The AI debate is still mostly framed as a contest of models, with more parameters, faster releases and better demos getting the most attention.
But that way of thinking misses the real limit on AI in 2026, which is that progress no longer depends on making another general-purpose model. Instead, it depends on electricity, advanced chips and data centers. The sector needs more physical capacity and fewer models.
This change is already affecting regular users. AI tools that used to be easy to find are now behind paywalls, and free tiers have strict limits on how much you can use them. Furthermore, safety systems are having a hard time keeping up with abuse.
While some may look at these outcomes as sudden greed or design accidents, they’re actually the visible effects of scarce power and limited compute.
How Will Physical Infrastructure Shape AI Progress?
- Physical Scarcity: Progress is limited by electricity, chips and data centers, not model design.
- Market Impact: Rising costs have moved free AI tools behind paywalls and strict usage limits.
- Safety Risks: Defensive AI and risk moderation are lagging because they are more compute-intensive than abuse.
- The Solution: Prioritizing fewer, stable models supported by massive “AI factories” and power grids.
Capital Is Flowing to Concrete and Silicon
Recent patterns in investments have made the above point obvious. According to Reuters, a group of companies, including BlackRock, Microsoft and Nvidia, agreed to buy Aligned Data Centers for about $40 billion, providing them with access to about 5 gigawatts of operating and planned capacity.
This level of electricity consumption rivals that of a small city. The investors aren’t paying for beta software. They’re paying for megawatts, land and access to the grid.
In addition, financial analysts at Morgan Stanley recently speculated that tech companies could spend as much as $400 billion within one year on AI infrastructure. And true to form, one of the biggest firms in the space, OpenAI, has made deals with Nvidia, AMD and Broadcom, in addition to its work on multiyear data centers. We can conclude that serious players are putting huge sums into so-called “AI factories” capable of continuously running models at a large scale.
Tracking this investment is important because it shows where the scarcity is. Models can be copied at near-zero cost once trained — power plants, chips and data centers cannot.
Why Fewer Models Would Actually Help
A lot of large language models on the market perform within a narrow band for most tasks. And releasing even more of them won’t solve the main problem: too many users competing for too little compute. If anything, every new model makes it harder to get the same GPUs and grid capacity.
A smaller number of well-maintained models, along with significant investments in infrastructure, would probably make things more reliable for everyone. It would also slow down the current cycle in which providers add new, flashy features and then limit access when usage goes up.
The immediate effect of this limit has been a change in pricing policies. For instance, Google and OpenAI have tightened limits on free access to flagship tools and expanded paid tiers. While some may say that these changes are about making money, in reality, they show how much it costs to run such energy-intensive systems on a global scale.
Each prompt uses specialized chips and data centers that consume a lot of power. When there isn’t enough supply, providers give priority to enterprise customers who pay for reliable access. Meanwhile, those relying on free versions, including students, creators and independent developers, are forced into narrower lanes. And this adjustment isn’t temporary; it’s how utility-style services work when there are physical constraints.
Safety Breaks When Infrastructure Falls Short
The same limits explained above are also why AI safety keeps lagging behind misuse. An obvious example is deepfake abuse, where detection systems need constant retraining and large-scale deployment, while attackers just rely on cheap voice cloning or replaying real media. Unfortunately, such tactics tend to spread more quickly than defenses can be updated.
According to an analysis cited by Deloitte, deepfake content volumes are rising into the millions each year, with estimated fraud losses reaching tens of billions of dollars in the US alone. The study demonstrated that detection accuracy often drastically decreases outside controlled environments, leaving platforms vulnerable.
As a result, some companies have resorted to restricting access to their tools behind paywalls. For example, the Guardian says that X limited access to Grok’s image features after abuse led to regulatory pressure across Europe. But lawmakers in the UK and Ireland called the move cosmetic, saying that payment barriers don’t stop harm.
Their criticism does, however, point out a more important problem: Safety becomes reactive when there isn’t enough computing power to run verification and moderation all the time. This is not because AI is inherently dangerous, but because defensive uses of AI grow more slowly than malicious ones.
AI isn’t good or bad by itself; it depends on how it’s used. When infrastructure is limited, however, bad actors can take advantage of the asymmetry. They can use cheap, one-time generation, while platforms have to pay for ongoing monitoring, detection and enforcement, making safety structurally reactive, even with the best intentions.
Infrastructure Shapes Trust and Privacy
Conversations about private and secret AI also fit into this framework. To process sensitive data without making it public, you need separate hardware, secure enclaves and dedicated capacity. These features depend on physical buildouts, not on clever prompts. Therefore, when infrastructure is limited, privacy shifts from a fundamental expectation to a premium option.
Content provenance systems such as C2PA illustrate the link. Verifying the origin and history of digital media across platforms requires consistent processing and storage. That consistency depends on well-provisioned data centers. Without them, trust remains fragmented.
Some argue that efficiency gains will ease pressure on hardware. But even though efficiency can help, history has a warning: When appliances that use less electricity became common, total power use went up because more people started using them. AI works similarly. As tools get better, the demand grows faster than the efficiency improvements can keep up.
The market consequences are already visible. Companies that control chip supply, long-term power contracts and data center real estate gain lasting advantages. Software-first players face tighter margins as compute costs dominate budgets.
For users, expectations must adjust. Reliable AI access increasingly resembles access to electricity or broadband: priced, metered and unevenly distributed when supply is tight.
The Real Choice Facing the Industry
The industry can continue releasing marginally different models while rationing access through limits and subscriptions. Or it can acknowledge reality and redirect focus. Fewer models, backed by huge investment in power generation, advanced chips and data centers, could improve the sector’s reliability and safety.
This path also has its downsides, however. Scarce and expensive computing will favor the more established companies that can pay for multibillion dollar training runs, leading to more models and data being centralized in the hands of a few firms.
At the same time, the flood of nearly identical models coming out to take advantage of the popularity of their more established counterparts has only added noise without really increasing capacity. The real challenge, then, isn’t just having fewer models but being able to distinguish between long-lasting systems backed by infrastructure and speculative releases that further tax already limited resources.
AI progress now depends less on what happens inside the model and more on what happens around it. Until power grids expand, chip supply deepens and data centers multiply, the experience for users will remain constrained.
The future of AI will not be decided by who trains the next model first, but by who secures the infrastructure to keep AI available, trustworthy and stable at scale.
