4 SEP 2026 — Across the last twelve months of disclosed AI infrastructure financings, 33 of 37 qualifying deals went to companies that had already raised capital before. Four were first financings. Read as a market, that is not a boom in AI infrastructure. It is a boom in the AI infrastructure companies that already exist.
The shape of the sample
Twenty-nine of the 37 rounds were above $100 million, or 78 per cent. The average round was $480.2 million and the median $275 million, and the gap between those two says a handful of very large deals are pulling the mean upward while the middle of the set still sits at growth-capital scale.
The median of $275 million is the better summary. It says the typical financing in this category is too large for a conventional venture round, and that the buyers are growth funds, sovereign investors and corporate balance sheets rather than the firms that write first cheques.
What the follow-on ratio does and does not measure
Eighty-nine per cent follow-on is striking, but it needs one caveat. A first financing is harder to observe than a later one; seed and Series A infrastructure rounds are often unannounced, so a disclosed-deal dataset undercounts them by construction.
So the honest reading is narrower than it first appears. Among financings large enough and public enough to be counted, almost all went to incumbents. Whether new companies are being funded quietly is a question this data cannot answer.
The 78 per cent above $100 million is the qualifier that matters. At that threshold you are counting growth rounds, which go to companies with revenue, so some of the high follow-on share is a property of the sample rather than a market signal.
Capital intensity explains most of it
The pattern is real anyway, because the cost structure is physical. An AI infrastructure company buys accelerators, leases or builds data centre space, and contracts for power; none of that can be done at small scale.
A software startup can test a hypothesis for a few hundred thousand dollars. A company proposing to operate accelerator capacity cannot demonstrate anything without a cluster, and a cluster is a nine-figure commitment before the first customer.
New entry is still possible. The entry ticket is just large enough that only the people already funding incumbents can write it, and that is a different market structure from the one venture capital usually means by a boom.
Depreciation compounds it. An accelerator fleet loses competitive value as the next generation ships, so the capital raised has to be earned back against a clock, and an incumbent with utilisation on an existing fleet is a better credit than a company proposing its first.
A second dataset points the same way
A narrower count of AI chip funding, covering 22 January to 18 August, found $5.37 billion across 12 disclosed rounds. The five largest took about 72 per cent of the capital, and only two reached a billion dollars.
The same set shows where within the category the money went: inference-focused companies took eight of the 12 deals and roughly 67 per cent of the capital. That is a market funding the part of the stack that has revenue attached, since inference is billed per request while training capacity is a cost until somebody rents it.
Two small datasets agreeing is not proof, and both are built from disclosed deals with the same blind spot. What they show together is that concentration appears whether the sector is cut by infrastructure or by silicon.
Where the money is going instead
The follow-on concentration is consistent with what has been visible in individual deals this year. We covered Nvidia agreeing to buy Hugging Face at about 86 times revenue and Equinix assembling an inference service from two partners' products rather than building one.
Both are the same move at larger scale. When a capability is expensive to build from nothing, the participants with capital buy or partner instead of funding a new company to try. The financing data is the venture-market shadow of that decision.
The consequence for anyone raising is unwelcome. A first-time AI infrastructure company competes against portfolios that already have revenue, contracts and a deployed fleet, in a year when four such companies were funded visibly.
What it means for this region
Southeast Asian AI infrastructure has been built by companies that are not startups: telecommunications incumbents, state-linked developers, and the regional arms of operators headquartered elsewhere. That looked like a regional peculiarity, and this data suggests it was the global pattern arriving early.
It also sets the realistic ambition. A new Southeast Asian entrant is not going to out-raise an incumbent on capital, so the competitive positions available are the ones capital does not decide: proximity to a specific market, a power or land position nobody else can get, or a regulatory footing that a foreign operator cannot match.
Those advantages are local rather than technical. That is why the region's AI capacity keeps being announced by companies whose main asset is a site rather than a technology.
The number worth tracking next
Watch the first-financing count rather than the total raised. Total capital into a sector goes up when a few very large companies raise again, which is what has happened, and tells you nothing about whether anyone new can start.
Four first financings in twelve months is a low enough figure that it will move visibly in either direction. If it stays at four while total capital rises again, the sector has finished forming companies and started consolidating them. If it rises, something has lowered the entry cost, and that would be worth understanding: shorter accelerator leases, a neocloud reselling capacity by the hour, or a sovereign programme underwriting first fleets.
One caution on all of the above: these are 37 disclosed deals, which is a small sample from which to draw a market-wide conclusion, and the compiler's inclusion rules determine most of what the figures show.
View author profile → · Editorial policy
