It's an infrastructure question: Reduce AI spend across the board.
Applied to the AI spend, the concentration of cost is not in the applications people are excited about. It is in the configuration underneath them.
AI is not a magic bullet. Organizations can't just throw money at it and expect a return on their investment. Yet that's what many companies are doing, exceeding their projected AI budgets without producing results that justify the cost overruns. Looking at where these costs are and aligning them with value is the work we need to be doing, says John Powers, former Global CIO of Deloitte.
Powers spent a career helping large organizations restructure their operations and effectively integrate new technologies. He developed a broad, deep understanding of how businesses successfully leverage emerging technology, and why they fail.
Machine learning and other AI technologies have been in use at the enterprise level for some time. In recent years, new AI-powered applications made it possible for everyone in an organization to use the technology. It seemed that almost any job or task could be improved with AI. This has led to a rush of implementation.
"It became a belief, almost a religion, that AI will make everything more efficient, AI will improve every job, and therefore everyone should have access to it."
When these new tools were deployed, they were primarily used on routine work that had already been automated to some extent, tasks like taking meeting notes or organizing a calendar. AI tools were a novelty for employees who quickly adopted them and started burning through tokens without a clear goal for how to create value. The enterprise as a whole behaved just like the individual employees.
"They spent wildly. So they have thought about this as it's cheap for some reason, and it might look cheap on a unit basis, but it's not cheap when you start to allow everyone to consume it because it's highly variable."
Powers correlates today's approach to AI spending with the early days of cloud computing. Instead of being stuck with fixed technology infrastructure, methodical step-function investment, and annual budgets that constrained overspending, the cloud offered flexibility, capability, and agility, but without sufficient discipline, overspending was an issue for many companies.
"We didn't put controls on how you used the cloud, and then people would use it in an inefficient way."
The AI tooling layer runs on the same commercial logic: consumption units bought against compute power, speed, and dedicated capacity. Databricks DBUs are a good example. The problem is not the pricing model itself; it is developing the discipline required to operate it.
"So you're buying it, but people don't really know how to size it, and they don't really know how to meter it or control it because they really wanna give the customer, the user of this, the best experience."
The result from a McKinsey survey is the number Powers keeps returning to.
"Ninety percent of organizations report that they're exceeding their IT and AI budgets because they do not have control over the variable costs associated with their AI implementations."
Overspending alone would be survivable if the return on investment more than covered it.
"So we've got a twofer, which is we're not creating the returns to absorb the excess. And we're not getting that value equation balanced here."
At the scale Powers describes, the wasted dollars are substantial.
"If ninety percent of the people are missing their budgets, we're talking about billions of dollars."
Powers advocates for a granular analysis of the compute costs associated with AI proliferation, because most organizations cannot yet see what they are buying.
"So what people really need to start doing is looking at all these AI costs very precisely, looking at line items now, not I spent X on AI."
When Powers breaks down AI costs, roughly half of expenditures come from infrastructure choices rather than value created by the application. Your AI bill is a function of how much data you move, how fast, and on what class of processor.
So costs scale in tandem with workloads created by components like streaming video, data coming off a satellite, a real-time market ticker, or surveillance images checked over and over. And because nobody wants to be the person who degrades the user experience, teams default to high-quality service. They allocate the most expensive configuration on the menu, dedicated infrastructure, the highest response-time tier available, and the most processing power they can buy. Powers' verdict on this configuration choice is unambiguous: "fifty percent of that cost is essentially completely out of control." The fix begins with questions that were skipped in the rush to deploy.
"Do you need it on all the time? Do you need them to scale up in a certain way? Do you need GPUs at all? How much information do you need to run? Can you run it in a batch process?"
Powers brings three decades of corporate restructuring experience to cost-cutting detective work. His instinct is to find the concentration point rather than trim everywhere.
"Every time I would restructure a business, I focus on the Pareto, the twenty percent that drives eighty percent of the value."
Applied to the AI spend, the concentration of cost is not in the applications people are excited about. It is in the configuration underneath them.
"This is the twenty percent that's driving eighty percent of the cost."
This reframing of the discussion tells a CIO where to spend limited capital. Governance conversations about individual application usage, while potentially arduous, are worth having. However, they do not touch the infrastructure layer that sets the price of every one of them.
"Haevek has just built a better mousetrap, is the best way to say it."
Powers sees the innovation Haevek brings to the AI stack as unglamorous but deeply impactful.
It's about timing as much as engineering. Haevek's team is experienced developing AI applications in complex production environments. They learned how to see the cost curve coming before most CIOs felt it. Haevek's Falcon Cloud Compute Platform was built to solve the heart of the problem rather than work around it. Falcon reduces how much processing power the same work requires, making the data movement in and out of analytics and AI applications materially more efficient.
The mechanism Powers describes is a downshift in resources rather than a compromise in experience.
"They're taking applications that people had continuous data exchange, and they've moved it down. You can go to a lower complexity processor, and you can move it to batch. And it moves so quickly now that the batch is gonna give the user the same experience as continuous."
For example, surveillance images being checked over and over, configured the way most teams configure the process, run continuously against dedicated top-tier GPUs, because the requirement was written as "real time" and nobody went back to define what that meant, what the real requirements were to reach it. They just designed for "as fast as possible."
There are three categories that set infrastructure costs: volume, speed, and processor class. Move the same job onto a lower-complexity processor on a batch cycle that completes fast enough to meet the spec and the cost associated with all three categories changes at once. The images still get checked, but at a reduced cost to the enterprise.
Powers cites real-world results visible in Haevek's production data. As an advisor to the organization, he has seen Haevek deliver cost reductions of roughly seventy to eighty percent while making the specified performance.
Of the bottom-line cost to deliver AI to the enterprise, 70-80 percent of the infrastructure compute bill reduces the cost of delivery across the board. Conservatively, it is one quarter of the total cost of AI tooling, the data processing layer, and the underlying cloud spend.
"That savings could end up being, just say rough and conservative, 25% of the average corporation's total all-in AI investment."
There is still work to be done assessing applications and workloads. Teams still have to sit down with users and establish what genuinely needs to be continuous rather than batched, and what performance the job actually requires. His argument is about sequence, not substitution; get the cost of delivering down for all applications, then assess how you can reduce the cost for an individual application. If the budget is blown up, it is worth addressing a 25 percent gap across the board.
His view of who feels this most acutely is specific: organizations that piloted AI, watched the bill accelerate, and now need to get back inside the number they committed to. For them, he argues, this is immediate.
"Still gotta work on the ROI but now you have some space and some time to breathe and start to look at how you've configured the infrastructure, virtual infrastructure, to support your AI ambition."
Want to see what that looks like against your own numbers? A Test Flight benchmarks Falcon on your workloads, in your environment, before anything changes.
Powers is clear that cost is the presenting symptom rather than the underlying cause.
"It is really ultimately a performance solution, and I think people won't necessarily recognize that because the first problem they have is cost."
He sees three distinct groups turning that performance into different things.
At the high-performance end, where machine learning is matching against images and video in near real time, the constraint has been reach. Efficiency changes the geography of what is deployable.
"So if you're talking about a city and you couldn't afford to deploy it to the outer outskirts of that city, now you've reduced the cost significantly, and you haven't degraded performance in any way, so you can push it out further."
More coverage, more information taken in, a more comprehensive picture in the same window.
"They're gonna provide a turbocharge to your processing speeds without increasing the cost to do that."
For organizations entering AI for the first time, the value is risk rather than reach. A lower cost of entry means more experiments that run faster, and, more importantly to Powers, a smaller chance of blowing a budget that has no slack in it. That changes their whole risk profile going in, and lets them plan, budget, and execute against a plan they can actually hold.
For the group stuck in the middle, deep into AI with more use cases than clarity, the effect is the most interesting. Change the cost profile and the ranking changes with it.
"So now suddenly you could see a totally different rack and stack of your use cases. And so that could turn many zero ROI projects to positive ROI. Now it has a twofold. It reduced my cost, but suddenly you're seeing, 'ah! this is where we're getting the ROI.'"
That is a strategy outcome, not a finance one. The organization that could not articulate where AI value lived gets a picture it can act on, and can point its energy at the twenty percent of use cases carrying the value.
It also addresses a mood Powers hears constantly, and which no efficiency pitch usually reaches.
"What I'm hearing from the companies I'm talking to is people are getting not just not excited about AI, but it's kind of a pushback. Like this is a waste of time. This is yet another ERP nightmare."
His term for the alternative is a "free restart" on the AI strategy, and where he lands is the reason he thinks any of this matters.
"Instead of it being a burden, AI becomes the accelerant we all thought it was."
Powers saves his most pointed argument for last, and he knows it is pointed.
"It is somewhat criminal that the people that are authorizing these companies to use the tools for AI broadly aren't going through the details line item by line item to make sure that people understand what optimization or what optimal is."
The reason Powers expects this to become mandatory rather than admirable is arithmetic. He is hearing from peers that AI will consume around half of enterprise IT budgets within a couple of years, and IT itself typically runs five to ten percent. A line item that large without a clear optimization story is not a number any CIO or CFO will be able to present to stakeholders unchallenged.
"I just don't think people cannot be in control. Haevek is an example of going at the core, which is important."
He is equally clear that going at the core does not finish the job. The user conversations, the batch versus continuous decisions, the performance requirements that were never actually specified, all of that still has to happen. What the compute layer does is buy the room to do it without missing the number in the meantime.
The cloud lesson was never that variable pricing is bad. Variable pricing is good, on one condition: that someone has done the work to connect spend to value. AI arrived before most organizations did that work, and the bill is now telling them so.
Powers' argument is that the fastest way back to control is not a governance program that takes a year. It is the layer where half the cost is already concentrated, and where the same workloads can run on lower-complexity processors, in batch, fast enough that nobody downstream notices anything except the invoice.
Falcon is the compute layer built for this. Bring your worst pipelines and take a Test Flight.