Interconnect Bottlenecks, and Broadening Semiconductor Pricing Power

By DripPublished

The gist

AI infrastructure demand is shifting pricing power from compute chips to the networking and capacity layers that now constrain deployment and monetization.

This week’s developments

Broadcom, Cisco, and Ciena Put the Interconnect Stack in the Spotlight

Broadcom said AI networking revenue more than doubled year over year, Cisco reported networking product orders up 40% with $9.3 billion in cloud-provider orders tied to AI infrastructure, and Ciena’s August 2026 survey found 88% of service providers say optical upgrades are urgently needed to meet AI SLAs. The latest signal is that the deployment bottleneck has now migrated into the fabric itself, with Goldman Sachs seeing AI networking TAM rising from about $15 billion to $154 billion by 2028, and co-packaged optics alone reaching roughly $91 billion.

That extends the value shift already underway toward optical and interconnect suppliers such as Coherent, Lumentum, Fabrinet, Credo, and Semtech, alongside the memory and packaging vendors that have been tightening supply for months. HBM, DRAM, and NAND inventories are reportedly down to about four weeks versus a normal eight to twelve, and SK Hynix has reportedly sold out 2026 HBM capacity, while TSMC, SK Hynix, Samsung, and ASE expanding packaging capacity, including in Malaysia, shows where the next leverage is concentrating. For operators, procurement is becoming cluster orchestration across optics, memory, and packaging; for vendors and investors, pricing power is moving further toward suppliers that control optical links, HBM allocation, and packaging slots.

Where should capital shift as AI networking becomes the bottleneck?

If you operate in this industry

  • The bottleneck has moved into optics, memory, and packaging, not GPUs.
  • Treat cluster buildout as supply-chain choreography; secure optical links, HBM, and packaging slots early or lose deployment speed.

Sources

If you sell into this industry

  • Budget is shifting to interconnect, optics, and capacity-constrained supply.
  • Push roadmaps and GTM toward AI networking and optical upgrades; the winners now own scarce slots, not just faster chips.

Sources

  • The Future of Compute Is Fungible — The Diligence Stack - By Creative Strategies, September 17, 2026

    Explains fungible compute, orchestration, and where supplier pricing power survives as buyers mix GPUs, custom silicon, and infrastructure.

  • Wafer supply bottleneck drives longer-term contracts — 디지털투데이, August 21, 2026

    Shows how wafer constraints are pushing longer LTAs, capacity-linked deals, and location-specific supply terms.

If you invest in this industry

  • AI value is migrating to interconnect and capacity bottlenecks.
  • Favor optics, interconnect, HBM, and packaging names; the demand signal is real, but supply scarcity will decide who captures margin.

Sources

Accelerator Scarcity Is Driving Semiconductor Pricing Power

Nebius raised selected on-demand AI cloud compute prices for the second time this year, after a May increase, underscoring that accelerator scarcity—not broad chip inflation—is setting pricing power. In the latest move reported by Reuters, CPU-only instances rose about 25% and memory about 41%, while Nvidia GPU lines moved less sharply: H100 instances were up roughly 17% and B300 instances about 21%, with B300 pricing rising from about $7.85 to $9.50 per hour. CEO Arkady Volozh said demand remains strong and framed the change as a mismatch between AI compute demand and available supply, not higher upstream chip costs.

The signal extends beyond one provider. AWS lifted EC2 Capacity Blocks pricing by about 20% on July 1, 2026 for accelerator capacity, while Reuters reported Huawei raised Ascend 950DT pricing above 250,000 yuan and Cambricon increased its next-generation 690 price by 20% to 30%. For operators, this means higher AI workload costs and greater value in utilization discipline and secured capacity. For vendors and investors, margin capture remains concentrated in premium accelerator-linked infrastructure, not the broader semiconductor stack.

Where should we invest or price for accelerator scarcity next?

If you operate in this industry

  • Accelerator scarcity, not chip inflation, is now your cost curve.
  • Lock capacity early and squeeze utilization; CPU/memory costs are rising, but GPU access is where pricing power and margin risk are concentrated.

Sources

If you sell into this industry

  • Premium accelerator-linked supply is where pricing power lives.
  • Shift roadmap and GTM toward scarce accelerator capacity, not broad semiconductor inputs; buyers will pay for secured supply and higher utilization.

Sources

If you invest in this industry

  • Scarcity is monetizing GPUs, not the wider semiconductor stack.
  • Favor accelerator and capacity owners; this validates premium AI infrastructure pricing, while non-accelerator chip inflation looks less investable.

Sources

Stay ahead in Semiconductor

Get the weekly Semiconductor brief in your inbox — the developments, what they mean by vantage, and what to do next.