Why AI is the Ace in the hand of the MAG7

In the race to dominate AI, there’s little question that the US leads the pack. It now controls around 75% of the total world’s AI compute power, trailed by China at 15% and Europe at 10%, according to the Tony Blair Institute for Global Change. That pack includes the Magnificent 7 (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla) while in China it’s the Dragon 7 who dominate (Tencent, Alibaba. JD.com, PDD Holdings, Baidu and NetEase).

These companies have deep pockets but they’ve also sought to finance their investment in data centres and energy infrastructure through bond sales, taking them from being cash-generating asset-light businesses into highly leveraged organisations dependent on bond buy backs. It’s reported they plan to spend £650bn in 2026 on building datacentres, for example, as well as investing in nuclear. But unlike the standard supply-and-demand economic model, these giants are utilising the circular economy to fund these enterprise in expectation of future demand. It’s a gamble, albeit a good one, which is seeing the MAG7 shore up their position for the long term.

Questions were raised in July, however, over just how viable that strategy was. US tech stocks tanked due to a sell-off triggered in Asian markets, leading the likes of Google and Meta to make minor sell-offs to maintain their AI and compute spend. Was this a correction of the market or an indication that this heavy investment is unsustainable? The jury is still out on that one. But what is clear is that these companies are now in it for the long haul and wield phenomenal power that is set to reshape the data centre market and with it cloud computing.

COLO and AI

The market dominance of these players means that when the majority of the S&P market become ready to move their projects into production and seek to scale, they will have to go to them for capacity. By then, they will have completely taken over the datacentre space. It will only be viable for heavily regulated industries to build their own private AI cluster while everyone else will need to use the hyperscalers who will likely fold in the colo / neo cloud provders into their control plane under their commercial frameworks, further cornering the market. It’s likely this will attract a premium charge which will clash with the new data (landing) zone strategy which prescribes an AI strategy must be built around governed workload placement. This ensures that each AI workload is legally allowed, economically sensible, operationally resilient and close enough to the data to be useful.

In some ways, AI makes the move to COLO inevitable. The 2026 State of the Data Center report reveals that nearly 70% of CIOs are now opting for COLO or hybrid environments for AI and machine learning (AI/ML) workloads such as chatbots, virtual assistants and web browsing because these increasingly demand a highly connected, power-dense setup. The report further makes the point that COLO is evolving beyond hosting to become an orchestration point that facilitates better connectivity over distributed enterprise environments. That’s because in a concentrated market place the COLO providers enable a level of reach that a traditional corporate datacentre would take years to achieve. This ability to pivot to COLO is hugely relevant in a world where inference will be distributed to be close to the where the data is created.

But it’s not just how we use datacentres that is set to change. How we consume that capacity is also significantly altering and with it the very dynamics of the business.

The corporate structure of these giants that has allowed them to be nimble enough to react to the market in this way because they are digital-only, have little legacy equipment and no real CIO/CTO. Instead, responsibilities are divided among other members of the leadership team. Unencumbered by this complexity and with a relatively flat management structure, these organisations are then not constrained by the same economic outcomes that typically drive technological investment. That lean model is also likely to become the norm.

Give and take

AI is expected to see massive culling among certain departments and that in turn has the power to collapse tech stacks. Fleets of AI agents will perform the adds, moves and changes previously carried out by IT, sweeping away the people, processes and complexity, with DevOps working with agents to accelerate development cycles and purge the business of legacy infrastructure. Cross-functional teams, governed by product managers, will then simply report to the CFO. As for the CIO/CTO? Those roles will be redundant.

The root cause of all of this change is the humble AI token. It finally provides a reliable way to compute expenditure versus productivity so that capacity equates to a measurable outcome. Compare that to the intangible runaway costs of Cloud and it’s clear to see that AI will make such spend truly accountable for the first time. As output per head increases due to more productivity, that should result in lower running costs for the business and token consumption will become the dominant narrative at board level.

The danger is of course that the cost of those tokens rests with the MAG7 who, through their cloud platforms, charge explicit per-token rates for input and output usage of large language models (LLMs). The MAG7 have a real problem right now in that they need to be able to demonstrate that all that heavy upfront investment is going to pay-off. A jittery stock market wants to see some evidence of sales growth from that spend within the next two years. So that does of course mean those companies will begin to exert their strangle hold over the market by exploring token charging models and potentially token costs at precisely the time when they will have become indispensable to the business.

How that could then play out is not yet clear. Customers could begin pulling AI from the cloud and going back to on premise to control spend. AI frontier providers who find themselves increasingly challenged by fluctuating GPU costs could elect to degrade their models to squeeze more queries through the same hardware stack, without any warning to customers downstream. Or we could see the MAG7’s grip loosen because the costs of doing business in the US are just too high. We’ve already seen Apple lobby for access to cheaper Chinese memory chips only to hit by a wall of import duties, for instance.

My guess would be that, with the CCP mobilising $28Trillion through their capital markets, we will see a wave of cheaper AI models and chips come onto the global market. These lower cost models that are already being favoured by US corporates and may well be where the win lies for cost-conscious AI shops. But the outlier is the lobbying capabilities of the MAG7 who could well still exert control over access to these cheaper models.