Qualcomm's AWS 1.6T Optical Deal: The Suppliers One Layer Down

On September 8, 2026, Qualcomm announced a multi-generational product collaboration with Amazon to supply AWS data centers with AI inference silicon and optical connectivity up to 1.6T. The effect may not stop at a share fight over optical DSPs; it could land one layer down, in foundry, analog front end and module assembly. [1]

The press release describes "advanced optical connectivity solutions, including links extending up to 1.6T," built on Qualcomm's own SerDes and optical DSP. [1][2] Qualcomm issued Amazon warrants over 25 million shares at $161.26, worth about $4 billion and expiring on September 3, 2036; 3.75 million shares vested immediately, with the rest vesting in tranches tied to commercial arrangements and purchases of up to $60 billion. [3]

Qualcomm closed at $174.09 that session, up 3.2% from September 4, on 25.5 million shares against 8.5 million the prior session; Broadcom rose 3.0%. Astera Labs, which sells connectivity silicon between accelerators, closed at $288.85, down 6.9%. [4]

How a 1.6T optical link is actually built

Inside a data center, traffic between servers and switches cannot stay on copper past a certain distance and has to be converted to light for the fiber. The part that does this is the optical transceiver: it turns electricity into light at one end and back at the other. 1.6T is the speed of a single module.

A module contains several pieces. The optical DSP handles encoding, error correction and distortion compensation, and the SerDes handles serialization. A laser emits the light, a driver drives it, and a transimpedance amplifier boosts the weak photocurrent at the receive end. Those parts sit on a photonic integrated circuit, and a contract manufacturer assembles the module.

What Qualcomm brings to this chain is the SerDes and the optical DSP. It has no silicon-photonics fab, no laser, no driver, no transimpedance amplifier and no module assembly.

AWS's existing 1.6T path runs on the Broadcom and Marvell optical DSP line. Qualcomm's is a second, independent architecture: the two run in parallel rather than replacing one another.

Why a second architecture lands one layer down

If AWS runs two 1.6T architectures side by side, module design, qualification and second-sourcing all get done twice. That is duplication rather than substitution, so wafer starts and module line demand rise instead of being reallocated inside one pool.

Because Qualcomm supplies only the DSP and SerDes, the second architecture's optical front end and manufacturing have to be bought outside. Those parts are indifferent to whose DSP sits next to them: silicon-photonics and SiGe wafers from a merchant foundry, drivers, TIAs and indium-phosphide lasers from a merchant analog supplier, assembly from a contract optical manufacturer.

Incremental wafer starts and module programs land on capacity that is already funded, converting to gross profit at each supplier's own rate: Tower has disclosed a 58% incremental gross contribution, MACOM about 57% gross margin and Fabrinet about 12%. [4][5]

Second-order companies that could be affected

MACOM (MTSI) supplies the analog front end Qualcomm lacks — drivers, transimpedance amplifiers and its own indium-phosphide lasers — into optical modules from 100G to 1.6T. Its 10-Q discloses that Data Center revenue in the quarter ended July 3, 2026 rose $61.8 million, or 81.5%, year over year, implying about $137.6 million of the $342.2 million reported for the quarter, or 40.2%. [6] Management said "those architectures still need, as you pointed out, drivers and TIAs," guided that business to roughly 35% sequential growth, and spent $61 million in the June quarter on a stake in IQE to secure indium-phosphide epitaxy. [7] A second architecture could add analog front-end and laser content, with data-center revenue and gross profit the landing points. The shares rose 2.2% on September 8; Lumentum rose 11.0%. [4]

Tower Semiconductor (TSEM) is the merchant foundry for the silicon-photonics and SiGe wafers both architectures need. Management disclosed that second-quarter silicon-photonics revenue grew over 60% sequentially to an annualized run rate above $680 million, targeting $1 billion in Q4-26; announced customer contracts represent approximately $1.3 billion of 2027 silicon-photonics revenue; and capacity is more than three times second-quarter shipments. [5] A second architecture would be added wafer starts rather than reallocated ones, with utilization and gross margin the landing points. IQE said on its September 7 call that it has signed a multiyear indium-phosphide agreement with Tower. [8] The shares fell 1.1% on September 8; Coherent rose 7.1% the same day. [4]

Fabrinet (FN) is a contract optical manufacturer, and Qualcomm has no module line of its own. The company disclosed FY26 revenue of $4.6 billion, up 36%, with four customers above 10% of revenue — Cisco at 20%, NVIDIA at 16%, Nokia at 11% and Amazon at 11%; fourth-quarter data-center revenue of $669 million was 51% of total, the exit run rate was $5.3 billion, and land and plans support $12.5-14 billion over time. [9] A second AWS 1.6T module family could become one more program for the same assembler to ramp, with revenue and gross profit dollars the landing points; at ~12% gross margin, revenue moves more than profit. The shares rose 2.2% on September 8. [4]

How to verify this

The near-term checkpoints are three earnings calls. Tower's third-quarter call in late October tests whether silicon photonics reaches the $1 billion run-rate guide and whether the 2027 contracted figure is raised. Fabrinet's fiscal Q1 2027 call in early November tests whether Amazon rises above 11% and whether its "new customers and programs" are a second hyperscaler architecture. MACOM's fiscal Q4 2026 call in November tests whether data-center revenue beats the sequential guide. Also watch whether Qualcomm names its foundry, laser and assembly partners.

The chain breaks if: Qualcomm's 1.6T links turn out to be vertically integrated or co-packaged optics with captive photonics; AWS substitutes the Qualcomm architecture for the incumbent rather than running both, leaving total volume unchanged and only reshuffling DSP share; module assembly goes to Celestica, an in-house line or Chinese module vendors; indium-phosphide scarcity, which IQE calls a bottleneck, caps total output so the second architecture competes for a fixed pool; or silicon-photonics capacity from UMC, GlobalFoundries and TSMC arrives fast enough to take the incremental wafer starts away from Tower.

On timing, partner names and design-in disclosures could appear over the next one to two quarters; revenue effects wait for Tower's fourth-quarter silicon-photonics ramp into its 2027 contracted base, and for MACOM's and Fabrinet's fiscal 2027.

This is only meant to surface transmission chains you may have overlooked - it is not a stock recommendation.

Sources

[1] Qualcomm press release (PR Newswire) · 2026-09-08 · https://www.prnewswire.com/news-releases/qualcomm-announces-multi-generational-product-collaboration-with-amazon-to-build-next-generation-ai-data-center-infrastructure-302871895.html

[2] The Next Web · 2026-09-08 · https://thenextweb.com/news/qualcomm-amazon-4bn-warrants

[3] CNBC · 2026-09-08 · https://www.cnbc.com/2026/09/08/qualcomm-amazon-data-center-infrastructure-deal.html

[4] Drillr database · 2026-09-08 · daily closes and volume, market capitalization, margins

[5] Tower Semiconductor Q2 2026 earnings call · 2026-08-04 · silicon-photonics run rate and 2027 contracts

[6] MACOM 10-Q · 2026-08-06 · year-over-year Data Center revenue change

[7] MACOM fiscal Q3 2026 earnings call · 2026-08-06 · data-center guidance and the IQE stake

[8] IQE H1 2026 results call · 2026-09-07 · multiyear indium-phosphide agreement with Tower

[9] Fabrinet fiscal Q4 2026 earnings call · 2026-08-17 · customer mix and data-center revenue