The seven companies building most of the world's AI infrastructure are on track to spend $863 billion in capital expenditure this year, 88 per cent more than last year, and analysts estimate roughly two-thirds of that, some $550 billion, is AI-related. If you were waiting to hear that the industry is finally pulling back, this isn't that number. But buried inside it is a more useful figure than the headline total, and it says something a Malaysian SME committing its own, much smaller, AI budget this quarter would do well to sit with.
Research from Exponential View, the newsletter run by Azeem Azhar's team, dug into a detail most coverage of these numbers skips entirely: when capital spending actually starts producing anything. Money spent on AI infrastructure doesn't hit a company's earnings the moment it's spent. While a data centre or chip cluster is being built, the cost sits on the balance sheet as "construction in progress," and only starts depreciating, only starts counting as a real cost of doing business, once the asset is actually ready for use. Across the four hyperscalers that disclose this figure, assets not yet in service now total $315 billion, up from $281 billion just one quarter earlier. That's not spending. That's a queue.
Money spent is not money working
The clearest illustration is Meta's. A dollar Meta spends on AI infrastructure today now waits, on average, around 1.7 years before it goes into service, a full year longer than in FY2024. Put differently, for every dollar Meta commits this year, only about a third of it will actually be doing anything by the end of the year (not a bad problem to have, if you're Meta, but a telling one all the same). The rest is sitting in a warehouse, on a truck, in a permitting queue, or half-wired into a data hall somewhere, generating no output and, for the moment, no return. Other hyperscalers show the same pattern to varying degrees, and none of this is a scandal or a sign of mismanagement; building physical infrastructure at this scale always has lead times. It is simply a reminder that "AI capital spending" and "AI capability in production" are two different numbers, and the gap between them has been growing, not shrinking, even as the total gets larger.
Before you sign off on any AI vendor contract this quarter, ask a second question alongside the price: when does this actually go live, doing the specific task you're buying it for, in the hands of the people who'll use it day to day? Treat the signed contract and the working deployment as two separate milestones with two separate dates, and hold the vendor to both.
Why this should worry you more than it worries them
A company the size of Meta or Microsoft can absorb an 18-month wait between spending and payoff without much difficulty; they have the balance sheet, the patience of long-term investors, and enough parallel projects running that some are always further along than others. A Malaysian SME committing even a modest sum to an AI project does not have that luxury, and the mistake I see most often is treating the invoice date as the finish line, as though value arrives the moment the purchase order is signed. It doesn't. The gap between paying for a capability and actually getting useful, adopted, day-to-day work out of it is real at every scale, and at the SME scale it's the difference that decides whether the project is remembered as a win or quietly written off eighteen months later.
What Most SMEs Actually Face
It is fair to say that most SMEs do not really consider huge budgets for AI relevant to them when their main concern is probably the marketing and sales investment needed for their business. But most business realise AI is here to stay, and spending money on it is inevitable. However, what do they spend it on and how do they start? What constitutes good AI investment that gives positive returns? Perhaps knowing what it is NOT and the mistakes others are making in this area would be instructive.
Why This Matters for Malaysian SMEs
Most SME AI budgets here are measured in thousands of ringgit, not billions of dollars, which sounds like it should make the deployment gap irrelevant, but it makes it more dangerous, not less, because there's no reserve of parallel projects to absorb a slow one. If you're budgeting for an AI initiative this quarter, build the deployment timeline into the business case explicitly rather than assuming it collapses into the purchase date: who owns getting it into daily use, by when, and what does "actually working" look like as distinct from "successfully installed." A vendor happy to talk about the tool's capabilities should be equally happy to commit to a going-live date and be measured against it, and if they aren't, that reluctance is itself useful information.
Want the next one of these in your inbox?
Subscribe on the Insights page for practical perspectives on AI in Malaysia, no hype, no spam.
What This Means for You
The $863 billion figure will keep making headlines as evidence that the AI build-out is accelerating, and it is. But the more useful number sitting inside it, the growing pile of capital that's spent but not yet working, is a caution against measuring your own AI progress by how much you've committed rather than by what's actually running. Spending on AI and using AI are not the same event, and treating them as though they happen on the same day is how budgets quietly disappear into projects nobody can point to a result from.