AI goes beyond chatbots: how AI agents, data centers and energy will shape the next stage of the technology race

The AI race is entering a new stage as leading technology companies move beyond chatbots toward AI agents that can handle tasks, use different tools and work with less human supervision. At the same time, the rapid expansion of AI is driving demand for data centers, advanced chips and energy, turning the technology race into a much broader infrastructure challenge. In this weekly analysis, Samir Hajibayli looks at what these shifts mean for investors, businesses and the emerging AI markets of Central Asia and the Caucasus.

Samir Hajibayli, Baku city, VC at Caucasus Ventures, LinkedIn

For most people, the difference between one leading AI model and another is becoming harder to see. A faster answer or a better benchmark may matter to engineers, but it rarely changes everyday life for most people. This week made the direction easier to understand. OpenAI introduced dots, persistent agents meant to keep working across projects. Anthropic and Google launched models designed for longer tasks. The product names will change, but the direction is clear: we are moving from asking AI for an answer to giving it responsibility.

That is where technology becomes personal. We forgive a chatbot for being wrong because we remain in control. An agent that opens files, sends messages, edits records or writes code enters our working life. Then intelligence alone is not enough. We need to know whether it understands its limits, admits when it is unsure and stops before making a real mistake. It also connects stories that first look unrelated. A White House meeting about AI safety, NVIDIA’s hardware controls, Anthropic’s IPO plans, Starship’s flight and Google’s chips in orbit all point to the same question. As AI takes on more responsibility, who sets its limits, who supplies its energy and who pays when it fails?

When AI Stops Waiting for Instructions

OpenAI’s dots were the clearest symbol of the change. A dot can stay with a project, move across work tools, research information, prepare documents and build software. OpenAI says Codex and ChatGPT Work already reach more than 35 million people each week. The number matters because something that feels experimental today could quickly become a normal way of working for millions of people.

The failed moments in the live demonstration were more revealing to me than the polished ones. The system struggled several times to deliver voice updates. A wrong chatbot answer is inconvenient. A wrong action can reach a colleague, alter a record or continue while the user is away. This is why the old question, which model is smartest, is losing some of its value. We should ask how often the whole job is completed correctly, how much supervision is needed and what happens when the system makes a mistake.

This also changes how I look at AI startups. A simple product built around a popular model can disappear when the model company adds the same feature. A company becomes stronger when it understands a real workflow, knows what it can access, keeps track of its actions and earns the user’s trust. Anthropic’s Claude Sonnet 5.5 and Google’s Gemini 4 Argon point in this direction as well. Their benchmark scores and technical limits will keep changing. The more important ambition is to build systems that remain useful through a long chain of work. We do not choose a colleague because that person won a quiz. We choose someone whose judgment we trust when the situation does not follow the script.

For investors, this changes how they evaluate AI companies. Fast growth is less convincing if every new customer requires more human supervision or creates a larger chance of a costly error. I would rather see a narrow agent with clear limits and repeatable results than a general assistant that promises to do everything. The company that wins may be the one that removes uncertainty from a specific job, even if its model is not leading the public rankings.

Safety Is Also a Business Decision

Once AI can act, safety moves from a distant policy debate into everyday product design. On 29 September, Google, Anthropic, Meta, OpenAI, xAI and NVIDIA met at the White House and signed a voluntary commitment built around internal checks, outside evaluation and board oversight. The agreement is useful because it creates a common expectation. Its weakness is equally clear: it sets no penalty for a company that ignores it. Voluntary promises are easiest to keep when they do not slow growth.

NVIDIA offered a more concrete answer. Its safety platform can isolate an agent in milliseconds when the agent crosses a defined boundary. That protection matters, but speed does not solve the central problem. A machine can detect that a rule was broken. People still decide what the rule should be, who may change it and whether commercial pressure justifies an exception. Safety therefore lives in the technology and in the institution around it. The guardrail is only as credible as the people who control it.

Most people will never read a governance agreement or inspect a hardware safeguard. They will meet this issue when an employer lets an agent act inside their inbox, a bank lets one review a financial decision or a public institution uses one to process a request. In those moments, safety means something simple: people should know when AI acted, be able to question the result and reach a human who is accountable.

Anthropic makes this tension especially visible. The company has built much of its identity around responsible AI while preparing for an IPO that could value it above 2 trillion dollars. I do not read that as proof of hypocrisy. I read it as a reminder that good intentions operate inside powerful incentives. Once investors, employees and markets expect rapid growth, choosing to slow down becomes expensive. The real test of an AI company’s safety culture comes when a limit costs revenue, delays a launch or gives a competitor time to catch up. Safety matters most when it requires a real sacrifice.

AI Is Becoming an Energy Story

At first, the AI race felt almost weightless. We saw software on a screen and rarely thought about what sat behind it. Jensen Huang’s five-layer explanation makes the hidden structure easier to see: applications depend on models, models depend on data centers and chips, and all of them depend on energy. The International Energy Agency expects data centers to use slightly more electricity in 2030 than Japan uses today. That is the fact I keep returning to because it turns AI from a software trend into an industrial one.

The two space stories from this week belong in that same picture. Starship reached orbit and delivered satellites. A few days later, a separate Falcon 9 mission carried Google’s experimental AI chips into orbit. Google wants to learn whether future computing systems could use the more constant solar energy available in space. This is still an experiment, not a working orbital data center. Space also does not provide effortless cooling, and equipment there must survive radiation and operate without easy repair. The important signal is that leading technology companies are searching beyond ordinary power grids for the next source of computing capacity.

The human impact will not arrive only through a better assistant on a phone. It may arrive through a new transmission line, pressure on local water supplies, construction noise or a higher electricity bill. A data center can strengthen a grid, create skilled work and bring computing capacity closer to local companies. It can also consume scarce resources while most of the value leaves the country. The difference comes down to how the project is designed, who its customers are and what the surrounding economy receives in return.

The Choice Facing Caucasus and Central Eurasia

Our region has already noticed the opportunity. Kazakhstan is promoting a Data Center Valley, Uzbekistan has started a large DataVolt project and Azerbaijan is presenting its unused generation capacity as an advantage for AI infrastructure. These announcements show that governments understand the value of energy and location. They do not yet prove that the projects will have customers, reliable access to advanced chips or a lasting effect on the local technology sector.

As an investor from the region, I think our common risk is copying the most visible part of a global trend. When a new technology wave arrives, attention quickly moves to the largest facility, the biggest announcement and the newest label. Yet an expensive building is not a strategy. Before capital goes into a data center, someone should be able to explain who will buy the capacity, how the grid will improve and what local companies will gain from it.

This matters because capital in our region is scarce. Every large infrastructure bet competes with grids, schools, transport and other businesses for money and political attention. We do not have the luxury of treating a data center as a monument to ambition. Its value should be judged by the demand it serves, the capability it leaves behind and the risks it places on the public.

We also do not need to train a frontier model to have a meaningful place in this market. Regional founders can build useful agents for local languages, trade, banking, energy, agriculture and public services. Infrastructure can support those companies, while those companies create local demand for infrastructure. Rules matter as well. Banks, ministries and operators of critical systems should know who can stop an agent, inspect its decisions and report an incident. Environmental costs such as water use, emissions and noise should be measured before a project is approved, not after complaints begin.

Taken together, this week’s launches, safety talks and infrastructure plans show AI asking for a larger place in the world. It wants access to our work, our data, our electricity and eventually our trust. For Caucasus and Central Eurasia, the sensible response is to decide where we have a genuine advantage and where we need discipline. We should build infrastructure when the customers and economics are real, help founders solve problems that global platforms overlook and treat safety as a condition of adoption.

Over the next decade, success will depend on whether AI makes people more capable without leaving them less in control. If productivity rises while workers, consumers and communities carry hidden risks, adoption will eventually meet resistance. Our regional strategy should therefore be judged by a simple outcome: does it create useful local capability while keeping responsibility visible? That is the position worth building toward.