AI Servers Have a New Bottleneck: Money
Original Article By SemiVision Research [Reading time: 8 mins]
AI Servers Have a New Bottleneck: Money
Taiwan’s manufacturers are discovering that an artificial-intelligence boom can strain balance-sheets as well as factories.
For the past two years, the hunt for bottlenecks in artificial intelligence has resembled a game of industrial whack-a-mole. First there were too few advanced GPUs. Then HBM memory became scarce. Next came shortages of advanced packaging, substrates, power equipment, cooling systems and, increasingly, electricity itself.
Now another constraint is emerging—one that does not sit inside a server rack.
It sits on the balance-sheet.
In a recent conversation, bankers in Taiwan described an unusual feature of the AI-server boom. The island’s biggest electronics manufacturers have been borrowing heavily to support production, and at some domestic banks the credit capacity allocated to large OEM groups is becoming increasingly stretched. In some cases, much of the exposure that banks are comfortable extending to individual groups has already been used.
This does not mean Taiwan’s banks have run out of money. Far from it. Large, creditworthy technology firms are still able to raise substantial loans. The more interesting point is that banks do not lend according to liquidity alone. They must also manage concentration risk: how much they are prepared to lend to one borrower, one corporate group or one industry.
AI is beginning to test those limits.
That matters because the economics of AI hardware differ markedly from the way investors often imagine them. A booming order book may be wonderful for revenues, but it can be demanding for cash. Before an AI server is delivered, somebody has to finance the components, inventory, assembly, testing and logistics needed to build it.
For Taiwan’s giant contract manufacturers, that somebody is often themselves.
The expensive middleman
Consider the journey from an order to a completed AI rack.
A hyperscaler places an order with an ODM. Production then begins. Circuit boards, memory, storage, networking equipment, power supplies, cooling hardware, connectors, cables and mechanical components have to be secured. Factories assemble and test the machines. Entire racks may undergo burn-in and system-level validation before they can be shipped.
The cash-flow sequence is awkward.
Money leaves the manufacturer before money returns from the customer.
That is hardly unique to AI hardware. What makes the current cycle different is the extraordinary value of the equipment moving through the system.
Traditional servers were already expensive. Modern AI systems are in another category altogether. Accelerators, high-bandwidth memory, high-speed networking and sophisticated power and cooling systems have dramatically increased the value of each rack.
Multiply that by thousands of systems and the sums become formidable.
A manufacturer that is enjoying extraordinary revenue growth may therefore also require extraordinary amounts of working capital. Inventory rises. Receivables rise. Suppliers still expect to be paid. The gap has to be financed.
This is one reason the AI-server boom can create a counter-intuitive outcome:
The faster the business grows, the more money the manufacturer may need to borrow.
Not every GPU sits on the ODM’s balance-sheet
There is an important complication. Not all AI-server programmes are financed in the same way.
Some customers provide expensive components—particularly accelerators—directly to manufacturers under consignment or customer-supplied-material arrangements. In such cases the ODM does not have to finance the full value of the GPU itself.
That can make a large difference.
A rack containing extremely expensive accelerators may generate a large headline revenue number under one commercial arrangement while requiring much less working capital under another.
But consignment does not make the financing problem disappear.
Manufacturers may still need to fund printed-circuit-board assemblies, memory, networking components, power-delivery equipment, cooling hardware, mechanical parts, intermediate inventory, labour and logistics.
And every customer structures procurement differently.
This means that two companies with apparently similar AI-server revenues can have radically different capital requirements.
For investors, that distinction is becoming increasingly important.
Revenue is not cash
The income statement has long been the easiest way to tell the AI-server story.
Orders rise. Revenue rises. AI becomes a larger percentage of sales. The share price follows.
The cash-flow statement tells a less glamorous story.
Suppose an ODM generates an additional $5bn of AI-server sales. To support that business it may have to build substantially more inventory. Its accounts receivable may increase. It may have to pay suppliers weeks or months before collecting from customers.
The result can be spectacular revenue growth accompanied by substantial cash absorption.
There is nothing inherently worrying about this. If customers are reliable, inventory moves on schedule and payments arrive, working capital eventually turns back into cash.
But the speed of that conversion matters enormously.
Imagine two manufacturers each adding $5bn in annual sales. One requires $500m of incremental working capital to do so; the other requires $1.5bn.
Their revenues may look identical.
Their economics are not.
This is why investors should increasingly examine measures that have received less attention during the AI frenzy: inventory, receivables, payables, short-term borrowings, operating cash flow and the cash-conversion cycle.
The financial architecture of procurement matters too. Customer-owned components, supplier payment terms and consignment inventory can materially change how much capital a manufacturer must deploy for every dollar of revenue.
The next important AI-server KPI may therefore be neither shipments nor margins.
It may be capital efficiency.
Banks have limits too
This brings the Taiwanese bankers back into the story.
Banks can have ample deposits and liquidity yet still be reluctant to extend another large loan to a familiar borrower.
Banking is partly an exercise in avoiding excessive concentration.
A lender must ask how much money it has already extended to an electronics group, how much of that exposure is unsecured, how much of its corporate loan book is tied to similar businesses and what would happen if the AI cycle suddenly weakened.
Taiwan is unusually exposed to these questions because its manufacturing groups occupy such a central position in the global technology supply chain.
Foxconn, Quanta, Wistron, Wiwynn and other manufacturers are not simply shipping more servers. They are expanding factories, moving production geographically, buying equipment and supporting ever larger AI-infrastructure programmes.
If several giant manufacturers approach banks simultaneously asking for more working-capital facilities, the constraint may eventually become institutional rather than monetary.
The bank still has money.
It simply does not want too much of that money exposed to the same borrowers.
That is a more precise way of understanding the current situation than saying that Taiwan’s banks have been “borrowed dry”.
It is also more consequential.
Credit allocation itself is becoming part of the AI supply chain.
The rich get richer
If so, the biggest manufacturers have another advantage.
AI-server manufacturing has traditionally been described as an operational contest. Suppliers compete on procurement, engineering expertise, yield, production ramps, rack integration and delivery.
Increasingly, they will also compete on financial capacity.
Imagine a hyperscaler awarding several billion dollars of new business to a supplier. Having the factory and engineers is not enough. The manufacturer must also finance the production cycle.
For a smaller company, even a highly profitable order could create an uncomfortable requirement for cash.
Large ODMs have deeper pockets, stronger banking relationships, greater access to bond and equity markets and more negotiating power with suppliers. They can also spread financing across more institutions and jurisdictions.
In other words, their balance-sheets are becoming industrial assets.
This could reinforce consolidation in AI-server manufacturing.
Scale has always mattered in electronics. But AI raises the financial penalty for being small.
A supplier may possess the technical ability to build a product and still lack the financial capacity to build enough of it.
Going American, at a cost
Another source of capital demand is geography.
Taiwanese manufacturers are rapidly expanding their footprints outside China, particularly in Mexico and the United States, as customers seek production closer to North American data centres.
Such localisation is usually discussed through the lens of geopolitics. It is also a financing problem.
New factories require land, buildings and equipment. Once operating, they require local inventory, employees and supplier networks. Working capital has to follow production abroad.
That does not mean American or Canadian credit lines are themselves running out. There is not enough public evidence to make such a claim.
Overseas factories can be financed through a mixture of parent-company capital, local borrowing, offshore facilities, intercompany loans, bonds and retained earnings.
But the broader point remains: every step towards a more geographically diversified AI supply chain requires more capital somewhere.
Resilience is expensive.
From chips to credit
The progression of AI bottlenecks now tells an interesting story about the maturity of the boom.
Early shortages appeared at the semiconductor level: GPUs, HBM and advanced packaging.
Later constraints appeared at the system level: networking, power and cooling.
Then came infrastructure bottlenecks: transformers, grid connections and electricity.
A financing constraint would mark yet another stage.
It would suggest that the AI build-out has grown large enough that the difficulty is no longer merely producing the hardware. The challenge is financing the enormous quantity of hardware sitting between semiconductor fabs and data centres.
Every physical bottleneck has a financial twin.
Inventory needs capital.
Receivables need capital.
Factories need capital.
Overseas expansion needs capital.
Suppliers need capital.
Data centres themselves need staggering amounts of capital.
Viewed this way, the AI industry is not simply assembling a new computing architecture. It is constructing an enormous financing chain alongside the physical supply chain.
And the larger the physical system becomes, the more important the financial one will be.
Watch the balance-sheet
For investors, the implication is straightforward.
Tracking AI-server revenue remains useful, but it is no longer sufficient.
The next round of winners may not simply be those with access to GPUs, sufficient factory space or expertise in liquid cooling.
They may be manufacturers that can convert orders into cash most efficiently—and that still have enough financial headroom to accept the next enormous order.
That puts a different set of figures on the watch-list:
inventory growth, receivable days, operating cash flow, short-term borrowing and the cash-conversion cycle.
Above all, investors should pay attention to credit capacity.
The AI industry has become accustomed to discovering shortages in unexpected places. Two years ago the scarce resource was advanced packaging. More recently it has been electricity.
The next one may be less glamorous, but no less important.
It may simply be the amount of money a bank is still willing to lend.
International banks such as DBS and JP Morgan are also providing working capital financing to the Taiwanese ODMs. Sources of financing is not limited to domestic Taiwanese banks.