Over the past two months, Chinese AI startup DeepSeek has sent shockwaves through the technology industry with the release of two large language models. Competitive dynamics among U.S. technology issuers may have changed as a result, and downstream industries that use these technologies may rethink their deeper AI strategies.

Net beneficiaries have yet to emerge, though a few indicators of the potential credit impact on U.S. technology issuers have appeared, and S&P Global Ratings will be examining these in this report. The sector is likely to see disruption, particularly in three key areas--training efficiencies, AI adoption, and competition. But ultimately, we see DeepSeek's groundbreaking technology as a normal and healthy part of the sector's growth story.

Deep Changes Ahead For Tech Industry

DeepSeek challenges long-held assumptions about AI development and could reshape the technology industry's competitive dynamics. It claimed extraordinary cost efficiency in its model development, stating it trained one model for approximately $5.6 million using only 2,000 chips--a fraction of the 16,000 chips typically employed by industry leaders. Instead of using massive infrastructure, DeepSeek uses innovative training techniques.

Perhaps most impressively, it used NVIDIA's Parallel Thread Execution, an intermediate instruction set, which allowed DeepSeek engineers to make fine-grained optimizations, bypassing NVIDIA's CUDA GPU programming software, widely seen as a key moat for the company. The industry already understood many of these techniques, but DeepSeek implemented them in combination to achieve a highly efficient outcome. We expect the leading model makers to research and adopt these techniques to the extent their capabilities will allow.

The $5.6 million figure may be modest as a measure of the model's cost, and DeepSeek acknowledges it excludes earlier-stage research and experiments. Still, Anthropic CEO Dario Amodei characterized DeepSeek's achievements as "an expected point on an ongoing cost reduction curve" in AI development. DeepSeek's claims throw into question the conventional wisdom that developing increasingly powerful AI models requires more and more computational power--the so-called scaling law that has driven massive investments in data centers and high-end chips. By claiming to achieve comparable performance to leading models at a fraction of the cost, DeepSeek pushes the industry toward a greater focus on training efficiency and away from raw computing power.

For U.S. technology issuers, the credit implications aren't cut and dried. The very training efficiency gains that sparked a recent plunge in the share price of companies like NVIDIA Corp.could accelerate AI adoption, as described by the Jevons Paradox: the principle that making a resource more efficient often increases its total consumption. As AI becomes more cost-effective, we expect broader AI implementation across industries, potentially driving greater aggregate demand for computing resources despite lower per-model requirements.

If training efficiencies advance far enough, we believe it would drive value away from infrastructure--where it mainly exists today due to the scarcity and importance of GPUs and high-end components. Instead, value would move toward platforms and software, where participants must currently pay a hefty price for access to the infrastructure, leaving their return on investment sparse.

In a world of falling training costs and expanding AI adoption, infrastructure investment priorities would shift from training to "inference," or deploying AI to respond to real cases. This trend could slow the growth in demand for the highest-end GPUs and open opportunities for custom silicon, CPUs, and other less expensive AI accelerators.

DeepSeek's models also increase the prospects for edge AI in personal computers and smartphones. This could catalyze an upgrade cycle for device original equipment manufacturers and open even more use cases.

Finally, DeepSeek's innovations may also reduce barriers to entry in AI development by reducing training costs. This opens the field to more players and increases competition.

Role models and the role of models

One controversial aspect of DeepSeek's methodology is its use of model distillation--a process by which a student model asks a teacher model hundreds of thousands or millions of questions and uses the answers as part of its training data. While DeepSeek has acknowledged using distillation on open-source models from Meta and Alibaba in the past (which the open source community typically supports), OpenAI says it has evidence that DeepSeek also distilled its proprietary models, which would violate its terms of service.

Distillation may be a way for followers to catch up quickly, but we haven't seen much evidence that it can push the frontier forward. It also creates two problems for the industry that we will closely watch: Who will invest to train the next frontier model if it can be distilled a few months later? And can distillation be effectively limited?

Still, the success of DeepSeek's open-source approach may boost the competitiveness of open-source models against proprietary ones. With that boost, the commoditization of foundational models would likely accelerate.

DeepSeek's success also raises questions about the effectiveness of export controls on advanced semiconductors. That DeepSeek achieved these breakthroughs despite limited access to cutting-edge U.S. chips suggests that raw computing power may be less critical than previously thought. This development, combined with China's demonstrated AI competitiveness, suggests that further restrictions might provoke China to accelerate domestic semiconductor development.

Table 1

Leading Indicators: Frontier Modelers, Hyperscalers, Picks And Shovels

For now, we have not taken any rating actions on U.S. technology companies based on the DeepSeek developments. Many of our hypotheses require more time to interrogate, and the magnitude of such change is difficult to judge. Furthermore, many potentially challenged issuers have good cushion within the ratings to absorb a modestly weaker competitive position. We already have a favorable view of the business profiles of the potential beneficiaries.

That said, indicators we're watching include comments from the leading model makers, spending patterns of the hyperscale data center providers, and guidance from leading AI semiconductor companies--the so-called "picks and shovels" of AI investment. We will also pay close attention to insights from frontier model makers OpenAI, Anthropic, Meta, and Alphabet regarding how the competitive landscape is changing.

Spending plans from hyperscale data center providers Microsoft Corp., Alphabet Inc., Meta Platforms Inc., Amazon.com Inc., and Oracle Corp. will also reveal oncoming shifts in AI. These plans are downstream from the model makers; therefore, if model makers keep their training plans as closely held secrets, their intentions will eventually show up in hyperscale spending.

Guidance from Microsoft's and Meta's recent earnings suggest that their trajectory has not yet changed. Microsoft's fiscal third- and fourth-quarter capital expenditure (capex) guidance is similar to the second quarter's, which indicates about 60% growth adjusted for leases (40% cash only) for the fiscal year ending in June. It also said that fiscal 2026 capex will increase but at a slower pace than fiscal 2025. Furthermore, the mix of investing in short-lived assets such as data center equipment will increase relative to long-lived assets such as land and buildings, which is pushing up fiscal 2025 capex. This means that spending on data center equipment will be stronger than headline 2026 capex growth will suggest.

Meta also maintained its capex guidance, which suggests 68% growth at the midpoint of the range. CEO and founder Mark Zuckerberg said, "I continue to think that investing very heavily in capex and infra is going to be a strategic advantage over time. It's possible that we'll learn otherwise at some point, but I just think it's way too early to call that." We think he left open the possibility for some capex consolidation in the future, but that massive infrastructure will continue to provide Meta a competitive advantage.

Other major signs that we will monitor are results and guidance for leading AI semiconductor beneficiaries. In order of importance, these are NVIDIA, memory makers (SK Hynix Inc., Micron Technology Inc., and Samsung Electronics Co. Ltd.), and custom chip makers (Broadcom Inc. and Marvell Technology Inc.). These indicators will provide insight into the how the mix of hyperscale spending is changing.

Finally, the progression of the $500 million Stargate Project and OpenAI's reported effort to raise $40 billion at a $340 billion valuation are points to watch. If these efforts are delayed or derailed, it could indicate that the increased focus on training efficiency we expect is taking hold.

Related Research

• AI Brief: DeepSeek Is A Catalyst For Innovation, Feb, 4, 2025