بایگانی دسته بندی ها: یادداشت ها

This article is part six of an ongoing series – AI Power Plays – that explores the fiercely competitive AI landscape, where tech giants and startups battle for dominance while navigating the delicate balance of competition and collaboration. In this final article, we focus on the ongoing ‘dominant design’ battle among Language Models (LMs) such as ChatGPT, Gemini, and DeepSeek and consider three scenarios how AI LMs might evolve.

The AI LM Inflection Point

AI LMs are approaching an inflection point. Rapid advances in model architecture (DeepSeek’s V3 and R1), computing power (NVIDIA’s Project DIGITS democratising access to AI infrastructure), and autonomous agency (OpenAI’s Deep Research enabling AI to autonomously explore and analyse web information) have propelled AI to the forefront of business agendas, national policies, and everyday life. But the path forward is anything but settled.

The recent breakthroughs in AI LMs reflect fundamental design choices about how AI LMs are built and deployed. We predict that over time, different ‘dominant designs’ of AI LMs will emerge, based on fundamental design choices and the use cases where LMs are applied. Understanding these parameters is crucial for grasping future scenarios for AI LMs.

Design Choices Shaping AI LM Futures

Large vs Small LMs. The relationship between model size and capability is undergoing a fundamental transformation. What began as a simple correlation – “bigger is better” – has evolved into a more nuanced interplay of architecture, efficiency, and specialised expertise. This dichotomy is reflected in many of the design parameters we highlight below.

Scale and Training Cost. Related to size, the general trend for LMs has been towards large models that contain tens of billions of parameters and are trained with vast datasets. This suggests a massive upfront investment in model training and hardware infrastructure and implies a future dominated by tech giants and government-backed corporations. However, technological breakthroughs such as those heralded by DeepSeek may upend this trend and open the door for resourceful startups, academic institutions, open-source communities to enter the market with smaller and specialised LMs. Access to high-quality data for model training may become a key differentiator for LMs.

Operating Costs are shaped by the efficiency of AI hardware and the complexity of user queries. Use cases dominated by simple queries do not require heavy processing, whereas agentic LMs with deep reasoning abilities will be more costly to operate. This has implications for the cost of adding new users and applicable revenue models. We see different use cases emerging, addressed by differently designed LMs that are operated by different players.

Proprietary vs Open LMs. Currently, the LM landscape features a mix of proprietary (e.g., OpenAI’s GPT and o series, Google’s Gemini, Anthropic’s Claude) and more open LMs (e.g., Meta’s LLaMA, NVIDIA’s NVLM, Mistral’s Pixtral, DeepSeek’s R1) made available through Hugging Face. The two approaches represent radically different approaches to the LM business. Whereas proprietary LMs tend to be large and premium ones, models with less restrictive licenses have been adopted by many players and applied to a wide range of use cases. Often the same players offer both proprietary and more open LMs (e.g., Google’s Gemini and BART LMs).

Model architecture and training process: brute force vs efficiency: The Mixture-of-Experts (MoE) architecture exemplifies the evolution from brute force to efficiency. Consider DeepSeek R1: while housing 671 Billion parameters (B), it activates only about 37B for any specific task, while matching the performance of larger models. The future lies not in sheer size, but in the intelligent orchestration of specialised experts. The trend toward efficiency extends beyond architecture to the training process. Low-Rank Adaptation (LoRA allows models to adapt to new domains without significant computational overhead). Through model distillation or quantisation techniques, the capabilities of Models can be preserved while reducing resource requirements.

Regulation. The regulation of AI LMs ranges from tightly controlled regimes that prioritise national security and sovereignty to more open, market-driven regimes that foster rapid innovation and global collaboration. The regulatory dimension shapes strategic dependencies, military applications, government investment, technological and data sovereignty, cultural norms, and trade policies.

The extent of AI regulation directly influences whether a single dominant design will emerge or whether multiple regional standards will coexist. Tightly regulated AI LM environments promote fragmentation, sovereignty-driven divergence, and government-imposed standards, reducing the likelihood of a globally unified design. Flexible regulations foster convergence, enabling a few powerful firms or open-source communities to establish dominant designs through competitive selection.

In the long term, AI LM regulation will determine whether AI LMs follow the trajectory of global technological convergence (as seen in internet protocols and semiconductors) or regional divergence (as seen in telecom standards and cybersecurity models). Beyond technical aspects, the future of AI LM dominant designs will also be shaped by the balance between regulatory intervention, geopolitical constraints, and market forces.

Deployment Architectures shape how AI systems operate and interact with users. This includes both physical architecture (where processing happens) and logical architecture (how systems are controlled and moderated).

For physical architecture, the choice largely depends on computing requirements. Large models typically demand cloud deployment in centralised data centres, creating provider dependencies but simplifying deployment. Edge computing runs smaller models on local devices, offering autonomy but facing computational limits. Apple Intelligence demonstrates an emerging hybrid approach: specialised models handle simple tasks locally while routing complex operations to the cloud, suggesting future systems may emphasise intelligent resource distribution over centralisation.

The logical architecture determines how model capabilities are accessed and controlled. Model providers (e.g., OpenAI, DeepSeek) enforce strict moderation through safety classifiers. Cloud platforms (e.g., Microsoft Azure, Amazon Bedrock, Nebius) offer flexible controls over model behaviour within platform guidelines. Self-managed deployments through rented compute or local installations provide complete control over model boundaries – crucial for enterprises handling sensitive data or requiring specialised behaviours. These deployment choices shape market dynamics and innovation patterns. While integrated cloud solutions attract enterprises seeking reliability, self-managed deployments appeal to those prioritising autonomy. Regional deployments using hybrid infrastructures serve specific market and regulatory needs.

From Design Choices to scenarios

These parameters interact and influence each other, creating the conditions for different possible futures. Based on how these deployment patterns interact with fundamental design parameters—scale of investment, proprietary versus open approaches, and regulatory frameworks—we see three scenarios emerging: (1) Corporate-Led Standardisation, (2) Decentralised Innovation, and (3) Geopolitical Fragmentation. Each scenario arises from specific dynamics—such as the amount of up-front investment required, control dynamics, technology maturity, implementation patterns and operational costs—and provides insight into how AI might evolve over the next decade. By understanding these scenarios, business leaders, policymakers, and technologists can better prepare for whichever future becomes dominant.

Scenario 1: Corporate-Led Standardisation

Dominant design choices at play:

• Upfront Investment: Large
• Deployment Architecture: Cloud
• Proprietary vs open: mainly proprietary
• Operational Costs: High

In this scenario, well-funded technology giants who control or partner with cloud platforms—think Google, Microsoft, OpenAI, and a few others—take the lead in building and maintaining AI technologies. Because the up-front costs are immense, only these behemoths can afford to invest.

These dominant firms would offer AI capabilities via cloud platforms (subscription and API access). Businesses, governments, and individuals gain access to state-of-the-art (SOTA) models but remain heavily dependent on corporate providers, such as Microsoft’s Azure, Amazon’s Bedrock, and various providers available on Hugging Face and OpenRouter. These providers enforce strict validation and safety controls through their official deployments, maintaining tight governance over how their models are used.

As running inference of large AI models eats up large amounts of computing power, a few players who can manage these costs will set the price for accessing AI technologies. Smaller organisations may be priced out, limiting competition and reinforcing a cycle where AI remains an elite tool controlled by a few firms.

An emblematic use case: Industry and Enterprise-specific AI Copilots where large corporations in finance, healthcare, and legal industries rely on AI copilots for tasks like financial analysis or healthcare diagnostics. These systems would be standardised, secure, and integrated with existing enterprise software.

In essence, Scenario 1 paints a world where scale and control of computing power win the day. While it delivers highly efficient, well-tested AI solutions, it risks locking businesses into proprietary ecosystems that limit choice, hamper competition, and concentrate profits and power at the top. The true moat in this scenario isn’t just money – it is the integration of specialised hardware, vast data centres, and proprietary training methods. Like oil refineries of the digital age, these AI factories require both enormous capital and deep technical expertise to operate efficiently.

Scenario 2: Decentralised Innovation

Dominant design choices at play:

• Upfront Investment: Small
• Deployment Architecture: Edge
• Proprietary vs open: mainly open
• Operational Costs: Low

In this scenario, AI development is driven by vibrant open-source communities and a diverse range of stakeholders rather than a handful of dominant tech companies. Consider how recent breakthroughs by DeepSeek challenged conventional wisdom: using clever training approaches rather than brute force computing power, smaller teams proved they could match tech giants’ capabilities. This suggests a future where innovation comes from unexpected places, as tools and knowledge become more widely accessible by large audiences instead of a few large players.

In this scenario, research collectives, universities, non-profits, startups, open-source linchpins such as Hugging Face, and even some large corporations—such as Meta and Alibaba, which integrate AI into their existing platforms without commercialising AI technologies—collaborate actively. They share new model architectures, training datasets, and software tools through public repositories, fostering transparency and generative innovation by large, distributed developer communities.

With affordable specialised hardware and efficient model training techniques, even small teams can develop and refine AI models (e.g., Mistral and DeepSeek). This accessibility fosters a culture of rapid experimentation, democratising technological and use case innovation.

Open-source projects constantly iterate, pivoting quickly with each new breakthrough. For instance, a new training algorithm discovered by a small research lab can be rapidly adopted worldwide. This fosters a decentralised model of progress—faster, but potentially more chaotic.

Instead of sending data to cloud platforms, most processing takes place on local data centres and installations. This approach reduces reliance on cloud services and can lead to substantial cost savings. Edge-based AI also enhances privacy by keeping sensitive data local, making it particularly valuable in scenarios where confidentiality is key. Users have full control over model behaviour and validation parameters, enabling more flexible and customised deployments, while taking on greater responsibility for safety and governance.

An emblematic use case: Local AI Assistants where individuals run personal AI assistants on their phones and computers (e.g., Apple’s M-series laptops and NVIDIA’s Project DIGITS), without relying on centralised servers. These local AI tools learn from personal data privately, respecting user privacy and control.

On the flip side, Scenario 2 may come with challenges around standardisation. Without a powerful central authority, ensuring consistent security, data governance, and reliability becomes more difficult. Still, this vision highlights an exciting possibility: AI technology that is truly of the people, by the people, and for the people—grassroots, inventive, and broadly accessible.

Scenario 3: Geopolitical Fragmentation

Dominant design choices at play:

• Upfront Investment: Small
• Deployment Infrastructure: Mix of Cloud and Edge
• Proprietary vs open: combination
• Operational Costs: Medium

Unlike the first two scenarios, which emphasise either corporate dominance or grassroots innovation, Scenario 3 places governments at the centre of AI development where nations develop, adopt and customise models, creating relatively Balkanised regional AI ecosystems to safeguard national interests and technological sovereignty.

Medium-sized and large countries in particular tend to want to avoid overreliance on foreign corporations, especially where it comes to strategic technologies. To preserve a degree of technological sovereignty, countries may promote open-source standards not only in LMs but also in related technologies (e.g., RISC-V for microprocessor architectures, Open Computing Project for data centre hardware). They may promote investment in local cloud infrastructures and fine-tune open models to align with regional priorities. This way, different regions may promote distinct “flavours” of AI that reflect their unique characters. DeepSeek’s R1 model, for instance, demonstrates deep understanding of both classical traditions (Tang Dynasty poetry) and contemporary cultural dynamics (Baidu Tieba and RedNote social networks), while Claude and Grok models excel at parsing complex social dynamics on platforms like Reddit and 4chan (from meme culture to community-specific discourse patterns). This could herald a future where regional AI ecosystems diverge, supporting different languages, ethical frameworks, and security protocols.

An emblematic use: France public authorities have introduced its sovereign LM: Albert and it is progressively expanding within the country’s public administration. The system aims to reduce reliance on foreign technologies and reinforce national control over sensitive data. Today, it assists administrative advisors in responding to citizen inquiries with reliable information. Albert is also embedded within the government’s secure messaging system. Albert serves as an API-based infrastructure, providing computational resources and machine learning algorithms for public institutions developing AI-powered solutions. However, the tax authorities prefer to develop their own LM and to avoid using Albert for sensitive data.

For nations pursuing technological sovereignty, Scenario 3 could provide strategic autonomy and localised innovation. But it also risks deepening divisions between regions, making global cooperation on AI ethics, safety, and research more difficult.

Conclusion: Preparing for the AI Worlds Ahead

The three scenarios are not mutually exclusive. We expect that multiple dominant designs will co-exist, each optimised for different use cases and constraints. The different dominant design parameters are also not mutually exclusive and often interact. It is virtually guaranteed that technological breakthroughs will continue to emerge and upend different scenarios and their technological and use case drivers. The evolution will be iterative, and dominant designs will shift over time. We further expect that open-source communities and commercial providers will co-exist in a dynamic equilibrium: corporations continue to adopt open-source breakthroughs, and open-source projects benefit from corporate-funded infrastructure (e.g., LLaMA and DeepSeek models running on Groq servers in Saudi Arabia or on Nebius servers in Finland).

What do these scenarios mean for business leaders, policymakers, and innovators charting their paths today?

Anticipate Power Shifts. In a corporate-led world, forging strong alliances with tech giants and maintaining sufficient capital reserves for AI solutions will be essential. In a decentralised innovation logic, adaptability, open-source collaborations, and edge-based solutions become key. Meanwhile, in a fragmented globe, the ability to understand and comply with diverse national regulations will become a prerequisite to success.

Balance Innovation with Governance. Whichever direction AI takes, companies must keep one eye on short-term performance gains and the other on long-term ethical and regulatory obligations. Stakeholders need to champion responsible data use, equity, and security, or risk public backlash and legal scrutiny.

Balance AI Investments. Given the unpredictability of breakthroughs and the fluid nature of regulations, spreading resources across multiple strategies—corporate partnerships, open-source initiatives, and strategic national collaborations—helps hedge against sudden disruptions.

No matter which paths AI LMs take, and they will be several, AI’s influence on business, society, and global politics is set to intensify. The key question isn’t just who will own the dominant AI designs—it is how we can guide AI’s development to serve the broadest possible set of human interests. By understanding potential AI LM futures, stakeholders can better position themselves while working toward an AI ecosystem that benefits all society.