We published this article in February 2025, with the word "revolutionizing" in the title, in the middle of the loudest week the AI industry has ever had. Eighteen months later the honest thing to do is not to quietly delete it. It is to score it.

Here is the scorecard. The prediction that cheap Chinese models would break the AI oligarchy was wrong. The prediction that the price of machine intelligence was about to fall off a cliff was right, and understated. Those two things sound contradictory. They are not, and the gap between them is the only part of the DeepSeek story that still matters to anyone who hires people for a living.

This piece has been rewritten from the ground up in July 2026. It covers what the $5.6 million figure actually measured, what DeepSeek costs today and what that does to screening arithmetic, the data-residency problem that almost every 2025 take on this subject skipped, and the specific reason a recruiting team will never see this technology the way the hype cycle promised: not through a GPU, but through a line item on a software invoice.

1. The $5.6 million number, and what it actually bought

The famous figure is real, and it was almost universally misreported. DeepSeek's own technical report puts the training of DeepSeek-V3 at 2.788 million H800 GPU hours, which at an assumed rental rate of $2 per GPU hour works out to $5.576 million - DeepSeek-V3 Technical Report. That breaks down as 2,664K hours of pre-training, 119K for context extension and just 5K for post-training.

Read the next sentence in that report, though, because it is the one nobody quoted. The figure covers the final training run only. It explicitly excludes prior research, ablation experiments on architectures and algorithms, and the data work that made the final run possible. It also excludes the cluster itself. Renting is not owning, and DeepSeek did not rent. A $5.6 million electricity-and-time bill on hardware you already own is a fundamentally different claim from "we built a frontier model for the price of a Bay Area house."

So the correct reading was never "frontier AI now costs $5 million." It was "the marginal cost of the final run of a very good model is small relative to what everyone assumed, if you have already paid for the cluster and the research." That is still a genuinely important result. It is just not the one that moved the market.

2. The market called it a revolution, then kept spending

On 27 January 2025, Nvidia fell 17% to close at $118.58 and shed $589 billion in market capitalisation, the largest single-day value destruction in US corporate history - CNBC. The thesis behind the sell-off was simple: if models can be trained this cheaply, nobody needs this many GPUs.

The thesis was wrong, and it was wrong fast. Nvidia closed fiscal 2026 with $215.9 billion in revenue, up 65% year on year, and then posted $81.6 billion in a single quarter (Q1 FY2027, up 85%) - NVIDIA. The five largest hyperscalers are on track to spend somewhere in the region of $700 billion on capital expenditure in 2026 - CNBC. Cheaper training did not reduce demand for compute. It expanded the number of things worth spending compute on, which is what usually happens when a critical input gets cheaper.

This matters for recruiters for one non-obvious reason. The 2025 story was "AI is about to become free, so small teams will get the same weapons as the giants." What actually happened is that inference got absurdly cheap while the frontier kept getting more expensive. The floor collapsed. The ceiling did not. Your small agency did not get Google's capabilities. It got something arguably more useful: a per-token price so low that the cost of running AI across your pipeline stopped being a number worth thinking about.

3. What DeepSeek costs now, and the screening arithmetic

DeepSeek's current generation is V4, released in April 2026 in two sizes. The published API prices are $0.14 per million input tokens and $0.28 per million output tokens for V4-Flash, and $0.435 in / $0.87 out for the reasoning-heavy V4-Pro, both with a 1M-token context window and cache-hit input priced at a fraction of a cent - DeepSeek API docs. Note for anyone with an older integration: the legacy model names deepseek-chat and deepseek-reasoner are deprecated as of 24 July 2026.

Put those numbers against a recruiting task and the result is faintly ridiculous. A two-page CV runs roughly 1,200 to 1,500 tokens. Add a job description and instructions and call it 2,300 tokens of input per candidate, with a few hundred tokens of structured output coming back. Screening 1,000 candidates against a role on V4-Flash is therefore on the order of 2.3M input tokens and 0.3M output tokens: under 50 cents. Even on V4-Pro, with reasoning switched on, you are in the low single-digit dollars. Redo the arithmetic with your own numbers; the conclusion does not change.

This is the part of the 2025 prediction that came true, and then some. Model inference is no longer the constraint on AI screening, and it never will be again. Which forces an uncomfortable question: if the intelligence costs 50 cents, why does AI recruiting software still cost what it costs? The answer is that you were never paying for the intelligence. You are paying for the data, the integrations, the audit trail, the support contract and the legal liability. DeepSeek made the cheap part cheaper. The expensive parts did not move.

On raw capability, V4-Pro is close enough to the frontier that the gap rarely decides a screening use case: it posts 80.6 on SWE-Bench Verified against Claude's 80.8, while trailing the leaders on the hardest maths and agentic benchmarks - DataCamp. Reading a CV is not a frontier task. Almost any competent modern model can do it. That is precisely why model choice is the least interesting decision in your stack.

4. The problem every 2025 take skipped: where the data goes

Here is the sentence that should have been in the original version of this article, and was not. DeepSeek's own privacy policy states: "we directly collect, process and store your Personal Data in People's Republic of China" - DeepSeek Privacy Policy. The data covered explicitly includes your text input, prompts and uploaded files. A CV pasted into that box is not an abstraction. It is a named person's employment history and contact details, and often enough their health, ethnicity or union membership, which under GDPR are special categories.

Regulators did not wait. On 30 January 2025, three days after the Nvidia crash, Italy's Garante ordered an immediate and definitive limitation on the processing of Italian users' data by DeepSeek, citing storage in China and Article 32 on security of processing - Bird & Bird. South Korea's PIPC had the app pulled from the Apple and Google stores in February 2025 after finding that full prompt content was being transferred to Beijing Volcano Engine Technology, a transfer it judged unnecessary. More than a dozen EU jurisdictions opened their own probes - MIAI. China holds no GDPR adequacy decision, so there is no easy legal route around any of this.

The practical instruction for a recruiter is blunt, and it has not changed in eighteen months: do not put candidate data into the DeepSeek consumer app or its hosted API. Not because the model is bad. Because you would be exporting identifiable personal data to a jurisdiction with no adequacy decision, using a processor a European supervisory authority has already ordered to stop, without a defensible legal basis. That is not an AI ethics debate. It is a straightforward transfer-mechanism failure, and it is the kind that ends in a fine rather than a think-piece.

5. The model is not the company

This is the distinction that makes the open-source argument real rather than rhetorical, and it is the single most useful idea in this article.

DeepSeek V4 is published under the MIT licence. Both checkpoints, V4-Pro (1.6T total parameters, roughly 49B active) and V4-Flash (284B total, roughly 13B active), sit on Hugging Face with a 1M-token context and 384K max output - Hugging Face. MIT means you can run them commercially, fine-tune them, and build proprietary products on them, with no restriction beyond preserving the copyright notice.

Weights are numbers in a file. They do not phone home. The privacy problem in section 4 is not a property of the model, it is a property of DeepSeek's servers. Run the same weights somewhere else and the problem disappears. That is why DeepSeek models are offered through Microsoft Foundry and Amazon Bedrock in EU regions, alongside independent inference hosts, where the data stays in a region you chose under a contract you signed - Microsoft Foundry model catalog. Same model, same output, completely different legal posture.

So the honest version of "open source is the great equaliser" is narrower and more interesting than the 2025 version. It does not mean every HR department gets to train its own AI. It means a capable model can no longer be withheld from you by any single company or government, and that the price of the closed alternatives now has a permanent competitor sitting underneath it. That is a real structural change. It just shows up in your working life as pricing pressure rather than as empowerment.

6. What the cheap-AI dividend looks like on your desk

You are not going to self-host a 1.6-trillion-parameter mixture-of-experts model. Let us be adults about this. Serving V4-Pro at usable latency is a multi-GPU engineering problem with an on-call rota attached, and no recruiting team on earth should be solving it. Even V4-Flash, at 284B total parameters, is not something you run next to your ATS.

The dividend reaches you a different way: through the price of tools that already embed a model. This is the actual mechanism by which the 2025 prediction came true. AI screening stopped being an enterprise-only feature not because recruiters started renting H800s, but because inference got cheap enough that mid-market vendors could fold it into a normal seat price instead of selling it as a premium add-on. The entry end of the applicant tracking market now ships AI CV scoring inside a seat that costs less than a team lunch. Manatal is the clearest illustration of that shift, largely because it publishes a price at all, which is not the norm in this category.

The wider point stands regardless of which vendor you land on. The right question in 2026 is not "which model does this tool use." It is "where does my candidate data go, who is the processor, and what happens when the scoring is wrong." HeroHunt.ai sits at a different point on that spectrum, sourcing and contacting candidates autonomously rather than scoring a pipeline you already built, and it is worth pricing against a traditional ATS if your bottleneck is finding people rather than sorting them. Either way, what you are buying is the plumbing and the accountability. The intelligence, as section 3 showed, costs about 50 cents.

7. What cheap models do not fix

Every problem the original version of this article promised AI would solve is still unsolved, and none of them were ever about the price of a training run.

Bias did not go away, and it did not get cheaper to fix. A model trained on the internet learns the internet's hiring patterns. Making that model 50x cheaper makes biased screening 50x cheaper, which is not an improvement. Bias mitigation is a data and evaluation problem: you need labelled outcomes, adverse-impact testing across protected groups, and someone whose job it is to look. None of that got easier in 2026.

The compliance clock is still running, it just moved. Under the EU AI Act, AI used for recruitment, candidate filtering and evaluation is a high-risk Annex III system. Those obligations were due to apply from 2 August 2026, but the Digital Omnibus, given its final green light by the Council on 29 June 2026, deferred stand-alone Annex III high-risk obligations to 2 December 2027 - Gibson Dunn. You have roughly seventeen more months. Note what the delay does not do: it does not change the classification, and it has no bearing whatsoever on GDPR, which applied to your screening tool the entire time.

And your data is still your ceiling. Predicting who will succeed in a role requires knowing who succeeded before, which requires performance data most companies do not hold in any usable state. A cheaper model does not conjure the labels. It produces confident-sounding output over the same thin evidence, faster and for less money.

8. What to actually do

If you take one thing from the eighteen months since that Nvidia crash, take this: the model was never the hard part, and now it is not even the expensive part. Which means every hour spent choosing between DeepSeek, GPT and Claude for CV screening is an hour not spent on the things that decide whether AI screening works at your company.

Concretely. Assume the intelligence is free and commoditised, because at $0.14 per million tokens it effectively is, and make your decisions about everything else. Never send candidate data to a hosted endpoint whose privacy policy names a jurisdiction without an adequacy decision, and read the policy rather than the marketing page: DeepSeek's says China, in plain English, in the first section. If you genuinely want DeepSeek's economics, take the MIT weights and run them through an EU-region host under a contract, which is a real and boring option available today. And if you are a normal recruiting team with normal problems, buy AI inside a tool with a published price, a DPA and someone to shout at, then spend the energy you saved on the evaluation loop: sample the scores, check them against outcomes, and find out where the thing is wrong before it is wrong at scale.

The revolution in the title did happen. It just did not look like a Chinese lab dethroning Silicon Valley. It looked like the price of reading a CV falling to a rounding error, while every genuinely difficult part of hiring stayed exactly as difficult as it always was.

Written by Yuma Heymans (@yumahey), who built HeroHunt.ai and its AI Recruiter, and has spent the years since 2021 watching the price of a model matter far less to recruiting outcomes than what you feed it.

This article was substantially rewritten in July 2026. Model versions, API prices and regulatory positions in this field move fast: verify current details against the primary sources linked above before acting on them.