What if the AI hype is misleading and the business model changes?
How open-weight models and private AI infrastructure could reshape enterprise AI economics and create new opportunities for IT consulting.
Table of Contents
Select a screenshot or diagram to enlarge it.
AI is everywhere nowadays and major US firms make huge investments into AI infrastructure. According to Epoch AI (February 2026), combined capital expenditure by five major US hyperscalers could reach $770 billion in 2026 alone, if the current investment trend continues. Since the release of GPT-4 in March 2023, capital expenditure by these hyperscalers has increased roughly fourfold, so 4x. Behind these investments is the expectation (you can also call if hope if you want) that future AI demand will justify today’s enormous infrastructure spending, with customers generating sufficient revenue/profit to deliver a positive return on investment (ROI) on all those investments. We already see this nowadays when more and more AI offerings change from a fixed-price offering to so-called “pay-as-you-go”. In simple terms, you pay for the input and output tokens your AI models consume while executing tasks, agents or workflows. Earlier in 2026, some companies began introducing AI usage limits after initially announcing aggressive adoption strategies (statements like an engineer should use 3x his yearly salary in AI tokens etc.), as consumption costs proved harder to control than anticipated. While token costs seem to drop, companies still keep spending more.
Challenging the economics of large AI infrastructure investments
A large part of AI infrastructure seems to be financed with debt, sometimes even off-balance liabilities. The economics behind these investments only pay off if future demand makes sufficient use of those data centers and offerings to generate a positive ROI. However, more and more companies invest in their own AI infrastructure to avoid the unpredictable token costs. This means in plain English: They go buy servers, put 8 expensive high-end server GPUs inside, and then run their AI workloads locally using open source models, often from China.
Revolut is an interesting example of this larger shift that is beginning to happen. The fintech company claims to have reduced its AI coding-model costs by approximately 90% by deploying the open-weight GLM model from China’s Z.ai on dedicated GPU infrastructure (so self-hosted instead of cloud). Around 2,000 engineers are supported by this AI architecture, while commercial frontier models remain available for specialized tasks. Although these savings might not reflect the full infrastructure and operating costs (which will go into the millions of dollars for the hardware alone), the company shows how enterprises can rethink their AI deployment strategies, even on larger scales than your typical 10-person development team. The direction in a nutshell is: No more unpredictable “pay-as-you-go” token bills, as open models are good enough for certain enterprise workloads in a professional environment.
Despite the privacy, compliance and security benefits this raises the question of whether the massive investments into AI might just turn into a huge debt bubble? Well, this is a question for the companies investing in those data centers. The whole development has, however, another more practical implication that is sometimes missed in the whole discussion - At least I did not see a lot of posts and articles about it.
Impacts on the IT consulting industry
What happens to an IT consulting industry investing heavily in AI service offerings (going “all in on AI”, as some proclaim) if enterprises increasingly understand that they can run capable AI models on infrastructure they control, without paying expensive proprietary-model licensing fees? That is a very important question to ask, and just to address the elephant in the room: No, IT consulting will not be obsolete and advisory will not go away. However, it is likely that (consulting based) business models need to adapt to also work more and more with local AI infrastructure and/or with a mixture of both.
We can already see this inside the industry with two Big Four companies addressing this issue. Deloitte announced the launch of its practice in September 2026 an “Open Model Engineering Practice” (see link below) particularly targeting enterprise customers that want to deploy open models to have better control over their data & over their cost. Deloitte states: “This practice can help manage costs, support local requirements, and maintain greater control over data, intellectual property, and AI operations. We pair that flexibility with responsible engineering, strong governance, and a focus on measurable business results.” Also PWC is going ahead into this direction and recently published a guidance website with important things to consider for CISOs in the whole open-weight discussion (link below). This shows that consulting and advisory companies are widening their offerings, already going away from “cloud only” AI consulting, although PwC says that this is not a discussion for “open vs. closed”, which I agree with (btw. the PWC link has a great matrix - so I can recommend you check that one out!). I think that enterprise-level AI solutions will need both worlds, as offerings of big firms like Microsoft often do cost more (compared to self-hosting), but they integrate very well into existing enterprise stacks, they are ready to use out of the box, security is a strong focus and so on and so forth. Do you really want to develop your own “Copilot” and then find some way to integrate this into your productivity suite?! I am not sure. Those offerings are just too mature. That is the reality in my opinion.
So claiming that local AI will just wipe out cloud-based offerings would be naive in my opinion. But thinking that cloud-only services will stay 100% dominant might be the wrong bet too. Especially if the risk and compliance requirements are substantial - think of ERP systems, internal critical systems etc. - it is totally understandable that companies will invest into their own infrastructure and AI systems.
Scaling Effects and Learning Curves
However, running your own AI infrastructure is not as simple as downloading a model, buying a few GPUs and switching everything on. Operating AI models at enterprise scale requires expertise in GPU utilization, model optimization, inference performance, monitoring, cybersecurity and integration into existing IT landscapes. And that is where I see a substantial opportunity for the IT consulting industry. Companies need all those skills. Revolut obviously had those skills, but not every company has this.
Imagine a company currently spending hundreds of thousands of euros annually on AI services. Instead of immediately proposing another commercial AI product, a consultancy could assess the existing workloads, benchmark proprietary and open-weight models, and calculate the total cost of ownership across different deployment scenarios. And I am convinced this will happen more and more, replacing the older “let’s buy the SaaS” product automatism that was so common the last 20 years.
This evaluation process (maybe even next to the new SaaS offering) could result in a recommendation to remain entirely in the cloud, move some workloads into privately managed infrastructure, or operate a hybrid architecture. The important part here is the understanding that the decision is based on measurable business outcomes rather than existing vendor partnerships. Companies that can combine existing strong relationships with this emerging trend will have a clear edge in the space in my opinion.
Such advisory services might extend from the initial assessment to architecture design, implementation, security testing (very important!), governance and ongoing operations once the workloads are running fine.
But there is a catch, as explained: This requires genuine technical AI expertise.
Knowing how to configure existing AI offerings is, and will remain, valuable, but it is not the same as designing and operating large-scale inference infrastructure. Consulting companies will need engineers & architects who understand the underlying technology, can optimize performance vs. costs, and can show that their architectures work reliably in production conditions. Especially the last point is very important as over the last few years many AI projects failed before reaching production.
That is not necessarily something every consultant needs to learn. However, firms that want to widen their scope might have a look into this direction and start building up expertise through hackathons, sessions, conferences etc.
Conclusion
Does this mean the massive investments into AI infrastructure will turn into a bubble? Not necessarily. My personal opinion is no, but I am also not that deep into this topic. Lower per token costs might encourage even greater AI adoption than today, and many privately deployed models will still run on cloud infrastructure. The already high demand for computing power could therefore continue to increase. Companies like NVIDIA will stay on top anyways, as for them it’s not relevant who buys the hardware, as long as they can sell it.
But it does raise questions about where the long-term value will be captured and whether every investment will generate the returns currently anticipated. ROI calculation is a hot topic in AI, and many people want to kind of get around this topic. But remember, in the end, IT is here to serve the business, and not the other way around. So, ROI questions must be asked.
For the IT consulting industry & also more broadly for the consulting industry, I see this less as a threat and more as an opportunity to broaden its capabilities into this area.
The next “competitive advantage” might not be having the strongest partnership with a particular technology vendor that you worked with in the last 10 years. It might be having the technical expertise and independence to recommend the right AI architecture for each customer, even when that recommendation challenges established technology choices.
Perhaps the next phase of enterprise AI will be less about which model or tech stack a company uses and more about who can help them operate it securely, efficiently and economically?
This post reflects the personal opinion of the author, does not constitute consulting or professional advice, and is not affiliated with, endorsed by, or representative of any current or past employer, customer, or business partner