The rules are changing: what geopolitics and AI mean for tech investors

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

Most of my investing experience has taken place in a rising cycle. Liquidity was growing, valuations were expanding, and the next round often looked easier than the last. Watching this week’s diplomacy unfold made me realize how easy it is to treat those conditions as permanent. They are not. Technology still moves inside a world shaped by oil, interest rates, trade and political risk.

That is why I watched New York so closely, even though geopolitics is not the subject I naturally return to. The United Nations meetings brought Iran back to the centre of market attention. President Xi Jinping’s talks with President Donald Trump happened in Washington, not New York, but they belonged to the same story. Political decisions now reach technology companies through energy prices, supply chains, export rules and the cost of capital.

Geopolitics is part of technology investing

On 22 September, Brent crude closed at 99.25 dollars a barrel and West Texas Intermediate at 94.99 dollars. Prices moved as Saudi oil flows improved and traders assessed whether diplomacy with Iran could reduce risk around the Strait of Hormuz. Before the current conflict, roughly one fifth of global oil and liquefied natural gas passed through the strait. Oil near 100 dollars does not stay inside the energy market. It raises transport and production costs, can keep inflation and interest rates higher, and reduces the money investors are willing to place in risky assets.

The United States and China also agreed to lower tariffs on 30 billion dollars of non-sensitive goods in each direction, resume AI talks in November and create a channel for AI-related incidents. This was not a grand settlement between the two countries. It was still important because it showed that AI is now part of diplomacy and national security, alongside trade and military risk.

For me, the practical lesson is about underwriting. I now want to ask whether a company can survive 18 months without new equity, what happens if its cloud, energy or imported hardware costs rise by 20 percent, and how it performs after a 10 percent currency move. I also want to know whether one model provider, payment network or logistics corridor can stop the business. In an easy market, growth can hide weak unit economics. In a harder market, resilience becomes part of the product.

a16z is building more than companies

Andreessen Horowitz was the most active name in technology this week. On 22 September it announced the Horowitz Andreessen Academy, or HAA, a private, full-time school in San Francisco for young people leaving high school. It is not an accredited university and does not award a degree. The first one-year class is planned for autumn 2027, will be free, and will accept only 50 students. The project has 42 million dollars in backing, 10 major technology partners, more than 50,000 dollars in compute credits for each student, and a 5,000 dollar travel and research budget.

This subject is personal to me. During university I gave a presentation about a question that still bothers me: cars, phones and industries have changed beyond recognition, while a classroom from the nineteenth century can still look familiar today. That comparison is imperfect. Education has changed in access, curriculum and research. Yet the basic model of one teacher, a timetable, exams and a credential has remained remarkably stable.

HAA is therefore worth watching. Students will learn from operators and spend much of their time working on projects or in company placements instead of preparing for traditional tests. The optimistic case is that proof of work can become more valuable than a transcript. The hard question is whether the method is really better or whether 50 carefully selected people would succeed under almost any system. HAA may become an excellent talent accelerator, but that is different from replacing universities. We still need universities that fund basic research and build the theory behind tomorrow’s products.

Three days later, a16z launched Cosign, a curated network that shows who worked with whom, what they built and who endorses them. It already lists more than 69,000 open technology roles, while a small team uses AI to collect company data and filter spam. The opportunity is clear: trusted professional signals can help strong people who are outside famous networks. The danger is also clear: a reputation graph can turn existing insider circles into permanent advantage.

Meta is competing for the next interface

Meta’s announcements made the platform contest visible. It is bringing its Muse agent to smart glasses, says it will offer more than 100 styles of AI glasses by year end, and introduced the Muse Charm, a pocket device with a two-inch screen and 5G that opens Muse without first opening a phone app. Muse recorded 2.8 million downloads in its first 12 days in the United States and Canada. Meta’s shares rose more than 20 percent after the launch, adding over 200 billion dollars in market value by 22 September.

Those numbers do not prove that phones or apps are finished. Earlier dedicated AI gadgets failed to reach the mass market. They do show Meta’s distribution advantage. A large platform can enter a category and change its value overnight. It is too strong to say that Meta destroyed thousands of personal-assistant startups. It is fair to say that generic assistant products now have a much harder case to make. A startup needs its own distribution, data, regulated access or deep place inside a customer’s workflow.

Naval Ravikant has described AI agents as the beginning of the end of the iPhone’s dominance. He did not say Apple would disappear. His argument is that Apple could lose part of its control if people stop opening individual apps and start asking one agent to act for them. That is exactly what Meta is testing. The next interface may be a phone, glasses or a small wearable. The important change is that the agent, rather than the app store, may become the first point of contact.

Privacy is the cost hidden inside that convenience. Meta agreed to a 650 million dollar Illinois settlement in 2020 and a 1.4 billion dollar Texas settlement in 2024 over separate biometric-data claims, a combined 2.05 billion dollars. That is not the same as paying billions every year, but it gives users a reasonable basis for concern when a camera, microphone and personal agent are always close. Strong early downloads suggest that convenience often wins the first decision. Trust may decide whether the device becomes lasting infrastructure.

Anthropic shows why competition will not pause

Anthropic created another useful contradiction. CEO Dario Amodei recently asked the AI industry to slow the release of new capabilities, yet Anthropic launched Claude Opus 5.5 this week. The company says the model costs 40 percent less to run than its predecessor. It charges 4 dollars per million input tokens and 20 dollars per million output tokens, 20 percent below Opus 5. Anthropic also says outside groups tested the model and that it was about 85 percent less likely than two earlier systems to try to bypass containment in a dedicated internal evaluation.

This is not simply hypocrisy. A company can believe the whole industry should move more carefully while also believing that stopping alone would hand customers and talent to competitors. Anthropic argues that this release is both more capable and safer. The deeper point is that voluntary restraint is weak when the commercial reward for moving first is so large. Safety will need common standards and independent testing, not only promises from individual executives.

What this means for Caucasus and Central Eurasia

Our region cannot set the oil price, control relations between Washington and Beijing, or choose the speed of frontier AI. We still have choices. Higher oil can strengthen the external and public finances of exporters such as Azerbaijan and Kazakhstan. The same price can raise import bills and inflation pressure in energy-importing economies such as Armenia, Georgia, the Kyrgyz Republic and Tajikistan. Even exporters are not automatic winners, because volatility raises risk premiums and easy commodity income can delay reform. The IMF has listed tighter financial conditions and renewed geopolitical tension among the main risks for the Caucasus and Central Asia.

For founders and investors, I take four practical lessons from the week. First, test survival as seriously as upside: longer runways, currency stress and platform dependence belong in every investment case. Second, use cheaper global AI models but own a local advantage in data, distribution, trust or workflow; a thin wrapper can disappear with the next model release. Third, borrow the logic of HAA without copying its price tag: small regional fellowships, working operators, paid company placements, diaspora mentors and public project results could connect education to employment. Fourth, treat privacy as a product feature before wearables and agents become normal, not as a rule to add after a scandal.

The common thread this week was not one company or one device. It was the speed at which old assumptions can stop working. The underwriting model, the university, the professional network, the app store and even an AI safety promise are all being tested. Adaptation does not mean chasing every launch. It means building companies and institutions that can change before the outside world makes the decision for them. The question for our region is simple: which of our systems are designed to learn fast enough?