DeepSeek ВЫЛОЖИЛ то, что прятал 1,5 года. Это меняет цены на ИИ!

Oct 26, 2026

Read: ~6m · You save: 9 min

DeepSeek Released the Engine It Had Hidden for 1.5 Years: What It Means for AI Prices

DeepSeek open-sourced the engine it hid for 1.5 years. Its cheap long-context tricks could change AI prices. Here is how it works and how to use it in Claude Code.

On October 20, DeepSeek released the inference engine behind its models and published figures on the economics of its data centers. The company had kept the engine private for about 18 months. According to the company's documentation, its operations are profitable at current prices. This article summarizes the disclosed figures, the technical methods the engine uses, and the setup process for running DeepSeek with Claude Code.

Background

In April 2025, DeepSeek stated that it would not open-source the engine because it is directly tied to the company's data centers. The company also said that a small team could not, at that time, maintain the engine together with the rest of its stack.

On October 17, 2026, DeepSeek released model V4.1 Flash. On October 20, it released the engine for that model under the name DFINER. Within one week, the repository had 22,000 stars on GitHub and 2,500 forks. The model weights were also published on Hugging Face.

The repository includes a document with answers to common questions. It describes how the system works, how it achieves its efficiency, how GPUs are used, and the cost economics of the approach.

Cost and Margin Figures

DeepSeek reports the following figures:

One cluster processes 155 trillion tokens per day, or nearly 1.8 billion tokens per second.
Each GPU generates about $38 per hour of revenue at discounted prices.
The estimated market cost of one such GPU-hour is $5 to$ 10.
The resulting margin is about 80%, before research and other operating costs.

According to DeepSeek, its models remain among the lowest-cost models available.

Technical Methods

The article identifies four main techniques in the engine. A fifth component, a tokenization project named Resype, is described separately.

1. Separation of Reading and Writing

A standard model passes a request through all of its layers. The model in question has 40 layers. Output begins only after the computation reaches the final layer.

DeepSeek divides the layers into two groups:

The first 20 layers process the input only (reading).
The upper 20 layers generate the output (writing).

Because the input is not processed by the upper layers, reading a long prompt costs about half as much. Reading and writing run as separate workloads, so a single writer does not wait on the reader.

The engine applies the same logic to hardware: reading and writing are assigned to separate GPU groups. The smallest self-hosted deployment is 64 GPUs for reading and 64 GPUs for writing.

2. Mixture of Experts

The model contains 750 billion parameters and 384 specialized experts, such as experts for code and for text. For each token, only six experts are activated. The rest of the parameters are not used for that token.

The experts are distributed across GPUs, with approximately seven experts per card.

Code requests are the most demanding. They make up less than half of all requests, but they are the heaviest by size. Experts for code therefore become overloaded during peak load.

DeepSeek keeps spare expert replicas. When a primary expert is saturated, a copy handles the load. According to the company, the most loaded GPU runs only 9% slower than the average GPU.

3. Speculative Decoding

Generating text one word at a time is slow. A small, fast draft model proposes the next five words. The large model then checks all five in one pass.

The standard approach has a cost. If the third word is wrong, the fourth and fifth words are discarded, but the time spent on them is already consumed.

DeepSeek trained the draft model to estimate its own confidence in each word. Low-confidence drafts are discarded early, when needed. As a result:

The number of verification checks fell by 30%.
The number of accepted words stayed nearly the same.
The full pipeline runs 4% to 7% faster.

4. Context Caching

Continuing a conversation requires the model to process the prior context again. In Claude Code, the cache is kept for one hour. According to the author's usage data, 97% of tokens over the past month were spent re-reading earlier context in the same conversations. In long sessions, a re-read can involve about one million tokens.

DeepSeek stores this context on disk for longer periods:

3 days for agent workloads.
7 days for standard chat.

The company uses 100 TB of storage for this purpose. When a user returns to a conversation, the model recalculates only the last 128 tokens instead of re-reading the full history.

Cost comparison:

Item	Price
Re-reading a conversation (DeepSeek), per million tokens	$0.006
New input tokens (o3), per million tokens	About 50 times higher than the DeepSeek re-read price

According to the source, the cost of returning to a long task after a day is almost negligible on DeepSeek.

Resype

Before a prompt reaches the GPU, it must be split into tokens. For very large inputs, this step causes a noticeable delay. Resype splits the text into segments before it reaches the GPU, which distributes the workload and speeds up processing.

Hardware

The engine runs on two GPU types:

Nvidia Blackwell
Huawei Ascend 950

Nvidia H100 is not supported. According to Bloomberg, DeepSeek plans to use at least 160,000 accelerators for a new data center. The article notes that the company has begun using Chinese components.

The following components were not released in the open-source package:

The conversation memory storage (100 TB).
The load balancers.

The published code includes the engine itself. These two components remain proprietary.

Setup With Claude Code

The article describes a three-step process to run DeepSeek as the model behind Claude Code. The Claude Code interface stays the same. Only the underlying model changes.

Create an API key. Generate the key in the DeepSeek dashboard. A top-up of a few dollars is sufficient for testing. Payment covers only the usage.
Configure nine lines. Take the nine configuration lines from the supporting materials. Add the DeepSeek endpoint address, the API key, and the model name. Enclose the values in quotation marks. The configuration works in the standard macOS Terminal.
Launch Claude Code. Start Claude Code in the project folder as usual. DeepSeek then handles the requests.

The author plans to test both agents on the same task in a follow-up. The task is to analyze competitors for a healthy-food office delivery service. It includes collecting competitor prices and customer complaints and writing conclusions on market entry. The comparison will focus on the total token usage and whether the savings are real.

Author's Assessment

The author presents the following views. They are opinions, not verified findings.

Over the past two years, AI companies have spent heavily. The author describes Claude as a company with substantial losses. The author states that OpenAI showed profits at one point, and that OpenAI has recently raised prices to $500.
DeepSeek's position is that the lowest-cost provider of comparable intelligence will attract users.
Anthropic is expected to adopt DeepSeek's architecture, based on the author's expectation that the approach works fundamentally differently and will be copied.
The author expects that DeepSeek's principles will shape agents, Claude, and Codex within the next year.
The author compares the situation to Chinese automakers, whose cars now compete with BMW models at similar prices. The author expects Chinese AI models to follow a similar path.
For coding tasks in production environments, the author states that DeepSeek does not yet match Claude Code and is insufficient for production projects. For standard tasks, the author considers DeepSeek suitable.

Introduction and the Claim of a Profitable DeepSeek

DeepSeek published its engine and cost data, claiming its inference is profitable and cheaper than competitors.

On October 20 DeepSeek released data it had kept private for about 1.5 years, including figures on its data centers.
The speaker says the earlier claim that DeepSeek runs at a loss is wrong; he says it operates at a large profit.
DeepSeek's model is cheap, and the video says it has been among the cheapest AI models for a long time.
The engine and techniques released on October 20 are expected to shape agents, Claude and Codex over the next year.

Cluster Throughput and Margins

One cluster processes about 155 trillion tokens per day, and the speaker estimates the economics per GPU.

One cluster processes about 155 trillion tokens per day, roughly 1.8 billion tokens per second.
The speaker says Claude spends tens of billions on compute and calls Anthropic a very loss-making company.
Per the video, at discounted prices each DeepSeek GPU earns about 38 dollars per hour.
The estimated hourly cost of one such GPU is 5 to 10 dollars, which the speaker says leaves about an 80% margin.
He notes that research and other costs exist but says DeepSeek's hardware efficiency is unmatched.

Why Long Context Is Expensive

Long conversations are expensive because models re-read the whole context on each turn; DeepSeek caches this context.

The speaker says 97% of his tokens in the last month went to re-reading the same conversation.
Each time a user sends a message, the agent recalculates the conversation, especially after more than an hour.
Claude's cache lasts 1 hour, according to the speaker.
He says the highest reported figure is nearly 1 million input tokens re-read per session.
Average Claude Code and Codex usage is dominated by this re-reading cost.

Layer Split and Mixture of Experts

DeepSeek's first technique splits the transformer into reader and writer layers, so the reading of long prompts is about half as expensive.

The model has 40 layers, which the speaker compares to a building.
The first 20 layers are readers that process the prompt; the top 20 layers are writers that generate the answer.
Writers never read the user's text directly, so reading a long prompt costs about half as much.
The model is built from 384 experts, and each token activates only 6 of them.
The model has 750 billion parameters, but only a small part runs for each token.

Expert Placement and Spare Experts

Distributing experts across GPUs and adding spare expert replicas removes bottlenecks.

Analogy: one picker walking the warehouse is slow; 64 pickers with their own shelves work instantly.
The 384 experts are spread across GPU cards, roughly 7 experts per card.
Code experts receive the heaviest load, even though code is less than half of all requests.
DeepSeek keeps spare expert copies that spin up when one fills up.
The busiest card runs only 9% slower than the average card.

Speculative Decoding With a Draft Model

A small draft model proposes five tokens at a time, and the large model verifies them. DeepSeek makes the draft model estimate its own confidence.

Output is generated one word at a time, which is slow, so a draft model proposes the next five words.
Previously, if the third word was wrong, the fourth and fifth were discarded even though time was spent on them.
DeepSeek trained the draft model to estimate its confidence and drop uncertain drafts early.
Verification passes fell by 30%, and the accepted word count stayed nearly the same.
The combined pipeline is reported to be about 4 to 7% faster.

Memory Cache: The Most Useful Technique

DeepSeek stores conversation memory on disks so returning to old chats costs very little.

Claude Code re-reads the whole conversation after an hour or two, and the speaker calls this a major unsolved problem.
DeepSeek keeps memory on separate disks: 3 days for agents and 7 days for regular chat.
DeepSeek reportedly has a 100 TB-class storage system for this, as stated in the video.
On return, the model recomputes only the last 128 tokens instead of the whole history.
The re-read cost is about 0.006 dollars per million tokens, versus about 3 dollars for a new request, a roughly 50x difference.

Chinese AI Competition Analogy

The speaker compares Chinese AI to Chinese cars that caught up with European brands, and notes DeepSeek uses Huawei hardware.

He compares DeepSeek's trajectory to Chinese cars that were once mocked and now match BMW.
DeepSeek already uses Huawei chips alongside Nvidia.
The speaker says DeepSeek aims to offer similar quality at several times lower cost.
The video contains a call to comment and subscribe; this is excluded from the summary.

Tokenization Offload: ReSype

DeepSeek ReSype splits text into pieces before it reaches the GPU, so long requests process faster.

The fifth technique is a project called DeepSeek ReSype, which splits incoming text into chunks before it reaches the GPU.
Tokenization of very long requests is described as a noticeable pause, which the chunking reduces.
This distributes the load so the model responds faster.

Hardware and What Stayed Closed

The engine officially runs on Nvidia Blackwell and Huawei Ascend 950 GPUs. DeepSeek kept the memory cache and load balancer closed.

The engine officially supports two GPU types: Nvidia Blackwell and Huawei Ascend 950.
The previous Nvidia H100 generation is not supported.
Per Bloomberg, DeepSeek plans a new data center with at least 160,000 accelerators.
Not open-sourced: the conversation memory cache (100 TB-class storage) and the load balancer.
The smallest self-hosted deployment is 64 cards for reading and 64 for writing, which suits large companies.

Setting DeepSeek Up in Claude Code

The speaker gives a three-step guide to using DeepSeek as the model behind Claude Code.

Step 1: create an API key in the DeepSeek dashboard; topping up a couple of dollars is enough for testing.
Step 2: take the nine configuration lines from the video materials, add the DeepSeek endpoint, the model name, and the key in quotes.
The settings also work in a standard Mac terminal.
Step 3: run Claude Code in the project folder as usual; the DeepSeek model handles the reasoning instead.
The speaker plans a follow-up video comparing one task run on two agents, measuring tokens spent and actual savings.

Conclusion: Cheaper Intelligence Wins

DeepSeek pursues low prices while OpenAI raises prices; DeepSeek is good for standard tasks but not yet for production code.

The speaker says OpenAI raised prices to as much as 500 dollars.
He says DeepSeek is ideal for standard tasks but not yet strong enough for production projects.
He predicts more Chinese models will follow this low-cost strategy.
The full cost comparison between DeepSeek and Claude Code is promised in a separate follow-up.