DeepSeek AI: The Chinese AI That Shocked Silicon Valley

In January 2026 a Chinese lab published a paper saying it had trained a frontier-class model for around six million dollars, released the weights under a permissive licence, and let anyone download them. Nvidia's share price took its largest single-day loss in history shortly after.

The story was mostly true and almost entirely misunderstood, and the gap between those two things is the most useful thing to understand about DeepSeek.

The $6 Million Figure, Handled Honestly

The number was real, it came from DeepSeek's own paper, and it did not mean what nearly every headline said it meant.

What it measured was GPU hours for the final successful training run of V3. What it explicitly excluded: all prior research, the salaries of the people who did it, every failed or abandoned run that preceded the successful one, data acquisition and cleaning, and — most significantly — the capital cost of the GPU cluster itself, which credible estimates put in the hundreds of millions.

So it was not "a frontier model for six million dollars." It was "the electricity and compute for the last run cost six million," which is a genuinely impressive engineering result and a completely different claim from the one that moved markets.

Being precise about this matters, because the real achievement gets undersold by the exaggeration. DeepSeek demonstrated that architectural and training efficiency could substitute for a meaningful fraction of raw spending — that the frontier was not purely a function of who had the biggest cheque. That holds up. "Anyone can build GPT-4 for the price of a house" does not, and repeating it makes you easy to dismiss.

Where The Models Actually Stand In 2026

DeepSeek V4 is the current flagship, released in April 2026 — not the February date that circulated in early rumours. It shipped in two variants, which is the more interesting part: V4-Pro, a very large mixture-of-experts model aimed at quality-sensitive reasoning, and V4-Flash, a much smaller-activation model built for fast, cheap serving. The split is a direct acknowledgement that most real workloads do not need the biggest model, and that serving cost is where open weights actually win.

DeepSeek R1 remains the model that made the reputation. It was trained with reinforcement learning applied directly, without the supervised fine-tuning stage that was assumed to be a prerequisite, and it produced long visible chains of reasoning that it would check and correct mid-answer. Watching it argue with itself and reverse a wrong conclusion was, at the time, genuinely novel to see in an open model.

R2 has still not shipped. This is worth stating clearly because it has been "imminent" in coverage for well over a year. Reporting attributes the delay to the CEO being unsatisfied with performance, alongside a failed training run on non-Nvidia hardware that forced a reversion. Treat any confident R2 release date you encounter as speculation.

The R1-Distill family is the underrated part of the lineup. These are smaller models trained to imitate R1's reasoning behaviour, and several punch far above their size class. If you want reasoning capability on hardware you actually own rather than a datacentre, this is where to look.

The Thing Nobody Else Offers

Every comparison of DeepSeek against ChatGPT and Claude gets stuck arguing about benchmarks, which is the least decision-relevant thing about it.

Benchmark parity moves constantly and rarely survives contact with your actual workload. What does not move is the structural difference: DeepSeek's weights are downloadable and the others' are not. You can put the model on your own hardware, behind your own firewall, with no outbound connection, and keep it running unchanged for as long as you like. No vendor can deprecate it, reprice it, change its behaviour under you, or read what you send it.

For most casual users this is worth nothing. For a law firm, a clinic, a defence contractor, or anyone under data residency obligations, it is the only thing that matters, and no amount of benchmark superiority from a hosted API substitutes for it. The relevant trade-offs are covered further in what happens to your data when you use AI tools.

That is the honest framing. DeepSeek is not "the free ChatGPT." It is the one you can own.

When To Actually Reach For It

Use DeepSeek when you are doing maths or reasoning-heavy work, when API cost at volume is a real constraint on what you can build, or when the data genuinely cannot leave your infrastructure. The last of these is not a preference — it is a requirement or it is not, and if it is, the decision is already made.

Do not use DeepSeek when you want a polished daily assistant with a mature ecosystem, when you need the integrations and enterprise tooling the incumbents have spent years building, or when you are putting confidential client material into a hosted chat window under a jurisdiction you have not thought carefully about. The ChatGPT vs Claude comparison is the more relevant read if that is your situation.

On the privacy question, resist both available scripts. "It is Chinese, therefore unsafe" is lazy — the open weights are auditable in a way no closed Western model is, which is a genuine security advantage. "It is no different from any other cloud AI" is also wrong, because jurisdiction and legal recourse genuinely differ. The accurate position: the hosted service deserves the caution you would give any foreign cloud provider, and the self-hosted version sidesteps the question entirely. Pick the version that matches your risk, rather than arguing about the company.

Getting Started

If you just want to try it, the web interface at deepseek.com needs no account for basic chat. Put the same non-trivial reasoning problem to it that you would give ChatGPT — a real one from your own work, not a riddle — and read the chain of thought rather than just the answer. That comparison will tell you more in ten minutes than any benchmark table.

If you are building, the API uses a standard OpenAI-compatible format, so swapping it into existing code is usually a base URL and key change rather than a rewrite. Price it against your current provider at your actual volume; the gap is normally the entire reason to switch.

If privacy is the driver, pull the weights from Hugging Face and run them through Ollama or similar. Start with a distilled variant rather than the full model unless you have serious GPU memory available — people routinely try to run the flagship on a laptop, conclude the model is bad, and are actually just out of VRAM.

Some grounding worth having first: how AI actually works explains why the reasoning chains behave the way they do, and what a large language model is covers the architecture that makes the efficiency story possible. For coding specifically, the AI coding assistants comparison places DeepSeek against the tools built for that job.

What It Changed

The lasting effect of DeepSeek was not a model. It was a repricing of assumptions.

Before it, the working consensus was that frontier capability required capital that only a handful of companies possessed, and that open weights would trail closed models by years. Neither survived intact. Efficiency turned out to be a genuine lever, the gap between open and closed narrowed to months, and export controls demonstrably did not prevent the result they were designed to prevent.

What remains unresolved is whether this is the beginning of genuinely distributed AI development or simply a wider oligopoly with a few more members. Two years on, it looks more like the second than the first — which is still a better outcome than the alternative, and considerably less than the revolution announced at the time.

Sources: SitePoint on the V4 release, Fello AI on R2's status, Layer3Labs R2 explainer.