Broadcom CEO Hock Tan told Reuters that capital — not memory chip supply — is the principal bottleneck slowing AI-driven technology growth, pushing back against the view that high-bandwidth memory (HBM) and DRAM constraints gate the next leg of AI infrastructure build-out.

The framing cuts against a consensus that has shaped 2024 and 2025 earnings calls. Memory makers, hyperscaler buyers and sell-side analysts have spent 18 months debating whether HBM3e and HBM4 capacity at SK hynix, Samsung Electronics and Micron Technology will keep pace with Nvidia and AMD accelerator shipments. Tan instead points the bottleneck upstream: financing for silicon, advanced packaging, networking and the buildings themselves.

Why Tan's framing matters

Broadcom has become the largest non-Nvidia beneficiary of hyperscaler AI capex. The company sells custom AI accelerators — the XPUs and ASICs powering workloads at Google, Meta and ByteDance — and the Ethernet and PCIe switching silicon that ties AI cluster fabrics together. If customers cannot fund the full stack — chips, CoWoS-L advanced packaging from TSMC, racks, power infrastructure and data-center shells — order books slip regardless of how much HBM SK hynix ships.

The memory camp's counter-argument

Samsung, SK hynix and Micron have each argued that HBM supply will remain structurally tight through 2026, with HBM4 qualification and 12-Hi stack production absorbing leading-edge DRAM wafer capacity that would otherwise feed client and server DIMMs. SK hynix is widely regarded as the leading HBM3e supplier to Nvidia, while Samsung and Micron chase qualification slots for HBM4. Tan's position does not reject that view — it subordinates it. In his telling, memory is a solvable problem once funding is secured.

The supply-chain ripple effect

If capital becomes the binding constraint, second-order risk migrates downstream. ASML, Applied Materials and Lam Research face order-book exposure tied to foundry expansion plans at TSMC, Samsung Foundry and Intel Foundry Services. Foundry capex, in turn, depends on hyperscaler and ASIC buyers raising the cash to pull through planned process nodes — including TSMC's N3 and N2 — that feed AI accelerator die.

Advanced packaging amplifies the picture. CoWoS-L capacity at TSMC has drawn the most supply-chain commentary as a bottleneck for AI accelerator shipments since 2023; capacity expansions are in flight but remain constrained. If capital tightens, those expansion orders stall.

Competitive and geopolitical context

Tan delivers this message against a backdrop where U.S. export controls on advanced AI silicon to China — tightened in 2022 and 2023 — have redirected some hyperscaler spend toward domestic inference capacity. The CHIPS Act continues to subsidize onshore fab construction at TSMC Arizona, Samsung Texas and Intel Ohio. Capital availability in this environment depends on a mix of corporate debt, sovereign and customer financing, and equity markets.

Broadcom itself has signaled interest in pieces of Intel's manufacturing and design operations, a capital-intensive move that would itself depend on the financing environment Tan now describes as the limiting factor.

What to watch next

Upcoming earnings calls from Broadcom, Marvell Technology and ASIC peers will test whether Tan's view holds through 2026. Watch also for customer-financing structures — vendor-financing deals, off-balance-sheet data-center vehicles and hyperscaler debt issuance — that would either validate or undercut the capital-bottleneck thesis.