The debate around artificial intelligence has changed quickly. Some AI leaders warn of catastrophic risks, President Donald Trump is publicly pushing back against parts of that argument, and the companies building frontier AI are still racing to secure enormous amounts of computing power. At the same time, organisations outside the technology industry are starting to build their own private AI infrastructure. That makes me wonder whether we’re asking the wrong question. The immediate issue is not only whether AI could become dangerous on its own. It is also who owns the systems, who controls the infrastructure, who decides how AI is used, what data the systems can access, and who is accountable when those systems affect real people.
AI safety matters, but so do ownership, infrastructure, data control, commercial power, vendor dependence and accountability. We should be able to discuss all of those risks at the same time.
1. The AI safety argument is getting louder
When I first started writing about this subject, one of the most striking things was the change in tone coming from the AI industry itself. Companies competing to build bigger and more capable models were also warning that future systems might become difficult to control.
That argument has become even more public. Anthropic chief executive Dario Amodei has argued for stronger safety measures around advanced AI, while OpenAI chief executive Sam Altman has supported stronger oversight and coordination. Other researchers and industry leaders have also warned that increasingly capable systems may create risks that are difficult to manage if safeguards fall behind capability.
The opposing argument has become louder too. On 14 September 2026, President Donald Trump publicly dismissed fears about rogue AI as a “hoax” and argued that heavy regulation could damage American competitiveness. The wider disagreement now includes people who want stronger safeguards and people who believe exaggerated risk claims could slow innovation or strengthen the largest incumbents. Associated Press, 14 September 2026
These are very different positions. I don’t think the disagreement proves that one side is automatically right and the other automatically wrong. It does show that AI safety is no longer simply a technical debate. It now involves economics, regulation, infrastructure, competition and questions about who gets to decide what “safe AI” actually means.
2. The companies warning about AI are still building huge amounts of compute
One reason I remain cautious about treating the debate as a simple choice between safety and progress is that the largest AI companies are not stepping away from the technology. They are expanding the physical infrastructure required to build and run it.
OpenAI says its Stargate programme was created to secure 10 gigawatts of AI infrastructure in the United States by 2029, and in April 2026 it said it had already surpassed that target. That is a huge commitment to the physical infrastructure behind AI. OpenAI, April 2026
Microsoft is also expanding options for organisations that want more control over where AI runs. Its Sovereign Cloud and Azure Local services include options to run AI training or inference locally, so sensitive information can remain under tighter organisational control. Microsoft Sovereign Cloud
Anthropic is taking a different route, but the scale is just as significant. In April 2026, the company announced agreements for multiple gigawatts of compute through Google and Broadcom, and a separate agreement with Amazon for up to 5 gigawatts of new capacity. Anthropic said the Amazon agreement represented more than $100 billion in expected AWS technology spending over ten years. Anthropic and Amazon
That does not prove that these companies are being insincere when they talk about safety. A company can genuinely worry about a technology while still developing it. But it does show the tension at the centre of the debate. The organisations warning about advanced AI are also racing to secure the infrastructure that will determine who can build and deliver it at scale.
3. Latham & Watkins is a major sign that private AI is becoming mainstream
The development that caught my attention most recently did not come from an AI laboratory. It came from the legal industry.
Latham & Watkins, one of the world’s largest law firms, has bought Nvidia GPU servers to build and run AI systems in-house. According to the Financial Times, the servers are located in leased secure data-centre space and are being used by the firm’s technology teams for open-weight models and other AI workloads. Financial Times, September 2026
The important part is not simply that a law firm bought some GPUs. Latham still uses external services such as ChatGPT, Claude and Gemini. The firm is not abandoning cloud AI. It is creating another option alongside it.
That matters because a global law firm handles highly sensitive information. Client strategy, litigation material, transaction documents, legal advice and commercially confidential data all create obvious reasons to want more control over where AI processing happens.
Use cloud AI when the cloud is the right tool. Use private infrastructure when confidentiality, control, economics or vendor independence make private deployment more appropriate.
Latham’s move also supports a wider point I have been making about AI ownership. Organisations do not need to train frontier models from scratch to control more of their AI stack. They can own more of the surrounding infrastructure, choose open-weight models, control where information is processed and reduce dependence on a single external provider.
4. Vendor independence is becoming a serious business issue
The private AI discussion has moved beyond hobbyists running models on gaming PCs. Large organisations are now asking practical questions about where information goes, who can retain it, what happens if a provider changes its terms, and how much of the AI stack they should control themselves.
Microsoft is answering that market with sovereign and local deployment options. Latham is answering it by buying its own servers. Other organisations are negotiating stronger contractual protections with model providers or using several providers so they are not tied to one platform.
The technical solutions differ, but the underlying concern is similar. Businesses increasingly want the advantages of advanced AI without giving up every part of the infrastructure, data flow and operating model to somebody else.
5. The biggest risk is not only a machine deciding to become dangerous
I understand why people are worried about autonomous AI. If future systems can plan, write code, use tools, copy information across systems, persuade people or improve their own capabilities with less human supervision, those risks deserve serious research and governance.
But I think the science-fiction version of the debate can overshadow risks that already exist. AI does not have to become conscious to change the balance of power in society. It only has to become useful enough that companies, governments and institutions depend on it for important decisions and workflows.
That is where human responsibility becomes central. An AI model does not independently decide that a company should monitor employees, that an organisation should deny somebody a service, that a bank should tighten lending rules, or that a public body should automate a decision. People and institutions decide to connect models to those processes.
So my concern is not that we should ignore technical safety. My concern is that we should not reduce AI safety to a story about a machine suddenly waking up and attacking humanity—ownership, deployment, permissions, incentives, data access and accountability matter just as much.
The question is not only whether the model is safe in isolation. It is whether the whole system around the model is designed, authorised and supervised responsibly.
6. Genuine AI risk and commercial self-interest can coexist
People often frame the AI debate as a choice between two positions. Either advanced AI poses extraordinary risks and requires stronger controls, or AI companies are exaggerating those risks for commercial reasons.
I think reality can be more complicated. Advanced AI can create genuine risks in cybersecurity, fraud, manipulation, privacy, critical infrastructure and automated decision-making. Those risks deserve careful research and proportionate controls.
At the same time, regulation can affect competition. A rule that requires expensive evaluation, insurance, cybersecurity certification, specialist legal teams, and recurring external audits may be manageable for Microsoft, OpenAI, Google, or Anthropic, but much harder for a small company to afford.
This does not mean we do not need controls. It means how we design regulations is important. Safety rules should lower real risks without making it so only the biggest organisations can take part.
7. Dependency may be a more immediate danger than extinction
Imagine that artificial intelligence never becomes conscious. Imagine it never independently decides to destroy humanity. We could still create a major social and economic dependency if most useful machine intelligence is delivered through a small number of companies.
AI is becoming a writing assistant, researcher, coding tool, translator, analyst, automation layer, customer-service system, search interface and personal knowledge tool. In many organisations, it is becoming part of the basic digital environment through which people work.
If all of those capabilities come from somebody else’s servers, the provider can change the model, price, usage limits, retention rules, availability and product policies. Organisations can negotiate contracts, but they remain dependent on infrastructure they do not own.
That is why the Latham story matters to me. The firm is not rejecting cloud AI. It is reducing the risk of total dependence by adding privately controlled compute to its architecture.
8. This is why I am building my own private AI system
When I say I want to build my own AI, I don’t mean training a frontier language model from scratch. I mean running a capable open-weight model on hardware I control, connecting it to my own approved knowledge and building the surrounding retrieval, access, automation and governance services myself.
For the first version, I’m taking a practical approach. I plan to start with a Mac Studio M1 Ultra with 128GB of Unified Memory. It is older hardware, but it gives me a large shared memory pool and very high memory bandwidth without forcing me to spend heavily on the newest machine before I have proved the architecture works.
My thinking is simple: build it, test it, measure it, and show that it solves a real problem. If the system needs more power later, I can upgrade the hardware then.
I would rather find out the real limits with good existing hardware than spend thousands more based on guesses.
9. The private AI stack is bigger than the model
The model is just one part of the system. I also need a runtime to load it, a retrieval layer for current information, an embedding model to capture meaning, a vector database for searchable knowledge, access controls to limit information, automation to link workflows, and logging to track important actions.
This is what interests me most about local AI. The intelligence works with the data where it is, instead of always sending the data to someone else’s service.
10. RAG gives a local model current knowledge
One downside of a local language model is that its built-in knowledge can become outdated over time. But that does not mean you have to retrain the model every time something changes.
Retrieval-Augmented Generation, or RAG, keeps current information outside the model weights. Documents are indexed separately, and when a user asks a question, the system retrieves the most relevant material and supplies it to the model as context.
This separation between reasoning and current knowledge is very helpful for business systems. Policies, procedures, project documents, and approved information can change without retraining the main language model.
11. Private AI does not mean ungoverned AI
Owning more infrastructure gives an organisation more control, but it also gives it more responsibility. A local model can still leak information through poor permissions. It can still hallucinate. It can still be manipulated by prompt injection. An AI agent can still cause damage if it is given excessive access to email, files, code or production systems.
A strong private AI system needs access control, data protection, model tracking, patch management, retrieval permissions, logging, human review, and clear rules about which tools the model can use.
The real choice is not between “Big Tech AI is dangerous” and “local AI is safe.” The real difference is how much control and visibility the organisation has over the whole system.
12. My preferred future is hybrid AI
None of this means I want to stop using ChatGPT, Claude, Gemini or other advanced cloud systems. Cloud AI can be excellent, and some workloads suit large hosted models and elastic infrastructure better.
I think a hybrid architecture is stronger. Public and low-risk work can use the strongest suitable cloud model. Normal business information can use approved enterprise AI services with appropriate contracts and controls. Sensitive work can move to private inference and private RAG. Highly sensitive workloads can remain inside tightly controlled infrastructure.
That is also why I find Latham & Watkins so interesting. A major law firm is effectively demonstrating the same principle at a much larger scale. It can keep using external AI services while also owning the infrastructure needed for workloads where control and flexibility matter more.
13. So will AI kill humanity?
I do not think anyone can honestly settle the long-term existential-risk debate today. Researchers who see dangerous capabilities should be taken seriously, and governments should not ignore credible evidence simply because regulation is inconvenient.
But I also don’t think society should focus so much on a future superintelligence that it overlooks the power structures being created around AI right now.
Artificial intelligence is already being integrated into companies, public services, finance, healthcare, research, education and the software people use every day. People are already making decisions about who owns the infrastructure, who can access the models, what data they can see, and what actions they are allowed to take.
That is the main point I keep returning to. AI does not take away human responsibility. As AI becomes more powerful, human oversight becomes even more important because the technology can amplify whatever goals, permissions, and incentives we set.
The real danger may be that humanity becomes dependent on AI systems it doesn’t own, cannot inspect, doesn’t fully understand, and cannot truly control.
That is why I want to own more of the technology stack. Not because local AI solves every problem or because cloud AI is always bad, but because digital independence, infrastructure choice, and human accountability are now important parts of the AI safety discussion.
Sources and further reading
- Associated Press: Trump calls AI risks a hoax
Associated Press, 14 September 2026 - Associated Press: AI leaders debate coordinated safeguards and slowdown
Associated Press, September 2026 - Financial Times: Latham & Watkins buys Nvidia servers to set up in-house AI systems
Financial Times, September 2026 - Latham & Watkins: Reprint covering the firm’s in-house AI server strategy
Latham & Watkins PDF - OpenAI: Building the compute infrastructure for the Intelligence Age
OpenAI, April 2026 - Anthropic: Anthropic and Amazon expand compute collaboration
Anthropic, April 2026 - Anthropic: Anthropic expands Google and Broadcom compute deal
Anthropic, April 2026 - Microsoft: Sovereign Cloud and local AI deployment
Microsoft Sovereign Cloud
AI GOVERNANCE & RESPONSIBLE AI
AI governance starts with the whole system.
Responsible AI is not only about the model. Organisations also need to understand ownership, data access, infrastructure, human oversight, retrieval, permissions and how AI fits into real services and decisions.
Explore AI Governance
