Generative AI value chain winners & losers

In short

• The recent release of DeepSeek’s R1 model has shaken tech stocks, with shares of some market participants (e.g., NVIDIA) down ~20% since January 27, 2025.
• R1 is largely open, on par with leading proprietary models, appears to have been trained at significantly lower cost, and is cheaper to use in terms of API access, all of which point to an innovation that may change competitive dynamics in the field of Generative AI.
• IoT Analytics sees end users and AI applications providers as the biggest winners of these recent developments, while proprietary model providers stand to lose the most, based on value chain analysis from the Generative AI Market Report 2025–2030 (published January 2025).

Why it matters

• For suppliers to the generative AI value chain: Players along the (generative) AI value chain may need to re-assess their value propositions and align to a possible reality of low-cost, lightweight, open-weight models.
• For generative AI adopters: DeepSeek R1 and other frontier models that may follow present lower-cost options for AI adoption.

Background: DeepSeek’s R1 model rattles the markets

DeepSeek’s R1 model rocked the stock markets. On January 23, 2025, China-based AI startup DeepSeek released its open-source R1 reasoning generative AI (GenAI) model. News about R1 quickly spread, and by the start of stock trading on January 27, 2025, the market cap for many major technology companies with large AI footprints had fallen drastically since then:

• NVIDIA, a US-based chip designer and developer most known for its data center GPUs, dropped 18% between the market close on January 24 and the market close on February 3.
• Microsoft, the leading hyperscaler in the cloud AI race with its Azure cloud services, dropped 7.5% (Jan 24–Feb 3).
• Broadcom, a semiconductor company specializing in networking, broadband, and custom ASICs, dropped 11% (Jan 24–Feb 3).
• Siemens Energy, a German energy technology vendor that supplies energy solutions for data center operators, dropped 17.8% (Jan 24–Feb 3).

Market participants, and specifically investors, reacted to the narrative that the model that DeepSeek released is on par with cutting-edge models, was supposedly trained on only a couple of thousands of GPUs, and is open source. However, since that initial sell-off, reports and analysis shed some light on the initial hype.

The insights from this article are based on

Generative AI Market Report 2025-2030

A 263-page report on the enterprise Generative AI market, incl. market sizing & forecast, competitive landscape, end user adoption, trends, challenges, and more.

---

Download a sample to learn more about the report structure, select definitions, select market data, additional data points, and trends.

Already a subscriber? View your reports here →

DeepSeek R1: What do we know until now?

DeepSeek R1 is a cost-efficient, cutting-edge reasoning model that rivals top competitors while fostering openness through publicly available weights.

1. DeepSeek R1 is on par with leading reasoning models. The largest DeepSeek R1 model  (with 685 billion parameters) performance is on par or even better than some of the leading models by US foundation model providers. Benchmarks show that DeepSeek’s R1 model performs on par or better than leading, more familiar models like OpenAI’s o1 and Anthropic’s Claude 3.5 Sonnet.
2. DeepSeek was trained at a significantly lower cost—but not to the extent that initial news suggested. Initial reports indicated that the training costs were over $5.5 million, but the true value of not only training but developing the model overall has been debated since its release. According to semiconductor research and consulting firm SemiAnalysis, the $5.5 million figure is only one element of the costs, leaving out hardware spending, the salaries of the research and development team, and other factors.
3. DeepSeek’s API pricing is over 90% cheaper than OpenAI’s. No matter the true cost to develop the model, DeepSeek is offering a much cheaper proposition for using its API: input and output tokens for DeepSeek R1 cost $0.55 per million and $2.19 per million, respectively, compared to OpenAI’s $15 per million and $60 per million for its o1 model.
4. DeepSeek R1 is an innovative model. The related scientific paper released by DeepSeekshows the methodologies used to develop R1 based on V3: leveraging the mixture of experts (MoE) architecture, reinforcement learning, and very creative hardware optimization to create models requiring fewer resources to train and also fewer resources to perform AI inference, leading to its aforementioned API usage costs.
5. DeepSeek is more open than most of its competitors. DeepSeek R1 is available for free on platforms like HuggingFace or GitHub. While DeepSeek has made its weights available and provided its training methodologies in its research paper, the original training code and data have not been made available for a skilled person to build an equivalent model, factors in defining an open-source AI system according to the Open Source Initiative (OSI). Though DeepSeek has been more open than other GenAI companies, R1 remains in the open-weight category when considering OSI standards. However, the release sparked interest in the open source community: Hugging Face has launched an Open-R1 initiative on Github to create a full reproduction of R1 by building the “missing pieces of the R1 pipeline,” moving the model to fully open source so anyone can reproduce and build on top of it.
6. DeepSeek released powerful small models alongside the major R1 release. DeepSeek released not only the major large model with more than 680 billion parameters but also—as of this article—6 distilled models of DeepSeek R1. The models range from 70B to 1.5B, the latter fitting on many consumer-grade hardware. As of February 3, 2025, the models were downloaded more than 1 million times on HuggingFace alone. 
7. DeepSeek R1 was possibly trained on OpenAI’s data. On January 29, 2025, reports shared that Microsoft is investigating whether DeepSeek used OpenAI’s API to train its models (a violation of OpenAI’s terms of service)—though the hyperscaler also added R1 to its Azure AI Foundry service.

Understanding the generative AI value chain

GenAI spending benefits a broad industry value chain. The graphic above, based on research for IoT Analytics’ Generative AI Market Report 2025–2030 (released January 2025), portrays key beneficiaries of GenAI spending across the value chain. Companies along the value chain include:

• The end users – End users include consumers and businesses that use a Generative AI application.
• GenAI applications – Software vendors that include GenAI features in their products or offer standalone GenAI software. This includes enterprise software companies like Salesforce, with its focus on Agentic AI, and startups specifically focusing on GenAI applications like Perplexity or Lovable.
• Tier 1 beneficiaries – Providers of foundation models (e.g., OpenAI or Anthropic), model management platforms (e.g., AWS Sagemaker, Google Vertex or Microsoft Azure AI), data management tools (e.g., MongoDB or Snowflake), cloud computing and data center operations (e.g., Azure, AWS, Equinix or Digital Realty), AI consultants and integration services (e.g., Accenture or Capgemini), and edge computing (e.g., Advantech or HPE)
• Tier 2 beneficiaries – Those whose products and services regularly support tier 1 services, including providers of chips (e.g., NVIDIA or AMD), network and server equipment (e.g., Arista Networks, Huawei or Belden), server cooling technologies (e.g., Vertiv or Schneider Electric).
• Tier 3 beneficiaries – Those whose products and services regularly support tier 2 services, such as providers of electronic design automation software providers for chip design (e.g., Cadence or Synopsis), semiconductor fabrication (e.g., TSMC), heat exchangers for cooling technologies, and electric grid technology (e.g., Siemens Energy or ABB)
• Tier 4 beneficiaries and beyond – Companies that continue to support the tier above them, such as lithography systems (tier-4) necessary for semiconductor fabrication machines (e.g., AMSL) or companies that provide these suppliers (tier-5) with lithography optics (e.g., Zeiss)

Winners and losers along the generative AI value chain

The rise of models like DeepSeek R1 signals a potential shift in the generative AI value chain, challenging existing market dynamics and reshaping expectations for profitability and competitive advantage. If more models with similar capabilities emerge, certain players may benefit while others face increasing pressure.

Below, IoT Analytics assesses the key winners and likely losers based on the innovations introduced by DeepSeek R1 and the broader trend toward open, cost-efficient models. This assessment considers the potential long-term impact of such models on the value chain rather than the immediate effects of R1 alone.

Clear winners

End users

• Why these innovations are positive: The availability of more and cheaper models will ultimately lower costs for the end-users and make AI more accessible.
• Why these innovations are negative: No clear argument.
• Our take: DeepSeek represents AI innovation that ultimately benefits the end users of this technology.

GenAI application providers

• Why these innovations are positive: Startups building applications on top of foundation models will have more options to choose from as more models come online. As stated above, DeepSeek R1 is by far cheaper than OpenAI’s o1 model, and though reasoning models are rarely used in an application context, it shows that ongoing breakthroughs and innovation improve the models and make them cheaper.
• Why these innovations are negative: No clear argument.
• Our take: The availability of more and cheaper models will ultimately lower the cost of including GenAI features in applications.

Likely winners

Edge AI/edge computing companies

• Why these innovations are positive: During Microsoft’s recent earnings call, Satya Nadella pointed out that “AI will be much more ubiquitous,” as more workloads will run locally. The distilled smaller models that DeepSeek released alongside the powerful R1 model are small enough to run on many edge devices. While small, the 1.5B, 7B, and 14B models are also comparably powerful reasoning models. They can fit on a laptop and other less powerful devices, e.g., IPCs and industrial gateways. These distilled models have already been downloaded from Hugging Face hundreds of thousands of times.
• Why these innovations are negative: No clear argument.
• Our take: The distilled models of DeepSeek R1 that fit on less powerful hardware (70B and below) were downloaded more than 1 million times on HuggingFace alone. This shows a strong interest in deploying models locally. Edge computing manufacturers with edge AI solutions like Italy-based Eurotech, and Taiwan-based Advantech will stand to profit. Chip companies that specialize in edge computing chips such as AMD, ARM, Qualcomm, or even Intel, may also benefit. Nvidia also operates in this market segment.

Note: IoT Analytics’ SPS 2024 Event Report (published in January 2025) delves into the latest industrial edge AI trends, as seen at the SPS 2024 fair in Nuremberg, Germany.

Data management services providers

• Why these innovations are positive: There is no AI without data. To develop applications using open models, adopters will need a plethora of data for training and during deployment, requiring proper data management.
• Why these innovations are negative: No clear argument.
• Our take: Data management is getting more important as the number of different AI models increases. Data management companies like MongoDB, Databricks and Snowflake as well as the respective offerings from hyperscalers will stand to profit.

GenAI services providers

• Why these innovations are positive: The sudden emergence of DeepSeek as a top player in the (western) AI ecosystem shows that the complexity of GenAI will likely grow for some time. The higher availability of different models can lead to more complexity, driving more demand for services.
• Why these innovations are negative: When leading models like DeepSeek R1 are available for free, the ease of experimentation and implementation might limit the need for integration services.
• Our take: As new innovations come to the market, GenAI services demand increases as enterprises try to understand how to best utilize open models for their business.

Neutral

Cloud computing providers

• Why these innovations are positive: Cloud players rushed to include DeepSeek R1 in their model management platforms. Microsoft included it in their Azure AI Foundry, and AWS enabled it in Amazon Bedrock and Amazon Sagemaker. While the hyperscalers invest heavily in OpenAI and Anthropic (respectively), they are also model agnostic and enable hundreds of different models to be hosted natively in their model zoos. Training and fine-tuning will continue to happen in the cloud. However, as models become more efficient, less investment (capital expenditure) will be needed, which will increase profit margins for hyperscalers.
• Why these innovations are negative: More models are expected to be deployed at the edge as the edge becomes more powerful and models more efficient. Inference is likely to move towards the edge going forward. The cost of training cutting-edge models is also expected to go down further.
• Our take: Smaller, more efficient models are becoming more important. This lowers the demand for powerful cloud computing both for training and inference which may be offset by higher overall demand and lower CAPEX requirements.

EDA Software providers

• Why these innovations are positive: Demand for new AI chip designs will increase as AI workloads become more specialized. EDA tools will be critical for designing efficient, smaller-scale chips tailored for edge and distributed AI inference
• Why these innovations are negative: The move toward smaller, less resource-intensive models may reduce the demand for designing cutting-edge, high-complexity chips optimized for massive data centers, potentially leading to reduced licensing of EDA tools for high-performance GPUs and ASICs.
• Our take: EDA software providers like Synopsys and Cadence could benefit in the long term as AI specialization grows and drives demand for new chip designs for edge, consumer, and low-cost AI workloads. However, the industry may need to adapt to shifting requirements, focusing less on large data center GPUs and more on smaller, efficient AI hardware.

Likely losers

AI chip companies

• Why these innovations are positive: The allegedly lower training costs for models like DeepSeek R1 could eventually increase the total demand for AI chips. Some referred to the Jevson paradox, the idea that efficiency leads to more demand for a resource. As the training and inference of AI models become more efficient, the demand could increase as higher efficiency leads to lower costs. ASML CEO Christophe Fouquet shared a similar line of thinking: “A lower cost of AI could mean more applications, more applications means more demand over time. We see that as an opportunity for more chips demand.”
• Why these innovations are negative: The allegedly lower costs for DeepSeek R1 are based mainly on the need for less cutting-edge GPUs for training. That puts some doubt on the sustainability of large-scale projects (such as the recently announced Stargate project) and the capital expenditure spending of tech companies mainly earmarked for buying AI chips.
• Our take: IoT Analytics research for its latest Generative AI Market Report 2025–2030 (published January 2025) found that NVIDIA is leading the data center GPU market with a market share of 92%. NVIDIA’s monopoly characterizes that market. However, that also shows how strongly NVIDA’s faith is linked to the ongoing growth of spending on data center GPUs. If less hardware is needed to train and deploy models, then this could seriously weaken NVIDIA’s growth story.

Other categories related to data centers (Networking equipment, electrical grid technologies, electricity providers, and heat exchangers)

Like AI chips, models are likely to become cheaper to train and more efficient to deploy, so the expectation for further data center infrastructure build-out (e.g., networking equipment, cooling systems, and power supply solutions) would decrease accordingly. If fewer high-end GPUs are needed, large-capacity data centers may scale back their investments in associated infrastructure, potentially impacting demand for supporting technologies. This would put pressure on companies that provide critical components, most notably networking hardware, power systems, and cooling solutions.

Clear losers

Proprietary model providers

• Why these innovations are positive: No clear argument.
• Why these innovations are negative: The GenAI companies that have collected billions of dollars of funding for their proprietary models, such as OpenAI and Anthropic, stand to lose. Even if they develop and release more open models, this would still cut into the revenue flow as it stands today. Further, while some framed DeepSeek as a “side project of some quants” (quantitative analysts), the release of DeepSeek’s powerful V3 and then R1 models proved far beyond that sentiment. The question going forward: What is the moat of proprietary model providers if cutting-edge models like DeepSeek’s are getting released for free and become fully open and fine-tunable?
• Our take: DeepSeek released powerful models for free (for local deployment) or very cheap (their API is an order of magnitude more affordable than comparable models). Companies like OpenAI, Anthropic, and Cohere will face increasingly strong competition from players that release free and customizable cutting-edge models, like Meta and DeepSeek.

Analyst takeaway and outlook

The emergence of DeepSeek R1 reinforces a key trend in the GenAI space: open-weight, cost-efficient models are becoming viable competitors to proprietary alternatives. This shift challenges market assumptions and forces AI providers to rethink their value propositions.

1. End users and GenAI application providers are the biggest winners.

Cheaper, high-quality models like R1 lower AI adoption costs, benefiting both enterprises and consumers. Startups such as Perplexity and Lovable, which build applications on foundation models, now have more choices and can significantly reduce API costs (e.g., R1’s API is over 90% cheaper than OpenAI’s o1 model).

2. Most experts agree the stock market overreacted, but the innovation is real.

While major AI stocks dropped sharply after R1’s release (e.g., NVIDIA and Microsoft down 18% and 7.5%, respectively), many analysts view this as an overreaction. However, DeepSeek R1 does mark a genuine breakthrough in cost efficiency and openness, setting a precedent for future competition.

3. The recipe for building top-tier AI models is open, accelerating competition.

DeepSeek R1 has proven that releasing open weights and a detailed methodology is helping success and caters to a growing open-source community. The AI landscape is continuing to shift from a few dominant proprietary players to a more competitive market where new entrants can build on existing breakthroughs.

4. Proprietary AI providers face increasing pressure.

Companies like OpenAI, Anthropic,and Cohere must now differentiate beyond raw model performance. What remains their competitive moat? Some may shift towards enterprise-specific solutions, while others could explore hybrid business models.

5. AI infrastructure providers face mixed prospects.

Cloud computing providers like AWS and Microsoft Azure still benefit from model training but face pressure as inference moves to edge devices. Meanwhile, AI chipmakers like NVIDIA could see weaker demand for high-end GPUs if more models are trained with fewer resources.

6. The GenAI market remains on a strong growth path.

Despite disruptions, AI spending is expected to expand. According to IoT Analytics’ Generative AI Market Report 2025–2030, global spending on foundation models and platforms is projected to grow at a CAGR of 52% through 2030, driven by enterprise adoption and ongoing efficiency gains.

Final Thought:

DeepSeek R1 is not just a technical milestone—it signals a shift in the AI market’s economics. The recipe for building strong AI models is now more widely accessible, ensuring greater competition and faster innovation. While proprietary models must adapt, AI application providers and end-users stand to benefit most.

Disclosure

Companies mentioned in this article—along with their products—are used as examples to showcase market developments. No company paid or received preferential treatment in this article, and it is at the discretion of the analyst to select which examples are used. IoT Analytics makes efforts to vary the companies and products mentioned to help shine attention to the numerous IoT and related technology market players.

It is worth noting that IoT Analytics may have commercial relationships with some companies mentioned in its articles, as some companies license IoT Analytics market research. However, for confidentiality, IoT Analytics cannot disclose individual relationships. Please contact compliance@iot-analytics.com for any questions or concerns on this front.

More information and further reading

Are you interested in learning more about Generative AI?

Generative AI Market Report 2025-2030

A 263-page report on the enterprise Generative AI market, incl. market sizing & forecast, competitive landscape, end user adoption, trends, challenges, and more.

Download the sample to learn more about the report structure, select definitions, select data, additional data points, trends, and more.

Already a subscriber? View your reports here →

Related articles

You may also be interested in the following articles:

Related publications

You may also be interested in the following reports:

Subscribe to our newsletter and follow us on LinkedIn to stay up-to-date on the latest trends shaping the IoT markets. For complete enterprise IoT coverage with access to all of IoT Analytics’ paid content & reports, including dedicated analyst time, check out the Enterprise subscription.