Analyst(s): Olivier Blanchard
Publication Date: February 7, 2025
Olivier Blanchard, Research Director at The Futurum Group, shares his insights on what DeepSeek’s disruption signals for the US tech sector. While many focused on how China’s progress in model training may threaten incumbent US competitors, the bigger signal we should focus on is how a confluence of efficiency improvements in both model training and AI-forward semiconductor performance could make large model training accessible to an increasingly broader market.
Key Points:
- Chinese startup DeepSeek recently sent ripples across the tech sector with its release of R1 – its own open-source version of a GPT-like reasoning model, which managed to outperform incumbent competitors, allegedly at a fraction of their development and training costs. While many initial claims had not yet been confirmed or debunked at the time, the announcement nonetheless spooked markets, triggering a selloff of A-list stocks ranging from AI-forward semiconductor companies to AI solutions providers.
- While the market overreacted to the news and appears to have overindexed DeepSeek’s ability to disrupt incumbent players in the AI space – let alone critical semiconductor vendors – in the long term, it may nonetheless be onto something: The AI ecosystem is going through an evolution that could reframe the assumptions that currently inform future investment and ROI calculations.
- DeepSeek has highlighted a throughline that has already begun to alter the trajectory of the AI industry at large: Steady, rapid gains in both AI hardware and AI software efficiency are already lowering compute costs and resource requirements for training and inference workloads.
Overview:
Disruption from the DeepSeek R1 announcement clarified a throughline already impacting the trajectory of the AI industry: Steady, rapid gains in both AI hardware and AI software efficiency, both of which already point to a lowering of compute costs and resource requirements for AI model training and inference workloads. These efficiency improvements come from both ends of the AI ecosystem simultaneously: Model training improvements (software) and processor improvements (hardware).
On the software/AI model training side, many of the models that required significant hardware and data center resources to train just a year ago have already become so efficient that they can now be trained at a fraction of the cost and a fraction of the resources. (Some can already run on AI-enabled devices such as AI PCs and Mobile devices.) This is opening the door to a much more “hybrid” and federated model – one that combines cloud, edge, and on-device compute to deliver a cost and resource-optimized full stack of training and inference ecosystem. Full-stack technology vendors such as Lenovo and Dell have already begun to steer their business strategies toward this hybrid AI model.
On the hardware side, more efficient, higher-performance semiconductors also contribute to the space’s overall training and inference efficiency improvement trajectory, again helping achieve better results faster with fewer resources. NVIDIA’s Blackwell announcements confirm this shift, with Blackwell processors promising to deliver significant ROI based on their ability to handle more compute than previous processor generations and work faster while requiring fewer resources. Moving down into the edge and device segments, Qualcomm, Intel, AMD, and MediaTek, among others, continue to drive NPU-enabled endpoint solutions processors toward increasingly AI-focused performance that enables organizations to run those increasingly efficient training and inference workloads locally.
As AI becomes increasingly critical to the fabric of business operations, this simultaneous combination of steady, continued hardware and software efficiency improvements is likely to lower costs, improve operational ROI, and open the door to more competition and innovation in the AI model training layer, and among AI-enabled services vendors. Being able to do less with more – and specifically train large language and large mixed models faster and with fewer resources – could give rise to a new crop of AI startups whose aim will be to compete against incumbents such as OpenAI, Meta, Google, and others. It could also arm enterprises with the ability to develop and train their own secure, in-house, proprietary AI platforms and agents quickly and cost-effectively and rely less on cloud-based subscription AI solutions and services.
An example of how the device ecosystem vendors may be able to build custom, in-house AI solutions more easily now is Samsung, whose own custom on-device agentic AI play was outlined at its recent Galaxy Unpacked launch of the new Galaxy S25 series. Samsung is leveraging a custom version of Qualcomm’s Snapdragon 8 Elite mobile SOC, dubbed “Snapdragon 8 Elite for Galaxy,” to run and train its own secure, local, hyper-personalized agent directly on S25 devices. And while Galaxy devices, being part of the Android ecosystem, also leverage Google’s Cloud-based Gemini AI solutions to enable a broad range of features, Samsung’s proprietary on-device Galaxy AI agentic solution points to the type of specialized, value-add, market-differentiated layers of on-device and edge-based agentic AI solutions and experiences that will increasingly complement more ubiquitous and general-use cloud-based options.
As another example, Lenovo has also recently been sharing demos of enterprise-class agents for business environments that will add to its overall value proposition and potentially generate new revenue streams for its full stack of products and services. Expect a lot more of those types of in-house AI products and services launching in 2025 and beyond from enterprise players that may not have been traditionally associated with AI innovation in the way that OpenAI, Microsoft, Meta, and Google have.
The lesson here is that the combination of increasingly efficient models and processors we are seeing across the AI space is already lowering barriers of entry for startups, enterprise solutions vendors, and mid-sized ISVs. It is also accelerating the training, development, and deployment of AI models and enabling more companies to build their own AI solutions at a fraction of the cost of doing so just a year ago.
But does this spell bad news for semiconductor vendors in the data center space? The short answer is no.
For starters, building the global AI infrastructure is going to be a massive lift, and more data centers and supporting infrastructure will need to be built. Second, a lot of the aging hardware needs to be upgraded to deliver ROI to data center operations and keep cloud providers competitive. This speaks to both net new shipments of AI-enabling semiconductors and a healthy refresh cycle for an already significant existing install base. Those two parallel growth vectors remain strong, with the latter benefiting the most from AI efficiency improvements in hardware and software and delivering a longer runway for sustained growth than the former.
The downward adjustment from very recent expectations set for just how much spend will be required to build and equip data centers for the age of AI is not particularly significant. The sums of money thrown around following the announcement of Project Stargate in mid-January 2025 may have unnecessarily inflated expectations by assigning 2024 cost estimates relative to AI compute needs to a far more cost-efficient and resource-efficient 2030 cost table. But as increasingly efficient AI training and inference expand to the edge and to endpoint devices, what we start to see is that spend that might have initially been intended for the data center layer of the overall AI layer cake is likely to move into the edge and endpoint device layers.
The full report is available via subscription to Futurum Intelligence—click here for inquiry and access.
Futurum clients can read more in the AI Devices Intelligence Portal, and nonclients can learn more here: AI Devices Intelligence IQ.
About the Futurum AI Devices Practice
The Futurum AI Devices Practice provides actionable, objective insights for market leaders and their teams so they can respond to emerging opportunities and innovate. Public access to our coverage can be seen here. Follow news and updates from the Futurum Practice on LinkedIn and X. Visit the Futurum Newsroom for more information and insights.
Author Information
Research Director Olivier Blanchard covers edge semiconductors and intelligent AI-capable devices for Futurum. In addition to having co-authored several books about digital transformation and AI with Futurum Group CEO Daniel Newman, Blanchard brings considerable experience demystifying new and emerging technologies, advising clients on how best to future-proof their organizations, and helping maximize the positive impacts of technology disruption while mitigating their potentially negative effects. Follow his extended analysis on X and LinkedIn.
