The artificial intelligence boom has a bottleneck problem, and it is making the biggest technology companies even harder to catch.

That is the argument from Daniel Newman, CEO of Futurum Group, who points to shortages in compute power and energy capacity as the twin forces reshaping the AI market. In his assessment, the companies that can secure scarce hardware and power are the ones that will dominate the next wave of AI deployment, while everyone else gets left behind.

Newman’s framing is blunt: the AI infrastructure crunch is not a temporary hiccup. It is a structural condition that rewards incumbents with deep pockets, long-term supply contracts, and existing data-center footprints. The rest of the industry — startups, mid-size firms, academic labs — is left to fight over what remains.

The compute shortage is real and getting worse

The demand for graphics processing units (GPUs) and specialized AI accelerators has far outpaced supply for two years running. Nvidia’s H100 and newer B100 chips are the gold standard for training and running large language models, and they are effectively allocated months or years in advance. Big tech companies — Microsoft, Google, Amazon, Meta — have the purchasing power and the relationship leverage to secure those allocations. Smaller players often find themselves on waiting lists or paying premiums on secondary markets.

Newman highlights compute power as the first major pinch point. Without enough GPUs, a company cannot train competitive models at scale. It cannot run inference fast enough to deliver a good user experience. It cannot iterate quickly. Compute becomes the gatekeeper, and big tech holds the keys.

This is not a problem that goes away when new fabrication plants come online. TSMC and Samsung are building new capacity, but the lead times are measured in years. Meanwhile, the appetite for compute is growing faster than the supply. Every major AI lab wants more chips than they can get, and the allocation decisions made today will determine who leads in 2026 and 2027.

Energy capacity is the hidden throttle

Less visible but equally important is the energy constraint. Training a single frontier model can consume as much electricity as a small town. Running inference at scale requires power that is always on, always cheap, and preferably low-carbon. Data center operators are already struggling to find enough power in key markets, especially in Northern Virginia, Silicon Valley, and parts of Europe.

Newman flags energy capacity as the second pillar of the crunch. Big tech companies are signing power purchase agreements directly with utilities, buying land near renewable energy sources, and even exploring nuclear options. Google and Microsoft have both inked deals to purchase power from advanced small modular reactors. Amazon has acquired data center campuses with dedicated power infrastructure.

Smaller firms cannot make those kinds of commitments. They lease space in colocation facilities and pay market rates for power. When power becomes scarce or expensive — and it is becoming both in many regions — their margins shrink or their operations stall. The result is a two-tier market where access to affordable, reliable energy is a competitive advantage that money alone cannot buy quickly.

Who wins and who loses in this environment

The winners are the hyperscalers: the cloud providers that can build their own chips, design their own data centers, and negotiate for power at utility scale. They are also the companies that can afford to pre-order $10 billion worth of GPUs without blinking. For them, the infrastructure crunch is a moat. It keeps competitors out and gives them time to build the next generation of AI products.

Newman’s analysis suggests that the crunch also benefits companies that own unique data sets or have entrenched distribution channels. But those advantages only matter if the compute and energy are there to support them. The fundamental gate remains infrastructure.

The losers are a long list: AI startups that rely on rented compute, enterprise IT departments trying to deploy AI on-premises, research universities without corporate partnerships, and any organization that hoped to build its own large models from scratch. They face higher costs, longer wait times, and a shrinking pool of available resources.

This dynamic is already visible in the market. The most successful AI startups have been either acquired by big tech (Microsoft’s deal with Inflection AI) or have negotiated deep partnerships for compute (Anthropic’s tie-up with Google and Amazon, OpenAI’s relationship with Microsoft). Standalone AI companies that try to go it alone are finding the infrastructure costs prohibitive.

Is this consolidation inevitable?

Some argue that the crunch will ease as more efficient chips arrive and energy grids expand. Newman does not dispute that possibility, but he warns that the easing will be gradual and uneven. The infrastructure advantages that big tech builds today — multi-year contracts with chipmakers, ownership of power substations, custom cooling designs — are durable. They cannot be replicated in a quarter or two.

There is also a regulatory dimension. Governments are starting to treat AI compute as a strategic resource, with export controls on advanced chips and incentives for domestic manufacturing. This adds another layer of complexity. Companies that can navigate both the technical and political bottlenecks have an edge.

Newman’s bottom line is that the AI boom is not a level playing field. The infrastructure crunch is a feature, not a bug, of the current market structure. It concentrates capability in the hands of a few firms that have the capital, the relationships, and the long-term vision to secure compute and energy at scale.

That may be good for shareholders of those big tech companies. For the rest of the industry, it means the bar to entry keeps rising. And for anyone who hoped that AI would democratize access to advanced technology, the crunch is a sobering reminder that hardware and power still matter as much as algorithms.

As Futurum Group’s Newman puts it, the winners of the AI race are not necessarily the ones with the best ideas. They are the ones with the best access to the stuff that makes ideas run.