage per query depends entirely on the type of model you choose.

According to independent academic benchmarks from late 2025 and 2026 measuring actual energy footprint in Watt-hours (Wh) per query, DeepSeek models consume anywhere from 10% to 500% of the electricity used by top American models depending on whether you are using a standard chat model or a reasoning model.

1. Standard Models: DeepSeek-V3 vs. GPT-4o & Gemini

When comparing typical chat and coding assistants handling standard, direct queries, DeepSeek’s architectural efficiency shines.

• Google Gemini 2.5: ~0.24 Wh per query (Highly optimized infrastructure)  

• OpenAI GPT-4o: ~0.42 Wh per query  

• DeepSeek-V3: ~3.51 Wh per query  

For basic text generation, DeepSeek actually draws more baseline raw power per query than highly mature, scaled-down American flagships like GPT-4o or Gemini. However, it operates at a massive advantage compared to "heavyweight" models like GPT-4.5 (6.72 Wh) or dense open models like Llama 3.1-405B (1.99 Wh - 6.91 Wh), using roughly 50% of the energy of those massive architectures.

2. Reasoning Models: DeepSeek-R1 vs. OpenAI o1 & o3

The narrative changes entirely with reasoning models (the AIs that display a "thinking..." process). These models don't just vomit out an immediate answer; they generate thousands of internal words to double-check their logic before replying. This makes them massive energy hogs.