OpenAI and the infrastructure race: what changes for SMEs

AI is hungry, and the main course is infrastructure

There is a conversation that rarely reaches the boardrooms of small and medium businesses, but that determines how much they will pay for AI, when they will gain access to certain tools, and how fast the market will evolve. That conversation is called compute, raw computational capacity, and OpenAI is at the center of it.

In recent months, OpenAI has been negotiating large-scale deals to expand its data center infrastructure, increase access to high-performance chips (mainly NVIDIA GPUs), and secure energy contracts to support the explosive growth in demand for its models. The move is not merely operational, it is strategic. And it repositions the company not only as a provider of language models, but as an AI infrastructure platform for the global enterprise market.

What is at stake in this negotiation

OpenAI already operates on Microsoft Azure infrastructure, the result of an investment that exceeds US$13 billion since 2019. But that dependency has limits, both technical and commercial. The current expansion signals that the company wants more autonomy over its own compute layer, plus capacity to serve very high-volume enterprise contracts without bottlenecks.

In practice, this involves negotiations with energy suppliers, data center builders, and chip manufacturers. Project Stargate, announced in early 2025 in partnership with SoftBank, Oracle and others, foresees up to US$500 billion in investment in AI infrastructure in the United States over four years. Part of that effort directly fuels OpenAI's ability to scale its products.

Why does this matter beyond the US? Because availability of compute directly affects the price and latency of models offered via API, and those models power countless tools that Brazilian SMEs already use, from customer service platforms to content generation and data analysis systems.

The silent bottleneck that affects SMEs

When computational capacity is scarce, the consequences show up in price and queue times. Large corporations with enterprise contracts have guaranteed SLAs, higher usage limits, and priority access to newer models. SMEs operating on the free tier or basic plans feel the difference: token limits, usage restrictions during peak hours, and a longer gap between a new model's release and broad access.

This is not hypothetical. In 2023, when ChatGPT adoption exploded, OpenAI paused new Plus plan subscriptions for weeks, precisely due to infrastructure limits. In 2024, the launch of GPT-4o with voice capabilities was initially restricted to a select group of users, with a gradual rollout determined in part by server capacity.

The current expansion aims to fix exactly that kind of friction. More compute means shorter queues, potentially more stable prices, and a product roadmap that can move faster.

What changes in practice for businesses using AI

Earlier access to more advanced models

With greater installed capacity, OpenAI can accelerate the release of more powerful GPT versions and autonomous agents to the general market, not only to enterprise partners. For SMEs that integrate AI into their flows via API or third-party platforms, this means more compute available without necessarily increasing cost per token.

Competitive pressure on prices

The race for infrastructure is not exclusive to OpenAI. Google (with its TPU infrastructure and Gemini), Amazon (Bedrock plus Trainium chips) and Meta (with open-source models running on its own infrastructure) are all investing heavily. That competition tends to put downward pressure on prices over time, which is good news for businesses that consume AI as an operational input.

Reliability as a decision factor

For an SME that automates critical processes, such as customer service, proposal generation, or lead screening, the stability of the underlying infrastructure is not a technical detail, it is operational risk. A provider with greater compute capacity tends to offer more resilience, less downtime and stronger SLAs. When evaluating AI suppliers, the question "where does this run and with what guarantees?" should be on the table.

The strategic reading for Brazil

In Brazil, SMEs still face a dual barrier, the adoption curve for AI and the cost of tools priced in dollars. The good news is that global infrastructure expansion tends to reduce the cost per unit of compute, which will eventually translate into more affordable plans and local solutions built on cheaper APIs.

The bad news is that companies that wait for the "right moment" to adopt AI will keep losing ground to competitors that are already learning, failing and adjusting their workflows now. Infrastructure is being built. The window to gain competitive advantage, however, does not wait for construction to finish.

At FM Solutions & Consulting, we have seen Brazilian SMEs, as well as Italian and American ones, underestimate the cost of not acting. Compute is being solved elsewhere. The question is what you will build when it reaches you.