Hock Tan: AI's Biggest Problem Isn't Chips — It's Cash

The most feared balance-sheet operator in the semiconductor industry just told the world what the AI boom is actually running out of. And it isn't chips.

Broadcom CEO Hock Tan, speaking on a panel in Singapore on Monday, said that "capital will be the biggest constraint facing the company as AI development continues." The Reuters dispatch, reported by Jun Yuan Yong, ran just three paragraphs — but the line that matters is the one Tan aimed at the industry's favourite excuse. Wafer shortages? Memory shortages? Broadcom, he said, is overcoming both. The real ceiling is money.

Tan rarely hands the industry a free thesis statement, so when he does, it's worth reading twice. Here's the twice-read version: the man whose company just arranged $42 billion in financing so a customer could lease his own chips is warning that the AI buildout's bottleneck has moved from the fab floor to the bank.

Photo: Broadcom CEO Hock Tan. Speaking at a panel in Singapore on Monday, Tan said capital — not memory chips — is the biggest constraint on AI growth. Greg Bezat / Bezat Video Group (CC BY-SA 4.0), via Wikimedia Commons.

What Tan actually said in Singapore

The Reuters report is spare, so let's be precise about what it contains and what it doesn't. Tan said three things:

First: capital will be the biggest constraint facing Broadcom as AI development continues. Second: the company is "overcoming wafer shortages and memory shortages" — his words, via the Reuters paraphrase — but capital remains the binding limit. Third, and most interesting: he name-checked the frontier labs as proof the economics work, saying "frontier labs Anthropic and OpenAI have shown that they can generate significant revenue from the latest models that have been deployed for AI personal assistants."

Read that third point carefully, because it's doing quiet heavy lifting. Tan isn't just complaining that growth is expensive. He's making a lender's argument: the revenue exists, the demand is real (AI personal assistants, not benchmarks), and the constraint is who can marshal the capital to serve it. That's the sound of a CEO who finances other people's data centers for a living.

The memory shortage he just waved away

Here's where Tan's confidence gets spicy, because everyone else in the industry is still screaming about memory.

We pulled the memory figures ourselves, and they're brutal. DRAMeXchange's September contract data, reported by Seoul Economic Daily on October 8, has DDR4 8-gigabit DRAM at $26 a unit — up from $6.30 a year earlier, a fourfold jump. NAND flash is worse: 128-gigabit multi-level cell went from $3.80 to $30.60, an eightfold spike. TrendForce's estimates, via DatacenterDisk's October 11 analysis, have DDR5 64GB server modules doubling in price and DDR4 climbing 60–80%.

The structural cause is no mystery. Samsung, SK Hynix and Micron — who together control over 95% of DRAM production — are diverting wafers to high-bandwidth memory for AI accelerators, which earns three to five times the revenue per wafer of conventional DRAM. TrendForce reckons HBM takes about 18% of the top three's wafer input now, rising to 22% by end of 2026 and 30% by end of 2027. SK Hynix's own July guidance says the shortage could persist past 2030.

The consumer market is already bleeding. IDC reported on October 8 that worldwide PC shipments collapsed 20.1% year-on-year in the third quarter, to 62.7 million units — memory that used to go into laptops is being redirected to AI data centers at margins consumer PCs can't touch. Micron shuttered its Crucial consumer RAM and SSD brand late last year to chase enterprise demand. Meritz Securities estimates Samsung's third-quarter DRAM operating margin at a staggering 78.5% — roughly ten points above Nvidia's.

So when Tan says Broadcom is "overcoming" the memory shortage, translate that honestly: he's overcoming it for Broadcom. The company doesn't fab memory; it designs the networking and custom AI chips (Google's TPUs run on Broadcom silicon) that sit next to it. His supply chain is contracts, not commodities. For everyone else — laptop makers, gamers, hospitals buying servers — the shortage is very much not overcome. It's the difference between being the bank and being the customer.

Photo: Broadcom's headquarters in Irvine, California. The chip giant has become one of the biggest financiers of the AI buildout — including a $42 billion arrangement letting Anthropic lease its own chips. Coolcaesar (CC BY-SA 3.0), via Wikimedia Commons.

He would know — Broadcom just put $42 billion on the table

This is the part of the story Tan didn't need to say out loud, because he already did it with money. As we reported, Broadcom lent Anthropic $42 billion so the AI lab could lease Broadcom-designed chips — effectively financing its own demand. It's the most aggressive vendor-financing move the chip industry has ever seen, and it tells you exactly how Tan sees the constraint: not "can we build the chips," but "who pays for them, and on what terms."

Tan has spent more than a decade running the industry's most disciplined acquisition machine — Avago into Broadcom, CA Technologies, Symantec's enterprise business, the $69 billion VMware deal — almost always with leverage. When a man whose career is a masterclass in using other people's capital tells you capital is the constraint, he's speaking his native language. This isn't a warning from the sidelines. It's the house view of the casino.

The debt machine doesn't care about chips

Tan's Singapore line lands in the middle of a financing frenzy that our reporting has been tracking all month. The AI industry's $113 billion debt machine — the stack of loans, leases and convertibles funding the data-center buildout — is the single most leveraged bet in technology history. SoftBank alone is sitting on a $25 billion margin loan against its Arm stake, a $40 billion bridge facility due in March 2027, and roughly $65 billion in cumulative OpenAI exposure, as we detailed last week.

Against that backdrop, "capital will be the biggest constraint" stops sounding like a complaint and starts sounding like a description of the business model. Every hyperscaler, every lab, every sovereign fund is bidding for the same finite pool of financing to build the same generation of data centers. The chips will get built — TSMC and ASML report earnings this week (our preview of the two reports that will audit the whole boom), and their order books are the least of anyone's worries. The question is who can afford to keep buying.

Why Anthropic and OpenAI matter to his argument

Tan's choice of examples was deliberate, and it deserves a raised eyebrow along with the applause. Anthropic and OpenAI "have shown that they can generate significant revenue" — his words — from AI personal assistants. That's the revenue proof lenders need to keep the debt machine fed.

But it's worth saying the quiet part: these are two labs famous for spending more than they make. Anthropic burned through $4.2 billion in 2025, as we reported, while targeting a $2 trillion valuation. OpenAI's revenue run-rate is the most argued-over number in Silicon Valley. Tan is citing revenue, not profit — and in the AI economy, revenue is the easy part. The hard part is the part he named first: capital.

Still, his framing is shrewder than it looks. If the revenue is real and growing — and the personal-assistant boom suggests it is — then the constraint really is financing, not demand and not supply. That's a bull case wearing a bear costume. Tan isn't saying the boom ends. He's saying the boom belongs to whoever can fund it. And he happens to be one of the funders.

Our take: the bottleneck moved from the fab to the balance sheet

Let's be blunt. Tan is right, but he's right in the specific way a banker is right: the binding constraint on AI growth in 2026 is no longer technical. The models work. The chips work — his company designs some of the best ones. What's scarce is the willingness of capital markets to keep underwriting hundred-billion-dollar buildouts on revenue projections that are, charitably, still being written.

The memory shortage is real, it's structural, and SK Hynix thinks it lasts past 2030 — but it's a pricing problem, not an existential one. Capital is the existential one. If the debt machine seizes up, the fabs keep running and the data centers stop getting built. That's the world Tan is describing, and he described it on a Monday morning panel in Singapore with the casualness of a man who has already placed his bets.

The test of his thesis comes this week: ASML reports Wednesday, TSMC on Thursday. If their numbers confirm the order books are fine — and every indication says they will — then Tan's point stands proven. The chips are coming. The only question left is who can pay for them.

Sources: Reuters (Jun Yuan Yong, Singapore; edited by David Stanway), October 12, 2026; TrendForce wafer-allocation estimates via DatacenterDisk, October 11, 2026; DRAMeXchange contract pricing via Seoul Economic Daily, October 8, 2026; Meritz Securities Samsung margin estimates; SK Hynix July 2026 guidance; IDC Q3 2026 PC shipment data, October 8, 2026; TechPulse USA's own reporting on the Broadcom–Anthropic arrangement, the $113B AI debt machine and SoftBank's balance sheet. All quotations from Hock Tan are verbatim from the Reuters dispatch.

— TechPulse USA, Singapore and New York desks

About TechPulse USA: TechPulse USA is an independent technology newsroom covering AI and tech across the US, Europe, Canada and Australia. Every story is researched from multiple reputable sources and every fact is verified before publication — no filler, no spam.

Update, Oct. 12: the capital constraint, Exhibit A

Forty-eight hours after Tan spoke, the market delivered his point in a single financing. Bloomberg reported on Oct. 12 that Bain Capital-owned Bridge Data Centres is in talks to more than double an existing loan to about $6.6 billion — among the largest data-center borrowings in Asia — to fund its Malaysia, Thailand and India hyperscale build-out, with the expanded deal possibly launching as early as November. The same reporting says the company's separate Thailand loan plan has stalled, around the time Bangkok suspended 49 data-center projects over energy usage. Capital, not chips, is deciding who builds. That's Tan's thesis in one deal.

Related: Bain's Bridge seeks $6.6B debt as Thailand freezes 49 builds; Waymo just borrowed $5 billion — its first-ever debt