DeepSeek and Huawei challenge Nvidia’s AI software advantage

DeepSeek has released programming tools developed with Huawei’s support for Ascend AI chips, strengthening a Chinese alternative to Nvidia’s software ecosystem. The September 30, 2026 announcement could make selected AI workloads easier to move between platforms. Its wider effect will depend on production reliability, hardware supply and total cost. For the Philippines, the clearest opportunities are in engineering, electronics, assembly and testing, and businesses that try these tools still carry Data Privacy Act duties for any personal data involved.

News analysis · Reporting checked October 2, 2026. Reuters reported the partnership on September 30, citing DeepSeek’s announcement. The public software repositories corroborate technical collaboration. They do not establish an exclusive commercial agreement or a complete departure from Nvidia.

The development matters because AI competition extends beyond designing a powerful processor. Developers need practical ways to use it. Better software can make an alternative chip more attractive, but a downloadable library alone cannot establish dependable, competitively priced AI services.

What did DeepSeek and Huawei actually release?

The new infrastructure addresses computation and communication, two essential parts of large AI systems. Training builds a model; inference uses that model to answer requests. Both involve repeated mathematical operations, and large deployments need many processors to exchange data efficiently.

• DeepGEMM-Ascend handles matrix computation. Its official repository dates the initial release to September 30 and identifies Ascend 950 support. It adapts DeepSeek’s existing library with matching programming interfaces and explicitly acknowledges Huawei’s engineering assistance.
• DeepEP-Ascend handles communication between processors. The project documentation describes training and inference support, including the movement of work among the specialized components of mixture-of-experts models.
• TileLang adds an upstream Ascend 950 backend. A pull request merged into the main TileLang repository on September 29 introduces a native backend for Ascend 950 processors. This kernel-programming project provides a higher-level way to write performance-sensitive operations and also supports Nvidia, AMD and other hardware, so treating it as an exclusively Huawei language would miss part of its value.

This is an expansion of an existing effort. The earlier TileLang-Ascend project records an open-source release in September 2025. The current news concerns the coordinated software release and newer upstream support, rather than the first occasion on which DeepSeek-related code could run on Huawei hardware.

How serious is the challenge to Nvidia’s software advantage?

Nvidia’s CUDA platform encompasses compilers, libraries and development tools. A company choosing an AI system also considers whether its engineers can debug problems, find working examples and keep applications running after updates. Existing code and experience therefore influence purchasing alongside processor specifications.

Analysis: When alternative hardware can use familiar interfaces and higher-level tools, developers may spend less effort rewriting selected operations. That can lower one barrier to adopting Ascend and make a second platform more practical. It does not establish identical behavior, performance or support across every model.

The release documentation contains an important qualification. DeepEP-Ascend’s reported benchmarks used a Huawei proof-of-concept hardware-development package with manual configuration, rather than a publicly released commercial version. The project points to a commercial release planned for around October 15, 2026; that availability had not been established when this article was last checked on October 2. Some functions remain experimental or incomplete, and validation is specified for Ascend 950DT.

A fast individual kernel is useful evidence about that operation. It is not an independent, end-to-end comparison of model quality, training time, service uptime or operating cost. The important next test is whether customers can reproduce good results using a supported commercial configuration over sustained workloads.

Open-source availability helps others inspect and improve code. Procurement teams still need to establish hardware availability, security maintenance, integration support and the cost of recovering from failures.

What could change in global technology markets?

The most direct competitive pressure is in China. Better domestic tools give Chinese AI companies another option when access to foreign hardware is restricted or uncertain. If that option works reliably, more engineering effort and customer investment could accumulate around it. Adoption elsewhere would also require lawful availability, distribution, support and compelling economics.

Export controls remain part of the picture. In January 2026, the US Bureau of Industry and Security announced case-by-case review for exports of Nvidia H200, AMD MI325X and similar chips to China under specified conditions. That is more complicated than a blanket assertion that all Nvidia hardware is banned. Earlier US controls also covered high-bandwidth memory and chipmaking technologies, citing national-security concerns.

Analysis: Restrictions can make advanced technology harder to obtain while also strengthening incentives to build substitutes. This software announcement does not settle which effect will dominate. Manufacturing, memory, packaging, electricity and interconnects can still constrain a system even when its programming tools improve.

Nvidia is continuing to develop its own platform. Its CUDA release notes describe ongoing software and hardware support work. Its second-quarter fiscal 2027 results, released August 26, 2026, showed a large global data-center business and stated that its outlook did not assume any data-center compute revenue from China. These are company disclosures, not a response to the September 30 release.

The commercial repercussions could reach cloud operators, memory makers, network suppliers and packaging companies as customers diversify. But a software release alone cannot establish future market shares, profit margins or share prices. Cheaper computation could also encourage more AI use, so a rival gaining workloads does not necessarily mean the total infrastructure market shrinks.

Who supplies AI chips and the components around them?

The companies below occupy different parts of the supply chain. A designer, wafer foundry, memory producer and outsourced assembly-and-test provider are not interchangeable competitors. “Main markets” describes representative customers and applications, not a market-share ranking or a guarantee that every product is available in the Philippines.

AI chip and accelerator designers

These companies design the processors and the software that AI developers program against.

Foundries, memory, assembly and test

These companies manufacture, supply memory for, package or test the chips that designers create.

These are examples rather than a complete supplier map. TSMC and SMIC manufacture customers’ designs; memory makers supply another essential component; assembly-and-test providers package or validate devices. A supplier’s global capabilities must not automatically be attributed to every factory it operates.

How might everyday consumers feel the effects?

Lower-cost AI services are possible, but conditional. If competition and better utilization reduce the cost of useful computation, providers could offer cheaper subscriptions, higher usage limits or more capable features at the same price. Whether savings reach customers depends on competition and providers’ pricing decisions.

There are costs pulling in the other direction. Reuters reported on September 10 that memory shortages were driving up prices for Chinese AI accelerators, citing people familiar with pricing. Hardware purchase price also leaves out electricity, cooling, porting, maintenance and spare capacity. “Open source” does not mean the service is free to operate.

More choice could improve resilience. Applications with viable infrastructure alternatives may be less exposed to one supplier’s shortages or policy changes. A Philippine business could eventually have more cloud or on-premises options. That benefit requires working services and support, rather than merely an available code repository.

Fragmentation could also create friction. Different regional rules and software stacks may mean different features, prices or performance depending on where a service runs. Maintaining several versions can cost providers money and slow updates.

For an ordinary user, the first visible change is more likely to be an AI app’s price, limits or responsiveness than a new chip on a retail shelf. There is no demonstrated direct link from this release to cheaper phones, laptops or gaming cards. Chip origin also does not determine an app’s privacy or accuracy; assess the operator, data handling and deployment. See our guide to AI business accountability in the Philippines.

Where is the realistic Philippine opportunity?

The Philippines has an existing semiconductor and electronics base to build on. Amkor says its Philippine factories provide a full range of assembly and test services. Analog Devices identifies assembly and test operations in Cavite. IMI’s test and systems engineering, alongside electronics manufacturing, is another relevant but different capability.

These operations do not establish that the country fabricates leading-edge Ascend or Nvidia processors, or that Philippine plants perform every advanced packaging process offered by their parent companies. No DeepSeek–Huawei supply contract with a Philippine company has been confirmed for this article.

Analysis: four practical openings deserve attention.

1. Assembly, testing and qualification. More chip architectures and system variants could create demand for test development, reliability work and failure analysis. Advanced AI packages would require the appropriate equipment, processes and customer qualification.
2. Power and supporting electronics. AI servers also need power conversion, sensing, control, connectivity and well-engineered boards. Existing analog and electronics skills offer possible entry points around the main accelerator.
3. Software and systems engineering. Philippine teams could help customers port workloads, evaluate models, integrate systems and support deployments across several platforms. Demonstrating useful performance and reliability would matter more than promising universal compatibility.
4. Industrial and edge products. Manufacturers could pursue customer-approved designs for inspection equipment, controllers and AI-enabled devices where thermal management, lifecycle support and dependable operation are valuable.

These are potential business directions, not announced orders or job forecasts. The OECD’s December 2024 study of the Philippine semiconductor ecosystem identified infrastructure and regulatory bottlenecks, workforce development and the wider business environment as areas needing action. Winning higher-value work requires addressing them as well as promoting investment.

Proposed projects also need careful labels. An August 2026 Philippine Information Agency report quoted the Bases Conversion and Development Authority describing the New Clark City Pax Silica project as a semiconductor and advanced-manufacturing industrial ecosystem, not a data center. It should not be counted as an operating AI-chip factory or a confirmed customer for this partnership.

Finally, location alone does not remove export-control obligations. Philippine firms considering restricted technologies or counterparties need qualified, transaction-specific compliance advice. The opportunity is to become a capable, dependable supplier within applicable rules.

What should Philippine businesses and developers do before adopting these tools?

Treat the new tools as an engineering option to test, not as a compliance shortcut. If personal data is involved, the Data Privacy Act of 2012 (Republic Act No. 10173) applies whichever chip or provider runs the model, and the Philippine business that decides how the data is processed stays accountable. See our Data Privacy Act guide for the wider framework.

What the law requires (binding statute):

• You remain responsible after a transfer. Section 21 makes each personal information controller responsible for personal information transferred to a third party for processing, “whether domestically or internationally,” and requires contractual or other reasonable means to provide a comparable level of protection (RA 10173, Sec. 21).
• Outsourcing needs safeguards. Section 14 allows a controller to subcontract processing only if it ensures proper safeguards for confidentiality, prevents unauthorized use and complies with the Act; the processor must comply too (RA 10173, Sec. 14).
• Security must fit the risk. Section 20 requires reasonable and appropriate organizational, physical and technical measures, taking into account the nature of the data, the risks, the organization’s size and current best practices (RA 10173, Sec. 20).
• Foreign providers can be covered. Section 6 extends the Act to processing outside the Philippines that relates to Philippine citizens or residents where the entity has specified links to the country (RA 10173, Sec. 6).

What the regulator expects (NPC guidance): NPC Advisory No. 2024-04, dated December 19, 2024, applies the Data Privacy Act to AI systems that process personal data, including during training and testing. It calls for transparency to data subjects, privacy impact assessments, mechanisms for meaningful human intervention in automated decisions, bias monitoring and data minimization. It also states that when processing is outsourced, controllers remain accountable for their processors. The advisory is regulator guidance applying the Act, not a separate AI statute.

Practical steps (best practice, not legal requirements):

1. Map the data flow first. Running open-source libraries such as DeepGEMM-Ascend on infrastructure you control does not by itself send data to DeepSeek or Huawei. Using a hosted AI service or cloud operated abroad does send data to a third party. Record which applies, what personal data moves and where it is processed.
2. Test with non-personal data. Benchmark porting effort, accuracy and cost on synthetic or de-identified workloads before production data touches a new platform. DeepEP-Ascend’s own documentation says its published results used a proof-of-concept hardware kit with manual configuration.
3. Assess before going live. Complete a privacy impact assessment before an AI system processes personal data in production, as the NPC advisory expects.
4. Put provider terms in writing. Cover processing instructions, security, subprocessors, processing locations, retention and deletion, breach notification and whether your data may be used for training. Our AI vendor due diligence checklist lists the questions.
5. Treat export controls as a supply and contract risk. US policy on advanced chips for China changes; in January 2026 the Bureau of Industry and Security moved H200- and MI325X-class exports to case-by-case review under conditions. These are not Philippine laws, but they can affect whether hardware can be bought, supported or resold. Check supplier terms, end-use declarations and counterparties, and get transaction-specific trade-compliance advice.
6. Keep an exit path. Code written in a layer that compiles for more than one backend, such as TileLang, can reduce dependence on a single chip vendor.

Where to go and what to do first: the National Privacy Commission administers and enforces the Data Privacy Act and publishes its advisories and circulars. For a specific export-control question, rely on the issuing authority’s published rules and a qualified trade-compliance adviser. Your first action: list every AI workflow that would send personal data to an outside provider and confirm that each one has a written agreement covering the points above.

What evidence should readers watch next?

Three developments would make the commercial significance clearer:

• Production readiness: a supported commercial software and firmware configuration, followed by reproducible deployments
• Complete-workload results: comparable model quality, latency, uptime, energy use and total cost under disclosed conditions
• Actual supply and customers: delivered hardware, documented adoption and qualified suppliers, rather than roadmaps alone

One plausible outcome is a useful second platform for selected workloads. Another is improving software held back by hardware or operating costs. A third is greater regional adoption accompanied by more integration and compliance overhead. These are scenarios, not forecasts; several could occur together.

Frequently asked questions

Has DeepSeek stopped using Nvidia?

The September 30 announcement does not establish a complete exit. It confirms new Ascend-oriented tools and technical collaboration with Huawei.

Does TileLang replace all of CUDA?

No. It is a kernel-programming layer with support for multiple hardware platforms. CUDA includes a wider collection of tools and libraries.

Will this immediately lower Philippine gadget prices?

There is no demonstrated immediate effect. Any consumer benefit would depend on subsequent deployment, costs, availability and providers’ pricing.

Is there already a Philippine supplier deal?

No DeepSeek–Huawei contract with a Philippine supplier has been confirmed for this article. The local opportunities discussed here are analysis.

Can a Philippine business send customer data to a foreign AI provider?

Section 21 of the Data Privacy Act contemplates processing by third parties abroad, but the business stays accountable and must use contracts or other reasonable means to ensure comparable protection. NPC Advisory No. 2024-04 also expects a privacy impact assessment for AI systems that process personal data.

Bottom line

DeepSeek and Huawei have made a concrete software contribution to an alternative AI-computing ecosystem. Its importance will grow if customers can turn the public code into reliable, economical production systems.

For consumers, watch the quality and cost of services rather than assuming cheaper hardware. For Philippine industry, the practical goal is to win qualified engineering, electronics, assembly and test work while developing more advanced capabilities. Commercial results will depend on customers, investment and execution.

Sources

• Reuters (via MarketScreener), September 30, 2026: DeepSeek partners with Huawei on chip programming tools
• DeepGEMM-Ascend repository (DeepSeek) and DeepEP-Ascend repository (DeepSeek)
• TileLang pull request #3308: Ascend 950 backend
• US Bureau of Industry and Security, January 13, 2026 license review policy
• Nvidia second-quarter fiscal 2027 results
• Republic Act No. 10173, Data Privacy Act of 2012 (National Privacy Commission)
• NPC Advisory No. 2024-04 on AI systems processing personal data
• OECD, Promoting the growth of the semiconductor ecosystem in the Philippines (2024)
• Philippine Information Agency, August 10, 2026: BCDA on the Pax Silica project

Sources and method: This analysis draws on Reuters reporting, public project documentation, company disclosures, Philippine industry sources and National Privacy Commission materials. Company benchmarks and future plans are attributed; this article does not present independent hardware testing.

Featured image: illustrative hardware photograph, not a Huawei Ascend processor or DeepSeek facility. Photo by Umberto on Unsplash, used under the Unsplash License.

Sources rechecked as of: October 2, 2026.