Daniel J. Jacobs • Fractional & Interim CIO and CISO • 7 July 2026
After Oracle’s warning and Meta’s capacity resale, what runaway AI infrastructure costs mean for a board, and how to govern your cloud and AI spend.
This week, the two biggest names in cloud told you, in their own filings and share prices, that AI infrastructure costs have become a board-level problem. Oracle used its annual report to warn investors about every way its data-centre bet could fail, having lifted capital spending to 55.7 billion dollars in a single year, up from 21.2 billion the year before, with 90 to 95 billion planned for next year. Its shares are down about 40 per cent over the past month, as investors turn cautious on the cost of the build-out. On the same day, Meta announced it would build a cloud business to resell the AI capacity it has overbought, and its shares jumped 9 per cent on the news. Meta plans to spend as much as 145 billion dollars on infrastructure this year.
Read those two stories together, and the message is clear. The companies building the AI economy are now openly nervous about the cost, and investors reward any sign that someone is getting the bill under control. Your board is not spending $ 145 billion. But it is carrying the same risk in miniature: AI infrastructure costs, folded into the cloud bill, compounding faster than anyone in the business is actually governing. The hyperscalers’ problem is your problem with fewer zeroes, and the fix is the same. Someone has to own the number and treat cloud cost management as a board discipline rather than a procurement chore.
Why AI infrastructure costs just became a board issue
Because AI infrastructure cost stopped being a back-office technology line and became a market signal. When Oracle lists “overbuilding,” “stranded capacity” and “customers may not pay their bills” as headline risks, and when Meta’s stock rises simply because it found a way to resell what it over-purchased, cost discipline has become the thing the market is watching. That framing flows downhill. The mid-market CFO who reads that Oracle is down 40 per cent on spending fears will, sooner or later, ask what your own cloud and AI bill is doing, and whether anyone can defend it.
This is not a reason to panic or to rip anything out. It is a reason to treat cloud spend the way you treat any other material cost of goods: measured, owned, and challenged on a schedule. Andreessen Horowitz made the underlying point back in 2021, before the AI surge made it acute. Studying the top 50 public software companies, it found committed cloud spend averaging half of their cost of revenue and, on its own model, estimated as much as 100 billion dollars of market value suppressed by those cloud margins. Its summary has aged well: “You’re crazy if you don’t start in the cloud; you’re crazy if you stay on it.” AI has simply raised the stakes on the second half of that sentence.
What does an ungoverned cloud bill actually cost?
More than the invoice, because the drift is invisible until someone looks. Starting in the cloud is the right call for almost every growing company. You avoid buying a kit to sit idle, you move fast, and you pay for flexibility you genuinely use. Then the workload settles, growth slows, an AI pilot quietly spins up a cluster nobody switches off, and you keep paying a premium for elasticity you no longer draw on. In the exit-readiness operating models I have built, the cloud line was rarely the one anyone had thought to challenge. The CIO is measured on uptime, not unit economics. The CFO sees a single line called “cloud” and books it as fixed. So it compounds, unowned, until an event like this week’s makes someone finally ask what it is buying.
That drift is a form of what I would call Strategic Technology Debt: a decision made years ago for good reasons that now constrains your options and drains margin without ever tripping an audit. AI accelerates it, because AI workloads are expensive, easy to start, and hard to attribute. The board that waits for the finance director to raise it has already lost a year of margin.
Is everyone really pulling workloads back off the cloud?
No, and this is where the headlines mislead. The Barclays 2024 CIO Survey found that 83 per cent of CIOs plan to repatriate at least some workloads, up from 43 per cent in 2020, and that figure is repeated as if a mass exodus is under way. Read the four words that matter: “at least some.” IDC’s 2024 work, “Assessing the Scale of Workload Repatriation,” found that most companies are moving some workloads back, while full repatriation remains rare. Even the famous cases prove it: when Dropbox moved the majority of its workloads onto its own infrastructure, its gross margins rose from 33 to 67 per cent over three years, a gain Dropbox credited to that infrastructure optimisation alongside higher revenue rather than to repatriation alone, and it kept plenty in the cloud. What is actually happening is not an exodus. It is disciplined, selective pruning by companies that finally started treating cloud spend as a cost to manage.
The clearest public numbers come from 37signals, the company behind Basecamp, which published its cloud exit in full (reported by the data-centre title DCD). It cut its annual cloud bill from 3.2 million dollars to 1.3 million after leaving AWS and Google Cloud, a saving close to 2 million in 2024, with the hardware paying for itself inside a year. Those savings are real, and on the right workload, they are large. But the honest version, which 37signals is candid about, is that the figure does not carry the hardware refreshes, the operations people, or the power and cooling. Load it fully, and it still wins for 37signals, because their workload is the ideal candidate: large, steady, predictable, running all day every day. Load it for a business with a thin technology team and a demand curve that moves every quarter, and it often does not.
The four questions a board should ask before it moves anything
Whether the answer is to repatriate, optimise, or leave well alone, the same four questions settle most of the decision. If you cannot answer them, you are not ready to decide, and that is the finding itself.
First, what does this specific workload cost us fully loaded today, in the cloud, and what would it cost to run ourselves, refreshes and people included, over five years? Not the compute line. The whole number.
Second, is the workload the right shape? Steady, predictable and always-on favours bringing it home. Spiky, seasonal, or still-growing, and especially an experimental AI workload whose demand you cannot yet predict, favours leaving it where it is.
Third, do we have the capability to run it, or a credible costed plan to get it? Five years cloud-first, and the muscle to run your own infrastructure has wasted away. Rebuilding it is a real line in the case, and the one companies forget until they are committed.
Fourth, what does the move do to our future options? A move that saves money but locks you into one site, one vendor, or a capability you cannot reverse has traded one form of Strategic Technology Debt for another.
A board that can answer those four for a given workload can make the call with confidence. A board that cannot has just found its blind spot, which is worth more than a rushed decision in either direction. Notice that none of the four is about the technology. This is a commercial decision that happens to run on servers.
How a fractional CIO gets ahead of this
This is exactly the kind of question that does not need a full-time hire but does need someone senior who has worked on infrastructure economics before. The analysis is not hard because it takes a large team. It is hard because it takes judgement: building the fully loaded number honestly, knowing which workloads are candidates and which are traps, and being willing to tell a sponsor who has read this week’s headlines, “no, not that one.” When I have run this for boards preparing a business for sale, ambition has never been the problem. Producing the honest number was.
For a private-equity-backed business, it is even sharper, because gross margin is the number that drives valuation. Getting AI infrastructure costs and cloud spend governance right before an exit process is not housekeeping. It is a lever on the price. That is the work a fractional or interim technology leader does well: sit in the board conversation, run the technology due diligence on the actual numbers, shape the IT operating model that would run any repatriated workload, and give a straight answer rather than a migration everyone regrets. It is why PE operators increasingly want a technology voice in the room who has held the mandate before.
If you want a fast read on whether your technology function could even carry a decision like this, the Technology Health Check is a sensible place to start before you commission any deeper analysis.
Oracle and Meta just made cloud cost a boardroom word for a week. The businesses that benefit are not the ones that panic. They are the ones that use the moment to put a name against the number, run it honestly, and decide, workload by workload, with a spreadsheet rather than a headline.
Frequently asked questions
Why are AI infrastructure costs suddenly a board issue in 2026?
Because AI has pushed cloud and infrastructure spending to levels the market now scrutinises. In the week to 1 July 2026, Oracle warned in its annual report that its AI data-centre spending (55.7 billion dollars in FY2026, up from 21.2 billion) carried serious risks; its shares are down about 40 per cent over the past month, while Meta’s rose 9 per cent on news it would resell excess AI capacity. When cost discipline moves share prices, boards start asking about their own bill.
What is the fastest way to get control of cloud costs?
Put one accountable owner against the cloud line, then cost your largest workloads fully loaded, both ways, over five years. Good cloud cost management starts there: most companies cannot produce that number, and producing it is where the savings and the risks both surface.
Should we repatriate workloads to save money?
Sometimes, on specific workloads, by a lot. 37signals cut its bill from $ 3.2 million to $ 1.3 million per year. But full repatriation is rare and usually unwise; the dominant pattern is selective pruning. Repatriate large, steady, predictable workloads; keep spiky, seasonal or experimental AI workloads in the cloud.
Is cloud cost governance a board or an IT job?
It is a board decision with a technical execution. Cloud and AI spend is now a material cost-of-goods line that affects gross margin and valuation. The board should own the decision framework; the technology function executes it.
How does AI change the cloud cost question?
AI workloads are expensive, easy to start, and hard to attribute, so they accelerate ungoverned spend. They are also less predictable, which usually argues for keeping experimental AI workloads in the cloud until their demand curve is understood, rather than committing capital to run them yourself.
By Daniel J. Jacobs, fractional and interim CIO and CISO, Starkhorn.
Sources: Gizmodo on Oracle’s FY2026 10-K (1 Jul 2026); CNBC on Meta’s cloud business (1 Jul 2026); Andreessen Horowitz, “The Cost of Cloud, a Trillion Dollar Paradox” (2021); Data Center Dynamics on 37signals (2024); Barclays 2024 CIO Survey via EE Times; IDC (2024).