Broadcom's Hock Tan: capital, not chips, is AI's next bottleneck

WIRE · SEMICONDUCTORS · GLOBAL · Data as at 13 Oct 2026

Broadcom chief executive Hock Tan said on 12 October 2026 that capital, not chips, memory or power, may become the biggest constraint on the AI build-out. Speaking on a panel with Intel chief executive Lip-Bu Tan at DBS's Asia Future Forward Conference, he said the wafer squeeze should ease by 2028, memory could look very different a year from now, and demand for compute should still outrun supply into 2029, probably 2030. On the semiconductors layer, that is the largest custom-chip seller telling buyers the stuck step is moving off the fab and onto the balance sheet.

What happened

Why cash cares

For two years the stuck step on this map has been physical: wafers, packaging, HBM, then power. Tan's claim is that each of those gets solved on a known clock, and the next limit is who will fund the halls before the revenue arrives. That matters for the chip sellers because they sit inside his "$10bn a gigawatt" slice. Broadcom, Nvidia, the memory makers (Samsung, SK hynix, Micron) and the foundry get paid from that slice, and only when someone writes the cheque for the racks.

His own numbers show who can carry the coupon. A frontier lab earning about $30bn per gigawatt against about $10bn of yearly cost keeps roughly $20bn, so in his words the labs "can afford to pay whatever coupons you like to charge." The open-weight camp, losing about $70bn a year on his estimate, cannot. Enterprises buying tokens without a return, which he said is most of them, sit further back still.

Our arithmetic, not a company figure. Under 2.5 GW at about $10bn a GW a year puts under $25bn a year of Anthropic's spend in the pool that chips, memory, networking and halls share. That is the cash Broadcom's custom-chip business is chasing, and it is why the financing behind those racks, such as the $60B AI-chip package Broadcom's banks began gathering on 2 October, is now part of the chip story.

Figure: The bind moves to money. Revenue and cost per gigawatt as Hock Tan stated them, with his supply clock. The under-$25bn pool is our arithmetic.

The second-order read

Who gets paid. Chip, memory and foundry sellers, as long as frontier labs keep earning and lenders keep funding. Tan's split says the lab, not the vendor, keeps most of the value per gigawatt.

Who waits. Buyers that need debt and lack frontier-lab revenue: open-weight model builders, enterprise projects still looking for a return, and build-outs whose financing is not closed.

Who carries the risk. Lenders and landlords funding halls ahead of revenue, more so if long bond yields keep rising. Tan said demand could slow if adoption disappoints or models stop getting much better.

Stuck step. Signed, funded orders. A wafer slot freed up in 2028 does not pay anyone if the buyer cannot raise the money for the rack.

Two cautions

These are a chief executive's on-stage estimates, not filed figures. The $200bn token spend, the $30bn-per-GW revenue and the $10bn-per-GW cost are his numbers, built partly on OpenRouter usage data.

Chip shares fell on 13 October: Samsung and SK hynix each dropped more than 7% in Seoul and ASML more than 3% in Amsterdam. Reports tied the sell-off to both the financing warnings and ChangXin Memory's claim of a denser DRAM cell, so do not read it as a verdict on Tan's remarks alone.

What would change the view

• Chip makers' Q3 order books showing buyers cancelling or pushing out for funding reasons, not supply.
• The $60B Broadcom-linked financing closing, or stalling, at syndication.
• Frontier-lab revenue growth slowing enough to dent Tan's $30bn-per-GW figure.

What to watch

• TSMC's Q3 results on 15 October: whether the packaging and advanced-node commentary still reads as supply-bound.
• Debt costs for hyperscalers and labs: long bond yields were the trigger for the question on stage.
• Memory pricing into 2027: Tan's "a year from now" line is a direct test of how long the HBM squeeze lasts.

Read: Semiconductors, What Is Cash Capex? The Cash That Leaves Before AI Hours, and Broadcom's banks start raising $60B for AI chips

Get the Wire. A free brief every day, here and in your inbox. Subscribe free →

Not investment advice. Do your own research. The under-$25bn-a-year pool and the roughly $70bn open-weight shortfall are our arithmetic from Hock Tan's stated figures.

References

1. CNBC, "Intel, Broadcom CEOs warn capital crunch could become AI's biggest bottleneck", 12 Oct 2026. Capital quote, wafers easing by 2028 or 2029, demand ahead of supply into 2029-2030, Lip-Bu Tan on cost of capital.
2. Computer Weekly, "Capital, not chips, may be the real brake on the AI build-out", 12 Oct 2026. $200bn token spend vs $150bn revenue, frontier vs open-weight split, Anthropic $65bn on under 2.5 GW, $30bn per GW, $10bn per GW cost.
3. The Business Times, "Companies still in search of AI use cases with worthwhile ROI, says Broadcom's Hock Tan", 12 Oct 2026. Panel context, enterprise ROI remarks, capital as the next bottleneck.
4. Seoul Economic Daily, "Chinese Memory Maker's Tech Claim Sinks Samsung, SK hynix Shares", 13 Oct 2026. Samsung and SK hynix each down more than 7%, CXMT 4F² DRAM claim.
5. Invezz, "ASML, European chip stocks fall on CXMT threat and AI funding fears", 13 Oct 2026. ASML down more than 3%, sell-off tied to financing warnings and CXMT.
6. Daily Silicon, "TSMC Eyes Sixth Straight Record Profit", 13 Oct 2026. TSMC Q3 results due Thursday 15 October.