What Founders Should Know About AI Infrastructure

AI conversations often focus on the latest models, breakthroughs and product launches. But beneath the headlines, a quieter battle is taking place: the race to build and control the infrastructure that powers AI.

In the recent episode of Nothing Ventured, a podcast led by our founder and CEO Aarish Shah, he sits down with Meryem Arik, co-founder of Doubleword to discuss insights about AI economics, inference and the future of the industry. For founders, the discussion highlighted several trends that could shape both the cost and availability of AI over the coming years.

The era of cheap AI may not last

Many founders have become accustomed to rapidly improving AI capabilities at relatively low cost. However, current pricing may not reflect the true economics of the market.

Providers such as OpenAI and Anthropic are effectively subsidising token usage while competing for market share. As demand continues to grow and infrastructure costs rise, those economics may become harder to sustain.

For founders building AI-powered products, this raises an important question: what happens when token costs increase?

Businesses with strong unit economics and clear AI strategies will be better positioned than those relying on permanently low inference costs. The sooner companies understand their AI cost base, the more resilient they’ll be if pricing changes.

Open-source models are closing the gap

For many startups, choosing an AI model often starts with the most recognisable names. But that may not always be the most cost-effective option.

Open-source models have improved dramatically over the past two years and can often deliver comparable performance at a fraction of the cost. While frontier models still maintain an advantage in some areas, the gap is narrowing.

For founders, this creates more flexibility. Instead of defaulting to a single provider, businesses can increasingly evaluate models based on performance, cost, privacy requirements and deployment needs.

As AI becomes more embedded in products and workflows, these decisions could have a significant impact on margins.

Inference may matter more than training

When people talk about AI, they often focus on model training. Yet training happens once, while inference happens every time a user interacts with a model.

This distinction is becoming increasingly important.

The ability to run models efficiently, reliably and at scale may prove to be one of the most valuable capabilities in the AI ecosystem. Companies that can optimise inference costs and performance will have a significant advantage as adoption grows.

For startups building AI products, understanding inference is no longer just a technical consideration. It’s becoming a commercial one.

Why sovereign inference is gaining attention

One of the more interesting ideas discussed was sovereign inference.

Rather than focusing solely on creating domestic AI models, countries may benefit more from ensuring they have the infrastructure needed to run AI systems within their own borders.

The logic is straightforward. Access to AI could become as strategically important as access to energy, communications or cloud infrastructure. If geopolitical tensions increase, countries that rely entirely on external providers could find themselves vulnerable.

While this may seem distant from the day-to-day concerns of most founders, it reflects a broader trend: infrastructure is becoming a strategic asset.

What this means for founders

The biggest takeaway is that the future of AI will not be determined by model quality alone.

Infrastructure, compute availability, energy requirements and inference costs are increasingly becoming competitive advantages. As the market matures, founders will need to think beyond which model performs best today and consider the long-term economics of the technology they are building on.

The winners of the next phase of AI may not simply be those with the smartest models. They may be the businesses that best understand the infrastructure underneath them.

For founders building AI-enabled products, that conversation is worth paying attention to now rather than later.