An image of a purple whale in a castle's moat. Hyper realistic
DeepSeek proves that the moat was never real to begin with. (Image generated locally with flux1-dev-bnb-nf4-v2.safetensors)

It has been impossible to avoid the DeepSeek discussion over the past couple of weeks. The release of DeepSeek-R1, along with reports of potential data exfiltration from OpenAI, has sparked a new round of debate about AI and the fragility of existing moats in the industry. (I feel like this article is just a small drop in a vast ocean of content that has been created over the past week) It has been quite the trip watching mainstream media and even Donald Trump congratulating DeepSeek on their ability to create a comparable AI with much less compute.

While AI companies have relied on massive compute resources and proprietary data to maintain their competitive edge, DeepSeek is demonstrating that a determined group can challenge this dominance with models that run locally, offering privacy and flexibility that centralized services struggle to match. The leaked Google memo was right: “We Have No Moat, and Neither Does OpenAI”

Why building models isn’t enough

The core training techniques and transformer architectures behind today’s AI models are generally not trade secrets, and honestly, for good reason. While we see significant innovation happening within private research labs, many of the foundational breakthroughs in AI—such as transformers—originated from open academic collaboration. However, as AI has become more commercially valuable, key advancements are increasingly being developed behind closed doors. Research on diffusion models and scaling laws, for example, was pioneered by private institutions rather than open research communities.

Yet, even as AI development becomes more privatized, the ability to maintain a competitive edge remains fragile. One of the biggest reasons for this is model distillation, the process of training a smaller model to replicate the performance of a larger one. Companies investing billions in training frontier models are finding their advantage erased in months as open-weight alternatives start to come out of the woodwork. This shows that the real competitive edge in AI is not in the models themselves but in the ecosystems surrounding them.

Microsoft has acquired evidence that there was an Azure instance that they believe was potentially being used to exfiltrate a large amount of data to Chinese users. The fact that OpenAI and the services hosting its models did not take explicit steps to prevent this sort of abuse is, frankly, embarrassing. Given the billions of dollars Microsoft has invested in OpenAI, this oversight highlights just how fragile any AI company’s moat truly is. Not only are you giving these outside researchers the keys to the car, you're also letting them see the blueprints, the tools, and the CNC machine files too.

The hosting and compute advantage is dwindling

For now, companies like Nvidia and Microsoft enjoy a significant advantage in the AI landscape: they have the infrastructure to build, host, and serve massive, compute-intensive models to millions of users. For now, this requires competitors to either build their own expensive hosting infrastructure or rely on these same cloud giants for access. The hosting and compute advantage has been central to the dominance of AI leaders, as it allows them to deliver state-of-the-art performance at scale.

But this moat may be fake too. Innovations in model efficiency are rapidly changing the landscape. Smaller, more efficient models are proving that cutting-edge performance no longer requires the same massive compute resources. These models can be fine-tuned and deployed with significantly lower overhead, opening the door for competitors, smaller companies, and even individual developers to bypass the need for centralized hosting entirely.

A huge part of the appeal of DeepSeek is that engineers can get a capable reasoning model running on their machines to help with code in a safe, and privacy-forward way. For the enthusiasts, they are even running the entire 671B parameter model on their home computers [1] [2]. This is a tremendous challenge to OpenAI’s entire business model—if users can get comparable performance from locally hosted models, the incentive to pay for centralized AI services diminishes almost entirely.

The future of AI will not be defined by who can host the biggest models but by who can create the most value for users -- both locally and in the cloud. As models get smaller and hosting advantages fade, companies will need to focus on building robust ecosystems, designing exceptional user experiences outside of the chatbox, and creating tools that solve real-world problems.

Exceptional UX is still the key to AI

As I pointed out in a previous post, AI UX is a challenging and nascent space. Many AI products today focus too much on the raw capabilities of their models without considering how users actually interact with them. Poor UX can render even the most powerful AI models frustrating or inaccessible.

If we look at OpenAI's moat in this space, I would argue that their UX actually has quite a big advantage over competitors. Both their Whisper service and text-to-speech systems are top of the line and play a massive role in why users would choose them over a competitor. Features like assistants and scheduled tasks also play a critical role in democratizing AI, making it more useful beyond simple chat interactions.

Hyper-personalization as a moat

AI personalization today is still largely constrained by static model weights, making it difficult to create truly individualized experiences. Tools like GPT’s custom instructions and retrieval-augmented generation (RAG) have allowed users to guide AI behavior and inject external context, but they don’t fundamentally alter the model’s way of thinking. This is why AI-generated content often feels generic, much like this blog post probably does (I am using AI to help write it, of course).

Perhaps the future of personalization lies in fine-tuning becoming a cheap, accessible commodity. Instead of relying on clever prompting or heavy post-editing, AI will naturally generate text, code, or content exactly as the user would without needing extra intervention. As models become smaller and more efficient, users will be able to create personalized AIs with custom weights that truly think like them, talk like them, code like them. This level of hyper-personalization will eliminate the need for constant course correction we engage in now and make AI feel more like an extension of oneself rather than a tool that needs managing.

A new frontier entirely?

Some experts argue that relying solely on LLMs for progress toward general intelligence may be misguided. Gary Marcus of NYU recently appeared on CNBC, sharing a skeptical take on OpenAI:

I think the final problem is that this software, large language models, just isn't that reliable. We all know by now that it hallucinates, makes stuff up, makes weird errors and so forth and that's a problem. Somebody, could be OpenAI, could be somebody else, could think of something that is much more useful than LLMs, but I don't think that's going to happen tomorrow. In the meantime, OpenAI is burning money, so we may see their valuation plummet.

His skepticism raises an important question: If LLMs aren't the final destination, what comes next? New AI architectures may need to emerge as competitors to transformer-based models if we are to hope to reach AGI this decade.

Future opportunities in the AI space

Just as a thought experiment, here are a few places where AI companies can extend their reach beyond the realm of agentic LLM applications:

  • Hyper-personalization at scale: Smaller, more efficient models that allow for fine-tuned, personalized applications and AI experiences. Whether it’s a lawyer needing an AI fine-tuned to specific case law or a musician training an AI collaborator on their past works, the future of AI is about adapting to individual needs, not just generic responses.

  • Cross-AI tooling: Allow users to do more with the models they choose. Applications like LM Studio, ComfyUI, and LangChain are huge here, providing essential tooling for developers looking to incorporate multiple local and remote models into a single workflow.

  • Robotics and real world AI: Tools like Nvidia Cosmos may be the bridge we need to take the transformer architecture to the next step. By creating these "world foundation models" we can more accurately represent reality, leading to less hallucinations and better AI outcomes overall.

  • Transformers in other spaces: Nvidia DLSS shows that you can take the same transformer architecture that powers LLMs and use it in a variety of different use cases that aren't strictly language related. There are a ton of untapped use-cases to be found here.

The collapse of moats in a post-singularity world

Maybe it's even worth considering that this collapse of moats could be a sign of things to come. If we are to suspend disbelief and imagine a near future with AGI/ASI, then a future wherein no businesses have significant moats around their IP could potentially arise. Competition could be reduced to a raw battle of compute allocation. The only meaningful differentiator is how much processing power an individual, company, or nation can throw at a problem.

This creates an economic structure where the winners are not necessarily the most innovative but rather the best-resourced. A startup with a novel idea could briefly hold an advantage in the market, but the moment a larger player decides to replicate it with greater AGI compute, that lead vanishes in an instant. AI would dissolve entire software industries into a state of constant flux, where no position is secure, and no breakthrough lasts longer than the time it takes for a competitor's AI system to recreate it.

I don't suspect that will be the case for quite some time, but it is interesting to think about.

Conclusion

AI’s competitive landscape is shifting rapidly. The moats of the past—hosting and compute—are eroding as models become smaller and more efficient. The real advantage will belong to those who build ecosystems, integrations, and exceptional user experiences that make AI indispensable.

Companies that fail to evolve beyond model-building will find themselves in a race to the bottom, where pricing pressure and open-weight alternatives render their technology commoditized. The future of AI won’t be defined by who has the biggest model—it will be shaped by those who understand how to make AI truly useful, intuitive, and seamlessly embedded into everyday life.