On This Page8 sections
Key Takeaways
- France has advanced a proposal, not activated a new download tax. The National Assembly adopted a budget amendment on October 30, 2026; the legislative process remains unfinished.
- The proposed rate is €0.055 per gigabyte, with content providers bearing the proposed liability rather than individual internet subscribers.
- 510 GB × €0.055 = €28.05, but that calculation is an uncapped scenario—not an established charge for downloading a DeepSeek model.
- Actual exposure would depend on transferred bytes, the responsible entity, revenue caps and the final rules. Reporting on the amendment confirms that it could still change or disappear.
Updated October 31, 2026. The central issue for open-source AI is whether distributing a free model could create an additional cost for the organization delivering its files.
What France Actually Voted For
The official voting record identifies amendment I-1358, proposed by Laurent Lhardit, during the first reading of France’s 2027 finance bill.
It passed with 11 votes in favor, 10 against and 14 abstentions. Those figures describe this particular amendment vote, not approval of the completed budget by the whole legislature.
The budget’s legislative dossier still records the National Assembly’s first-reading stage. There is therefore no basis for treating the amendment as an operational tax or promising a confirmed collection date.
For businesses, the practical distinction is between modeling a possible expense and recognizing an existing tax obligation. This development supports the former.
How the Proposed Internet Data Tax Would Work
The amendment text calls the measure a Taxe sur les producteurs de données numériques, or tax on digital data producers.
| Element | Proposed treatment |
|---|---|
| Tax base | Data transmitted by content providers toward French internet users |
| Rate | €0.055 per GB |
| Proposed payer | Content provider/data emitter |
| Measurement | Internet access providers, supervised by Arcep and the tax administration |
| Reporting | Quarterly, aggregated and anonymized |
| CDN delivery | Outgoing CDN traffic is expressly included |
| Revenue caps | Wording refers to 15% of French turnover or 2% of total turnover |
| Implementation | Technical measurement rules would be defined by decree |
The text prohibits requiring records of the data’s content or destination. It does not name an open-source exemption or a small-business turnover threshold. The interaction of the two caps should not be converted into a definitive tax formula without further clarification.
Why Supporters Want It—and Why the Government Objects
Supporters present traffic as a measurable indicator of digital activity and expect the measure to encourage compression. Their projected annual revenue is approximately €10 billion, according to Les Numériques’ coverage. That is a proposal-stage forecast, not a validated collection result.
The official parliamentary debate records two substantial objections:
- Cost shifting: Budget rapporteur Philippe Juvin warned that powerful technology companies could pass the burden to advertisers.
- Measurement and scope: Public accounts minister David Amiel questioned implementation and argued that approximately half the projected burden could fall outside the largest technology companies, including on smaller French providers.
Lhardit disputed the scale of the domestic impact. These were competing political assessments, not settled incidence studies. Both the government and the rapporteur opposed the amendment.
The broader economic question is straightforward: bytes do not measure profit. A large, freely distributed research file and a profitable advertising service can have very different revenues for the same transfer volume.
Would Downloading DeepSeek Really Generate €28.05 in Tax?
The official DeepSeek-V4.1-Flash-1017 file listing displays a repository size of approximately 510 GB. Its model card identifies the repository and model weights as MIT-licensed.
Using the displayed size as a hypothetical transferred volume gives:
510 GB × €0.055/GB = €28.05The arithmetic is correct. The conclusion requires additional assumptions:
- A user receives the full 510 GB.
- Those bytes fall within the eventual taxable measurement.
- The responsible provider is identified.
- Applicable caps or other final provisions do not reduce the amount.
Repository size alone establishes none of those conditions.
A smaller compatible model variant, a previously downloaded file or an interrupted transfer can produce a different network footprint. Conversely, repeated downloads can increase it.
The MIT label describes licensing permissions. It does not establish which infrastructure company would owe a hypothetical French tax.
What €0.055 per GB Means in Practice
The following are independent scenario calculations. They assume all listed bytes are in scope, use decimal units and ignore caps, exemptions, overhead and contractual allocation.
| Hypothetical delivery | Total transferred | Uncapped amount |
|---|---|---|
| One 10 MB image | 0.01 GB | €0.00055 |
| One 100 MB video | 0.1 GB | €0.0055 |
| One 5 GB file | 5 GB | €0.275 |
| One 100 GB game download | 100 GB | €5.50 |
| One 510 GB model download | 510 GB | €28.05 |
| 10,000 downloads of a 100 MB video | 1,000 GB | €55 |
| 1,000 full downloads of a 510 GB model | 510,000 GB | €28,050 |
For a first-pass scenario:
Uncapped amount in EUR = potentially in-scope transferred bytes / 1,000,000,000 × 0.055Keep GB and GiB separate. A GiB contains 1,073,741,824 bytes; substituting GiB values into a decimal-GB calculation introduces roughly a 7.4% difference. Preserve raw byte counts so the eventual measurement definition can be applied consistently.
These numbers are neither current invoices nor evidence that a platform will increase its prices by the same amount.
CDN Caching and Local Caching Have Different Effects
A CDN cache and a local cache sit at different points in the delivery path.
CDN cache: The file is already available near the user, so the origin server may avoid another transfer. However, the user still receives bytes from the CDN. Because the proposal expressly includes CDN output, an origin-bandwidth saving cannot automatically be treated as a corresponding reduction in the proposed tax base. This is an inference from the CDN provision.
Local cache: The file is already on the user’s machine. Reusing it can avoid another download of its contents. Hugging Face’s cache documentation explains how cached files are reused.
For example, ten complete transfers of a 10 GB file represent 100 GB of delivered file data. One transfer followed by nine local reuses represents 10 GB, excluding any metadata requests.
Hugging Face also documents Xet-based storage and transfer mechanisms. Their existence does not establish a fixed download saving for every client or file. Measure transferred bytes instead of assuming that deduplication eliminates network traffic.
Who Would Pay: DeepSeek, Hugging Face or the Delivery Provider?
The available evidence does not resolve liability for this particular delivery chain.
A useful operational review separates three roles:
| Role | Question to investigate |
|---|---|
| Model publisher | Who owns or publishes the model repository? |
| Hosting platform | Who operates the service through which the user requests it? |
| Storage or delivery provider | Whose infrastructure transmits the file bytes? |
These roles can belong to different organizations. Identifying the publisher from a repository page does not establish the tax treatment of every downstream transfer.
Likewise, an infrastructure provider passing a cost to a customer under a contract would be a separate question from the identity of the statutory taxpayer.
A statement that Hugging Face already owes €28.05 for every French DeepSeek download goes beyond what the vote, repository listing and proposed rules establish.
The Business Risk Depends on Revenue per Gigabyte
For product planning, traffic volume becomes more informative when compared with the revenue supporting it.
Consider two hypothetical services before caps and other adjustments:
| Scenario | Revenue | Delivered data | Uncapped amount | Amount as a share of revenue |
|---|---|---|---|---|
| Service A | €1,000 | 1,000 GB | €55 | 5.5% |
| Service B | €1,000 | 10,000 GB | €550 | 55% |
These are sensitivity examples, not forecasts of liability.
They illustrate why a free model mirror, a video subscription and a lightweight text application should not share one exposure estimate. Distribution intensity and monetization matter together.
An AI service’s GPU consumption is also a separate variable. Large model weights do not travel to every user when the model runs on a server. For a text API, investigate the actual request-and-response delivery path; for downloadable weights, investigate the model-file delivery path.
A blanket claim that all AI products face the same bandwidth exposure would obscure this distinction.
What Developers and Website Operators Can Do Now
The useful response is a modest measurement exercise, not an emergency infrastructure migration.
- Record delivered bytes by product and country. Keep the measurement method alongside the number; analytics location estimates are planning inputs, not an official tax determination.
- Separate origin traffic from user-facing delivery. Avoid combining them into one total that may count the same content twice.
- Distinguish full downloads, retries and partial transfers. A successful-download counter is not a byte meter.
- Preserve reusable local model files. Check that deployments are not repeatedly downloading unchanged weights into disposable environments.
- Offer appropriate file sizes. Smaller video exports or model variants can reduce delivery volume when they still meet the user’s requirements.
- Map responsibility and contracts. Document the publisher, hosting platform, CDN and billing relationships before assigning a potential expense.
- Maintain separate planning cases. Compare no enactment, enactment without material changes and a revised proposal with different scope or caps.
For AI video tools, model repositories and game distributors, the immediate deliverable should be a short exposure worksheet. It should distinguish observed bytes, assumed scope and unresolved legal treatment.
Questions the Final Rules Would Need to Resolve
Before converting a traffic scenario into a compliance process, operators would need reliable answers to questions such as:
- How is the responsible emitter identified when multiple customers share infrastructure?
- Which measurement layer determines taxable bytes?
- How are retransmissions, incomplete downloads and protocol overhead treated?
- How do the revenue caps interact, and at what entity or group level?
- What treatment applies to nonprofit distribution and projects without commercial revenue?
- How can providers reconcile or challenge measured volumes?
These are implementation questions, not claims that the draft already provides particular answers. An operational forecast should record uncertainty explicitly instead of silently filling these gaps.
Frequently Asked Questions
Is France already charging people for each gigabyte they download?
No. The development reviewed here is a proposed provider-side tax within an unfinished budget process. It does not establish a new charge on individual internet bills. Contemporaneous reporting distinguishes the vote from entry into force.
Would a free AI model necessarily remain free to distribute?
A zero purchase price does not eliminate hosting and transfer expenses. Whether this proposal would add a tax expense depends on the final law and responsible entity.
Does the €28.05 figure describe the price of DeepSeek?
No. It is the result of applying the proposed rate to a hypothetical 510 GB transfer. It is not a model license price or a confirmed download fee.
Should a website block French visitors now?
The reviewed evidence does not justify that response. Measure potential exposure and track the final legislation before making changes that restrict access.
Conclusion
France’s proposed internet data tax raises a serious question for services distributing large files: how should a volume-based cost interact with free access, subscriptions and open-source distribution?
The €28.05 DeepSeek example makes that question concrete, but it cannot settle liability, actual transfer volume or the effect of caps.
Review the 2027 budget dossier, measure user-facing delivery and build an explicitly hypothetical cost scenario before changing pricing or access. The next decisive evidence will be the final legal text and its implementation rules.
Continue Reading
More articles connected to the same themes, protocols, and tools.
Referenced Tools
Browse entries that are adjacent to the topics covered in this article.









