When a novel AI system can rapidly design useful new antibiotics like Halicin or engineer advanced chip layouts in hours that rival the best work of humans, we’re witnessing civilization-scale innovation as a real form of economic power. Hundreds of billions (and soon trillions) of dollars are currently flowing into AI R&D, from startups, big tech, and governments. Many of them are now sprinting toward superintelligence, which is basically defined as AI smarter (at least by a bit) than the most intelligent humans. Whoever controls the most powerful superintelligent models will literally hold the keys to future economies, scientific discovery, and social order.
Unfortunately, for now, today’s policymaking remains light-speed behind the technological development of AI. In 2023, Biden’s Executive Order 14110 instituted watermarking and a number of federal AI safety roles, but the follow-through was quite limited. And now, today, the White House under Trump unveiled “America’s AI Action Plan”, which is a bold pivot focusing on deregulation, rapid data-center expansion, and AI export bundles to allied nations.
The core recommendation: Create secure, full-stack AI export packages — hardware, software, standards — to anchor global model adoption to U.S. influence. It is essentially a geopolitical and economic gambit, not a working policy that will actually address the implications of how maintain some level of control over the key technical breakthroughs of superintelligence, when it arrives.
🔍 What’s at Stake
Scientific breakthroughs: AI is already accelerating drug discovery by years. The investment Just for Mitochondrial diseases has crossed $500M this year alone. More dramatic discoveries are likely when superintelligence arrives. The economic potential is almost certainly in the trillions.
Unexpected spin-offs: Quantum-safe optimization, climate model reductions, brain‑computer interface advances—none of which fit neatly into today’s regulatory boxes.
Global norms on open models vs. closed: Hugging Face CEO Clem Delangue warned this week that China’s open‑source models risk embedding state-driven censorship “cultural aspects…the Western world wouldn’t want to see spread”.
🚦 Forks in the Road Ahead: Open vs Closed AI Models
This is the defining dilemma of AI: Open models accelerate progress, but closed models consolidate power. Early evidence is clear: Open-source AI models can boost economic value and innovation, by enabling faster iteration, reproducibility, and a broader developer base. But openness comes with a steep geopolitical price: China now leads in open AI development, controlling 60% of global frontier models as of 2024 by some measures.
The reality is that state actors can co-opt Western breakthroughs overnight, at least in open settings. On the other hand, closed models may slow innovation and limit oversight but they can retain strategic control, keeping the most advanced capabilities behind corporate or national walls. The question isn’t whether openness or secrecy is better—it’s which risks we are prepared to absorb: Stagnation and concentration, or proliferation and misuse. This tradeoff now defines the fork in the road ahead:
Full-Steam Race No regulation; model innovation runs wild. Risk: Runaway power without checks. Global surveillance, identity manipulation, or misaligned AGI. The new AI Action Plan from the White House essentially puts us on this trajectory.
Carefully Coordinated Regime Binding export and compute caps, model provenance tracking, IP control, and multilateral audit frameworks—akin to nuclear treaties. Think: Global AI Marshall Plan calling for democratic compute and monitoring consensus.
Hybrid Path Strong national innovation, transparent oversight, and alliance-level tech guarantees—domestic push with democratic allied coordination that protects and controls technical advances in AI models to keep them out of the hands of unfriendly foreign governments and bad actors.
🔗 Key Policy Gaps
Export as influence: The new plan is about embedding U.S. standards through export bundles, not regulating domestic model releases .
Open-source bias: Without norms, Chinese models on open platforms may propagate censorship or propaganda .
No AGI governance: While massive investment is being spent on data centers and power generation, there are no mechanisms to regulate frontier model alignment on superintelligence and beyond.
🧭 Recommendations—Hedging Civilizational Risk
The core risk of uncontrolled open AI is this: Once a sufficiently advanced model is released, it cannot be recalled. And unlike traditional software, frontier models can enable catastrophic misuse with minimal modification. Researchers have already shown that large language models can generate viable biological weapon synthesis protocols, design novel pathogens, and construct autonomous cyberattack chains. These are capabilities once limited to nation-state labs are now one fine-tune away from public availability (RAND, Nov 2024; NTIA, Jan 2024).
The threat isn’t theoretical: Open models like LLaMA have been jailbroken and proven to work around safeguards within weeks of release. Without shared global standards for provenance, control, and auditing, we risk seeding the digital equivalent of unregulated nuclear material into the open internet. The recommendation section that follows assumes this reality—and asks what firms, enterprises, and governments must now do to stay ahead of the curve, without ceding the future to chaos or authoritarian dominance.
For AI Firms (OpenAI, Meta, Anthropic, etc.)
Implement secure-by-design: Watermarking + weight-signing + open audits.
Delay frontier model release until audit, licensing, and multi-party governance are in place (e.g., export as package, not leak).
Join international coalitions to standardize responsible openness.
For Enterprises
Demand model provenance, watermark verifiability, and supply-chain traceability.
Invest in hybrid infrastructures: cloud + on-prem control to hedge against AI ecosystem failure.
Insist on explainable alignment as a procurement standard.
For Governments
Embed AI export strategy within a democratic bloc—U.S., EU, Japan, UK—to enforce safety norms.
Mandate transparency: National oversight bodies to certify “open-yet-aligned” frontier models.
Prepare deterrence doctrine: credible threat of sanction or tech suspension against misaligned or weaponized AI use.
🏁 Conclusion
Those who follow my work know that I’m very much a tech positivist. But all innovation is a two-edged sword, and AI is likely the most powerful technology we’ve ever developed. It’s a watershed moment and we stand at a genuinely historic juncture: Build democracy-enabling superintelligence, or unleash power that could reshape societies without democratic control. The path ahead demands fusion. Bold innovation with ironbound governance, while we still can. If we balance speed with structure, we might just build the bright future that many of us are hoping for.
With the recent arrival of potent new AI-based models for representing knowledge, the methods enterprises use to manage data today is now faced with yet another major new transformation. I remember a few decades back when the arrival of SQL databases were a major innovation. At the time, they were both quite costly and took great skill to use well. Despite this, enterprises readily understood they were the best new game in town in which their most important data had to live, and so move they did.
Now vector databases and especially foundation/large language models (LLMs) have shifted the focus — in just a couple of short years — on the way organizations must now store and retrieve their own data . And we are also right back at the beginning of the maturity curve that most of us left behind a couple of decades ago.
While not everyone realizes this yet, the writing is now on the wall: Much of our business data now has to migrate again and be recontextualized into these new models. Because the organizations that don’t will likely be at a significant disadvantage, given what AI models of our data can deliver in terms of value.
Our Marathon with Organizational Data Will Continue With AI
The result, like a lot of technology disruption, will be an journey through a series of key stages of maturation. Each one will progressively enrich the way our organizations store, understand, and leverage our vast reservoirs of information using AI. This process will naturally be somewhat painful, and not all data will need to migrate. And certainly, our older database models aren’t going anywhere either. Butatthe core of this shift will be the creation of our own private AI models of organizational knowledge. These models must be carefully developed, nurtured, and protected, while also made highly accessible, with appropriate security models.
We’ve moved on from the early days of digital documents, capturing loosely structured data in primitive forms, to the highly structured revolutions introduced by relational and graph databases. Both phases marked a significant movement forward in how data is conceptualized and utilized within the enterprise. The subsequent emergence of JSON as a lightweight, text-based lingua franca further bridged the gap between these two worlds and the burgeoning Web, offering a structured yet flexible way to represent data that catered to the needs of modern Internet applications/services and also helped give rise to NoSQL, a mini-boom of a new database model that ultimately found a home in many Internet-based systems, but largely didn’t disrupt our businesses like AI will.
AI-Models Are A Distinct Conceptual Shift in Working with Data
However, the latest advancements in knowledge representation really do usher in a steep increase in technical sophistication and complexity. Vector databases and foundation models, including large language models (LLMs), represent a genuine quantum leap in how enterprises can manage their data, introducing unprecedented levels of semantic insight, contextual understanding, and universal access to knowledge. Such AI models are able to find and understand the hidden patterns that tie diverse datasets together. This ability can’t be understated and is a key attribute that emerges from a successful model training process. As such, it is one of the signature breakthroughs of generative AI.
Let’s go back to the unknown issues with AI in the enterprise. This uncertainty ranges from what the more effective technical and operational approaches are to picking the best tools/platforms and supporting vendors. This new vector- and model-based era is characterized by an exponential increase in not just the sophistication with which data is stored and interpreted, but in the very way it is vectorized, tokenized, embedded, trained on, represented and transformed. Each of these requires a separate set of skills and understanding, and very considerable compute resources. While this can be outsourced to some degree, this has many risks of its own, not least is that such outsourcers may not deeply understanding the domain of the business and how best to translate it into an AI model.
AI-Based Technologies for Enterprise Data
Vector databases, leveraging the power of machine learning to deeply understand and query enterprise data in ways that mimic human cognitive processes, offered us the first new glimpse into a new future. Going forward, the contextual understanding of our data will largely be based on these radical new forms that bear little resemblance to what came before. Similarly, foundation models like LLMs have revolutionized information management by providing tools that can seemingly comprehend, generate, synthesize, and interact with human language in a manner using neural nets, vast pre-trained parameter sets, and complex transformer blocks that each have a high learning curve to set up and create (using them, however, is very easy.) These technologies provide us with a new dawn of possibilities, from enhanced decision-making processes with unparalleled insights using all our available knowledge, to automating complex tasks with a nuanced understanding of language and context. But all these new AI technologies are generally not familiar to IT departments, which now have to make strategic sense of them for the organization.
Thus, this remarkable progress brings with it a large number of concerns and hurdles to make reality. First, the creation, deployment, and utilization of these sophisticated data models — at least with current technologies — entrails significantly higher costs compared to previous approaches to representing data, according to HBR. A real-world cost example: Google has a useful AI pricing page for benchmarking fundamental costs, which breaks down the various cloud-based AI rates, with grounding requests costing $35 per 1K requests,. Grounding — the process of ensuring that the output of an AI is factually correct — is probably necessary for many types of business scenarios using AI, and is thus a significant extra cost not required in other types of data management systems.
Furthermore, the computational resources, available time, and time required to develop and maintain such systems are also quite substantial. Moreover, the transition to these advanced data management solutions involves navigating a complex landscape of technical, organizational, and ethical considerations.
As enterprises stand on the cusp of this major new migration to AI, the journey ahead promises real rewards. It also demands careful strategizing, intelligent adoption, and I would argue at this early date, a lot of experimentation, prototyping, and validation. The phases of integrating vector databases and foundation models into the fabric of enterprise knowledge management will require a nuanced approach backed by rigorous testing, balancing the potential for transformative improvements against the practicalities of implementation costs and the readiness of organizational infrastructures to support such advancements.
That this is already happening, there is little doubt, based on my conversations with IT leaders around the world. We are witnessing the beginning of a significant shift in how enterprise knowledge is stored, accessed, and utilized. This transition, while demanding serious talent development and capability acquisition, offers an opportunity to redefine the very boundaries of what is possible in data management and utilization. The key to navigating this evolution lies in a strategic, informed approach to adopting these powerful new models, ensuring that an enterprise can harness its full potential while mitigating the risks and costs associated with such groundbreaking technological advancements.
Early Approaches For Private AI Models of Enterprise Data
Right now, the question I’m most often asked about enterprise AI is how best to create private AI models. Given the extensive concerns that organizations currently have about losing intellectual property, protecting customer/employee/partner privacy, complying with regulations, and giving up control over the irreplaceable asset of enterprise data to cloud vendors, there is a lot of searching around for workable approaches that produce cost-effective private AI models that produce results while minimizing the potential downsides and risks of AI.
As part of my current research agenda on generative AI strategy for the CIO, I’ve identified a number of initial services and solutions from the market to help with creating, operating, and managing private AI models. Each has their own pros and cons.
Services to Create Private AI Models of Enterprise Data
PrivateLLM.AI – This service will train an AI model on your enterprise data and host it privately for exclusive use. They specialize in a number of vertical and functional domains including legal, healthcare, financial services, government, marketing, and advertising.
Turing’s LLM Training Service – Trains large language models (LLMs) for enterprises. Turing uses a variety of techniques to improve the LLMs they create, including data analysis, coding, and multimodal reasoning. They also offer speciality AI services like supervised fine-tuning, reinforcement learning from human feedback (RLHF), and direct preference optimization (DPO), which helps optimizing language models to adhere to human preferences.
LlamaIndex – A popular way to connect LLMs to enterprise data. Has hundreds of connectors to common applications and impressive community metrics (700 contributors with 5K+ apps created.) Enables use of many commercial LLMs, so must be carefully evaluated for control and privacy issues. Make it very easy to use Retrieval-Augmented Generation (RAG), a way to combine a vector database of enterprise information with pass-through to a LLM for targeted but highly enriched results, and even has a dedicated RAG offering to make it easy.
Gradient AI Development Lab – This is an end-to-end service for creating private LLMs. They offer LLM strategy, model selection, training and fine-tuning services to create custom AIs. They specialize in high security AI models and offer SOC2, GDPR, and HIPAA certifications and guarantees enterprise data “never leaves their hands.”
Datasaur.AI – They offer an LLM creation service that provides customized models for LLM development including using vector stores to provide enterprise-grade domain-specific context. They offer a wide choice of existing commercial LLMs to build on as well, so care must be taken to create a private LLM instance. They are more platform-based than some of the others, which makes it easier to get started, but may limit customization downstream.
Signity – Has a private LLM development service that is optimized more to specific data science applications.However, they can handle the whole LLM development process, from designing the model architecture, developing the model and then tuning it. They can create custom models using PyTorch, TensorFlow and many other popular frameworks.
TrainMy.AI – A service that enables enterprises to run an LLM on a private server using retrieval augmented generation (RAG) for enterprise content. While it is more aimed at chatbot and customer service scenarios, it’s very easy to use and allows organizations to bring in vectorized enterprise data for RAG enhancement into a conversational AI service that is entirely controlled privately.
NVIDIA NeMo – For creating serious enterprise-grade custom models, NVIDIA, the GPU industry leader and leading provider of AI chips, offers an end-to-end “compete” solution to creating enterprise LLMs. From model evaluation to AI guardrails, the platform is very rich and is ready to use, if you can come up with the requisite GPUs.
Clarifai – Offers a service that enterprise can quickly use for AI model training. It’s somewhat self-service and allows organization to set up models quickly and continually learn from production data. Has pay-as-you-go pricing and can train pre-built, pre-optimized models of their own already pre-trained with millions of expertly labeled inputs, or you can build your own model.
Hyperscaler LLM Offerings – If you trust your enterprise data to commercial clouds and want to run your own private models in them, that is possible too and all the major cloud vendors offer such capabilities including AWS’s SageMaker JumpStart, Azure Machine Learning, and Google Cloud offers private model training on Vertex AI. These are more for IT departments wanting to roll their own AI models and don’t produce business-ready results without technical experience, unlike many of the services listed above.
Cerebras AI Model Services– The maker of the world’s largest AI chip also offers large-scale private LLM training. They take a more rigorous approach with a team of PhD researchers and ML engineers that they report will meticulously prepare experiments and quality assurance checks to ensure a predictable AI model creation journey to achieve desired outcomes.
Note: If you want to appear on this list, please send me a short description to dion@constrellationr.com.
Build or Customize an LLM: The Major Fork in the Road
Many organizations, especially those unable to maintain sufficient internal AI resources, will have to decide whether to build their own AI model of enterprise data or carefully use a third party service. The choice will be tricky. For example, OpenAI now offers a fine-tuning service, for example, that allows enterprise data to augment how GPT-3.5 or 4 produces domain specific data. This is a slippery slope, as there are many advantages to building on a high capability model, but many risks, including losing control over valuable IP.
Currently, I believe that the cost of training private LLMs will continue to fall steadily, and that service bureaus will increasingly make it turn-key for anyone to create capable AI models while preserving control and privacy. The reality is, that most enterprises will have a growing percentage of their knowledge stored and accessed in AI-ready formats, and the ones that move their most strategic and high value data early are likely to be the most competitive in the long run. Will vector databases and LLMs become the dominant model for enterprise knowledge? The jury is still out, but I believe they will almost certainly become about as important as SQL databases are today. But the main point is clear: It is high time for most organizations to proactively cultivate their AI data-readiness.
On Digital Strategy | Dion Hinchcliffe 