Last week, a client asked me on WeChat: "We've implemented ERP, and all our business data is in the system. Can we apply for that high-quality dataset in Baiyun District?"
I replied to him: First, look and see whether there is any document from the national, provincial, or municipal competent authorities that states the name of your unit in black and white.
He searched for a long time, but didn't find it. That's the end of this matter.
On November 6 at 21:52, the Guangzhou Baiyun District Government Services and Data Management Bureau posted on the district government portal website the "Notice on Launching Applications for Several Measures to Support the High-Quality Development of the Artificial Intelligence and Data Industry in Baiyun District, Guangzhou"; the same page attached five .docx thematic guidelines. The application window is November 9 to November 17, 2026 (Guangzhou Baiyun District Government Services and Data Management Bureau, issued 2026-11-06, pursuant to Yunfu Ban [2026] No. 65). The "VI. Application Time" section of the guidelines is very strict: overdue applications are deemed withdrawn, and the next year will no longer accept supplementary applications for the previous year.
Having read all five guidelines, the first thing worth mentioning is not the amount, but the application logic.
Four special topics require a name list; one special topic requires five certificates
The five guidelines are divided into three groups: computing power; high-quality datasets and trusted data spaces under the data element market; large model launch filing and algorithm filing under core algorithm models.
The latter four share a common point: you must first have been recognized by others. High-quality datasets require that, in the same period, they be assessed by national, provincial, or municipal competent authorities (typical case lists, pilot trial lists, chain leader entities, competition awards); trusted data spaces require being selected for a pilot list at the municipal level or above; large model launch filing requires completing, for the first time, national-level filing for the launch of generative artificial intelligence services; algorithm filing requires completing, for the first time, national-level algorithm filing for domestic deep synthesis services, with the type being 'generative synthesis' and the status normal.
In the "IV. Application Materials" section of the four guidelines, the most critical item is not the project summary written by the enterprise itself, but the recognition document, selection notice, list announcement, or screenshot of the public notice page—the screenshot is also required to show the issuing unit, list name, publication date, applicant unit name, and its sequence number.
Computing power vouchers are the only exception; they are calculated based on spending: if the cost of purchasing computing power or token services from non-affiliated parties during the same period exceeds RMB 1 million, a subsidy of 10% of the actual service purchase cost is granted, and for OPC and other micro and small startups applying for the first time, this is raised to 15%; the total annual amount of computing power usage subsidies redeemed shall not exceed RMB 10 million (Guangzhou Baiyun District Government Services and Data Administration, Application Guidelines (Special Topic on Strengthening Computing Power Infrastructure Support), 2026-11-06; the basis is the actual tax-exclusive service purchase cost during the application period; RMB 10 million is the upper limit for the district's total annual redemption amount, and differs from the denominator used for the enterprise subsidy ratio).
The 'Compute Power and Token Service Procurement and Usage Aggregation Table' to be submitted lists five types of supporting documents: procurement contracts/orders, invoices, bank payment vouchers and receipts, service provider settlement statements, and usage logs or system records. The guidelines require that the subject, amount, and period of each item of material correspond to one another; fees paid by means other than bank transfer and without bank payment records shall not be included. There is also a series of exclusions: non-computing-power cloud resources such as storage, network, security, database, and CDN; ordinary CPU computing power; hardware purchases; data center leasing and electricity charges; on-site operations and maintenance and labor outsourcing; and third-party large model licensing or one-time buyout.
Where the amount is set is where the pitfall lies
| Special Topic | grade | Threshold | Upper limit for a single entity |
|---|---|---|---|
| High-quality dataset | National-level 1 million / Provincial-level 500,000 / Municipal-level 200,000 | Must be the lead organization, and the certification document must state the name of this organization | 2 million per unit per year; for the same dataset, the higher amount applies without duplication. |
| Trusted Data Space | National-level 1 million / Provincial-level 700,000 / Municipal-level 300,000 | To be selected for the list of pilots at or above the municipal level, the entity must be a lead enterprise. | For the same space, choose the higher one and do not duplicate. |
| Large Model Launch Filing | Up to 1 million | First national-level launch filing, no fewer than 2 market application cases, and actual completed contract value exceeding RMB 5 million | Did not receive similar funding support from other regions |
| Algorithm Filing | Up to 200,000 | First National-Level Domestic Deep Synthesis Service Algorithm Filing | The same algorithm for filing only needs to be reported once. |
| Computing power voucher | 10%; first-time micro and small startups 15% | Annual fees for purchasing computing power/token services exceed 1 million yuan | The total annual disbursement amount for the entire district is 10 million yuan. |
Table note: The amounts, ratios, and thresholds all come from the five thematic guidelines on the same page of the Baiyun District People's Government portal notice dated 2026-11-06, "II. Support Content and Award Standards" and "III. Application Conditions." The amounts of 100/50/20 ten thousand yuan and 100/70/30 ten thousand yuan are assessment tier brackets, not cumulative items; competition awards are converted at 100% for the Special Prize and First Prize, 80% for the Second Prize, and 60% for the Third Prize. The guideline example states, "the provincial Second Prize shall not exceed 400,000 yuan" (500,000 × 80%). All amounts are upper limits; Section V, Notes (2) of the guideline explicitly states that actual disbursement is subject to the control of the total annual fiscal budget.
For the large model launch filing topic, "actual completed contract amount exceeding RMB 5 million" is neither the contract amount nor the invoiced amount. Annex 4-10 of the guidelines is very detailed: the tax-exclusive amount invoiced for that contract during the application period is used as the calculation base for the RMB 5 million threshold; receipts are based on bank receipts; if invoicing and receipts are inconsistent, the smaller value is taken; voided or red-reversed invoices are not counted; framework agreements and letters of intent are not counted; if invoiced and receipts are not less than 30% of the contract amount, it is deemed to have been performed.
Take one example: for a contract worth 3.2 million yuan, with 1.8 million yuan invoiced during the reporting period and 1.5 million yuan received in the bank, the amount that can be counted toward the “actual completed contract amount” is 1.5 million yuan. Reaching 5 million yuan relies on several actual payments received, not a polished framework agreement.
Which tables in Odoo do these materials fall into
The question customers ask most often is 'Can these be exported directly from these digital systems?' Yes, but only if the documents were created in a verifiable way from the start.
The five sets of vouchers for compute vouchers correspond in Odoo 19 CE to purchase orders (purchase.order), vendor bills (account.move), payments (account.payment), settlement statement attachments, and external usage logs. The native approach is sufficient: collect under the same analytic account, generate payments from bills, attach settlement statements and logs to the same project, and any one voucher can trace back to the other four. What often needs supplementing is not functionality, but the lack of unified numbering at the outset—invoices numbered by financial transaction, contracts by sales, logs by cloud vendor order—so at filing time you can only manually stitch together Excel.
The detailed schedule in the large-model filing special topic tests the matching between invoice lines and receipts: account.move.line issues invoices transaction by transaction, account.payment is linked to the corresponding bill, and the aging report can show which payment has not been received. One situation we have encountered is that the contract amount signed is large but only 30% of the payment is received; the enterprise fills out the form based on the contract amount, and the review is returned as soon as the bank receipt is checked. The guide's instruction to "take the smaller value" is precisely to prevent this.
The two data-related application forms ask about another matter. For high-quality datasets, you need to fill in the dataset type, data modality, volume (TB), composition of data sources (enterprise internal business data, authorized operation of public data, public data openness, collaborative collection across the industrial chain, external procurement), as well as the launch time and call or download status of the application scenario; for trusted data spaces, you need to fill in the number of participating construction units, the number of data product catalogs available for circulation, and the total number of calls.
These fields are not copywriting questions. If you check “enterprise internal business data,” you must clearly state which table and field it comes from and who has the authority to modify it; if you fill in “total number of calls,” there must be logs that can be replayed. What manufacturing can truly present is usually material master data, process parameter databases, and equipment maintenance and quality inspection records—provided that “one code per item, one file per customer” has been implemented. One honest caveat: this round is assessment-result-based awards and subsidies, not data governance subsidies. Without an approved list, no matter how standardized the system is, it cannot be submitted; if you are on the list but cannot clearly explain the data lineage, you may still be questioned during the public announcement period.
There are two more general provisions: the same investment, order, or achievement may not be used to apply repeatedly for similar projects at the municipal and district levels; in principle, district-level fiscal support to the same enterprise in the same year shall not exceed its economic contribution in Baiyun District in the previous year. The guidelines also state that 'in principle, the support funds applied for in the current year will be disbursed in the following year'—do not count this RMB 1 million in this year's fourth-quarter cash flow.
Before the window closes on the 17th, don't reverse the order
Between today and November 17, first take stock of the list; do not start writing materials yet. Make a list of all national/provincial/municipal recognition documents, public announcement lists, filing numbers, and award certificates under the company's name since January 1, 2026, and check item by item whether the level of the issuing authority and the unit name are consistent with the business license (if there has been a name change, also find the change registration materials). If the list is empty, directly abandon the four assessment-type topics.
If using computing power vouchers, complete a five-document alignment drill before November 13. The condition is that the annual purchase cost of computing power/token services actually exceeds 1 million yuan and all payments are made via bank transfer. The action is to pull together the purchase order, invoice, payment receipt, settlement statement, and usage log, and check the entity, amount, and period against the same contract number; exclude storage/CDN/CPU/hardware/data center/labor outsourcing/model buyout. If an invoice cannot be matched to a bank payment receipt, do not fill in this item.
If you are going through the large-model launch filing, recalculate the 'actually completed contract amount' according to the guideline's definition by November 14. The criteria are a first national-level launch filing and cases from no fewer than 2 non-affiliated customers. For each transaction, take the smaller of the 'invoiced amount excluding tax' and the 'bank receipt amount' and sum them; confirm that the cumulative amount exceeds RMB 5 million (not including the threshold itself), then check whether the amount received under each contract reaches 30% of the contract amount. If it falls just short, do not use framework agreements to make up the difference; the guideline explicitly says they are not counted.
Don't reverse the order of the three steps above: first determine the list, then match the vouchers, and only then fill in the system.
—— Basis: the 2026-11-06 notice on the Baiyun District People's Government portal website and the five thematic guides on the same page. Amounts and terms are quoted from the original text of the guides; the sample calculations are illustrative examples and are not actual application data for any enterprise. The media's reference to 'a maximum of RMB 10 million per single item' does not appear in the main text of the notice or guides and is not to be used as a basis for application.
